acceptodds
Under review as a conference paper at ICLR 2027

SABER: Subject-Adaptive and Boundary-Aware Deep Multi-modal Clustering Network for Infant Group-Level Cortical Parcellation

Abstract

Constructing a group-level multi-modal cortical parcellation map for infants during the first two postnatal years is essential for studying early brain development and identifying disease risks from structural and functional perspectives. However, infant data are noisy and high-dimensional, making machine learning prone to fragmented and unstable parcellations, which motivates the use of deep learning. Yet, deep learning remains underexplored for this task because jointly modeling associative and complementary structural-functional information is challenging, while existing unsupervised learning objectives insufficiently address group-level parcellation practical demands, including improving cross-subject adaptability and anatomical consistency while preserving spatial continuity. To this end, we propose SABER, the first Subject-Adaptive and Boundary-aware DEep multi-modal clusteRing network for infant group-level cortical parcellation. SABER employs a two-stage fusion strategy combining multi-layer graph mutual learning with vertex-level modality weighting to capture structural-functional associations and complementarity. It further introduces two tailored objectives: the Dynamic Spatial-Weighted Generalized Homogeneity Loss improves cross-subject adaptability by optimizing parcellation homogeneity over a large set of individual-level features, while incorporating dynamic spatial weighting mechanism to prevent fragmentation; the Local Gradient-Weighted Active Boundary Loss incorporates local gradient-based boundary evidence to enhance sensitivity to potential biological boundaries. Experiments on the BCP dataset show that SABER outperforms existing SOTA methods and published parcellations across evaluation metrics and downstream tasks. Spatial null model analysis further indicates that the resulting parcellations capture biologically meaningful organization rather than random spatial patterns. Ablation studies demonstrate the plug-and-play ability of both losses and show that improved cross-subject adaptability benefits downstream task performance. https://anonymous.4open.science/r/SABER2027.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Infant group-level cotical parcellation, Multi-modal, Deep learning

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_7HUsOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents a task-driven multimodal deep learning framework for infant cortical parcellation. Its main contributions lie in incorporating cross-subject homogeneity, spatial continuity, and local boundary consistency into the optimization of group-level infant cortical parcellation. Overall, the paper is well structured, the illustrations are clear, and the experimental results are convincing. I consider this a highly competitive work.

Critical strengths
  1. The two specially designed loss functions are well motivated by the practical requirements of the task and are supported by relatively comprehensive experimental evidence.
  2. The contributions are not entirely dependent on the specific architecture of SABER, as the proposed loss functions demonstrate a certain degree of transferability and can be incorporated into other clustering frameworks.
Critical weaknesses
  1. The authors claim to use the Hausdorff distance to evaluate the anatomical alignment between the predicted parcellation boundaries and Brodmann area boundaries. However, the actual metric appears to be a one-sided average nearest-boundary distance rather than the standard Hausdorff distance. Consequently, the improvement in anatomical alignment brought by LBoundary in Table 3 may not necessarily be attributed entirely to more accurate boundary localization. It could also be partially explained by changes in the distribution or complexity of the predicted boundaries.
  2. Several major technical components of SABER are closely related to existing approaches. Specifically, multi-layer graph mutual learning and vertex-level weighting can be viewed as a combination of established multimodal fusion techniques. The core idea of LDGHomoL_{DGHomo} extends existing homogeneity objectives to individual-level features while incorporating a heuristic dynamic spatial constraint. Similarly, LBoundaryL_{Boundary} extends the existing Active Boundary Loss to cortical parcellation with additional local gradient-based weighting. Although these designs are well motivated and experimentally effective, it remains unclear to what extent they introduce fundamentally new insights from a general machine learning perspective. I would encourage the authors to further clarify the essential differences between the proposed loss functions and existing spatial regularization and boundary optimization methods.

3.Some minor errors may cause confusion for readers:

  1. In the legend at the upper-left corner of Figure 1, the authors appear to have mixed up the subscripts, resulting in inconsistent labeling of the FCN and MSN modalities.
  2. The paper reuses the notation PiP_i to represent two entirely different probability distributions, which may lead to ambiguity and confusion.
Questions for authors

The proposed model ultimately produces a single age-independent group-level cortical parcellation map covering infants aged 0–800 days, which represents a substantial developmental span. Considering that cortical structural and functional organization may differ significantly between infants at two months and two years of age, I would like to understand whether LDGHomoL_{DGHomo} truly learns a parcellation map that generalizes well across the entire infant population, or whether it instead produces a compromise solution that is better suited to certain developmental stages represented in the dataset.

