52:I[191445,["/_next/static/immutable/chunks/42yzn81wlo94m.js","/_next/static/immutable/chunks/2bm6_f2j__8vm.js","/_next/static/immutable/chunks/3i5397s96cu0a.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/2f4lnnvm42_vp.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2888ntn3nn8dg.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/41cng7f4va9y-.js","/_next/static/immutable/chunks/2r_ddychp2ui5.js","/_next/static/immutable/chunks/1g9smhz0fks9x.js","/_next/static/immutable/chunks/2o71xrq9b0yu6.js","/_next/static/immutable/chunks/3qs9vw68v38q9.js","/_next/static/immutable/chunks/1rm5ppn9o7oew.js","/_next/static/immutable/chunks/2hbgrpnh1fmap.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/3ol7h3r_z2pit.js","/_next/static/immutable/chunks/3skty2nb9a-2-.js","/_next/static/immutable/chunks/3oekgz5eyywh-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/25buewscho2w5.js","/_next/static/immutable/chunks/3u4carzbf5fig.js","/_next/static/immutable/chunks/0ewa-4giwctvp.js","/_next/static/immutable/chunks/0fopioox00_wn.js","/_next/static/immutable/chunks/17sbo0v8uw3y7.js","/_next/static/immutable/chunks/05727lidoh_64.js","/_next/static/immutable/chunks/0b0jg9r2u87k8.js","/_next/static/immutable/chunks/459bt34i9w_4u.js","/_next/static/immutable/chunks/2xi3oa1rgz4zi.js","/_next/static/immutable/chunks/0nuoj_60cpek_.js","/_next/static/immutable/chunks/25x_rwsdb74m3.js","/_next/static/immutable/chunks/0p2ndf8a03w3d.js","/_next/static/immutable/chunks/3rpw8gofk-lvz.js","/_next/static/immutable/chunks/0yqqzq4qo78e-.js","/_next/static/immutable/chunks/1lvgxqj71539d.js","/_next/static/immutable/chunks/412rzi0r7ty5b.js","/_next/static/immutable/chunks/2m_mu4_r75f07.js","/_next/static/immutable/chunks/084wa77nfb_vx.js","/_next/static/immutable/chunks/3mt6t4u0f9pd2.js","/_next/static/immutable/chunks/1lgj52kxs241t.js","/_next/static/immutable/chunks/2_9cwii8mx0pv.js","/_next/static/immutable/chunks/4579_u3pt_bl4.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/3du8tbb5hc0uo.js"],"WorkTabProvider"] 53:I[649892,["/_next/static/immutable/chunks/42yzn81wlo94m.js","/_next/static/immutable/chunks/2bm6_f2j__8vm.js","/_next/static/immutable/chunks/3i5397s96cu0a.js","/_next/static/immutable/chunks/2x03jtn4ys_wx.js","/_next/static/immutable/chunks/2f4lnnvm42_vp.js","/_next/static/immutable/chunks/16-tr_iilxrhb.js","/_next/static/immutable/chunks/2888ntn3nn8dg.js","/_next/static/immutable/chunks/2xqpa5c7ogclt.js","/_next/static/immutable/chunks/41cng7f4va9y-.js","/_next/static/immutable/chunks/2r_ddychp2ui5.js","/_next/static/immutable/chunks/1g9smhz0fks9x.js","/_next/static/immutable/chunks/2o71xrq9b0yu6.js","/_next/static/immutable/chunks/3qs9vw68v38q9.js","/_next/static/immutable/chunks/1rm5ppn9o7oew.js","/_next/static/immutable/chunks/2hbgrpnh1fmap.js","/_next/static/immutable/chunks/0tfe231gghziz.js","/_next/static/immutable/chunks/3ol7h3r_z2pit.js","/_next/static/immutable/chunks/3skty2nb9a-2-.js","/_next/static/immutable/chunks/3oekgz5eyywh-.js","/_next/static/immutable/chunks/1e0jyw4aho8bn.js","/_next/static/immutable/chunks/2ri0u_hhdt48v.js","/_next/static/immutable/chunks/25buewscho2w5.js","/_next/static/immutable/chunks/3u4carzbf5fig.js","/_next/static/immutable/chunks/0ewa-4giwctvp.js","/_next/static/immutable/chunks/0fopioox00_wn.js","/_next/static/immutable/chunks/17sbo0v8uw3y7.js","/_next/static/immutable/chunks/05727lidoh_64.js","/_next/static/immutable/chunks/0b0jg9r2u87k8.js","/_next/static/immutable/chunks/459bt34i9w_4u.js","/_next/static/immutable/chunks/2xi3oa1rgz4zi.js","/_next/static/immutable/chunks/0nuoj_60cpek_.js","/_next/static/immutable/chunks/25x_rwsdb74m3.js","/_next/static/immutable/chunks/0p2ndf8a03w3d.js","/_next/static/immutable/chunks/3rpw8gofk-lvz.js","/_next/static/immutable/chunks/0yqqzq4qo78e-.js","/_next/static/immutable/chunks/1lvgxqj71539d.js","/_next/static/immutable/chunks/412rzi0r7ty5b.js","/_next/static/immutable/chunks/2m_mu4_r75f07.js","/_next/static/immutable/chunks/084wa77nfb_vx.js","/_next/static/immutable/chunks/3mt6t4u0f9pd2.js","/_next/static/immutable/chunks/1lgj52kxs241t.js","/_next/static/immutable/chunks/2_9cwii8mx0pv.js","/_next/static/immutable/chunks/4579_u3pt_bl4.js","/_next/static/immutable/chunks/3-y1wg0q2kip1.js","/_next/static/immutable/chunks/3du8tbb5hc0uo.js"],"WorkHeader"] 54:T1a02,Proposal: Energy-Based Models for Brain Age Estimation 1. Overview Scientific rationale & importance: Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk1,2,3. Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects4. This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models5. Facilities, approvals, and timeline: MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months. Data sharing commitment: All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements. 2. Introduction Background & rationale: Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)1, which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners). Energy-Based Models (EBMs)5 model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods. Research questions & hypotheses: H1: EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors. H2: EBM latent spaces encode less scanner-related variance than conventional embeddings. H3: EBMs retain or surpass standard models in predictive accuracy. Impact statement: If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible. 3. Methods Sample Datasets: To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification. Inclusion: quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data. Exclusion: missing covariates, poor segmentation, or acquisition artefacts. Materials and Procedures Preprocessing Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain. Brain extraction and bias-field correction. Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter. Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible. Models Input: single-channel T1w volumes Output: chronological age (regression). Loss: smooth L1 (Huber) or MAE; auxiliary site-adversarial head Baselines SFCN (Simple Fully Convolutional Network): canonical brain-age architecture as baseline. SFCN + Conformal Prediction: split-conformal calibration on validation folds to provide marginal prediction intervals. Proposed DenseNet EBM: 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness. Optional invariance: gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects. Evaluation Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal). Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal. Hardware & Environment 4× NVIDIA RTX A5000 (24 GB each; total 96 GB), Software: Python, PyTorch, MONAI/TorchIO, scikit-learn. Reproducibility Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance. Planned Analyses Prediction performance: MAE, RMSE, R². Uncertainty calibration: expected calibration error (ECE), coverage vs. nominal quantiles. Scanner robustness: Adjusted silhouette width (ASW) and η² by site. Interpretability: Saliency maps (GradCAM). Ethics and Data Management All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0. 4. Budget (Total = $6500 USD) CategoryDescriptionCost (USD) Compute resources, storage & management GPU-hours for EBM training and baselines 2000 Biobank Data Access Fees UKBiobank, NAKO, SHIP 2000 Dissemination open-access fees, conference registration fee 2000 Overhead & contingency Unexpected compute or data-transfer costs 500 Total 6500 USD 5. References Cole, James H., et al. "Brain age predicts mortality." Molecular psychiatry 23.5 (2018): 1385-1392. Cole, James H., and Katja Franke. "Predicting age using neuroimaging: innovative brain ageing biomarkers." Trends in neurosciences 40.12 (2017): 681-690. Franke, Katja, et al. "Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters." Neuroimage 50.3 (2010): 883-892. Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668. Song, Yang, and Diederik P. Kingma. "How to train your energy-based models." arXiv preprint arXiv:2101.03288 (2021).55:T1ad8,Proposal: Energy-Based Models for Brain Age Estimation 1. Overview Scientific rationale & importance: Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk1,2,3. Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects4. This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models5. Facilities, approvals, and timeline: MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months. Data sharing commitment: All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements. 2. Introduction Background & rationale: Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)1, which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners). Energy-Based Models (EBMs)5 model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods. Research questions & hypotheses: H1: EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors. H2: EBM latent spaces encode less scanner-related variance than conventional embeddings. H3: EBMs retain or surpass standard models in predictive accuracy. Impact statement: If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible. 3. Methods Sample Datasets: To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification. Inclusion: quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data. Exclusion: missing covariates, poor segmentation, or acquisition artefacts. Materials and Procedures Preprocessing Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain. Brain extraction and bias-field correction. Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter. Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible. Models Input: single-channel T1w volumes Output: chronological age (regression). Loss: smooth L1 (Huber) or MAE; auxiliary site-adversarial head Baselines SFCN (Simple Fully Convolutional Network): canonical brain-age architecture as baseline. SFCN + Conformal Prediction: split-conformal calibration on validation folds to provide marginal prediction intervals. Proposed DenseNet EBM: 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness. Optional invariance: gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects. Evaluation Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal). Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal. Hardware & Environment 4× NVIDIA RTX A5000 (24 GB each; total 96 GB), Software: Python, PyTorch, MONAI/TorchIO, scikit-learn. Reproducibility Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance. Planned Analyses Prediction performance: MAE, RMSE, R². Uncertainty calibration: expected calibration error (ECE), coverage vs. nominal quantiles. Scanner robustness: Adjusted silhouette width (ASW) and η² by site. Interpretability: Saliency maps (GradCAM). Ethics and Data Management All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0. 4. Budget (Total = $6500 USD) CategoryDescriptionCost (USD) Compute resources, storage & management GPU-hours for EBM training and baselines 2000 Biobank Data Access Fees UKBiobank, NAKO, SHIP 2000 Dissemination open-access fees, conference registration fee 2000 Overhead & contingency Unexpected compute or data-transfer costs 500 Total 6500 USD 5. References Cole, James H., et al. "Brain age predicts mortality." Molecular psychiatry 23.5 (2018): 1385-1392. Cole, James H., and Katja Franke. "Predicting age using neuroimaging: innovative brain ageing biomarkers." Trends in neurosciences 40.12 (2017): 681-690. Franke, Katja, et al. "Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters." Neuroimage 50.3 (2010): 883-892. Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668. Song, Yang, and Diederik P. Kingma. "How to train your energy-based models." arXiv preprint arXiv:2101.03288 (2021). 56:T5dd5,{"type":"doc","content":[{"type":"heading","attrs":{"id":"8c28b8c9-ba5d-4a7c-a810-d157cb4f36df","data-toc-id":"8c28b8c9-ba5d-4a7c-a810-d157cb4f36df","textAlign":null,"level":1},"content":[{"type":"text","text":"Proposal: Energy-Based Models for Brain Age Estimation"}]},{"type":"heading","attrs":{"id":"daa53e13-088c-4086-bb5d-8c6d4f8a05bb","data-toc-id":"daa53e13-088c-4086-bb5d-8c6d4f8a05bb","textAlign":"left","level":2},"content":[{"type":"text","text":"1. Overview"}]},{"type":"paragraph","attrs":{"id":"54b6e928-499a-4bd9-882b-7ac91934466f","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Scientific rationale & importance:"},{"type":"hardBreak"},{"type":"text","text":"Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk"},{"type":"text","marks":[{"type":"superscript"}],"text":"1,2,3"},{"type":"text","text":". Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects"},{"type":"text","marks":[{"type":"superscript"}],"text":"4"},{"type":"text","text":"."},{"type":"hardBreak"},{"type":"text","text":"This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models"},{"type":"text","marks":[{"type":"superscript"}],"text":"5"},{"type":"text","text":"."}]},{"type":"paragraph","attrs":{"id":"66f55b7d-0a7a-488e-9b5f-c66386b6eb9d","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Facilities, approvals, and timeline:"},{"type":"hardBreak"},{"type":"text","text":"MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months."}]},{"type":"paragraph","attrs":{"id":"7256aaee-f6b9-4940-849e-ba9a4bcee6ad","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Data sharing commitment:"},{"type":"hardBreak"},{"type":"text","text":"All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements."}]},{"type":"heading","attrs":{"id":"57bff2c2-8ac8-4a74-bf28-d132a0603f0b","data-toc-id":"57bff2c2-8ac8-4a74-bf28-d132a0603f0b","textAlign":"left","level":2},"content":[{"type":"text","text":"2. Introduction"}]},{"type":"paragraph","attrs":{"id":"62cd449b-13ad-4d21-928a-8e666394bd29","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Background & rationale:"},{"type":"hardBreak"},{"type":"text","text":"Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)"},{"type":"text","marks":[{"type":"superscript"}],"text":"1"},{"type":"text","text":", which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners)."}]},{"type":"paragraph","attrs":{"id":"a5984008-699e-4072-883b-dfca04bda16d","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Energy-Based Models (EBMs)"},{"type":"text","marks":[{"type":"bold"},{"type":"superscript"}],"text":"5"},{"type":"text","text":" model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods."}]},{"type":"paragraph","attrs":{"id":"3eab8b5b-046d-4dd1-bb97-f15ba5b96645","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research questions & hypotheses:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"03ee3b41-4ecf-478e-8f4a-172e48181179","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"H1:"},{"type":"text","text":" EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"fd36076c-7031-49c7-9c80-80f5de6dc62e","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"H2:"},{"type":"text","text":" EBM latent spaces encode less scanner-related variance than conventional embeddings."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2069df1f-89bd-4f7b-a6e2-c6756b3e95c5","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"H3:"},{"type":"text","text":" EBMs retain or surpass standard models in predictive accuracy."}]}]}]},{"type":"paragraph","attrs":{"id":"0709f7df-586d-4542-8de0-4d9693a92df1","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Impact statement:"},{"type":"hardBreak"},{"type":"text","text":"If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible."}]},{"type":"heading","attrs":{"id":"71004ee8-bacd-4777-8b74-5b1fda386681","data-toc-id":"71004ee8-bacd-4777-8b74-5b1fda386681","textAlign":"left","level":2},"content":[{"type":"text","text":"3. Methods"}]},{"type":"heading","attrs":{"id":"35ccffe0-b253-4a1a-b019-cf38887fbb6f","data-toc-id":"35ccffe0-b253-4a1a-b019-cf38887fbb6f","textAlign":"left","level":3},"content":[{"type":"text","text":"Sample"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"dabebc34-0ca5-4ea7-a5b1-6a319d6e16da","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Datasets:"},{"type":"text","text":" To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"748ba500-7352-44c3-9814-11d0674fc726","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Inclusion:"},{"type":"text","text":" quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"0a3e8475-30dc-484e-98f6-bc47d27060bc","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Exclusion:"},{"type":"text","text":" missing covariates, poor segmentation, or acquisition artefacts."}]}]}]},{"type":"heading","attrs":{"id":"f84f6e7f-4019-4cd8-816b-67effc41d057","data-toc-id":"f84f6e7f-4019-4cd8-816b-67effc41d057","textAlign":null,"level":3},"content":[{"type":"text","text":"Materials and Procedures"}]},{"type":"heading","attrs":{"id":"7b754b2a-9786-4f2b-9f0f-2abfd2d6628d","data-toc-id":"7b754b2a-9786-4f2b-9f0f-2abfd2d6628d","textAlign":null,"level":3},"content":[{"type":"text","text":"Preprocessing"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2c8a531c-6994-47e9-9a8b-0435bdb3ac29","class":null,"textAlign":null},"content":[{"type":"text","text":"Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1803ed42-c58b-4916-bb51-0615db787a58","class":null,"textAlign":null},"content":[{"type":"text","text":"Brain extraction and bias-field correction."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"82762017-060f-408b-b443-f7ab199a2663","class":null,"textAlign":null},"content":[{"type":"text","text":"Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"567564d2-31f7-4b25-8b74-ecdff8c21d20","class":null,"textAlign":null},"content":[{"type":"text","text":"Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible."}]}]}]},{"type":"heading","attrs":{"id":"c8b7d250-2915-4004-8528-1936b85f66aa","data-toc-id":"c8b7d250-2915-4004-8528-1936b85f66aa","textAlign":null,"level":3},"content":[{"type":"text","text":"Models"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"4155e441-9d62-4350-9255-c95d99ca264a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Input:"},{"type":"text","text":" single-channel T1w volumes"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"9a51cc45-ff8e-4e64-ad80-62550689d30b","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Output:"},{"type":"text","text":" chronological age (regression)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2929205a-db16-409a-a538-799cacdbf30a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Loss:"},{"type":"text","text":" smooth L1 (Huber) or MAE; auxiliary site-adversarial head"}]}]}]},{"type":"heading","attrs":{"id":"e1859471-1ae5-4370-8021-56a66f859244","data-toc-id":"e1859471-1ae5-4370-8021-56a66f859244","textAlign":null,"level":3},"content":[{"type":"text","text":"Baselines"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"50df35b0-894f-40b4-b716-fc4a48b122af","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"SFCN (Simple Fully Convolutional Network):"},{"type":"text","text":" canonical brain-age architecture as baseline."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"405e077d-2827-4870-bf87-3f1095f4523b","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"SFCN + Conformal Prediction:"},{"type":"text","text":" split-conformal calibration on validation folds to provide marginal prediction intervals."}]}]}]},{"type":"heading","attrs":{"id":"2e8ba557-6a83-43c7-9b87-03b601b8fd6a","data-toc-id":"2e8ba557-6a83-43c7-9b87-03b601b8fd6a","textAlign":null,"level":3},"content":[{"type":"text","text":"Proposed"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"94dae09b-f591-457b-b72b-8a40423ae673","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"DenseNet EBM:"},{"type":"text","text":" 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"13fcd54c-9b61-49e8-a323-18433d014112","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Optional invariance:"},{"type":"text","text":" gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects."}]}]}]},{"type":"heading","attrs":{"id":"35b59b31-962f-4b2e-9cfe-3d12dd86bb4f","data-toc-id":"35b59b31-962f-4b2e-9cfe-3d12dd86bb4f","textAlign":null,"level":3},"content":[{"type":"text","text":"Evaluation"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"993419b6-2573-4b4e-859e-471168f0d918","class":null,"textAlign":null},"content":[{"type":"text","text":"Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1c1dfa30-dac5-4a55-a945-962213a74cfc","class":null,"textAlign":null},"content":[{"type":"text","text":"Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal."}]}]}]},{"type":"heading","attrs":{"id":"654de9eb-8a40-45d4-86fa-469a9a0e48b5","data-toc-id":"654de9eb-8a40-45d4-86fa-469a9a0e48b5","textAlign":null,"level":3},"content":[{"type":"text","text":"Hardware & Environment"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"5900b057-d1db-4e75-ac41-0568896d5049","class":null,"textAlign":null},"content":[{"type":"text","text":"4× "},{"type":"text","marks":[{"type":"bold"}],"text":"NVIDIA RTX A5000"},{"type":"text","text":" (24 GB each; total 96 GB), "}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c431ed4c-fa5f-42c2-9fc8-903121868b20","class":null,"textAlign":null},"content":[{"type":"text","text":"Software: Python, PyTorch, MONAI/TorchIO, scikit-learn."}]}]}]},{"type":"heading","attrs":{"id":"fc27d2a2-b166-44ba-a669-80db3a8eb124","data-toc-id":"fc27d2a2-b166-44ba-a669-80db3a8eb124","textAlign":null,"level":3},"content":[{"type":"text","text":"Reproducibility"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"41779e3e-2b3e-4d06-be95-333bce5c31dc","class":null,"textAlign":null},"content":[{"type":"text","text":"Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance."}]}]}]},{"type":"heading","attrs":{"id":"f2aa22d6-9e44-453a-bd9a-e1bf2c18e1ba","data-toc-id":"f2aa22d6-9e44-453a-bd9a-e1bf2c18e1ba","textAlign":"justify","level":3},"content":[{"type":"text","text":"Planned Analyses"}]},{"type":"orderedList","attrs":{"start":1},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"5ed3d007-0921-4c7b-95de-60f4b944a9e1","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Prediction performance:"},{"type":"text","text":" MAE, RMSE, R²."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c96cc4c2-e8a5-422a-b426-fba4ee37cc0c","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Uncertainty calibration:"},{"type":"text","text":" expected calibration error (ECE), coverage vs. nominal quantiles."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c8034bb9-08cf-4387-be43-2ed2ece469af","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Scanner robustness:"},{"type":"text","text":" Adjusted silhouette width (ASW) and η² by site."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"163114ff-0639-4c87-8b72-bacb4b0c7186","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Interpretability:"},{"type":"text","text":" Saliency maps (GradCAM). "}]}]}]},{"type":"heading","attrs":{"id":"9b42b92b-48c6-46fc-852d-baee696d9bdd","data-toc-id":"9b42b92b-48c6-46fc-852d-baee696d9bdd","textAlign":null,"level":3},"content":[{"type":"text","text":"Ethics and Data Management"}]},{"type":"paragraph","attrs":{"id":"0fb02641-c5b7-4fd5-b316-9cd8c6fdba85","class":null,"textAlign":"justify"},"content":[{"type":"text","text":"All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0."}]},{"type":"heading","attrs":{"id":"45dfe073-8ab2-4233-856a-abe3b1dd6db8","data-toc-id":"45dfe073-8ab2-4233-856a-abe3b1dd6db8","textAlign":null,"level":2},"content":[{"type":"text","text":"4. Budget (Total = "},{"type":"text","marks":[{"type":"bold"}],"text":"$6500 USD"},{"type":"text","text":")"}]},{"type":"table","attrs":{"id":"704bbb27-e91f-4042-9421-8feed2221b90"},"content":[{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"87c7a563-37f6-4331-bc57-e31cdb4dfc39","class":null,"textAlign":null}}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"7aa71b38-890d-426b-9a60-ee2dceeb17e4","class":null,"textAlign":null}}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"71e8546b-6794-4c3e-813e-b40a44de1e57","class":null,"textAlign":null},"content":[{"type":"text","text":"CategoryDescriptionCost (USD)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"3f180030-1f36-4a69-bc81-3584ccba9a42","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Compute resources, storage & management"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"ab669c67-03f4-47b7-94bc-0fd096555b42","class":null,"textAlign":null},"content":[{"type":"text","text":"GPU-hours for EBM training and baselines"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"09d35367-3520-496b-af12-2315eda15a8e","class":null,"textAlign":null},"content":[{"type":"text","text":"2000"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"81a78735-cc9f-4a3a-96c3-b154f7678e04","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Biobank Data Access Fees"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"5e7748d3-c43b-42e8-ac39-7c6b22a12e29","class":null,"textAlign":null},"content":[{"type":"text","text":"UKBiobank, NAKO, SHIP"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"b6b08919-0a80-4b51-88fc-82156c3b6748","class":null,"textAlign":null},"content":[{"type":"text","text":"2000"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"d66c98ce-cd1b-4d86-9f76-b13b67745064","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Dissemination"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"c3a91185-be30-4bff-b65c-5e6166072646","class":null,"textAlign":null},"content":[{"type":"text","text":"open-access fees, conference registration fee"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"4c37d669-fb8f-4cd4-8819-95ee87d084dd","class":null,"textAlign":null},"content":[{"type":"text","text":"2000"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"9d949e96-cd01-496f-bb99-f828d846c346","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Overhead & contingency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"40257056-b729-48f7-9e54-6620cab34cc3","class":null,"textAlign":null},"content":[{"type":"text","text":"Unexpected compute or data-transfer costs"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"7b1c0969-1aee-4a0d-8ab0-294038e77cda","class":null,"textAlign":null},"content":[{"type":"text","text":"500"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"2099f4a8-1a6e-44cb-8a50-5e501b1c630c","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Total"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"bb61f687-a3ca-4102-8786-99f328fa1990","class":null,"textAlign":null}}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"468eccda-4c4d-446d-86df-a69375f620d4","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"6500 USD"}]}]}]}]},{"type":"heading","attrs":{"id":"a9fce6d4-a8f8-4fd5-991d-3f183e5604f4","data-toc-id":"a9fce6d4-a8f8-4fd5-991d-3f183e5604f4","textAlign":"left","level":2},"content":[{"type":"text","text":"5. References"}]},{"type":"orderedList","attrs":{"start":1},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f9e05b09-2502-4460-919f-eca1de0d4439","class":null,"textAlign":null},"content":[{"type":"text","text":"Cole, James H., et al. \"Brain age predicts mortality.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Molecular psychiatry"},{"type":"text","text":" 23.5 (2018): 1385-1392."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c9a54856-2994-4460-a6dd-d4e6108152a3","class":null,"textAlign":null},"content":[{"type":"text","text":"Cole, James H., and Katja Franke. \"Predicting age using neuroimaging: innovative brain ageing biomarkers.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Trends in neurosciences"},{"type":"text","text":" 40.12 (2017): 681-690."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"46e304ea-1ebf-4f55-8393-c4d7a6277388","class":null,"textAlign":null},"content":[{"type":"text","text":"Franke, Katja, et al. \"Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Neuroimage"},{"type":"text","text":" 50.3 (2010): 883-892."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"4ebc377d-d0d5-4662-aca2-ad089e5311c3","class":null,"textAlign":null},"content":[{"type":"text","text":"Kumari, LK Soumya, and R. Sundarrajan. \"A review on brain age prediction models.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Brain Research"},{"type":"text","text":" 1823 (2024): 148668."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1a81c68f-5dfc-4100-802e-1811858f8953","class":null,"textAlign":null},"content":[{"type":"text","text":"Song, Yang, and Diederik P. Kingma. \"How to train your energy-based models.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"arXiv preprint arXiv:2101.03288"},{"type":"text","text":" (2021)."}]}]}]},{"type":"paragraph","attrs":{"id":"9f6b49c8-8ea9-4fc0-b387-b5617260a1c8","class":null,"textAlign":null}}]}57:T1ad8,Proposal: Energy-Based Models for Brain Age Estimation 1. Overview Scientific rationale & importance: Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk1,2,3. Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects4. This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models5. Facilities, approvals, and timeline: MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months. Data sharing commitment: All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements. 2. Introduction Background & rationale: Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)1, which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners). Energy-Based Models (EBMs)5 model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods. Research questions & hypotheses: H1: EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors. H2: EBM latent spaces encode less scanner-related variance than conventional embeddings. H3: EBMs retain or surpass standard models in predictive accuracy. Impact statement: If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible. 3. Methods Sample Datasets: To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification. Inclusion: quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data. Exclusion: missing covariates, poor segmentation, or acquisition artefacts. Materials and Procedures Preprocessing Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain. Brain extraction and bias-field correction. Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter. Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible. Models Input: single-channel T1w volumes Output: chronological age (regression). Loss: smooth L1 (Huber) or MAE; auxiliary site-adversarial head Baselines SFCN (Simple Fully Convolutional Network): canonical brain-age architecture as baseline. SFCN + Conformal Prediction: split-conformal calibration on validation folds to provide marginal prediction intervals. Proposed DenseNet EBM: 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness. Optional invariance: gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects. Evaluation Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal). Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal. Hardware & Environment 4× NVIDIA RTX A5000 (24 GB each; total 96 GB), Software: Python, PyTorch, MONAI/TorchIO, scikit-learn. Reproducibility Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance. Planned Analyses Prediction performance: MAE, RMSE, R². Uncertainty calibration: expected calibration error (ECE), coverage vs. nominal quantiles. Scanner robustness: Adjusted silhouette width (ASW) and η² by site. Interpretability: Saliency maps (GradCAM). Ethics and Data Management All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0. 4. Budget (Total = $6500 USD) CategoryDescriptionCost (USD) Compute resources, storage & management GPU-hours for EBM training and baselines 2000 Biobank Data Access Fees UKBiobank, NAKO, SHIP 2000 Dissemination open-access fees, conference registration fee 2000 Overhead & contingency Unexpected compute or data-transfer costs 500 Total 6500 USD 5. References Cole, James H., et al. "Brain age predicts mortality." Molecular psychiatry 23.5 (2018): 1385-1392. Cole, James H., and Katja Franke. "Predicting age using neuroimaging: innovative brain ageing biomarkers." Trends in neurosciences 40.12 (2017): 681-690. Franke, Katja, et al. "Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters." Neuroimage 50.3 (2010): 883-892. Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668. Song, Yang, and Diederik P. Kingma. "How to train your energy-based models." arXiv preprint arXiv:2101.03288 (2021). 58:T5dd5,{"type":"doc","content":[{"type":"heading","attrs":{"id":"8c28b8c9-ba5d-4a7c-a810-d157cb4f36df","data-toc-id":"8c28b8c9-ba5d-4a7c-a810-d157cb4f36df","textAlign":null,"level":1},"content":[{"type":"text","text":"Proposal: Energy-Based Models for Brain Age Estimation"}]},{"type":"heading","attrs":{"id":"daa53e13-088c-4086-bb5d-8c6d4f8a05bb","data-toc-id":"daa53e13-088c-4086-bb5d-8c6d4f8a05bb","textAlign":"left","level":2},"content":[{"type":"text","text":"1. Overview"}]},{"type":"paragraph","attrs":{"id":"54b6e928-499a-4bd9-882b-7ac91934466f","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Scientific rationale & importance:"},{"type":"hardBreak"},{"type":"text","text":"Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk"},{"type":"text","marks":[{"type":"superscript"}],"text":"1,2,3"},{"type":"text","text":". Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects"},{"type":"text","marks":[{"type":"superscript"}],"text":"4"},{"type":"text","text":"."},{"type":"hardBreak"},{"type":"text","text":"This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models"},{"type":"text","marks":[{"type":"superscript"}],"text":"5"},{"type":"text","text":"."}]},{"type":"paragraph","attrs":{"id":"66f55b7d-0a7a-488e-9b5f-c66386b6eb9d","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Facilities, approvals, and timeline:"},{"type":"hardBreak"},{"type":"text","text":"MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months."}]},{"type":"paragraph","attrs":{"id":"7256aaee-f6b9-4940-849e-ba9a4bcee6ad","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Data sharing commitment:"},{"type":"hardBreak"},{"type":"text","text":"All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements."}]},{"type":"heading","attrs":{"id":"57bff2c2-8ac8-4a74-bf28-d132a0603f0b","data-toc-id":"57bff2c2-8ac8-4a74-bf28-d132a0603f0b","textAlign":"left","level":2},"content":[{"type":"text","text":"2. Introduction"}]},{"type":"paragraph","attrs":{"id":"62cd449b-13ad-4d21-928a-8e666394bd29","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Background & rationale:"},{"type":"hardBreak"},{"type":"text","text":"Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)"},{"type":"text","marks":[{"type":"superscript"}],"text":"1"},{"type":"text","text":", which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners)."}]},{"type":"paragraph","attrs":{"id":"a5984008-699e-4072-883b-dfca04bda16d","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Energy-Based Models (EBMs)"},{"type":"text","marks":[{"type":"bold"},{"type":"superscript"}],"text":"5"},{"type":"text","text":" model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods."}]},{"type":"paragraph","attrs":{"id":"3eab8b5b-046d-4dd1-bb97-f15ba5b96645","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Research questions & hypotheses:"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"03ee3b41-4ecf-478e-8f4a-172e48181179","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"H1:"},{"type":"text","text":" EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"fd36076c-7031-49c7-9c80-80f5de6dc62e","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"H2:"},{"type":"text","text":" EBM latent spaces encode less scanner-related variance than conventional embeddings."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2069df1f-89bd-4f7b-a6e2-c6756b3e95c5","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"H3:"},{"type":"text","text":" EBMs retain or surpass standard models in predictive accuracy."}]}]}]},{"type":"paragraph","attrs":{"id":"0709f7df-586d-4542-8de0-4d9693a92df1","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Impact statement:"},{"type":"hardBreak"},{"type":"text","text":"If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible."}]},{"type":"heading","attrs":{"id":"71004ee8-bacd-4777-8b74-5b1fda386681","data-toc-id":"71004ee8-bacd-4777-8b74-5b1fda386681","textAlign":"left","level":2},"content":[{"type":"text","text":"3. Methods"}]},{"type":"heading","attrs":{"id":"35ccffe0-b253-4a1a-b019-cf38887fbb6f","data-toc-id":"35ccffe0-b253-4a1a-b019-cf38887fbb6f","textAlign":"left","level":3},"content":[{"type":"text","text":"Sample"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"dabebc34-0ca5-4ea7-a5b1-6a319d6e16da","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Datasets:"},{"type":"text","text":" To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"748ba500-7352-44c3-9814-11d0674fc726","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Inclusion:"},{"type":"text","text":" quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"0a3e8475-30dc-484e-98f6-bc47d27060bc","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Exclusion:"},{"type":"text","text":" missing covariates, poor segmentation, or acquisition artefacts."}]}]}]},{"type":"heading","attrs":{"id":"f84f6e7f-4019-4cd8-816b-67effc41d057","data-toc-id":"f84f6e7f-4019-4cd8-816b-67effc41d057","textAlign":null,"level":3},"content":[{"type":"text","text":"Materials and Procedures"}]},{"type":"heading","attrs":{"id":"7b754b2a-9786-4f2b-9f0f-2abfd2d6628d","data-toc-id":"7b754b2a-9786-4f2b-9f0f-2abfd2d6628d","textAlign":null,"level":3},"content":[{"type":"text","text":"Preprocessing"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2c8a531c-6994-47e9-9a8b-0435bdb3ac29","class":null,"textAlign":null},"content":[{"type":"text","text":"Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1803ed42-c58b-4916-bb51-0615db787a58","class":null,"textAlign":null},"content":[{"type":"text","text":"Brain extraction and bias-field correction."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"82762017-060f-408b-b443-f7ab199a2663","class":null,"textAlign":null},"content":[{"type":"text","text":"Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"567564d2-31f7-4b25-8b74-ecdff8c21d20","class":null,"textAlign":null},"content":[{"type":"text","text":"Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible."}]}]}]},{"type":"heading","attrs":{"id":"c8b7d250-2915-4004-8528-1936b85f66aa","data-toc-id":"c8b7d250-2915-4004-8528-1936b85f66aa","textAlign":null,"level":3},"content":[{"type":"text","text":"Models"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"4155e441-9d62-4350-9255-c95d99ca264a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Input:"},{"type":"text","text":" single-channel T1w volumes"}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"9a51cc45-ff8e-4e64-ad80-62550689d30b","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Output:"},{"type":"text","text":" chronological age (regression)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"2929205a-db16-409a-a538-799cacdbf30a","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Loss:"},{"type":"text","text":" smooth L1 (Huber) or MAE; auxiliary site-adversarial head"}]}]}]},{"type":"heading","attrs":{"id":"e1859471-1ae5-4370-8021-56a66f859244","data-toc-id":"e1859471-1ae5-4370-8021-56a66f859244","textAlign":null,"level":3},"content":[{"type":"text","text":"Baselines"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"50df35b0-894f-40b4-b716-fc4a48b122af","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"SFCN (Simple Fully Convolutional Network):"},{"type":"text","text":" canonical brain-age architecture as baseline."