acceptodds
Under review as a conference paper at ICLR 2027

BrainNorm: Learning the Healthy Brain Representations from Structural MRI

Abstract

We introduce **BrainNorm**, a normative foundation model trained and evaluated on  66,000 T1-weighted structural MRI (T1w sMRI) scans. Trained exclusively on healthy subjects, BrainNorm learns a _Semantic Atlas Latent_ (SAL) space in which anatomically grounded parcel representations are jointly organized by anatomical identity and age. Through contrastive alignment with parcel-specific, age-conditioned healthy reference templates, SAL provides an explicit geometry of healthy anatomical variation that supports age-consistent template matching and represents _localized deviations_ directly in the latent space. Across multiple downstream cohorts, BrainNorm is evaluated on various tasks spanning age estimation, brain-age gap estimation, parcel identification, and single- & multi-disease classification. The resulting deviation patterns in SAL space enable zero-shot disease prediction from parcel-wise abnormalities and exhibit regional patterns that closely align with findings across neurodegenerative diseases reported in the clinical literature. Fine-tuning on healthy-only subjects from downstream cohorts further improves performance across various tasks. Across all classification tasks, linear probing on BrainNorm’s frozen embeddings outperforms nine baselines fine-tuned with end-to-end downstream supervision. Together, these results demonstrate that learning an anatomically structured, age-conditioned geometry of healthy brain variation yields transferable representations for detecting and characterizing departures from expected brain structure.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Normative Representation Learning, Foundation Model, T1w sMRI scans

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_58BDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes BrainNorm, a normative foundation model trained on approximately 66,000 healthy T1-weighted scans to learn a latent space for a semantic atlas.

Critical strengths

The paper is well written, and the model outperforms fully fine-tuned baselines on age estimation, parcel identification, and disease classification.

Critical weaknesses

However, several concerns remain regarding atlas construction, anatomical coverage, model validity, and figure accuracy.

Questions for authors
  • The rationale for combining the AAL3, JHU, and HO atlases to construct a unified brain atlas is unclear. Please explain why these atlases are considered complementary for achieving whole-brain coverage, whether this is an established practice, and whether the resulting atlas construction has been rigorously validated.
  • The authors should clarify why 218 anatomical parcels are sufficient to represent the entire brain volume. Please provide evidence that these parcels provide adequate anatomical coverage and resolution, and clarify whether the parcellation has been reviewed or validated by experienced neuroscientists or physicians.
  • The validity of the model’s claimed superiority is not sufficiently demonstrated. In Figure 4A, BrainNorm appears to underperform several other methods in terms of MAE, whereas the paper claims superior age-estimation performance compared with end-to-end supervised baselines across datasets. The authors should reconcile this apparent discrepancy and provide a more comprehensive experimental analysis, including clear comparisons, statistical significance, and an explanation of the evaluation protocol.
  • The legend in Figure 5B does not appear to correspond to the curves shown in the figure. In particular, the purple and green curves indicated in the legend cannot be identified in the plotted image. Please verify and correct the legend and figure.
3 comments
AuthorsOpenReview ↗

Response to Reviewer 58BD

We thank the reviewer for their time and constructive feedback, and for recognizing the paper’s writing quality and BrainNorm’s performance across the evaluated tasks. We address your concerns below and welcome all suggestions in refining the manuscript.


Q1,2: Regarding atlas combination and coverage

  • We thank the reviewer for raising this concern. Combining complementary anatomical atlases has precedent across neuroimaging applications [Halder et al., Smith et al., Dong et al., More et al.]. Following this rationale for BrainNorm, we integrated AAL3 for gray-matter regions, JHU for white-matter regions, and Harvard-Oxford for the ventricular regions as discussed in Appendix Sec A.3. Concurrently, on clinical note, the said construction was also validated by Neurosurgeon.
  • To assess the anatomical whole-brain coverage, we conducted a patch-coverage analysis across all datasets prior to the experimentation which reports:
DatasetScansAvailable Brain patchesCovered patchesCoverage (%)
UKB48,018229,045,860229,045,860100.000000
MCSA2,66212,724,36012,724,30899.999591
ADNI10,93552,155,06352,154,92899.999741
AIBL1,0845,156,0285,156,028100.000000
MIRIAD7083,342,6073,342,607100.000000
NIFD9624,652,6434,652,63899.999893
PPMI1,6638,041,3878,041,387100.000000
Overall66,032315,117,948315,117,75699.999939

Across datasets, we see nearly complete coverage of the whole-brain patches processed by BrainNorm. We have included this analysis in Appendix Table A2 of the revised manuscript.