6 comments

We sincerely thank the reviewer for taking the time to read our manuscript so carefully. The question raised is highly important and insightful, and has motivated us to further investigate the effectiveness of our proposed method.

Response to Weaknesses 1: Recomputation of the Anatomical Alignment Evaluation Metric

We agree that, based on its computational definition, the metric originally reported in our manuscript is a one-sided average nearest-boundary distance rather than the standard Hausdorff distance. We adopted this metric from Schaefer et al. (2018), who used it to assess the alignment between anatomical and parcellation boundaries and referred to it as the average Hausdorff distance. We therefore followed the same terminology. To address the reviewer's concern, we recalculated anatomical alignment using the standard symmetric Hausdorff distance.

Specifically, for each Brodmann area, we identified all parcels overlapping with it, merged these parcels, extracted the contour of the resulting region, and calculated the symmetric Hausdorff distance between this contour and the corresponding anatomical boundary.

---- Rebuttal table 1: Quantitative assessment of anatomical alignment using standard Hausdorff distance ----

MethodsArea 1,2 ↓Area 3,4 ↓Area 6 ↓Area 17,18 ↓Area 44,45 ↓Mean ↓
MV-Kmeans-20046.390225.013045.071790.713235.937148.6251 (22.3944)
Mv-SC-20020.98118.611912.854022.57689.053314.8154 (5.8955)
CGL-2009.38636.917320.888520.93648.304713.2866 (6.2754)
RML-2007.93785.119916.698911.260620.893112.3821 (5.7414)
SparseMVC-20031.856172.473124.890355.311998.838756.6740 (27.0386)
MSDIB-20014.706025.0130115.910213.51267.403535.3090 (40.6961)
SEABER-20012.53267.506712.931614.160210.191511.4645 (2.3598)
MV-Kmeans-30093.628875.431643.066087.761424.996164.9768 (26.5628)
Mv-SC-30020.22036.104711.817518.062110.561913.3533 (5.1385)
CGL-3008.462558.356516.528910.67297.792310.3626 (3.2360)
RML-30013.62846.079815.302910.304447.680518.5992 (14.8783)
SparseMVC-30023.00727.4877101.199411.849154.140539.5368 (34.8784)
MSDIB-3006.83555.216761.170615.067929.584723.5751 (20.6840)
SEABER-3005.64987.400714.127710.230810.93709.6692 (2.9360)
MV-Kmeans-40029.691524.730044.209412.275430.188428.2189 (10.2772)
Mv-SC-4009.25077.581710.854610.72979.05339.4940 (1.2073)
CGL-4009.654312.949510.14269.90219.776510.4850 (1.2428)
RML-4008.090923.664017.93469.793212.214914.3395 (5.7288)
SparseMVC-40023.2828130.541736.7251103.730054.242969.7045 (40.8425)
MSDIB-4004.727322.864835.823310.87859.904116.8396 (11.1984)
SEABER-4005.76097.90019.97889.00038.15688.0394 (1.4447)
MV-Kmeans-50029.691524.730043.066018.765630.188429.2883 (8.0287)
Mv-SC-50010.569327.58177.357511.56849.05339.2260 (1.6441)
CGL-5007.24387.13098.85269.12667.79238.0292 (0.8200)
RML-5005.41285.014415.30299.79328.76228.8571 (3.7165)
SparseMVC-50018.43797.048995.026452.987361.112146.9225 (31.4750)
MSDIB-5008.98776.178326.71618.21398.645411.7483 (7.5472)
SEABER-5006.44176.18166.84297.60246.98866.8114 (0.4881)

As shown in the Rebuttal table 1, SABER still achieves the best anatomical alignment among the compared methods. However, we also observe that the standard Hausdorff distance is highly sensitive to outliers and noise, as it is determined by the maximum nearest-boundary distance between the two sets of boundary vertices. Consequently, methods with fragmented parcellations tend to perform particularly poorly under this metric, as even a few spatially isolated boundary vertices can substantially increase the Hausdorff distance. In contrast, the average nearest-boundary distance is more robust to such outliers and may therefore provide a more representative assessment of the overall boundary alignment. We will incorporate these additional results and revise the corresponding descriptions in the manuscript accordingly.

[1] Alexander Schaefer, Ru Kong, Evan M Gordon, Timothy O Laumann, Xi-Nian Zuo, Avram J Holmes, Simon B Eickhoff, and BT Thomas Yeo. Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity mri. Cerebral cortex, 28(9):3095–3114, 2018.

