Comparative Study of Anatomical and Learned Features in AI Models for Structural Brain MRI
Boyang Yu, Miquel Lopez Escoriza, Long Chen, Arjun V. Masurkar, Narges Razavian, Carlos Fernandez-Granda
cs.CV · cs.AI
Quick verdict
Claims anatomical linear features match deep models on 7 MRI tasks; convincing, but ASP wins only on age estimation and equivalence rests on non-significance.
Summary
The paper systematically compares three feature paradigms for 3D T1-weighted brain MRI modeling across seven clinical tasks (Alzheimer's, Parkinson's and epilepsy diagnostics, tumor IDH mutation detection, brain-age estimation, tumor and stroke-lesion segmentation), using 18 public datasets totaling over 96,000 scans from ~80,000 participants. The paradigms are: (1) logistic regression on 178 FreeSurfer anatomical features (volumes, cortical areas, thickness); (2) supervised CNNs (ResNet-18/U-Net); (3) 3D ViT-Base models pretrained with masked autoencoding on 69,878 scans and finetuned via LoRA or linear probing. The central finding is that the linear anatomical model statistically matches the deep models on nearly all diagnostic comparisons — ASP ViT topped macro-AUC on an internal ADNI test and six external AD cohorts, but reached significance over CNN/anatomical baselines only on OASIS-2 — while U-Net stayed best on BraTS segmentation. On brain-age estimation with 62,890 training subjects, the proposed Anatomy Segmentation Pretraining (segmentation as auxiliary MAE-pretraining task via a prompt-based 39-region decoder) beats MAE-only pretraining (MAE 2.54 vs 2.89 years internal UKBB; 4.19 vs 5.20 on NACC); linear regression trails the ViTs by 1-2 years yet clearly beats the CNN. Limitations: equivalence claims rest on non-significant differences without equivalence tests or multiplicity correction; ASP's advantage is confined to age estimation; the internal age test set doubles as pretraining monitoring data; several cohorts are scan-heavy but subject-poor (MIRIAD 69 subjects). Code is public.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A linear logistic-regression model on 178 FreeSurfer anatomical features matches the diagnostic performance of supervised CNNs and MAE-pretrained ViT foundation models across seven clinical tasks on 18 public datasets (~80,000 participants).
- CNNs and pretrained ViTs implicitly capture anatomical information, with MAE embeddings becoming increasingly predictive of brain-region volumes during pretraining (Fig 4b), bypassing explicit feature extraction.
- Anatomy Segmentation Pretraining (ASP) — segmentation of 39 merged FreeSurfer regions as an auxiliary pretraining task with a prompt-based, skip-connection-free decoder — outperforms MAE-only pretraining in brain-age estimation (internal MAE 2.54 vs 2.89 years; NACC 4.19 vs 5.20), with each training component contributing (Table C10).
- Diagnostic performance on existing public structural MRI cohorts has reached a saturation point, and further progress requires substantially larger cohorts.
Strengths
Unmatched evaluation breadth with contamination awareness
18 public datasets (>96,000 scans, ~80,000 participants) across 7 tasks; seven external AD cohorts (NACC 4,926 down to MIRIAD n=69); participant-level splits (Fig 1); BrainIAC excluded from the AD task because its pretraining cohort overlaps the external test sets.
Unusually candid statistical reporting
Bootstrap 95% CIs (1000 resamples) with significance stated per comparison, often against the authors' own method: ASP significant over CNN/anatomical only on OASIS-2; U-Net remains best on BraTS; Pretrained ViT's Parkinson win significant only vs BrainIAC.
Well-motivated novel method with real ablations
Exemplary baseline diligence in Appendix B
Practical, costed takeaway
Weaknesses
Equivalence asserted from non-significance
ASP superiority is a single-task result
Internal age test set doubles as pretraining monitoring data
Thin external baselines in places
Cohort quirks and QC-by-failure
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks whether hand-crafted anatomical features still matter in the foundation-model era for structural brain MRI. Using 18 public datasets (>96,000 scans, ~80,000 participants), it compares three feature paradigms across seven tasks: Alzheimer's/Parkinson's/epilepsy diagnosis, tumor IDH mutation detection, brain-age estimation, and BraTS/ATLAS segmentation. The paradigms are logistic regression on 178 FreeSurfer v7.4.1 features ('Anatomical'), supervised CNNs (ResNet-18; U-Net for segmentation), and 3D ViT-Base pretrained by masked autoencoding on 69,878 scans (UKBB/HCP/CamCAN/IXI) then adapted via linear probing or LoRA. A fourth model, ASP ViT, adds segmentation of 39 merged FreeSurfer regions as an auxiliary pretraining task. Four claims are advanced: (1) linear anatomical features match sophisticated deep models diagnostically; (2) deep models implicitly encode anatomy during pretraining; (3) ASP amplifies this and wins age estimation; (4) public-cohort diagnostic performance is saturated. Claims 1-2 are well supported; claim 3 holds only for age estimation; claim 4 is honestly framed as conjecture.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Do deep learning models beat simple anatomical features for brain MRI diagnosis?
Not in this study. Across 18 public datasets and ~80,000 participants, a logistic regression on 178 FreeSurfer features (regional volumes, cortical surface areas, thickness) statistically matched supervised CNNs and pretrained vision transformers on nearly all diagnostic tasks; the only significant advantage over those baselines appeared on one cohort (OASIS-2).
What is Anatomy Segmentation Pretraining (ASP)?
A pretraining scheme for 3D vision transformers that adds brain-region segmentation (39 merged FreeSurfer regions) as an auxiliary task to masked-autoencoding pretraining, using a lightweight prompt-based decoder without skip connections so anatomical knowledge lands in the encoder embedding. It improves brain-age estimation: mean absolute error 2.54 vs 2.89 years internally (UK Biobank) and 4.19 vs 5.20 years on the external NACC cohort.
How accurate is the brain-age model in this paper?
The best model (ASP ViT) reaches a mean absolute error of 2.54 years on an internal UK Biobank test set (n=6,950) and 4.19 years on external ADNI/NACC healthy cohorts. Errors grow for subjects over 80, who are underrepresented in training, and dementia patients receive consistently inflated age estimates.
Is the code and data available?
The training and evaluation code is public on GitHub (MIQUEL-LOPEZ-ESCORIZA/ai4brainmri). The 18 MRI datasets are publicly documented but most (UK Biobank, ADNI, NACC, AIBL, PPMI, etc.) require an application to the respective consortium; full replication also needs substantial GPU time (~10 days on 2 A100s for MAE pretraining, ~1 month on one RTX8000 for ASP).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Exact per-model AUC/Dice values for the diagnostic and segmentation tasks are presented only graphically (Figs 2-3); the text gives narrative rankings and selected MAE numbers, so per-task effect sizes beyond those quoted could not be audited.
- LoRA rank and per-task learning rates are not stated anywhere in the provided text (Appendix C describes the adaptation strategies but omits these values).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.