List of Papers Browse by Subject Areas Author List
Abstract
Deep learning on cortical surfaces faces a dilemma: capturing the complex topology of over 60 nomenclature-dependent sulci per hemisphere requires high-capacity models, yet the extreme scarcity of expert annotations (N = 62 subjects) inevitably causes overfitting. Standard supervised approaches fail to generalize on this data-scarce regime, particularly for variable and small sulci where topological ambiguity is high. To overcome this limitation, we introduce a Geometric-to-Semantic Spherical Transfer Learning framework. First, we leverage massive unlabeled data (UK Biobank with ≈ 30,000 subjects) to pre-train a spherical encoder using a locally-optimized strategy. By relying solely on continuous surface features (curvature and depth), the model encodes intrinsic topological priors. The relevance of this pre-training is confirmed by its ability to detect localized and rare topological traits, such as sulcal interruptions. The downstream labeling task, however, introduces extracted sulcal fundi (lines) as an explicit semantic input. To bridge this dimensional domain gap (from purely geometric to semantic) without causing catastrophic forgetting, these anatomical lines are integrated into the pre-trained backbone via a soft-initialized Topological Prior Injector. Our experiments demonstrate that this approach outperforms fully supervised baselines trained from scratch, achieving a mean Dice of 0.77. Crucially, a local analysis reveals that the self-supervised geometric priors yield the largest performance gains on variable and tertiary sulci (up to 14.8%), confirming that learning the cortex shape is highly beneficial for identifying its rarest parts.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3726_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/neurospin-projects/2026_stounsi_slabelin2
Link to the Dataset(s)
https://brainvisa.info/data/sulci_database/base_62/2019/
BibTex
@InProceedings{TouSae_GeometrictoSemantic_MICCAI2026,
author = { Tounsi, Saeb AND Chavas, Joël AND Gori, Pietro AND Frouin, Vincent AND Rivière, Denis AND Mangin, Jean-François},
title = { { Geometric-to-Semantic Spherical Transfer Learning for Cortical Sulci Labeling } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
Addressing the challenge of limited sulcal semantic labeling, this paper proposes a large-scale spherical self-supervised pre-training framework based on 30,000 unlabeled samples, integrating a Topology Prior Injector (TPI) for efficient geometry-to-semantics mapping that maintains robust generalization even under extreme data scarcity with only five labeled samples.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Established a powerful data-driven benchmark; by pretraining on 30,000 samples, the model achieves a Dice score of 0.77 even under extreme data scarcity with only 5 labeled samples.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The authors’ comparison between a spherical model pretrained on 30,000 real brains and a 2D foundation model (DINOv3), which relies on forced projections causing geometric distortions, is clearly unfair. It is suggested to introduce pretraining baselines from the medical imaging domain.
- One of the core contributions of the paper is the “soft-initialization” design of the TPI module to prevent catastrophic forgetting. However, the ablation study (Table 1) only compares the frozen state of the encoder, failing to validate the effectiveness of the initialization strategy itself.
- Is there a difference in the age distribution between the pretraining and downstream fine-tuning datasets? This is important because sulcal morphology, quantity, and width can all change with age.
- The visualization in Figure 3 appears inconsistent with the 1D sulcal line definition established in Figure 1.I suggest the authors provide a representative individual-level visualization that overlays the predicted 1D labels directly on the ground truth fundi lines to demonstrate the model’s precision in capturing complex folding patterns.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The comparison between a specialized spherical model and a 2D foundation model suffering from geometric distortions is somewhat unfair in terms of experimental design.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper focuses on the highly challenging task of cortical sulcal labeling. To address the issue of extremely scarce annotated data, the authors perform self-supervised pretraining on a large amount of unlabeled data to learn local geometric representations of the cortical surface. They then carefully design a fine-tuning strategy to transfer the pretrained representations to the sulcal labeling task, achieving significant performance improvements.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The paper presents a clear motivation and a well-defined objective, making it easy to understand. 2.Compared to training a fully supervised model from scratch, the paper innovatively adopts a pretraining-and-fine-tuning paradigm, offering an alternative solution to the problem and achieving superior performance.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The model’s fine-tuning process is somewhat complex, but acceptable.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
The paper adopts a VGG-style encoder and the Barlow Twins self-supervised paradigm. For future work, it would be worthwhile to explore more advanced model architectures (e.g., Transformers) as well as alternative self-supervised paradigms (e.g., masked autoencoders), to investigate how different architectural choices and pretraining paradigms may affect the final performance on the sulcal labeling task.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is comprehensive, and the experimental results effectively support the motivation behind the proposed method. It also demonstrates a reasonable degree of novelty.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes a spherical transfer learning framework to overcome the data-scarce issue when labeling high variability sulci. It decouples the sulci labeling into geometric representation learning and fine-tuning with topological prior injector. Geometric learning can be achieved with a large dataset in a non-supervised framework, which reduces the amount of labeled data needed.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Decoupling the sulci labeling into geometric learning and fine-tuning. 2.The topological prior injector is implemented in a learning way using the existing data for training.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The sensitivity of the method to the mis-traced sulci lines is not validated. 2.What is the comparison between the volume-based BrainVISA and the proposed method for sulci labeling is not discussed.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
It proposes a new try on the sulcal labeling using the surface based analysis, which can increase the analysis accuracy.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We sincerely thank all the reviewers and the Area Chair for their insightful comments and for recommending our paper for acceptance.