Response to Weaknesses 2: Clarification of Methodological Contributions

We agree that several building blocks in SABER are related to established techniques, and we do not claim that each component introduces an entirely new general-purpose machine learning primitive. Rather, our contribution lies in reformulating and tailoring these techniques to address three intertwined challenges in infant group-level cortical parcellation: cross-subject adaptability, spatial continuity, and sensitivity to potential biological boundaries.

Specifically, 'Dynamic Spatial-Weighted Generalized Homogeneity Loss' shifts the optimization objective of group-level parcellation from fitting group-averaged features to jointly accommodating the individual-level feature distributions of a large number of subjects, thereby explicitly accounting for substantial inter-subject variability in infants and aiming to obtain a more representative group-level parcellation. Moreover, instead of imposing a fixed spatial constraint uniformly across all parcels, it introduces parcel-specific adaptive spatial regularization, where each parcel's spatial weight is independently adjusted according to its fragmented status. This preserves spatial continuity without unnecessarily restricting parcel shapes, allowing anatomically plausible elongated areas to emerge.

Similarly, while 'Local Gradient-Weighted Active Boundary Loss' builds upon the Active Boundary Loss (Wang et al. 2022), its essential distinction lies in how boundary supervision is constructed and applied. Rather than determining boundaries solely based on potentially ambiguous feature differences in certain regions, it incorporates independent local-gradient-based boundary evidence from both structural and functional modalities to provide additional directional guidance for boundary localization. Meanwhile, it imposes stronger alignment constraints on high-probability boundaries while relaxing them in low-probability regions, thereby adapting the strength of boundary supervision to the reliability of the available boundary evidence.

Furthermore, our two-stage fusion strategy combines early intra-/inter-modal graph information exchange with late vertex-level modality weighting to capture both structural-functional associations and spatially heterogeneous complementary information. Importantly, these components are not merely combined to increase architectural complexity. Rather, they are designed to address complementary yet interrelated requirements of multi-modal cortical parcellation.

Our ablation studies further validate their respective contributions: individual-level homogeneity optimization improves cross-subject adaptability, dynamic spatial weighting effectively reduces parcel fragmentation, and incorporating independent local-gradient-based boundary evidence further improves anatomical alignment (Table 3 and Appendix Tables 10). Moreover, integrating the proposed losses into different existing clustering frameworks yields consistent performance gains (Table 4 and Appendix Tables 11–12), supporting their applicability beyond SABER.

Response to Weaknesses 3: Resolution of Notational Ambiguities

We have revised the manuscript accordingly to address the points raised.

[1] Chi Wang, Yunke Zhang, Miaomiao Cui, Peiran Ren, Yin Yang, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, and Weiwei Xu. Active boundary loss for semantic segmentation. In Proceedings of the AAAI conference on artificial intelligence, pp. 2397–2405, 2022.

Response to Question 1: Age-stratified Evaluation of Group-level Parcellation Applicability Across Entire Developmental Stages (1/3)

We thank the reviewer for this insightful suggestion, which motivated us to further investigate the effectiveness of the 'Dynamic Spatial-Weighted Generalized Homogeneity Loss' across different developmental stages. We conducted additional age-stratified evaluations at a parcellation resolution of 400. Specifically, the test set was divided into six age groups: 3M (0–144 days), 6M (145–223 days), 9M (224–318 days), 12M (319–410 days), 18M (411–591 days), and 24M (592–800 days), after which we conducted a two-part experimental evaluation:

i) We quantified the homogeneity of group-level parcellations generated by all competing methods across each age group in the test set, with all parcellations trained on the full training set (0-800 days), to first assess whether SABER achieves better age-stratified performance than other methods.

ii) We further conducted a more in-depth experiment by dividing the training set into the corresponding six age groups and independently training a model on each group to obtain age-specific group-level parcellations. We then evaluated each resulting parcellation across all six test age groups to further investigate whether a unified parcellation represents a compromise or can instead provide robust performance across different developmental stages.

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

SABER learns infant group-level cortical parcellations by fusing structural and functional features. Individual-level homogeneity, adaptive spatial weighting, and gradient-guided boundary alignment jointly promote representative, contiguous parcels. Evaluation on BCP includes parcellation metrics, spatial null models, downstream prediction, and ablations.