References:

  • Halder S, et al. “Prediction of brain-computer interface aptitude from individual brain structure”, Front Hum Neurosci, 2013.
  • Smith JL et al., Eagle-449: A volumetric, whole-brain compilation of brain atlases for vestibular functional MRI research. Sci Data. 2023
  • Z. Dong et al., "Brain-JEPA: Brain dynamics foundation model with gradient positioning and spatiotemporal masking," in Adv. Neural Inf. Process. Syst., 2024
  • More, S., et al. “Brain-age prediction: A systematic comparison of machine learning workflows”, NeuroImage, 2023
AuthorsOpenReview ↗

Response (Continued)


Q3: Regarding MAE plot

  • Thanks for raising this concern. We would like to clarify the reading of Fig 4A. The apparent confusion comes from the orientation of the MAE axis in the radar plot. All radar plots in the manuscript follow one convention: a point further from the centre means better performance. To follow this convention, the MAE axis is inverted, so a model plotted nearer the outer ring has a lower MAE. We have clarified this in the figure’s caption of the revised manuscript.
  • These results are also reported in the Supplementary Table A11. For clarity, we again report these results of MAE across cohorts below:
ModelUKBMCSAADNIAIBLMIRIADNIFDPPMI
3D-ViT4.239.064.48±0.164.48_{\pm0.16}5.24±0.375.24_{\pm0.37}8.15±4.438.15_{\pm4.43}6.53±0.606.53_{\pm0.60}9.33±0.919.33_{\pm0.91}
BrainIAC6.7618.495.60±0.065.60_{\pm0.06}5.23±0.385.23_{\pm0.38}5.03±1.285.03_{\pm1.28}5.97±0.725.97_{\pm0.72}9.39±0.689.39_{\pm0.68}
BrainMVP4.154.704.46±0.134.46_{\pm0.13}5.23±0.075.23_{\pm0.07}5.24±0.045.24_{\pm0.04}6.12±0.056.12_{\pm0.05}9.34±0.069.34_{\pm0.06}
BrainHarmony4.155.054.38±0.174.38_{\pm0.17}5.10±0.385.10_{\pm0.38}5.20±1.325.20_{\pm1.32}5.97±0.715.97_{\pm0.71}9.18±0.759.18_{\pm0.75}
RadFM4.835.553.87±0.243.87_{\pm0.24}3.96±0.443.96_{\pm0.44}4.49±1.074.49_{\pm1.07}5.83±0.775.83_{\pm0.77}6.53±0.596.53_{\pm0.59}
NeuroVFM4.255.933.85±0.033.85_{\pm0.03}5.04±0.065.04_{\pm0.06}5.18±0.175.18_{\pm0.17}5.98±0.025.98_{\pm0.02}9.23±0.249.23_{\pm0.24}
Neuro-JEPA4.646.243.98±0.063.98_{\pm0.06}4.99±0.134.99_{\pm0.13}5.32±0.055.32_{\pm0.05}6.11±0.066.11_{\pm0.06}9.26±0.139.26_{\pm0.13}
MedicalNet4.168.754.38±0.094.38_{\pm0.09}4.62±0.334.62_{\pm0.33}4.74±1.364.74_{\pm1.36}6.01±0.646.01_{\pm0.64}9.05±0.849.05_{\pm0.84}
y-Aware4.175.434.14±0.134.14_{\pm0.13}4.66±0.154.66_{\pm0.15}4.31±1.064.31_{\pm1.06}5.88±0.755.88_{\pm0.75}8.54±0.888.54_{\pm0.88}
BrainNorm (No FT)4.143.943.56±0.13\mathbf{3.56_{\pm0.13}}3.50±0.26\mathbf{3.50_{\pm0.26}}4.83±1.304.83_{\pm1.30}5.44±0.425.44_{\pm0.42}7.42±0.917.42_{\pm0.91}
BrainNorm (CN-only FT)4.693.663.72±0.203.72_{\pm0.20}3.72±0.203.72_{\pm0.20}3.23±1.18\mathbf{3.23_{\pm1.18}}4.35±0.62\mathbf{4.35_{\pm0.62}}5.44±0.55\mathbf{5.44_{\pm0.55}}
  • Regarding the evaluation protocol, as mentioned in Sec 4.3, all the foundation/ pretrained models considered as baselines were finetuned end-to-end for age prediction task. Whereas for BrainNorm, it is evaluated in two ways: (i) freeze the pretrained model weights and directly perform retrieval-based age estimation (No FT), or, (ii) finetune the model with only the CN subjects of the downstream cohorts using the same pretraining objective and then freeze the weights and perform retrieval-based age estimation (CN-only FT).
  • For statistical significance check, we report delta MAE, 95% confidence interval, and p-value in Appendix Fig. A2 of the revised manuscript for all downstream cohorts across all baselines for age prediction. On ADNI, BrainNorm without fine-tuning yielded lower MAE than all baselines (p < 0.05). With fine-tuning, BrainNorm achieved lower MAE than all pretrained baselines on AIBL, MIRIAD, NIFD and PPMI (p < 0.05). Numerous comparisons also reached p < 0.001, as detailed in the figure.


Q4: Regarding few-shot plot

Thanks for raising this concern. We would like to clarify the legend and plot in Fig 5B. For all the baselines, few-shot results are plotted as curves in shades of grey. Whereas, for BrainNorm, the few shot results are plotted as scatter plot, i.e., the circles indicate Finetuned+LP, triangles indicate Pretrained+LP, and, the color of these circles and triangles differentiates the representations, i.e., green indicate BrainNorm_M1 and purple indicate BrainNorm_M2. We have also updated the legend for better readability to avoid any confusion in the updated manuscript. Thank you for this suggestion.


We thank the reviewer again for the constructive assessment. We hope that our response will address your concerns. If you have any other questions, we will be happy to discuss them with you.