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"405e077d-2827-4870-bf87-3f1095f4523b","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"SFCN + Conformal Prediction:"},{"type":"text","text":" split-conformal calibration on validation folds to provide marginal prediction intervals."}]}]}]},{"type":"heading","attrs":{"id":"2e8ba557-6a83-43c7-9b87-03b601b8fd6a","data-toc-id":"2e8ba557-6a83-43c7-9b87-03b601b8fd6a","textAlign":null,"level":3},"content":[{"type":"text","text":"Proposed"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"94dae09b-f591-457b-b72b-8a40423ae673","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"DenseNet EBM:"},{"type":"text","text":" 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"13fcd54c-9b61-49e8-a323-18433d014112","class":null,"textAlign":"justify"},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Optional invariance:"},{"type":"text","text":" gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects."}]}]}]},{"type":"heading","attrs":{"id":"35b59b31-962f-4b2e-9cfe-3d12dd86bb4f","data-toc-id":"35b59b31-962f-4b2e-9cfe-3d12dd86bb4f","textAlign":null,"level":3},"content":[{"type":"text","text":"Evaluation"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"993419b6-2573-4b4e-859e-471168f0d918","class":null,"textAlign":null},"content":[{"type":"text","text":"Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal)."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1c1dfa30-dac5-4a55-a945-962213a74cfc","class":null,"textAlign":null},"content":[{"type":"text","text":"Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal."}]}]}]},{"type":"heading","attrs":{"id":"654de9eb-8a40-45d4-86fa-469a9a0e48b5","data-toc-id":"654de9eb-8a40-45d4-86fa-469a9a0e48b5","textAlign":null,"level":3},"content":[{"type":"text","text":"Hardware & Environment"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"5900b057-d1db-4e75-ac41-0568896d5049","class":null,"textAlign":null},"content":[{"type":"text","text":"4× "},{"type":"text","marks":[{"type":"bold"}],"text":"NVIDIA RTX A5000"},{"type":"text","text":" (24 GB each; total 96 GB), "}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c431ed4c-fa5f-42c2-9fc8-903121868b20","class":null,"textAlign":null},"content":[{"type":"text","text":"Software: Python, PyTorch, MONAI/TorchIO, scikit-learn."}]}]}]},{"type":"heading","attrs":{"id":"fc27d2a2-b166-44ba-a669-80db3a8eb124","data-toc-id":"fc27d2a2-b166-44ba-a669-80db3a8eb124","textAlign":null,"level":3},"content":[{"type":"text","text":"Reproducibility"}]},{"type":"bulletList","content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"41779e3e-2b3e-4d06-be95-333bce5c31dc","class":null,"textAlign":null},"content":[{"type":"text","text":"Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance."}]}]}]},{"type":"heading","attrs":{"id":"f2aa22d6-9e44-453a-bd9a-e1bf2c18e1ba","data-toc-id":"f2aa22d6-9e44-453a-bd9a-e1bf2c18e1ba","textAlign":"justify","level":3},"content":[{"type":"text","text":"Planned Analyses"}]},{"type":"orderedList","attrs":{"start":1},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"5ed3d007-0921-4c7b-95de-60f4b944a9e1","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Prediction performance:"},{"type":"text","text":" MAE, RMSE, R²."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c96cc4c2-e8a5-422a-b426-fba4ee37cc0c","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Uncertainty calibration:"},{"type":"text","text":" expected calibration error (ECE), coverage vs. nominal quantiles."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c8034bb9-08cf-4387-be43-2ed2ece469af","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Scanner robustness:"},{"type":"text","text":" Adjusted silhouette width (ASW) and η² by site."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"163114ff-0639-4c87-8b72-bacb4b0c7186","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Interpretability:"},{"type":"text","text":" Saliency maps (GradCAM). "}]}]}]},{"type":"heading","attrs":{"id":"9b42b92b-48c6-46fc-852d-baee696d9bdd","data-toc-id":"9b42b92b-48c6-46fc-852d-baee696d9bdd","textAlign":null,"level":3},"content":[{"type":"text","text":"Ethics and Data Management"}]},{"type":"paragraph","attrs":{"id":"0fb02641-c5b7-4fd5-b316-9cd8c6fdba85","class":null,"textAlign":"justify"},"content":[{"type":"text","text":"All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0."}]},{"type":"heading","attrs":{"id":"45dfe073-8ab2-4233-856a-abe3b1dd6db8","data-toc-id":"45dfe073-8ab2-4233-856a-abe3b1dd6db8","textAlign":null,"level":2},"content":[{"type":"text","text":"4. Budget (Total = "},{"type":"text","marks":[{"type":"bold"}],"text":"$6500 USD"},{"type":"text","text":")"}]},{"type":"table","attrs":{"id":"704bbb27-e91f-4042-9421-8feed2221b90"},"content":[{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"87c7a563-37f6-4331-bc57-e31cdb4dfc39","class":null,"textAlign":null}}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"7aa71b38-890d-426b-9a60-ee2dceeb17e4","class":null,"textAlign":null}}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"71e8546b-6794-4c3e-813e-b40a44de1e57","class":null,"textAlign":null},"content":[{"type":"text","text":"CategoryDescriptionCost (USD)"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"3f180030-1f36-4a69-bc81-3584ccba9a42","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Compute resources, storage & management"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"ab669c67-03f4-47b7-94bc-0fd096555b42","class":null,"textAlign":null},"content":[{"type":"text","text":"GPU-hours for EBM training and baselines"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"09d35367-3520-496b-af12-2315eda15a8e","class":null,"textAlign":null},"content":[{"type":"text","text":"2000"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"81a78735-cc9f-4a3a-96c3-b154f7678e04","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Biobank Data Access Fees"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"5e7748d3-c43b-42e8-ac39-7c6b22a12e29","class":null,"textAlign":null},"content":[{"type":"text","text":"UKBiobank, NAKO, SHIP"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"b6b08919-0a80-4b51-88fc-82156c3b6748","class":null,"textAlign":null},"content":[{"type":"text","text":"2000"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"d66c98ce-cd1b-4d86-9f76-b13b67745064","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Dissemination"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"c3a91185-be30-4bff-b65c-5e6166072646","class":null,"textAlign":null},"content":[{"type":"text","text":"open-access fees, conference registration fee"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"4c37d669-fb8f-4cd4-8819-95ee87d084dd","class":null,"textAlign":null},"content":[{"type":"text","text":"2000"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"9d949e96-cd01-496f-bb99-f828d846c346","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Overhead & contingency"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"40257056-b729-48f7-9e54-6620cab34cc3","class":null,"textAlign":null},"content":[{"type":"text","text":"Unexpected compute or data-transfer costs"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"7b1c0969-1aee-4a0d-8ab0-294038e77cda","class":null,"textAlign":null},"content":[{"type":"text","text":"500"}]}]}]},{"type":"tableRow","content":[{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"2099f4a8-1a6e-44cb-8a50-5e501b1c630c","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"Total"}]}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"bb61f687-a3ca-4102-8786-99f328fa1990","class":null,"textAlign":null}}]},{"type":"tableCell","attrs":{"colspan":1,"rowspan":1,"colwidth":null,"style":null},"content":[{"type":"paragraph","attrs":{"id":"468eccda-4c4d-446d-86df-a69375f620d4","class":null,"textAlign":null},"content":[{"type":"text","marks":[{"type":"bold"}],"text":"6500 USD"}]}]}]}]},{"type":"heading","attrs":{"id":"a9fce6d4-a8f8-4fd5-991d-3f183e5604f4","data-toc-id":"a9fce6d4-a8f8-4fd5-991d-3f183e5604f4","textAlign":"left","level":2},"content":[{"type":"text","text":"5. References"}]},{"type":"orderedList","attrs":{"start":1},"content":[{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"f9e05b09-2502-4460-919f-eca1de0d4439","class":null,"textAlign":null},"content":[{"type":"text","text":"Cole, James H., et al. \"Brain age predicts mortality.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Molecular psychiatry"},{"type":"text","text":" 23.5 (2018): 1385-1392."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"c9a54856-2994-4460-a6dd-d4e6108152a3","class":null,"textAlign":null},"content":[{"type":"text","text":"Cole, James H., and Katja Franke. \"Predicting age using neuroimaging: innovative brain ageing biomarkers.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Trends in neurosciences"},{"type":"text","text":" 40.12 (2017): 681-690."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"46e304ea-1ebf-4f55-8393-c4d7a6277388","class":null,"textAlign":null},"content":[{"type":"text","text":"Franke, Katja, et al. \"Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Neuroimage"},{"type":"text","text":" 50.3 (2010): 883-892."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"4ebc377d-d0d5-4662-aca2-ad089e5311c3","class":null,"textAlign":null},"content":[{"type":"text","text":"Kumari, LK Soumya, and R. Sundarrajan. \"A review on brain age prediction models.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"Brain Research"},{"type":"text","text":" 1823 (2024): 148668."}]}]},{"type":"listItem","content":[{"type":"paragraph","attrs":{"id":"1a81c68f-5dfc-4100-802e-1811858f8953","class":null,"textAlign":null},"content":[{"type":"text","text":"Song, Yang, and Diederik P. Kingma. \"How to train your energy-based models.\" "},{"type":"text","marks":[{"type":"italic"}],"text":"arXiv preprint arXiv:2101.03288"},{"type":"text","text":" (2021)."}]}]}]},{"type":"paragraph","attrs":{"id":"9f6b49c8-8ea9-4fc0-b387-b5617260a1c8","class":null,"textAlign":null}}]}59:T39d8,