Critical strengths
  1. The method addresses concrete parcellation requirements, and ablations quantify the contributions of individual-level optimization, spatial continuity, and boundary alignment (Tables 3-5). Transferring the losses to other networks strengthens their methodological value.
  2. Main parcellation and prediction evaluations use a subject-disjoint split of 148 training and 33 test infants. Multiple resolutions, spatial null controls, and downstream tasks provide complementary evidence (Appendices A-B).
Critical weaknesses
  1. Generalization evidence has a limited independent sample. The main test set contains 33 infants, while longitudinal identification uses 100 infants from the same 181-subject cohort (Appendices A.2, B.2). This necessarily includes atlas-training subjects, so identification accuracy does not independently establish generalization to unseen infants.
  2. The source of the boundary supervision is unclear. Section 2.2 explains how ProbB guides training, but does not specify the subjects used to construct the modality-specific probability maps. This prevents verification that this prior is independent of the test subjects.
Questions for authors
  1. What identification performance is obtained using only infants excluded from atlas construction?
  2. Exactly how are the ProbB maps generated, and are all contributing subjects restricted to the training partition?
4 comments

We sincerely thank the reviewer for taking the time to read our manuscript so carefully. The questions raised are highly detailed and deserve careful consideration.

Response to Weaknesses 1: Re-evaluation of the Longitudinal Individual Identification Task

We agree that including subjects who contributed to group-level parcellation construction could compromise the independence of this evaluation. To directly address this concern, we repeated the longitudinal identification experiments using only infants from the held-out test set who had multiple longitudinal scans.

Specifically, for each eligible subject, we randomly selected two longitudinal scans and assigned one to the target set and the other to the base set. We then constructed functional connectivity networks and morphological similarity networks from the corresponding scans. The upper-triangular elements of each network were extracted and flattened into one-dimensional vectors, which were then concatenated for longitudinal individual identification.

---- Rebuttal table 1: Quantitative results of longitudinal individual identification ----

MethodsLongitudinal Individual Identification (ACC, Functional ↑)Longitudinal Individual Identification (ACC, Structural ↑)
MV-Kmeans-2000.3727 (0.0407)0.9212 (0.0277)
Mv-SC-2000.3878 (0.0799)0.9181 (0.0304)
CGL-2000.3926 (0.0484)0.9121 (0.0163)
RML-2000.3742 (0.0445)0.9121 (0.0285)
SparseMVC-2000.3696 (0.0445)0.8636 (0.0338)
MSDIB-2000.3636 (0.0303)0.9212 (0.0337)
SEABER-2000.4272 (0.0566)0.9242 (0.0338)
MV-Kmeans-3000.4060 (0.0545)0.9666 (0.0163)
Mv-SC-3000.4060 (0.0453)0.9636 (0.0226)
CGL-3000.4303 (0.0503)0.9515 (0.0201)
RML-3000.4212 (0.0316)0.9606 (0.0163)
SparseMVC-3000.3363 (0.0458)0.9393 (0.0383)
MSDIB-3000.4212 (0.0369)0.9181 (0.0333)
SEABER-3000.4393 (0.0578)0.9787 (0.0194)
MV-Kmeans-4000.3939 (0.0635)0.9666 (0.0251)
Mv-SC-4000.4333 (0.0407)0.9636 (0.0326)
CGL-4000.4333 (0.0575)0.9787 (0.0194)
RML-4000.4266 (0.0543)0.9666 (0.0212)
SparseMVC-4000.3742 (0.0624)0.9606 (0.0138)
MSDIB-4000.4328 (0.0563)0.9468 (0.0267)
SEABER-4000.4515 (0.0416)0.9848 (0.0151)
MV-Kmeans-5000.4175 (0.0549)0.9575 (0.0201)
Mv-SC-5000.4242 (0.0542)0.9848 (0.0203)
CGL-5000.4424 (0.0363)0.9757 (0.0181)
RML-5000.4090 (0.0474)0.9818 (0.0309)
SparseMVC-5000.3787 (0.0433)0.9696 (0.0234)
MSDIB-5000.4484 (0.0353)0.9591 (0.0277)
SEABER-5000.4787 (0.0538)0.9909 (0.0138)
Glasser-1800.3787 (0.0769)0.9060 (0.0330)
Ours-1800.4151 (0.0742)0.9257 (0.0278)
Li-1050.3419 (0.0449)0.8590 (0.0386)
Ours-1050.3636 (0.0338)0.8863 (0.0343)

As shown in the Rebuttal table 1, SABER still achieves the best longitudinal identification performance across different parcellation resolutions and both modalities. These results further suggest that improving the cross-subject adaptability of group-level parcellations during early infancy helps preserve subject-specific identity information when constructing individual brain networks. We will incorporate these results into the revised manuscript.