Dear Reviewer 58BD,

Thank you once again for your thoughtful and constructive review. We sincerely appreciate the time and care you have taken in evaluating our work. We have updated the manuscript with your suggestions. As the discussion period is nearing its end, we would be very grateful if you could let us know if any concerns remain. We would be happy to clarify anything further.

Best Regards,

Authors

Reviewer_isz5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BrainNorm learns anatomically structured representations of healthy brain aging from large-scale structural MRI data. It encodes T1-weighted MRI scans into parcel-wise embeddings and compares them with age-conditioned healthy reference templates to estimate regional deviations. The authors evaluate these representations on several tasks, including brain-age estimation, anomaly detection, and disease classification, across multiple clinical cohorts. They also compare BrainNorm with classical normative modeling approaches and recent brain foundation models, assessing both direct transfer and performance after target-cohort adaptation.

Critical strengths
  • The proposed parcelization strategy has the merit of being easily interpretable in the context of normative modeling, which is not the case for other whole-brain approaches. The paper clearly argues that normative inference should ask where a subject deviates from the healthy population rather than merely asking whether the whole brain appears older.
  • The downstream clinical tasks are numerous and relevant, especially for monitoring the progression of neurodegenerative disorders. The evaluation across several independent cohorts is a strength and provides a more comprehensive assessment of transferability than evaluations restricted to a single dataset.
  • The pretraining data is broad, with approximately 48k UK Biobank subjects, and the evaluation covers multiple public cohorts, including ADNI, AIBL, MIRIAD, NIFD, and PPMI. This scale and diversity are valuable in neuroimaging, where limited sample sizes often make it difficult to assess the generalizability of learned representations.
  • The addition of traditional normative modeling baselines, particularly Hierarchical Bayesian Regression (HBR) using regional morphometric features, is an important improvement. It allows a more direct assessment of whether the proposed representation-learning approach provides benefits over established normative modeling methods, rather than only over deep learning baselines.
  • The comparison with recent brain foundation models, including NeuroJEPA, is also valuable. Evaluating frozen representations using linear probing and few-shot settings provides a more informative assessment of representation quality and data efficiency than relying exclusively on fully supervised fine-tuning.
  • The manuscript provides additional ablations of the text priors and a more complete description of the Parcel Encoder. These additions improve the transparency of the method and help clarify the respective roles of the text-based initialization and age-related regularization
Critical weaknesses
  • The zero-shot claim is insufficiently convincing. The proposed method relies on age-conditioned healthy templates to compute parcel-wise deviations, but the disease classification procedure also considers disease-specific anatomical priors to restrict the regions used for anomaly scoring. This introduces clinical knowledge into the prediction procedure and makes the setting different from fully unsupervised, disease-agnostic anomaly detection. More importantly, the direct-transfer performance without cohort-specific fine-tuning or disease-specific priors appears weak on some clinically relevant tasks, with performance close to chance in the reported ADNI-AD setting. The stronger results obtained after cognitively normal (CN)-only adaptation are interesting, but they demonstrate the value of target-cohort calibration rather than pure zero-shot transfer. These settings should be clearly distinguished, and the claims regarding zero-shot disease prediction should be revised accordingly.

  • The contribution of the proposed representation relative to classical normative modeling remains unclear. The addition of HBR is welcome, but the results reportedly show that HBR can outperform unadapted BrainNorm on some subject-level anomaly detection tasks. This raises an important question: what is the specific advantage of BrainNorm over conventional normative models? Is it improved detection under cohort shift, better parcel-level localization, or more transferable representations for downstream tasks? These are distinct contributions and should be evaluated and discussed separately. In particular, the results after CN-only adaptation should not obscure the performance of the unadapted model.

  • The baseline comparisons require careful interpretation. The inclusion of recent brain foundation models and few-shot evaluations strengthens the experimental section. However, differences in preprocessing, input representations, and adaptation protocols can still affect the conclusions. The manuscript should make explicit which comparisons use identical inputs and which use model-specific preprocessing, and distinguish frozen linear probing, few-shot adaptation, full fine-tuning, and CN-only calibration. These regimes measure different properties and should not be interpreted as directly interchangeable evidence of representation quality.

Questions for authors
  • How does zero-shot disease classification perform when all parcels are used to compute a disease-agnostic anomaly score, without CN-only fine-tuning or disease-specific anatomical priors? Could the authors report these results separately from prior-guided scoring and CN-only adaptation for all evaluated cohorts?

  • Given that HBR outperforms unadapted BrainNorm on some anomaly detection tasks, what is the main advantage of BrainNorm in the direct-transfer setting? Is there evidence that the learned representations improve disease detection, parcel-level localization, or robustness to cohort shift beyond what can be achieved with conventional regional morphometry and normative modeling?

8 comments
AuthorsOpenReview ↗

Response to Reviewer isz5

We thank the reviewer for the careful and constructive assessment of our work. We appreciate the detailed account of the paper’s strengths and the encouraging evaluation. We address your concerns below and welcome all suggestions in refining the manuscript.