Proposal: Energy-Based Models for Brain Age Estimation

1. Overview

Scientific rationale & importance:
Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk^1,2,3. Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects⁴.
This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models⁵.

Facilities, approvals, and timeline:
MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months.

Data sharing commitment:
All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements.

2. Introduction

Background & rationale:
Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)¹, which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners).

Energy-Based Models (EBMs)⁵ model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods.

Research questions & hypotheses:

H1: EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors.
H2: EBM latent spaces encode less scanner-related variance than conventional embeddings.
H3: EBMs retain or surpass standard models in predictive accuracy.

Impact statement:
If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible.

3. Methods

Sample

Datasets: To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification.
Inclusion: quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data.
Exclusion: missing covariates, poor segmentation, or acquisition artefacts.

Materials and Procedures

Preprocessing

Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain.
Brain extraction and bias-field correction.
Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter.
Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible.

Models

Input: single-channel T1w volumes
Output: chronological age (regression).
Loss: smooth L1 (Huber) or MAE; auxiliary site-adversarial head

Baselines

SFCN (Simple Fully Convolutional Network): canonical brain-age architecture as baseline.
SFCN + Conformal Prediction: split-conformal calibration on validation folds to provide marginal prediction intervals.

Proposed

DenseNet EBM: 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness.
Optional invariance: gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects.

Evaluation

Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal).
Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal.

Hardware & Environment

4× NVIDIA RTX A5000 (24 GB each; total 96 GB),
Software: Python, PyTorch, MONAI/TorchIO, scikit-learn.

Reproducibility

Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance.

Planned Analyses

Prediction performance: MAE, RMSE, R².
Uncertainty calibration: expected calibration error (ECE), coverage vs. nominal quantiles.
Scanner robustness: Adjusted silhouette width (ASW) and η² by site.
Interpretability: Saliency maps (GradCAM).

Ethics and Data Management

All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0.

4. Budget (Total = $6500 USD)

		CategoryDescriptionCost (USD)
Compute resources, storage & management	GPU-hours for EBM training and baselines	2000
Biobank Data Access Fees	UKBiobank, NAKO, SHIP	2000
Dissemination	open-access fees, conference registration fee	2000
Overhead & contingency	Unexpected compute or data-transfer costs	500
Total		6500 USD

5. References

Cole, James H., et al. "Brain age predicts mortality." Molecular psychiatry 23.5 (2018): 1385-1392.
Cole, James H., and Katja Franke. "Predicting age using neuroimaging: innovative brain ageing biomarkers." Trends in neurosciences 40.12 (2017): 681-690.
Franke, Katja, et al. "Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters." Neuroimage 50.3 (2010): 883-892.
Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668.
Song, Yang, and Diederik P. Kingma. "How to train your energy-based models." arXiv preprint arXiv:2101.03288 (2021).

5a:T39d8,

Proposal: Energy-Based Models for Brain Age Estimation

1. Overview

2. Introduction

Research questions & hypotheses:

H1: EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors.
H2: EBM latent spaces encode less scanner-related variance than conventional embeddings.
H3: EBMs retain or surpass standard models in predictive accuracy.

3. Methods

Sample

Datasets: To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification.
Inclusion: quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data.
Exclusion: missing covariates, poor segmentation, or acquisition artefacts.

Materials and Procedures

Preprocessing

Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain.
Brain extraction and bias-field correction.
Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter.
Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible.

Models

Input: single-channel T1w volumes
Output: chronological age (regression).
Loss: smooth L1 (Huber) or MAE; auxiliary site-adversarial head

Baselines

SFCN (Simple Fully Convolutional Network): canonical brain-age architecture as baseline.
SFCN + Conformal Prediction: split-conformal calibration on validation folds to provide marginal prediction intervals.

Proposed

DenseNet EBM: 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness.
Optional invariance: gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects.

Evaluation

Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal).
Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal.

Hardware & Environment

4× NVIDIA RTX A5000 (24 GB each; total 96 GB),
Software: Python, PyTorch, MONAI/TorchIO, scikit-learn.

Reproducibility

Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance.

Planned Analyses

Prediction performance: MAE, RMSE, R².
Uncertainty calibration: expected calibration error (ECE), coverage vs. nominal quantiles.
Scanner robustness: Adjusted silhouette width (ASW) and η² by site.
Interpretability: Saliency maps (GradCAM).

Ethics and Data Management

4. Budget (Total = $6500 USD)

		CategoryDescriptionCost (USD)
Compute resources, storage & management	GPU-hours for EBM training and baselines	2000
Biobank Data Access Fees	UKBiobank, NAKO, SHIP	2000
Dissemination	open-access fees, conference registration fee	2000
Overhead & contingency	Unexpected compute or data-transfer costs	500
Total		6500 USD

5. References

Cole, James H., et al. "Brain age predicts mortality." Molecular psychiatry 23.5 (2018): 1385-1392.
Cole, James H., and Katja Franke. "Predicting age using neuroimaging: innovative brain ageing biomarkers." Trends in neurosciences 40.12 (2017): 681-690.
Franke, Katja, et al. "Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters." Neuroimage 50.3 (2010): 883-892.
Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668.
Song, Yang, and Diederik P. Kingma. "How to train your energy-based models." arXiv preprint arXiv:2101.03288 (2021).