Response to Weaknesses 2: Clarification on the Data Used to Generate Boundary Probability Density Maps

A total of 299 longitudinal scans from 148 infants in the training set were used to construct the boundary probability density maps for both modalities, strictly following the local gradient method described in Wang et al. (2023). Next, we use the functional modality as an example to illustrate the procedure in detail. Specifically, for each scan, we first computed the first-order resting-state functional connectivity matrix by calculating pairwise Pearson correlations between the BOLD time series of all cortical vertices (for the structural modality, we instead calculated pairwise Pearson correlations among the four cortical attributes across all vertex pairs). We then calculated second-order correlation matrices by further correlating the connectivity profiles of different vertices, which enhances the detection of functional transitions between adjacent cortical regions. Subsequently, we performed row-wise gradient computation and row-wise watershed segmentation on the second-order correlation matrix, generating a binary boundary map for each vertex. These binary boundary maps were then averaged to obtain the final boundary probability density maps for a single scan. To establish more accurate inter-subject functional correspondence, we further performed a cortical surface registration using Spherical Demons, integrating individual functional boundary probability density maps as features. Specifically, individual boundary probability density maps were iteratively aligned to their population-level mean boundary probability density map, with the warped maps averaged to update the group-level representation until convergence (four iterations in our experiments). Finally, the boundary probability density maps for the two modalities were obtained, and the results are shown in Figure 2 and Appendix Figure 8.

[1] Fan Wang, Han Zhang, Zhengwang Wu, Dan Hu, Zhen Zhou, Jessica B Girault, Li Wang, Weili Lin, and Gang Li. Fine-grained functional parcellation maps of the infant cerebral cortex. elife, 12: e75401, 2023a.

Dear Reviewer,

We sincerely appreciate the time and effort you have devoted to reviewing our manuscript and providing constructive comments. We fully understand the considerable workload involved in the reviewing process and are truly grateful for your thoughtful evaluation.

Following your feedback, we have carefully revised our responses and provided additional experimental results and explanations to address the concerns raised. We would be very grateful if you could share whether these revisions and clarifications have adequately addressed your main concerns, or whether there are any remaining issues that you feel require further discussion. We would be more than willing to provide additional explanations or clarifications to address any unresolved questions.

Thank you once again for your valuable feedback, patience, and consideration. We sincerely appreciate your contribution to improving our work and your continued engagement in the discussion.

Best regards, The Authors

I appreciate the authors’ additional experiments and clarifications. Restricting longitudinal identification to held-out infants and confirming that both boundary probability maps were constructed exclusively from training subjects substantially address my original concerns. The exact number of infants included in the revised identification experiment should still be reported, and the description should clarify whether functional and structural features are evaluated separately or concatenated. These are remaining reporting issues rather than evidence of training–test overlap. I have taken the strengthened evaluation and clarifications into account, but my overall assessment remains unchanged, and I maintain my original score.

Reviewer_mNa2OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SABER, the first subject-adaptive and boundary-aware deep multi-modal clustering network for infant group-level cortical parcellation. Extensive experiments on the BCP dataset demonstrate that explicitly improving the cross-subject adaptability of group-level parcel- lations and incorporating local boundary evidence can enhance downstream task performance and sensitivity to potential biological boundaries, which is valuable for subsequent analyses and neurobi- ological interpretation. In addition, a simple yet effective two-stage fusion strategy further improves parcellation quality by more fully exploiting associative and complementary structural-functional information.

Critical strengths
  1. The proposed framework is well presented. The motivation of the proposed research makes sense. The paper, largely, is easy to follow.
  2. The proposed model is adequately evaluated. Several SOTAs have been reasonably compared with evidence to support the statements made in the paper.
  3. The topic itself is interesting to the audience.
Critical weaknesses
  1. In the contributions listed in the 1st section of the paper, the authors have pointed out that the proposed model contains two novel tailored losses to better match the practical demands of infant group-level cortical parcellation, improving the cross-subject adaptability of the group-level parcellation and enhancing its sensitivity to local potential biological boundaries. Then, in section 2.2, in the subsection 'Dynamic Spatial-Weighted Generalized Homogeneity Loss', the loss input can help maintaining spatial continuity while accommodating the anatomical prior of elongated cortical areas. However, it is not clear why / how the component can maintain spatial continuity.
  2. Similarly, when discussing 'Local Gradient-Weighted Active Boundary Loss', the paper constructs a predicted direction distribution Qi ∈R1×|N1(i)| via softmax over the KL divergences between the probability assignments of vertex I and its 1-ring neighbors. I wonder that, in the complicated cases, where there are a few areas of containing similar representations with close softmax outputs, the strategy to handle this difficult case is not thought through in this paper.
  3. The proposed framework has been evaluated over the BCP dataset (Howell et al., 2019). The dataset is old and limited. Several new datasets have been used in the public such as ALBERTs and BOBs datasets. The additional datasets can be used to measure the robustness and generalisation of the proposed model in this paper.
Questions for authors

Please address all the above concerns.