W1, Q1: Zero-shot disease classification

  • We thank the reviewer for raising this distinction. In the prior-guided zero-shot setting, the disease-relevant parcels are predefined exclusively from the clinical neurodegeneration literature. This setting uses external anatomical knowledge without training on disease examples or fitting a disease classifier. As mentioned in Sec 4.5.a (2) and Appendix Sec B.10, we call this setting as prior-guided zero-shot disease classification, consistent with the attribute-based zero-shot task (Lampert et al, Song et al.), where unseen classes are recognized through externally defined class attributes rather than labeled class examples. Prior-guided scoring thus provides a way to incorporate established anatomical knowledge into zero-shot disease prediction.
  • The deviation score-based performance as reported in Fig 4B and Appendix table A12, without CN-only finetuning is:
CohortTaskAnomaly AUROC (%)Prior-guided Anomaly AUROC (%)BAG-based AUROC (%)
ADNIAD/CN54.78±4.0454.78_{\pm4.04}64.61±3.0064.61_{\pm3.00}82.80±1.9582.80_{\pm1.95}
ADNIsMCI/pMCI47.33±4.1147.33_{\pm4.11}55.28±3.0455.28_{\pm3.04}73.59±5.0173.59_{\pm5.01}
AIBLAD/CN49.55±5.9849.55_{\pm5.98}61.24±6.1061.24_{\pm6.10}87.14±1.1987.14_{\pm1.19}
MIRIADAD/CN71.00±8.7471.00_{\pm8.74}81.89±7.2381.89_{\pm7.23}98.52±2.4798.52_{\pm2.47}
NIFDFTD/CN72.69±5.9572.69_{\pm5.95}75.46±6.6575.46_{\pm6.65}82.38±4.1382.38_{\pm4.13}
PPMIPD/CN49.60±2.4849.60_{\pm2.48}51.99±4.5351.99_{\pm4.53}55.03±4.5755.03_{\pm4.57}
ADNI+NIFD+PPMIAD/FTD/PD—61.31±1.5061.31_{\pm1.50}—
ADNI+NIFD+PPMIAD/FTD/PD/CN—63.30±1.0163.30_{\pm1.01}—
  • The anomaly-only score better separates disease from controls on MIRIAD and NIFD and is near chance on PPMI. The brain-age-gap score, which uses no priors or adaptation, gives 73.6–98.5 AUROC on the AD, MCI and FTD tasks. Furthermore, the prior-guided scoring, improves AUROC over anomaly-only on ADNI AD/CN, ADNI sMCI/pMCI, AIBL and MIRIAD (p < 0.001) and on NIFD (p < 0.05). Disease-relevant signal is therefore present under direct transfer, and the similarity-based top-K aggregation recovers only part of it. Cohort shift is a second likely factor, since CN-only calibration improves the same score.
  • PPMI’s performance is comparatively low in every setting. The fully fine-tuned baselines also reach only 55–61 AUC on this task (Table A13). Integrating multiple modalities (both imaging and non-imaging biomarkers) could be a key direction for future work to improve PD classification performance (Mitchell at al., Lehericy et al., Vasquez-Correa et al.).
  • CN-only fine-tuning uses training-set controls to adapt the normative representation without disease examples. Improvements in this setting demonstrate that adaptation of the healthy reference space can improve disease discrimination using the anomaly scores, without training on disease examples. As suggested, we report direct-transfer zero-shot prediction and prediction after CN-only adaptation separately in Fig. 4B and distinguish these settings in Sec. 4.5 of the revised manuscript.

References:

  • Lampert, et al, “Attribute-Based Classification for Zero-Shot Visual Object Categorization”, PAMI 2014
  • Song, et al, “Selective Zero-Shot Classification with Augmented Attributes”, ECCV 2018
  • Mitchell T et al., “Emerging Neuroimaging Biomarkers Across Disease Stage in Parkinson Disease: A Review”, JAMA Neurol, 2021
  • Lehericy S, et al., “The role of high-field magnetic resonance imaging in parkinsonian disorders: Pushing the boundaries forward”, Mov Disord, 2017
  • Vasquez-Correa et al., “Multimodal Assessment of Parkinson's Disease: A Deep Learning Approach”, IEEE J Biomed Health Inform, 2019
AuthorsOpenReview ↗

Response (continued)

W2, Q2: Comparison with HBR baseline

  • We thank the reviewer for raising this concern. Appendix Sec. B.11 distinguishes two complementary evaluations: subject-level anomaly scoring and linear probing of regional deviation features (without CN-only finetuning).
  • HBR achieves higher anomaly-score AUROC on several direct-transfer tasks. Following CN-only adaptation, BrainNorm achieves higher AUROC than adapted HBR on five of the six tasks, demonstrating that its learned healthy reference space can be effectively adapted using downstream cohort controls.
  • BrainNorm’s main demonstrated advantage lies in its parcel-wise deviation representations, even without CN-only finetuning. Appendix Table A10 compares linear probing of BrainNorm’s frozen deviation representations against linear probing of HBR’s concatenated regional morphometric deviation scores. BrainNorm achieves higher AUROC and balanced accuracy across all eight evaluated classification tasks as shown below:
TaskHBR BAccHBR AUROCBrainNorm BAccBrainNorm AUROCΔ BAccΔ AUROC
AD/CN [ADNI]82.2990.2586.8594.27+4.56+4.02
AD/CN [AIBL]83.5391.0685.9193.39+2.38+2.33
AD/CN [MIRIAD]88.4794.5792.7298.26+4.25+3.69
sMCI/pMCI [ADNI]67.4774.5571.9680.75+4.49+6.20
PD/CN [PPMI]55.8760.4259.0866.12+3.21+5.70
FTD/CN [NIFD]70.1475.4281.8289.15+11.68+13.73
AD/PD/FTD80.8592.8683.9897.01+3.13+4.15
AD/PD/FTD/CN68.6986.6470.3291.91+1.63+5.27

These results indicate that BrainNorm’s learned deviation representation retains disease-relevant information beyond that available in the evaluated HBR-derived features. The performance of BrainNorm’s frozen parcel-deviation representations further indicates that disease-relevant deviation structure is already encoded, although simple global anomaly aggregation may not fully capture this information. Together, these evaluations distinguish BrainNorm’s contribution over HBR in terms of better deviation representations.