5b:T1ad8,Proposal: Energy-Based Models for Brain Age Estimation 1. Overview Scientific rationale & importance: Brain age estimation from MRI is a key biomarker of neurobiological ageing and early disease risk1,2,3. Yet, most existing models, ranging from linear regression to deep CNNs, produce deterministic point estimates, lacking uncertainty quantification and sensitivity to scanner effects4. This project introduces Energy-Based Models (EBMs) as a novel probabilistic framework for brain-age regression, modelling the energy landscape over possible ages given neuroimaging features. EBMs are flexible, interpretable, and can capture multimodal, uncertainty-aware predictions, offering potential improvements over standard regression models5. Facilities, approvals, and timeline: MRI data are sourced from existing public and pay-walled repositories (AgeRisk, CamCAN, ADNI, HCP, UKBiobank, SHIP…). Analyses will be conducted on institutional HPC clusters and cloud GPUs associated with the Image Processing Laboratory at Universidad de Valladolid. The full project timeline is 9 months, with intermediate deliverables every three months. Data sharing commitment: All code and trained models will be made openly available on GitHub and ResearchHub upon project completion. No raw MRI images will be redistributed, adhering to data-use agreements. 2. Introduction Background & rationale: Brain age models provide an index of deviation between predicted and chronological age (the “brain age gap”)1, which has been linked to cognitive decline, psychiatric disorders, and accelerated neurodegeneration. Despite their clinical potential, current methods are typically black-box regressors that cannot express prediction uncertainty or handle distributional shift (e.g., different scanners). Energy-Based Models (EBMs)5 model data through an unnormalized energy function, assigning low energy to compatible image–age pairs and high energy to improbable pairs. This allows flexible conditional density modelling without restrictive Gaussian assumptions. EBMs can be trained with contrastive divergence or score matching and later used to derive calibrated conditional likelihoods. Research questions & hypotheses: H1: EBMs provide better-calibrated uncertainty estimates for brain age than deterministic regressors. H2: EBM latent spaces encode less scanner-related variance than conventional embeddings. H3: EBMs retain or surpass standard models in predictive accuracy. Impact statement: If EBMs outperform or even match standard models while offering improved generalizability and uncertainty calibration, they will redefine how continuous neuroimaging biomarkers are modelled, making individualised, uncertainty-aware brain health assessments feasible. 3. Methods Sample Datasets: To date, we have assembled over 6,000 publicly available T1-weighted MRI scans from healthy participants. For methodological advances, however, a substantially larger cohort (>20,000) spanning multiple scanners and vendors is desirable. Access to major biobanks such as NAKO, UK Biobank, or SHIP would help us achieve this scale. Furthermore, including datasets comprising individuals with specific clinical conditions could help reveal methodological limitations of these new models for diagnosis and risk stratification. Inclusion: quality-controlled scans, T1w images with available age, sex and intracranial volume (ICV) data. Exclusion: missing covariates, poor segmentation, or acquisition artefacts. Materials and Procedures Preprocessing Convert DICOM→NIfTI; reorient to RAS; intensity clipping; z-score within brain. Brain extraction and bias-field correction. Train-time augmentations: small affine, flips, Gaussian noise/blur, intensity jitter. Split strategy: site-/scanner-stratified k-fold CV; held-out test set from unseen scanners where possible. Models Input: single-channel T1w volumes Output: chronological age (regression). Loss: smooth L1 (Huber) or MAE; auxiliary site-adversarial head Baselines SFCN (Simple Fully Convolutional Network): canonical brain-age architecture as baseline. SFCN + Conformal Prediction: split-conformal calibration on validation folds to provide marginal prediction intervals. Proposed DenseNet EBM: 3D DenseNet backbone with energy-based head (EBM) for age regression; energy minimized at target age, with margin-based regularization to improve calibration/robustness. Optional invariance: gradient-reversal layer for site adversarial training and/or ComBat-GAM on penultimate features (train-only) to probe scanner effects. Evaluation Primary: MAE; secondary: RMSE, R², brainPAD bias vs. age (slope near 0 after correction), interval coverage (conformal). Robustness: performance by scanner/vendor; unseen-scanner generalization; site-classification AUC on learned features to quantify residual site signal. Hardware & Environment 4× NVIDIA RTX A5000 (24 GB each; total 96 GB), Software: Python, PyTorch, MONAI/TorchIO, scikit-learn. Reproducibility Pretrained weights, conformal calibration sets, and code to recompute metrics will be made available upon acceptance. Planned Analyses Prediction performance: MAE, RMSE, R². Uncertainty calibration: expected calibration error (ECE), coverage vs. nominal quantiles. Scanner robustness: Adjusted silhouette width (ASW) and η² by site. Interpretability: Saliency maps (GradCAM). Ethics and Data Management All datasets should be de-identified; no new participant recruitment. The study follows institutional ethics for secondary data use. Processed derivatives and scripts will be archived on a GitHub repository under CC-BY 4.0. 4. Budget (Total = $6500 USD) CategoryDescriptionCost (USD) Compute resources, storage & management GPU-hours for EBM training and baselines 2000 Biobank Data Access Fees UKBiobank, NAKO, SHIP 2000 Dissemination open-access fees, conference registration fee 2000 Overhead & contingency Unexpected compute or data-transfer costs 500 Total 6500 USD 5. References Cole, James H., et al. "Brain age predicts mortality." Molecular psychiatry 23.5 (2018): 1385-1392. Cole, James H., and Katja Franke. "Predicting age using neuroimaging: innovative brain ageing biomarkers." Trends in neurosciences 40.12 (2017): 681-690. Franke, Katja, et al. "Estimating the age of healthy subjects from T1-weighted MRI scans using kernel methods: exploring the influence of various parameters." Neuroimage 50.3 (2010): 883-892. Kumari, LK Soumya, and R. Sundarrajan. "A review on brain age prediction models." Brain Research 1823 (2024): 148668. Song, Yang, and Diederik P. Kingma. "How to train your energy-based models." arXiv preprint arXiv:2101.03288 (2021). 20:["$","$L52",null,{"children":["$","$L4",null,{"topBanner":["$","$L53",null,{"work":{"id":4501,"type":"$undefined","contentType":"preregistration","title":"Proposal: Energy-Based Models for Brain Age Estimation","slug":"proposal-energy-based-models-for-brain-age-estimation","createdDate":"2025-10-29T08:36:30.201753Z","createdByUserId":29736,"updatedDate":"2025-10-29T08:36:30.556012Z","publishedDate":"2025-10-29T08:36:30.201753Z","authors":[{"authorProfile":{"id":737014,"fullName":"Rafael Navarro González","firstName":"Rafael","lastName":"Navarro González","profileImage":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/29/blob_MqjpUP9","headline":"Brain Age, AI/ML, Magnetic Resonance Imaging ","profileUrl":"/author/737014","user":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"promotionalBalance":0,"fundingCredits":0,"balances":"$undefined","totalRsc":0,"hasCompletedOnboarding":false,"createdDate":"$undefined","moderator":false,"editorOfHubs":"$undefined","isModerator":false,"isFunder":false,"referralCode":"$undefined","authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0,"raw":29736},"description":"I work on MRI-derived brain biomarkers, with a focus on structural brain age. My interests include multimodal imaging biomarkers, innovative approaches to brain age modeling, and brain stimulation.","createdDate":"2021-05-23T20:57:10.648247Z","education":[{"id":7592,"name":"Universidad de Valladolid","year":{"label":"2026","value":"2026"},"major":"Biomedical Engineering","degree":{"label":"Doctor of Philosophy","value":"PhD"},"summary":"Biomedical Engineering PhD '26, Universidad de Valladolid","is_public":true,"university":{"id":7592,"name":"Universidad de Valladolid"}}],"twitter":"https://x.com/RanagoBrain","facebook":"$undefined","linkedin":"https://www.linkedin.com/in/navarrafa/","googleScholar":"https://scholar.google.es/citations?user=S8wiVBoAAAAJ&hl=es&oi=sra","orcidId":"https://orcid.org/0000-0002-2900-4107","isOrcidConnected":true,"isClaimed":true,"hIndex":4,"i10Index":4,"userId":"$undefined","isVerified":true,"isHubEditor":false,"editorOfHubs":[],"raw":{"id":737014,"is_public":true,"is_removed":false,"is_removed_date":null,"user":29736,"first_name":"Rafael","last_name":"Navarro González","created_date":"2021-05-23T20:57:10.648247Z","updated_date":"2026-04-01T22:34:36.021920Z","description":"I work on MRI-derived brain biomarkers, with a focus on structural brain age. My interests include multimodal imaging biomarkers, innovative approaches to brain age modeling, and brain stimulation.","h_index":4,"i10_index":4,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/29/blob_MqjpUP9","author_score":0,"university":null,"orcid_id":"https://orcid.org/0000-0002-2900-4107","openalex_ids":["https://openalex.org/A5103233331"],"education":"$20:props:children:props:topBanner:props:work:authors:0:authorProfile:education","headline":"Brain Age, AI/ML, Magnetic Resonance Imaging ","facebook":null,"twitter":"https://x.com/RanagoBrain","linkedin":"https://www.linkedin.com/in/navarrafa/","google_scholar":"https://scholar.google.es/citations?user=S8wiVBoAAAAJ&hl=es&oi=sra","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":0,"merged_with_author":null,"added_as_editor_date":null,"is_hub_editor_of":[],"is_hub_editor":false,"num_posts":1,"is_orcid_connected":true,"reputation":100,"reputation_v2":{"hub":{"id":75229,"name":"Radiology, Nuclear Medicine And Imaging","slug":"radiology-nuclear-medicine-and-imaging"},"score":1394,"percentile":0.2609444444444444,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]},"reputation_list":[{"hub":{"id":75229,"name":"Radiology, Nuclear Medicine And Imaging","slug":"radiology-nuclear-medicine-and-imaging"},"score":1394,"percentile":0.2609444444444444,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]},{"hub":{"id":75245,"name":"Psychiatry And Mental Health","slug":"psychiatry-and-mental-health"},"score":143,"percentile":0.03575,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]}],"suspended_status":{"probable_spammer":false,"is_suspended":false},"total_score":null,"wallet":null,"is_verified":true}},"isCorresponding":false,"position":"middle","raw":"$20:props:children:props:topBanner:props:work:authors:0:authorProfile:raw"}],"abstract":"$54","doi":"10.55277/researchhub.lxmm02n3","journal":"$undefined","topics":[{"id":33342,"name":"Imaging Biomarker","slug":"imaging-biomarker"},{"id":601919,"name":"Biomarkers","slug":"biomarkers"}],"formats":[],"license":"$undefined","pdfCopyrightAllowsDisplay":true,"figures":[],"versions":[],"metrics":{"votes":1,"adjustedScore":"$undefined","comments":0,"saves":0,"reviewScore":0,"reviews":0,"earned":0,"views":0},"unifiedDocumentId":7696941,"registeredReportId":null,"postType":"PREREGISTRATION","fundraise":"$undefined","note":{"id":20202,"access":"WORKSPACE","organization":{"id":3683,"memberCount":1,"userPermission":"$undefined","createdDate":"2021-10-22T21:59:24.383095Z","coverImage":"https://storage.prod.researchhub.com/organizations/cover_image/2021/10/22/org_image_29736_rafael-navarros-notebook.png","description":"","name":"Rafael