5 comments

We sincerely thank the reviewer for the insightful questions and for taking the time to carefully read our manuscript. We will do our best to address the reviewer's questions and concerns, and hope that our responses will help clarify our work.

Response to Weaknesses 1: Clarification on How the ‘Dynamic Spatial-Weighted Generalized Homogeneity Loss’ Maintains Spatial Continuity While Accommodating the Anatomical Prior of Elongated Cortical Areas

In fact, the key idea behind the 'Dynamic Spatial-Weighted Generalized Homogeneity Loss' for preserving spatial continuity is to introduce spatial regularization by considering both feature similarity and spatial distances during vertex-to-parcel assignment. This discourages spatially distant vertices from being assigned to the same parcel, thereby reducing fragmentation. However, a fixed spatial weight involves an inherent trade-off: an excessively large weight tends to produce nearly circular parcels, contradicting the anatomical prior that some cortical areas exhibit elongated shapes, whereas an excessively small weight may leave parcels fragmented. To address this issue, we introduce a dynamic parcel-specific spatial weighting mechanism. Specifically, relatively large spatial weights are initially assigned to all parcels to prevent fragmentation. As training progresses, each parcel's spatial weight is gradually reduced as long as its spatial continuity is maintained, allowing greater flexibility in its shape, but increased if fragmentation occurs. Eventually, the spatial weights of almost all parcels decrease to values close to zero (lower bound is set to 0.05), while only a very small number of parcels retain relatively large weights, as illustrated in Appendix Fig. 7. Furthermore, Appendix Fig. 8 shows that our resulting parcellation contains numerous elongated cortical areas, consistent with those observed in previously published parcellation maps.

Response to Weaknesses 2: Clarification on How the ‘Local Gradient-Weighted Active Boundary Loss’ Identifies Boundaries in Areas with Similar Feature Representations

First, we would like to clarify that the ‘Local Gradient-Weighted Active Boundary Loss’ does not identify boundaries solely based on the predicted direction distribution Qi\mathcal{Q}_i. Rather, even when neighboring vertices exhibit similar feature representations and consequently yield similar KL divergence values, it can still leverage the independent external boundary evidence provided by ProbB to determine the target direction distribution Pi\mathcal{P}_i, thereby explicitly guiding predicted boundaries toward regions with a high probability of containing anatomical boundaries. Specifically, for a predicted boundary vertex located in a region with similar feature profiles, we still identify the neighboring vertex closest to the high-probability boundary based on ProbB from both MSN and FCN modalities and use it to construct a one-hot target direction distribution Pi\mathcal{P}_i. Meanwhile, we apply softmax to the KL divergences between the cluster assignment distribution of the current vertex and those of its 1-ring neighbors to obtain the predicted direction distribution Qi\mathcal{Q}_i. Finally, a weighted cross-entropy loss encourages Qi\mathcal{Q}_i to align with the independently constructed Pi\mathcal{P}_i, with the strength of this alignment constraint determined by the actual boundary probability. In this way, the predicted boundary is progressively guided toward high-probability external boundary evidence. The alignment constraint is relaxed only when the boundary probability is low, in which case boundary optimization is primarily driven by the global homogeneity term.

Response to Weaknesses 3: Evaluation of SABER's Robustness and Generalization on an Independent External dHCP Dataset (1/3)

Following the reviewer's suggestion, we carefully examined the ALBERTs and BOBs datasets. However, both datasets provide only structural MRI modalities (T1w and T2w) without functional MRI data, making them unsuitable for directly evaluating our complete framework. Nevertheless, to further address the reviewer's concern, we conducted additional experiments on the dHCP neonatal dataset to evaluate the robustness and generalizability of SABER.