AuthorsOpenReview ↗

Response (continued)

W3: Evaluation Protocol

We thank the reviewer for highlighting the importance of distinguishing these evaluation settings. Our protocol follows BrainNorm’s normative objective: the encoder learns healthy variation through UKB pretraining and, optionally, adaptation using downstream training-set controls. We evaluate whether these representations support downstream prediction while keeping the encoder frozen. End-to-end supervised baselines provide relevant task-adapted comparisons because their encoders and prediction heads can both learn from downstream training labels, including CN and disease examples for classification.

  • To control differences arising from data selection and upstream preprocessing, all baselines are evaluated using the same preprocessed MRI scans and subject-level splits. Architecture-specific input transformations, such as resizing or cropping, are retained and documented in Appendix Sec. B.4. This protocol provides a common preprocessing basis while accommodating each model’s input requirements. Baseline encoders are subsequently optimized end-to-end on these inputs, allowing their representations to adapt to the downstream data.
  • Encoder adaptations: As described in Sec. 4.3, BrainNorm is evaluated with either UKB-pretrained weights (No FT) or weights adapted using downstream training-set CN subjects and the pretraining objective (CN-only FT). After pretraining or CN-only adaptation, the encoder and template mapper are frozen for downstream evaluation. All other baselines are fine-tuned end-to-end
    • For Disease classification: Baselines update their encoders and classification heads using diagnostic supervision. BrainNorm uses diagnostic labels only to train a linear probe on frozen parcel embeddings or parcel-deviation representations.
    • For Few-shot classification: Both approaches use the same k labeled training samples per class. Baselines update the encoder and classification head, whereas BrainNorm updates only the linear probe.
    • For Age prediction: Baselines are finetuned with end-to-end supervised age prediction, whereas BrainNorm estimates age through template retrieval without fitting an additional age-regression head.

These comparisons assess the downstream utility of BrainNorm’s frozen normative representations relative to task-adapted models.

Reviewer_fBFdOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

BrainNorm is a normative model trained on 38k healthy subjects from the UK Biobank. Each scan is encoded as a set of parcel embeddings (218 in total). Each parcel is defined based on an atlas. These embeddings, obtained from a patch and its MNI coordinates, are aligned to learnable templates that are age-aware and parcel-specific. Since brain age may not match patient chronological age, soft predictions are performed around an age window.

The learned space is tested on:

  • age estimation
  • parcel retrieval
  • zero-shot anomaly and brain age gap scoring
  • linear probing for neurodegenerative disorders (Alzheimer's disease, mild cognitive impairment, progressors, frontotemporal dementia and Parkinson's disease)
Critical strengths
  • The paper is pleasant to read.
Critical weaknesses

Concerning the technical motivation:

  1. Why should the parcel encoder learn the parcel identity? I don't understand the motivation for a representation that encodes which parcel it comes from. The parcel is known from the registration to the atlas. The model is in fact given a similar information as input: tokens are grouped by atlas overlap and carry MNI positional embeddings. The cross-parcel negatives then train the model to recover information it already has. This has three consequences on the evaluation: (i) Parcel retrieval is therefore not informative. The Categorical ablation reaches 99.99% Acc@1 on UKB. (ii) There is no downstream usage of the parcel-specific aspect of the embeddings (beyond retrieval). The anomaly score, the M2 features and the BAG all compare a parcel only with templates of the same given parcel. (iii) The natural baseline is regional brain age: a per-parcel age regressor on the same encoder. This ablation is missing.

Concerning the comparison with other baselines:

  1. The age-prediction evaluation ignores the brain-age literature. Brain-age prediction is an old problem, and none of the dedicated methods is compared, although several are cited. There is a benchmark available that could help.
  • On UKB, BrainNorm reaches an MAE of 4.14 years (Table A10). Published results on the same data are much lower:
    • SFCN: 2.14 years (Peng et al. 2021);
    • patch-based and whole-brain CNNs on the same MNI-registered UKB images: 2.1-2.6 years (Bintsi et al.,  2020);
    • a 3D CNN on non-linearly warped UKB images: 2.7-2.9 years (Dinsdale et al. 2021).
  • On PPMI controls, an SVR ensemble trained on independent data reports 4.3 years (Eickhoff et al.,  2021). BrainNorm reports 5.4-7.4 years.
  1. Several foundation-model baselines appear to be unadapted for the task or potentially not well fine-tuned. For example, BrainIAC (full FT) has r=0.01r = 0.01 on ADNI, and on PPMI, six of nine baselines have r≤0.26r \leq 0.26.