Navarro's Notebook","slug":"rafael-navarros-notebook"},"createdDate":"2025-10-28T21:52:38.464306Z","updatedDate":"2025-10-29T08:07:51.937599Z","title":"Proposal: Energy-Based Models for Brain Age Estimation","isRemoved":false,"post":{"id":4501,"slug":"proposal-energy-based-models-for-brain-age-estimation","contentType":"preregistration","documentType":"PREREGISTRATION","moderationStatus":"APPROVED","fundraise":{"id":186,"amountRaised":{"usd":0,"rsc":0},"goalAmount":{"usd":6500,"rsc":100544.1651305194},"contributors":{"numContributors":0,"topContributors":[]},"createdBy":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"totalRsc":0,"hasCompletedOnboarding":false,"moderator":false,"raw":null,"isModerator":false,"isFunder":false,"authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0},"status":"OPEN","goalCurrency":"USD","startDate":"2025-10-29T08:36:30.305200Z","endDate":"2025-11-28T08:36:30.305041Z","createdDate":"","updatedDate":"","postId":"$undefined","postTitle":"$undefined","postSlug":"$undefined","postImage":null,"reviewMetrics":"$undefined","nonprofit":"$undefined","raw":{"id":186,"amount_raised":{"usd":0,"rsc":0},"goal_amount":{"usd":6500,"rsc":100544.1651305194},"contributors":{"total":0,"top":[]},"status":"OPEN","goal_currency":"USD","start_date":"2025-10-29T08:36:30.305200Z","end_date":"2025-11-28T08:36:30.305041Z"}},"grant":"$undefined","doi":"10.55277/researchhub.lxmm02n3","topics":[{"id":33342,"name":"Imaging Biomarker","slug":"$undefined","imageUrl":"$undefined","description":"$undefined","namespace":"$undefined","category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":"$undefined","raw":{"id":33342,"name":"Imaging Biomarker"}},{"id":601919,"name":"Biomarkers","slug":"$undefined","imageUrl":"$undefined","description":"$undefined","namespace":"$undefined","category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":"$undefined","raw":{"id":601919,"name":"Biomarkers"}}],"authors":[{"authorId":737014,"userId":29736,"name":"Rafael Navarro González","raw":{"id":737014,"first_name":"Rafael","last_name":"Navarro González","user":29736}}],"contacts":"$undefined","image":"https://storage.prod.researchhub.com/uploads/posts/users/29736/c0128f8c-8c21-4171-8e8e-4b92ab829c73/chatgpt-image-28-oct-2025-23_10_29.png","raw":{"id":4501,"status":"APPROVED","authors":["$20:props:children:props:topBanner:props:work:note:post:authors:0:raw"],"hubs":["$20:props:children:props:topBanner:props:work:note:post:topics:0:raw","$20:props:children:props:topBanner:props:work:note:post:topics:1:raw"],"unified_document":{"fundraise":"$20:props:children:props:topBanner:props:work:note:post:fundraise:raw","grant":null},"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/29736/c0128f8c-8c21-4171-8e8e-4b92ab829c73/chatgpt-image-28-oct-2025-23_10_29.png","document_type":"PREREGISTRATION","slug":"proposal-energy-based-models-for-brain-age-estimation","doi":"10.55277/researchhub.lxmm02n3"}},"documentType":null,"proposalId":null,"registeredReportPrefill":null,"image":null,"previewImage":null,"topics":"$undefined","authors":"$undefined","raw":{"id":20202,"access":"WORKSPACE","latest_version":{"id":314556,"src":null,"created_date":"2025-10-29T08:36:05.854033Z","plain_text":"$55","json":"$56","note":20202},"organization":{"id":3683,"member_count":1,"user_permission":null,"created_date":"2021-10-22T21:59:24.383095Z","updated_date":"2025-01-29T18:19:23.303172Z","cover_image":"https://storage.prod.researchhub.com/organizations/cover_image/2021/10/22/org_image_29736_rafael-navarros-notebook.png","description":"","name":"Rafael Navarro's Notebook","note_created":true,"org_type":"PERSONAL","slug":"rafael-navarros-notebook","user":29736},"post":"$20:props:children:props:topBanner:props:work:note:post:raw","registered_report_prefill":null,"unified_document":{"id":7695787,"is_removed":false},"created_date":"2025-10-28T21:52:38.464306Z","updated_date":"2025-10-29T08:07:51.937599Z","document_type":null,"title":"Proposal: Energy-Based Models for Brain Age Estimation","created_by":29736},"content":null,"versionId":314556,"versionDate":"2025-10-29T08:36:05.854033Z","plainText":"$57","contentJson":"$58"},"previewContent":"$59","contentUrl":"https://storage.prod.researchhub.com/uploads/post_discussion/2025/10/29/RH-POST-PREREGISTRATION-USER-29736.txt","tips":[],"peerReviews":[],"image":"https://storage.prod.researchhub.com/uploads/posts/users/29736/c0128f8c-8c21-4171-8e8e-4b92ab829c73/chatgpt-image-28-oct-2025-23_10_29.png","enrichments":[],"moderationStatus":"APPROVED","raw":{"score":1,"user_flag":null,"authors":["$20:props:children:props:topBanner:props:work:authors:0:authorProfile:raw"],"id":4501,"created_by":{"id":29736,"author_profile":{"id":737014,"is_public":true,"is_removed":false,"is_removed_date":null,"user":29736,"first_name":"Rafael","last_name":"Navarro González","created_date":"2021-05-23T20:57:10.648247Z","updated_date":"2026-04-01T22:34:36.021920Z","description":"I work on MRI-derived brain biomarkers, with a focus on structural brain age. My interests include multimodal imaging biomarkers, innovative approaches to brain age modeling, and brain stimulation.","h_index":4,"i10_index":4,"profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/29/blob_MqjpUP9","author_score":0,"university":null,"orcid_id":"https://orcid.org/0000-0002-2900-4107","openalex_ids":["https://openalex.org/A5103233331"],"education":[{"id":7592,"name":"Universidad de Valladolid","year":{"label":"2026","value":"2026"},"major":"Biomedical Engineering","degree":{"label":"Doctor of Philosophy","value":"PhD"},"summary":"Biomedical Engineering PhD '26, Universidad de Valladolid","is_public":true,"university":{"id":7592,"name":"Universidad de Valladolid"}}],"headline":"Brain Age, AI/ML, Magnetic Resonance Imaging ","facebook":null,"twitter":"https://x.com/RanagoBrain","linkedin":"https://www.linkedin.com/in/navarrafa/","google_scholar":"https://scholar.google.es/citations?user=S8wiVBoAAAAJ&hl=es&oi=sra","claimed":true,"country_code":null,"created_source":"RESEARCHHUB","last_full_fetch_from_openalex":null,"two_year_mean_citedness":0,"merged_with_author":null,"added_as_editor_date":null,"is_hub_editor_of":[],"is_hub_editor":false,"num_posts":1,"is_orcid_connected":true,"reputation":100,"reputation_v2":{"hub":{"id":75229,"name":"Radiology, Nuclear Medicine And Imaging","slug":"radiology-nuclear-medicine-and-imaging"},"score":1394,"percentile":0.2609444444444444,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]},"reputation_list":[{"hub":{"id":75229,"name":"Radiology, Nuclear Medicine And Imaging","slug":"radiology-nuclear-medicine-and-imaging"},"score":1394,"percentile":0.2609444444444444,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]},{"hub":{"id":75245,"name":"Psychiatry And Mental Health","slug":"psychiatry-and-mental-health"},"score":143,"percentile":0.03575,"bins":[[0,1000],[1000,10000],[10000,100000],[100000,1000000]]}],"suspended_status":{"probable_spammer":false,"is_suspended":false},"total_score":null,"wallet":null,"is_verified":true},"balance":null,"balances":null,"created_date":"2021-05-23T20:57:10.640887Z","has_seen_first_coin_modal":true,"has_seen_orcid_connect_modal":true,"is_funder":false,"is_staking_opted_in":true,"is_suspended":false,"probable_spammer":false,"is_verified":true,"moderator":false,"reputation":100,"updated_date":"2025-01-29T18:15:42.933711Z","upload_tutorial_complete":false,"hub_rep":null,"time_rep":null},"created_date":"2025-10-29T08:36:30.201753Z","discussion_count":0,"document_type":"PREREGISTRATION","doi":"10.55277/researchhub.lxmm02n3","editor_type":"CK_EDITOR","full_markdown":"$5a","grants":[],"hubs":[{"editor_permission_groups":[],"hub_image":null,"id":33342,"is_locked":false,"is_removed":false,"name":"Imaging Biomarker","slug":"imaging-biomarker","namespace":null,"is_used_for_rep":false},{"editor_permission_groups":[],"hub_image":null,"id":601919,"is_locked":false,"is_removed":false,"name":"Biomarkers","slug":"biomarkers","namespace":"journal","is_used_for_rep":false}],"image_url":"https://storage.prod.researchhub.com/uploads/posts/users/29736/c0128f8c-8c21-4171-8e8e-4b92ab829c73/chatgpt-image-28-oct-2025-23_10_29.png","is_latest_version":null,"is_removed":false,"is_root_version":true,"note":"$20:props:children:props:topBanner:props:work:note:raw","peer_reviews":[],"post_src":"https://storage.prod.researchhub.com/uploads/post_discussion/2025/10/29/RH-POST-PREREGISTRATION-USER-29736.txt","preview_img":null,"registered_report_id":null,"renderable_text":"$5b","reviewed_by":null,"reviewed_date":null,"slug":"proposal-energy-based-models-for-brain-age-estimation","status":"APPROVED","title":"Proposal: Energy-Based Models for Brain Age Estimation","unified_document_id":7696941,"unified_document":{"id":7696941,"documents":[{"id":4501,"title":"Proposal: Energy-Based Models for Brain Age Estimation","slug":"proposal-energy-based-models-for-brain-age-estimation"}],"created_by":{"id":29736,"author_profile":{"id":737014,"first_name":"Rafael","last_name":"Navarro González","profile_image":"https://storage.prod.researchhub.com/uploads/author_profile_images/2025/01/29/blob_MqjpUP9"}},"reviews":{"avg":0,"count":0},"is_public":true,"is_removed":false,"document_type":"PREREGISTRATION"},"version_number":1,"updated_date":"2025-10-29T08:36:30.556012Z","bounty_type":null},"aiPeerReview":null,"linkedGrant":null,"isPublic":true},"metadata":{"id":7696941,"score":1,"topics":[{"id":33342,"name":"Imaging Biomarker","slug":"imaging-biomarker","imageUrl":"$undefined","description":"$undefined","namespace":null,"category":"$undefined","discussionCount":"$undefined","paperCount":"$undefined","subscriberCount":"$undefined","isLocked":"$undefined","isRemoved":"$undefined","isUsedForRep":false,"raw":{"id":33342,"name":"Imaging Biomarker","namespace":null,"slug":"imaging-biomarker","created_date":"2023-10-17T18:55:01.293562Z","is_used_for_rep":false}}],"metrics":{"votes":1,"comments":0,"saves":0,"reviewScore":0,"conversationComments":0,"reviewComments":0,"bountyComments":0},"fundraising":{"id":186,"amountRaised":{"usd":0,"rsc":0},"goalAmount":{"usd":6500,"rsc":100544.1651305194},"contributors":{"numContributors":0,"topContributors":[]},"createdBy":{"id":0,"email":"","firstName":"","lastName":"","fullName":"Unknown User","isVerified":false,"authorProfile":"$undefined","balance":0,"lockedBalance":0,"totalRsc":0,"hasCompletedOnboarding":false,"moderator":false,"raw":null,"isModerator":false,"isFunder":false,"authProvider":"$undefined","isStakingOptedIn":false,"balanceHistory":0},"status":"OPEN","goalCurrency":"USD","startDate":"2025-10-29T08:36:30.305200Z","endDate":"2025-11-28T08:36:30.305041Z","createdDate":"","updatedDate":"","postId":"$undefined","postTitle":"$undefined","postSlug":"$undefined","postImage":null,"reviewMetrics":"$undefined","nonprofit":"$undefined","raw":{"id":186,"amount_raised":{"usd":0,"rsc":0},"goal_amount":{"usd":6500,"rsc":100544.1651305194},"contributors":{"total":0,"top":[]},"status":"OPEN","goal_currency":"USD","start_date":"2025-10-29T08:36:30.305200Z","end_date":"2025-11-28T08:36:30.305041Z"}},"bounties":[],"activeBounties":0,"closedBounties":0},"contentType":"post"}],"rightSidebar":"$L5c","children":"$L5d"}]}]