Specifically, we included 395 neonates from the dHCP dataset, with postmenstrual ages ranging from 29 to 45 weeks. We divided the subjects into a training set of 300 and a test set of 95, and conducted two sets of experiments.

i) We directly transferred the parcellation maps constructed from the BCP training set to the dHCP test set for independent external validation.

ii) We directly constructed group-level parcellation maps using the dHCP training set and evaluated their performance on the dHCP test set using four complementary evaluation metrics, as well as downstream tasks.

---------------------------------------- Rebuttal table 1: BCP train set => dHCP test set, external validation ----------------------------------------

MethodsParcellation Homogeneity (Functional ↑)Parcellation Homogeneity (Structural ↑)Silhouette Coefficient (Functional ↑)Silhouette Coefficient (Structural ↑)
MV-Kmeans-2000.6437 (0.0936)0.3650 (0.0193)0.4112 (0.0347)0.2337 (0.0173)
Mv-SC-2000.7061 (0.0653)0.3520 (0.0224)0.4729 (0.0383)0.2257 (0.0176)
CGL-2000.7006 (0.0936)0.3519 (0.0206)0.4725 (0.0382)0.2257 (0.0173)
RML-2000.6913 (0.0854)0.3593 (0.0202)0.3898 (0.0279)0.1908 (0.0185)
SparseMVC-2000.5997 (0.1078)0.3382 (0.0181)0.2679 (0.0296)0.1539 (0.0104)
MSDIB-2000.6923 (0.0815)0.3656 (0.0203)0.4496 (0.0328)0.2244 (0.0213)
SEABER-2000.7161 (0.0761)0.3787 (0.0205)0.4831 (0.0351)0.2499 (0.0206)
MV-Kmeans-3000.6989 (0.0820)0.4032 (0.0181)0.4458 (0.0305)0.2505 (0.0203)
Mv-SC-3000.7566 (0.0547)0.4032 (0.0232)0.5224 (0.0366)0.2576 (0.0232)
CGL-3000.7579 (0.0566)0.4011 (0.0207)0.5215 (0.0371)0.2574 (0.0231)
RML-3000.7339 (0.0745)0.3975 (0.0209)0.4115 (0.0271)0.1991 (0.0203)
SparseMVC-3000.6175 (0.0990)0.3545 (0.0188)0.2588 (0.0231)0.1401 (0.0108)
MSDIB-3000.7374 (0.0711)0.4116 (0.0197)0.4827 (0.0320)0.2542 (0.0242)
SEABER-3000.7690 (0.0639)0.4190 (0.0215)0.5312 (0.0340)0.2832 (0.0245)
MV-Kmeans-4000.7394 (0.0730)0.4303 (0.0202)0.4737 (0.0284)0.2677 (0.0235)
Mv-SC-4000.7954 (0.0481)0.4357 (0.0227)0.5524 (0.0356)0.2713 (0.0248)
CGL-4000.7928 (0.0516)0.4263 (0.0221)0.5508 (0.0352)0.2729 (0.0267)
RML-4000.7624 (0.0671)0.4261 (0.0231)0.4353 (0.0251)0.2183 (0.0235)
SparseMVC-4000.6399 (0.0929)0.3642 (0.0181)0.2670 (0.0209)0.1512 (0.0108)
MSDIB-4000.7648 (0.0683)0.4324 (0.0234)0.5135 (0.0314)0.2803 (0.0272)
SEABER-4000.8034 (0.0564)0.4498 (0.0227)0.5623 (0.0339)0.3016 (0.0285)
Li-1050.6330 (0.0833)0.3383 (0.0214)0.4017 (0.0463)0.2359 (0.0156)
SEABER-1050.6318 (0.0947)0.3076 (0.0219)0.4123 (0.0408)0.1963 (0.0145)

As shown in the Rebuttal table 1, SABER demonstrates clear transfer advantages over the compared methods across all parcellation resolutions. The only exception is the published parcellation map proposed by Li et al. (2023), which achieves better performance in this setting. This is primarily because their parcellation map was constructed using the dHCP dataset itself, making it inherently better adapted to this particular population and age range.

[1] Mingyang Li, Xinyi Xu, Zuozhen Cao, Ruike Chen, Ruoke Zhao, Zhiyong Zhao, Xixi Dang, Kenichi Oishi, and Dan Wu. Multi-modal multi-resolution atlas of the human neonatal cerebral cortex based on microstructural similarity. NeuroImage, 272:120071, 2023.