  2. The experimental setups are also not comparable. Baselines are fine-tuned on a few hundred scans per fold, while BrainNorm uses 38k labelled UKB scans.

References: More et al., NeuroImage 2023 https://doi.org/10.1016/j.neuroimage.2023.119947 Peng et al., MedIA 2021 https://doi.org/10.1016/j.ebiom.2021.103600 Bintsi et al., MLCN 2020 https://arxiv.org/pdf/2008.12965 Dinsdale et al., NeuroImage 2021 https://doi.org/10.1016/j.neuroimage.2020.117401 Eickhoff et al., Brain Commun. 2021 https://doi.org/10.1093/braincomms/fcab191

Questions for authors
  1. Why does parcel retrieval drop to 67-74% Acc@1 on external cohorts (Table 1, no FT) when the atlas and positions are identical? Does this mean the embeddings are dominated by acquisition-related appearance?
  2. What happens without cross-parcel negatives, or with per-parcel age-regression heads in place of the template bank?
  3. Is any age-bias correction applied to the BAG?
  4. Can the baselines be trained on the same UKB data before transfer, as BrainNorm is?

For future work:

  1. Comparison needed with methods designed for well-established methods (see above)
11 comments
AuthorsOpenReview ↗

Response to Reviewer fBFd

We thank the reviewer for their time and constructive feedback. We are happy and encouraged that the reviewer felt the paper pleasant to read. We address your concerns below and welcome all suggestions in refining the manuscript.


We thank the reviewer for examining the motivation behind BrainNorm’s parcel-conditioned objective. BrainNorm aims to learn healthy reference representations indexed by anatomical region and age. The Parcel Encoder maps each atlas-defined set of image patches to a regional embedding, while the Template Mapper learns corresponding reference embeddings from age- and parcel-conditioned text features. The training objective is conditioned on age and parcel identity during pretraining. The resultant SAL space provides a set of templates which can act as expected references of healthy aging per parcel.


W1 and Q2: Role of parcel discrimination and cross-parcel negatives

  • The atlas specifies which image patches belong to each parcel, and the positional embeddings provide spatial information. These inputs establish anatomical correspondence. The training objective uses this information to organize the SAL space where image embeddings are aligned with parcel-specific healthy aging templates. Their correspondence is expressed through their similarities. Cross-parcel negatives discourage a parcel embedding from matching references for other anatomical regions, while within-parcel supervision shapes its similarity to references across age.
CohortObjectiveAge MAE ↓Pearson r ↑Parcel→Template Acc@1 (%) ↑Parcel→Template Acc@5 (%) ↑Template→Parcel Acc@1 (%) ↑Template→Parcel Acc@5 (%) ↑
UKBProposed loss4.140.8597.1799.8090.8799.02
UKBWithout cross-parcel negatives4.310.810.773.520.763.22
MCSAProposed loss3.940.8875.0289.5467.3385.19
MCSAWithout cross-parcel negatives4.750.860.793.500.622.80
  • We therefore evaluated the reviewer’s proposed ablation by excluding cross-parcel pairs from SigLIP while retaining within-parcel age supervision and AgeSoftCE. Removing cross-parcel negatives preserves good age-predictive information, while parcel retrieval falls below 1% Acc@1 in both directions. Thus, atlas-defined inputs and positional information alone do not ensure discriminative image-template alignment across parcels under the age-only pairing objective. The proposed loss achieves higher parcel alignment alongside age-sensitive matching. We have included this ablation in Appendix Table A3 and Sec B.5.1 of the revised manuscript.
  • The categorical ablation in Appendix Table A4 provides a complementary result. It achieves 99.99% UKB Text→Parcel Acc@1, but its age MAE is 10.05 years. On external MCSA, categorical initialization achieves 94.99% Text→Parcel Acc@1 but an MAE of 17.13 years. These results demonstrate that parcel identification alone is insufficient to obtain useful age-conditioned healthy references. The two evaluations test complementary properties of SAL.
  • Anatomical correspondence is used when comparing each image parcel with its corresponding parcel age-matched healthy reference. The resulting comparisons in SAL space provide regional anomaly scores, support disease-prior selection of anatomical regions, and produce parcel-wise deviation features for supervised disease classification. Each deviation remains linked to a named anatomical region.

Q1: Parcel retrieval on external cohorts

The atlas and positional embeddings specify anatomical correspondence across cohorts, while retrieval evaluates the learned alignment between image representations and parcel-specific, age-conditioned templates in SAL. Consequently, retrieval depends on both anatomical identity and alignment with the learned healthy aging references. Cohort differences in acquisition and demographic distributions can affect this alignment even when the spatial indexing is shared. The retrieval decline alone therefore cannot imply that acquisition-related appearance dominates the embeddings. This motivates the CN-only adaptation described in Sec. 4.3. Using the same pretraining objective, BrainNorm adapts its parcel representations and healthy reference templates to the CN subjects of the target cohort. The resulting improvements in parcel retrieval and several downstream evaluations across cohorts demonstrate the benefit of adapting the normative reference space to the target population using only CN subjects.

AuthorsOpenReview ↗

Response (continued)

W1(iii) and Q2: Regional age-regression comparison

Following the reviewer’s suggestion, we fitted parcel-wise ridge regressors on frozen BrainNorm embeddings using UKB training data and aggregated their predictions into a scan-level age prediction, and tested on UKB test and external MCSA test data.

