Geometric-to-Semantic Spherical Transfer Learning for Cortical Sulci Labeling
Saeb Tounsi, Joël Chavas, Pietro Gori, Vincent Frouin, Denis Rivière, Jean-François Mangin
cs.LG · cs.CV
Quick verdict
Promising transfer learning for sulci labeling: SSL on 30k UKB + soft-init TPI lifts Dice to 0.77 vs 0.72 from-scratch; small N=22 test, modest margin.
Summary
This paper addresses exhaustive cortical sulcus labeling (50 surface-consistent sulci) with only 62 expert-annotated subjects. The authors pre-train a VGG-style spherical convolutional encoder on ~30,000 UK Biobank subjects using a Barlow Twins objective modified for localization — no global average pooling, plus rotations, vertex dropout, and patch masking — relying only on curvature and depth. The representations are probed linearly, including on rare topological traits (central sulcus interruption; paracingulate sulcus presence). For fine-tuning, a soft-initialized 1x1 convolution (Topological Prior Injector) gradually admits the extracted sulcal-fundi channel into the pre-trained two-channel backbone, with differential learning rates. On a 22-subject held-out test, the full method reaches mean Dice 0.77±0.03 versus 0.72±0.05 for a from-scratch spherical U-Net, 0.70 for DINOv3-based dense prediction and 0.69 for a 2D ResNet-18. A frozen locally-pretrained encoder alone matches the from-scratch baseline (0.71±0.04), while frozen age/sex-supervised pre-training degrades to 0.58. Gains concentrate on small, variable sulci (up to −14.8% local error); with 5 training subjects, ESI is 0.24 versus 0.32. Limitations: the TPI itself is not ablated, the Wilcoxon paired unit is unspecified, and per-sulcus p<0.05 results lack multiple-comparison correction.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
4 strengths · 4 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 locally-optimized Barlow Twins SSL pre-training on ~30,000 UK Biobank subjects (curvature and depth only) yields spherical encoders whose frozen features already match a fully supervised from-scratch U-Net (0.71±0.04 vs 0.72±0.05 Dice).
- Soft-initialized injection of the sulcal-line channel into the pre-trained backbone (TPI), with differential learning rates, reaches mean Dice 0.77±0.03 and ESI 0.15±0.02 on a 22-subject held-out test, significantly above from-scratch training (p<0.01).
- The pre-training is sensitive to rare local topology, as shown by linear probes on central sulcus interruption (207 vs 207 cases) and paracingulate sulcus presence.
- Performance gains concentrate on small, highly variable sulci, with local error reductions up to −14.82% (superior postcentral sulcus).
- In extreme low-data regimes (5 training subjects), the transferred model degrades gracefully (ESI 0.24±0.045 vs 0.32±0.06 from scratch, Wilcoxon p<0.01).
Strengths
Comprehensive labeling scope with a surface-native label pipeline
Unlike prior spherical CNN works covering 6-13 sulci (Hao et al., Lyu et al.), this paper labels 50 surface-consistent sulci; Sec. 2.1 details a careful pipeline (TRACE fundi extraction, geodesic Voronoi projection bounded at 3.5 mm, order-6 icosahedral resampling, graph morphological closing) validated by expert QC.
Convincing ablation of the pre-training objective (Table 1)
The frozen locally-optimized SSL encoder reaches Dice 0.71±0.04, on par with the fully trained from-scratch baseline (0.72±0.05, p>0.05), while the frozen globally supervised (age/sex) encoder collapses to 0.58±0.07 — direct evidence that the local pretext preserves dense-prediction-relevant geometry that macroscopic targets destroy.
Pre-training validated on rare localized topological traits, not just demographics
Honest metric scoping and disciplined evaluation design
Weaknesses
The TPI — the paper's named contribution — is never ablated
Statistical reporting is under-specified for the headline claim
Baseline input configuration is ambiguous and one proxy is weak
Scope limits: one hemisphere, one pre-training domain, missing volumetric number in text
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets dense, nomenclature-compliant labeling of 50 surface-consistent cortical sulci on the spherical representation of the cortex, under an extreme annotation budget: 62 expert-labeled right hemispheres (public BrainVISA base_62). Its central claims are: (1) a locally-optimized Barlow Twins pre-training on ~30,000 UK Biobank subjects (curvature + depth only) yields representations whose frozen linear separability already supports labeling at the level of a fully supervised baseline; (2) a soft-initialized Topological Prior Injector (TPI) bridges the 2-channel geometric pre-training to the 3-channel semantic fine-tuning input (adding extracted sulcal fundi lines) without catastrophic forgetting; (3) the full framework reaches mean Dice 0.77±0.03 on a held-out test of 22 subjects, above a from-scratch spherical U-Net (0.72±0.05), 2D ResNet-18 (0.69±0.05), and DINOv3-based dense prediction (0.70±0.05); (4) gains concentrate on small, variable sulci (up to −14.82% local error, superior postcentral sulcus); (5) benefits are largest in low-data regimes (ESI 0.24±0.045 vs 0.32±0.06 at N=5 training subjects, Wilcoxon p<0.01).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 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
What problem does this paper solve?
Automatically labeling the cortical sulci (the brain's folds) normally needs expensive expert annotations; only 62 annotated subjects exist for the full 50-sulcus nomenclature. The paper shows that self-supervised pre-training on ~30,000 unlabeled UK Biobank cortical surfaces, followed by fine-tuning, makes accurate labeling possible in this data-scarce regime.
What is the Topological Prior Injector (TPI)?
It is a 1x1 convolution placed before the pre-trained encoder and initialized so that the two pre-trained input channels (curvature and depth) pass through unchanged while the new third channel (extracted sulcal fundi lines) starts with near-zero weights. This lets the network begin from its pre-trained input distribution and gradually admit the semantic line information during fine-tuning.
How much does the method improve over training from scratch?
On a 22-subject held-out test with 30 training subjects, mean Dice is 0.77±0.03 versus 0.72±0.05 for a from-scratch spherical U-Net. The improvement is larger for small, variable sulci (local error down by up to about 14.8%) and in very low data: with only 5 training subjects, the error-of-sulcal-identification is 0.24 versus 0.32 from scratch.
Is the code or data available?
Code is linked via an anonymous repository (anonymous.4open.science); the 62-subject expert-annotated dataset is publicly downloadable from brainvisa.info (base_62, 2019 release); the UK Biobank pre-training data requires an approved application (number 64984).
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.
- Figure bodies (Fig. 1, Fig. 2, Fig. 3) are images with no numeric content in the extracted text; all quantitative values in this review come from in-text statements and Table 1 — notably the Borne et al. volumetric baseline's score is never given as text.
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.