Reviewer #1 R1.W1 ( Comparison with DINOv3 ) We completely agree that this comparison is inherently asymmetric. As stated in the manuscript, DINOv3’s performance ‘highlights a structural domain gap between natural RGB images and distorted cortical topologies’. The cubemap projection used is the state-of-the-art for spherical-to-planar mapping, as it mitigates the polar singularities inherent to equirectangular projections while maintaining geometric consistency suitable for 2D convolutional kernels (even though it is not a conformal map). Our goal was not to claim absolute superiority, but to show that 2D projections create a structural bottleneck and a powerful foundation model fine-tuned on this projection still fails to match a surface-native approach. DINOv3 was used strictly due to lacking spherical foundation models. Spherical feature maps (curvature and sulcal depth) also differ drastically from the volumetric T1 MRI data on which some medical foundation models are trained. R1.W2 (TPI Ablation) Replacing soft initialization with random initialization yields Dice 0.74 ± 0.03 / ESI 0.17 ± 0.03, confirming the benefit of our strategy. This result was omitted from the original submission due to space constraints, as properly discussing the TPI’s contribution (which is most salient in the very low-data regime (5 subjects)) would have required an additional table and an extended analysis. We also note that monitoring TPI weights during training reveals an increase in the importance assigned to the sulcal lines channel. This confirms that soft initialization enables a progressive integration of semantic information into the pre-trained backbone. R1.W3 ( Age Distribution ) UK Biobank subjects are older adults (40–69 years), whereas the labeled cohort consists mostly of young adults (20–35 years). However, our self-supervised objective targets intrinsic topological features (local curvature and depth variations) rather than age-dependent traits like cortical thickness or sulcal width. Since sulcal topology, as defined by fundi lines, remains highly stable throughout adulthood, and our task is labeling rather than morphometry, this age gap does not hinder the transferability of the geometric priors. R1.W4 (Figure 3 vs. 1D definition) Figure 3 is a population-level visualization on a template brain: the Elocal metric is computed on the 1D sulcal lines and then aggregated across all test subjects, accompanied by statistical tests. We chose this representation over individual prediction overlays to provide a more globally informative picture of systematic gains and failures of the method across the full test set.
Reviewer #2 We sincerely thank Reviewer #2 for the positive assessment. Exploring Transformer-based architectures and masked autoencoder pre-training paradigms is indeed a natural and promising next step.
Reviewer #3 R3.W1 (Sensitivity to mis-traced sulcal lines) First, TRACE parameters were carefully calibrated through expert quality control, and the extracted lines were further refined via graph-based morphological operations (erosion and dilation) specifically designed to suppress false traces and fill small gaps. Second, the TPI is a learned module optimized end-to-end on lines that already carry inherent variability, meaning the network implicitly learns tolerance to imperfect inputs rather than relying on perfectly traced fundi. A formal robustness analysis under synthetic perturbations of the input lines would be a valuable extension and will be noted as future work. R3.W2 (Comparison with BrainVISA) Borne et al. [1] is the BrainVISA-based automatic labeling pipeline, already included as the volumetric baseline in Figure 2A and evaluated on the exact same dataset. The comparison therefore exists, and we will make this connection more explicit in the final version.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Reviewers agree the paper proposes a well-motivated self-supervised pretraining framework for sulcal labeling, with strengths in leveraging large unlabeled data, effective transfer learning design, and strong performance under extreme label scarcity. However, they note limitations including potentially unfair baseline comparisons, incomplete ablation/validation of key components, and insufficient analysis of data factors (e.g., age, noise sensitivity). Overall, the work is considered novel but needing more rigorous and fair evaluation.