CohortMethodAge MAE ↓Pearson r ↑
UKBTemplate retrieval4.140.85
UKBRegional ridge regression2.560.91
MCSATemplate retrieval3.940.88
MCSARegional ridge regression5.600.88

These results should be interpreted in the context of BrainNorm’s objective of learning a structured normative reference space that supports multiple downstream uses. Template retrieval estimates age directly from the learned parcel- and age-indexed references, without fitting an additional age-prediction head. Ridge regression introduces an explicitly supervised age-regression probe on the same frozen parcel embeddings. Its lower UKB MAE demonstrates that BrainNorm’s representations retain age information that can be extracted more accurately through a fitted regression head. Template retrieval achieves lower MAE on external MCSA, demonstrating its better generalizability. The regression experiment therefore evaluates an alternative age decoder, while BrainNorm is designed to estimate age directly from its learned normative references without fitting an additional head.


W2,Q5: Brain-age prediction baselines

  • We thank the reviewer for raising this concern. Our baseline selection reflects BrainNorm’s scope as a foundation model for 3D brain MRI whose generalization capabilities are evaluated across multiple downstream tasks. We therefore prioritized existing self-supervised foundation models and large-scale pretrained models trained on 3D brain MRI scans and designed to provide transferable representations. The methods cited by the reviewer were developed specifically for supervised age prediction task. The published MAEs were obtained under different training and evaluation protocols. Peng et al. used 12,949 UKB training subjects and 1,036 test subjects, while Bintsi et al. used 13,750 scans with an 5% test split, corresponding to ~688 test scans. Input representations also differ: Dinsdale et al. used 20 sampled middle slices, while the SVR ensemble used CAT-derived morphometric features. Direct numerical comparisons, therefore, require accounting for differences in data sizes, splits and input representations.
  • To provide a controlled assessment, we trained SFCN and a patch-based CNN using our UKB training split and evaluated them on the same 9,604 held-out UKB scans and external MCSA cohort.
MethodUKB-MAEUKB-Pearson rMCSA-MAEMCSA-Pearson r
SFCN3.670.826.590.81
patch-based CNN2.690.864.370.85
BrainNorm (supervised age-regression probe)2.560.915.600.88
BrainNorm (template-retrieval)4.140.853.940.88
  • Under this common evaluation protocol, BrainNorm’s supervised age-regression probe achieves the lowest UKB MAE, while direct template retrieval achieves the lowest external MCSA MAE among the compared methods. These results demonstrate the age information retained in BrainNorm’s frozen embeddings and its utility for age estimation without fitting an additional regression head. We have included these baselines in Appendix Sec B.8 of the revised manuscript.
  • On PPMI, supervised probing achieves an MAE of 4.66 years with the UKB-pretrained encoder and 4.62 years after CN-only encoder fine-tuning, with Pearson r = 0.86 in both settings. Both improve upon the corresponding template-retrieval results.
  • These additional experiments demonstrate competitive performance against dedicated age predictors and support BrainNorm’s broader objective of learning reusable normative representations.
AuthorsOpenReview ↗

Response (continued)

W3, W4, Q4: Regarding baseline finetuning

  • We clarify that all the age baselines were adapted by optimizing their encoders and age-prediction heads end-to-end using chronological-age supervision.
    • On UKB, eight of the nine baselines achieve Pearson correlations between 0.77 and 0.83. On ADNI, eight achieve correlations between 0.62 and 0.74. Several UKB-trained models also transfer effectively to external MCSA, including BrainMVP, BrainHarmony, RadFM and Neuro-JEPA, with correlations of 0.85–0.88. These results demonstrate successful age learning and transfer for multiple baselines under the reported protocol.
    • On PPMI, the low correlations of several baselines coexist with stronger results from RadFM (r = 0.76) and NeuroVFM (r = 0.48), showing variation across models within the same cohort. BrainIAC shows consistently weak age-prediction performance across our evaluated cohorts. Its low ADNI correlation therefore follows a broader pattern in our experiments and is not specific to our fine-tuning. For PPMI, age fine-tuning uses only 303 CN scans, with approximately 242 available for training and validation and 61 for testing per fold. This limited target-cohort training data may contribute to the weaker performance of several baselines.
    • BrainNorm brings age-structured representations and parcel-specific, age-indexed references learned from UKB, enabling direct age estimation before target-cohort adaptation. Its PPMI correlations of 0.83 without fine-tuning and 0.84 after target CN-only fine-tuning support the utility of this normative pretraining in a cohort with limited healthy training data.
  • As mentioned previously, we chose baselines including foundation models/ pretrained models that were pretrained on large-scale cohorts and tested for downstream task generalizability through finetuning. We clarify that the few hundred scans per downstream fold constitute the baselines’ target-cohort adaptation data using the released pretrained checkpoint. Before this adaptation, the baseline encoders had already undergone large-scale pretraining. As detailed in Appendix Sec. B.4, BrainIAC was trained and validated on 48,965 multiparametric brain MRIs, BrainHarmony was pretrained on 64,594 T1w brain scans, and NeuroJEPA on 1,551,862 multiparametric brain MRI scans. We also included broader medical foundation models, including NeuroVFM, pretrained on 5.24 million MRI and CT volumes, and RadFM, trained on 13 million 2D images and 615,000 3D scans. Other baselines like y-aware were pretrained on 10410^4 MRI scans, BrainMVP on 16,022 brain MRI scans, and MedicalNet on 23 medical datasets.
  • Regarding training on UKB before transfer, Fig 4A and Appendix Table A11 already includes UKB-supervised age training for the baselines followed by direct evaluation on external MCSA. BrainNorm outperforms these baselines on the external MCSA dataset.


Q3: Bias correction

No age-bias correction was applied to the BAG. As defined in Sec. 4.5, BAG is calculated directly as the estimated age minus the subject’s chronological age. In the zero-shot direct-transfer setting, the UKB-trained checkpoint is utilized unchanged, without target-cohort calibration, for BAG-based anomaly scoring. Following the cautions discussed by de Lange et al. and Butler et al., that corrections of predicted age may inflate apparent model accuracy, yielding high correlations and low errors even for weak models. Thus, we used the raw BAG as anomaly score.

References:

  • de Lange AG et al. “Mind the gap: Performance metric evaluation in brain-age prediction”, Hum Brain Mapp. 2022
  • Butler ER, et al. “Pitfalls in brain age analyses”, Hum Brain Mapp, 2021


We thank the reviewer again for the constructive assessment. We hope that our response will address your concerns. If you have any other questions, we will be happy to discuss them with you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the discussion

We sincerely thank all reviewers for their time, careful assessment, and constructive feedback. Their comments have helped us clarify BrainNorm’s normative foundation-model objective, strengthen the experimental evidence, and improve the manuscript’s presentation.

We appreciate the reviewers’ recognition of the paper’s strengths:

  • Anatomical interpretability: “The proposed parcelization strategy has the merit of being easily interpretable in the context of normative modeling.”
  • Normative motivation: The paper “clearly argues that normative inference should ask where a subject deviates from the healthy population rather than merely asking whether the whole brain appears older.” ,“a valuable contribution for normative modeling.”
  • Clinical relevance and transferability: “The downstream clinical tasks are numerous and relevant,” and evaluation across independent cohorts “provides a more comprehensive assessment of transferability than evaluations restricted to a single dataset.”
  • Scale and diversity: “This scale and diversity are valuable in neuroimaging, where limited sample sizes often make it difficult to assess the generalizability of learned representations.”
  • Representation evaluation: “Evaluating frozen representations using linear probing and few-shot settings provides a more informative assessment of representation quality and data efficiency than relying exclusively on fully supervised fine-tuning.”
  • Transparency: The additional ablations and methodological details “improve the transparency of the method and help clarify the respective roles of the text-based initialization and age-related regularization.”
  • Performance across tasks: “the model outperforms fully fine-tuned baselines on age estimation, parcel identification, and disease classification.”
  • Readability: “The paper is pleasant to read.” and “The paper is well written”

We believe our responses, additional experiments, and manuscript revisions address the major concerns raised. The principal additions and clarifications are:

  • Anatomical learning objective: We added cross-parcel-negative ablations for retrieval, age estimation, and downstream disease discrimination. These experiments support the complementary roles of anatomical specificity and age continuity in the learned healthy reference space.
  • Age estimation and baseline scope: We clarified that the main comparisons assess transferable representations from large-scale pretrained models. We additionally evaluated regional age-regression probes and dedicated supervised age predictors under a common protocol. These experiments distinguish the age information available in BrainNorm’s embeddings from its ability to estimate age directly through template retrieval, without fitting an additional age-regression head.
  • Zero-shot settings: We distinguished anomaly-only scoring, disease-prior-guided scoring, BAG-based scoring, and CN-only adaptation. We added results using all 218 parcels and qualified the abstract and contribution statements to identify prior-guided zero-shot classification.
  • Comparison with conventional normative modeling: We distinguished subject-level anomaly scoring from supervised probing of regional deviation representations. BrainNorm’s frozen deviation representations achieved higher balanced accuracy and AUROC than the evaluated HBR-derived features across all eight classification tasks.
  • Evaluation transparency: We clarified preprocessing, subject-level splits, encoder adaptation, downstream supervision, and BAG calculation with supporting explanations.
  • Anatomical coverage: We clarified that BrainNorm processes parcels from established source atlases independently, retaining their anatomical definitions. We provided quantitative evidence of near-complete foreground patch coverage, further details of the expert assessment, and precedents for selecting complementary parcels from multiple atlases.


We present BrainNorm, a normative foundation model whose central contribution is an anatomically structured healthy latent space by learning parcel-specific healthy aging templates as healthy references aligned with parcel-specific representations from T1w scans. Its objective is to support multiple downstream tasks through a shared normative representation. The resulting framework offers direct retrieval-based age estimation, brain-age gap estimation, parcel-wise anomaly scoring enabling prior-guided zero-shot disease classification, disease classification through linear probing frozen representations, and adaptation using healthy target-cohort controls. Together, these capabilities connect healthy-reference learning with regional assessment and data-efficient downstream prediction.


We sincerely hope that these revisions and clarifications address the reviewers’ concerns, and we would be grateful for their favorable consideration of our work. We thank all reviewers again for their thoughtful engagement and constructive feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.