Response to Weaknesses 3: Evaluation of SABER's Robustness and Generalization on an Independent External dHCP Dataset (2/3)

---- Rebuttal table 2: dHCP train set => dHCP test set, independent validation (four complementary evaluation metrics) ----

MethodsParcellation Homogeneity (Functional ↑)Parcellation Homogeneity (Structural ↑)Silhouette Coefficient (Functional ↑)Silhouette Coefficient (Structural ↑)Anatomical Alignment (Mean ↓)Spatial Continuity (Parcellation ↑)
MV-Kmeans-2000.6128 (0.0874)0.3796 (0.0224)0.3948 (0.0376)0.2463 (0.0189)14.3965 (5.2341)0.6893
Mv-SC-2000.7233 (0.0729)0.3786 (0.0231)0.4894 (0.0376)0.2484 (0.0183)4.3262 (1.2325)0.9894
CGL-2000.7218 (0.0864)0.3761 (0.0263)0.4863 (0.0394)0.2451 (0.0196)4.1286 (1.0395)0.9911
RML-2000.7070 (0.0767)0.3873 (0.0184)0.4664 (0.0338)0.2589 (0.0199)3.3284 (0.8950)0.9590
SparseMVC-2000.5137 (0.1266)0.3793 (0.0211)0.2008 (0.0303)0.2000 (0.0149)18.4326 (4.2367)0.4347
MSDIB-2000.6474 (0.0891)0.3901 (0.0187)0.4315 (0.0327)0.2603 (0.0226)6.1251 (3.9342)0.8315
SEABER-2000.7334 (0.0730)0.3929 (0.0202)0.5091 (0.0362)0.2785 (0.0218)2.9795 (0.6396)0.9938
MV-Kmeans-3000.6623 (0.0864)0.4268 (0.0196)0.4208 (0.0363)0.2731 (0.0262)12.6821 (4.3966)0.6921
Mv-SC-3000.7638 (0.0521)0.4210 (0.0227)0.5326 (0.0352)0.2632 (0.0228)3.4316 (0.8863)0.9876
CGL-3000.7668 (0.0552)0.4196 (0.0207)0.5332 (0.0367)0.2641 (0.0224)3.2896 (0.7695)0.9903
RML-3000.7584 (0.0656)0.4272 (0.0202)0.5078 (0.0325)0.2768 (0.0238)2.2320 (0.6567)0.9692
SparseMVC-3000.5436 (0.1154)0.3935 (0.0187)0.2169 (0.0238)0.1929 (0.0169)17.1573 (3.001)0.4585
MSDIB-3000.6808 (0.0801)0.4329 (0.0183)0.4536 (0.0277)0.2757 (0.0257)9.2669 (6.0611)0.8194
SEABER-3000.7786 (0.0619)0.4335 (0.0212)0.5432 (0.0337)0.3016 (0.0256)2.1149 (0.4557)0.9979
MV-Kmeans-4000.6987 (0.0749)0.4514 (0.0228)0.4363 (0.0298)0.2868 (0.0247)13.4726 (4.4112)0.6679
Mv-SC-4000.7991 (0.0477)0.4524 (0.0212)0.5547 (0.0341)0.2861 (0.0239)2.7796 (0.6533)0.9834
CGL-4000.7977 (0.0511)0.4478 (0.0223)0.5539 (0.0349)0.2846 (0.0261)2.6724 (0.5528)0.9893
RML-4000.7903 (0.0581)0.4597 (0.0211)0.5359 (0.0311)0.2994 (0.0289)2.2846 (0.4922)0.9701
SparseMVC-4000.5692 (0.1191)0.4146 (0.0199)0.2167 (0.0202)0.1839 (0.0176)20.6383 (7.2023)0.4495
MSDIB-4000.7107 (0.0741)0.4620 (0.0200)0.4730 (0.0265)0.2910 (0.0279)9.8055 (4.5456)0.8306
SEABER-4000.8079 (0.0556)0.4664 (0.0217)0.5646 (0.0331)0.3221 (0.0282)1.8153 (0.4133)0.9928
Li-1050.6330 (0.0833)0.3383 (0.0214)0.4017 (0.0463)0.2359 (0.0156)4.6682 (1.9553)1.0000
SEABER-1050.6347 (0.0955)0.3474 (0.0185)0.4186 (0.0401)0.2582 (0.0166)4.5045 (1.9720)0.9920

As shown in the Rebuttal table 2, SABER consistently outperforms the compared methods across different parcellation resolutions in terms of parcellation homogeneity, silhouette coefficient, anatomical alignment, and spatial continuity. Furthermore, when trained on the dHCP dataset, SABER also surpasses the published parcellation map proposed by Li et al. (2023).

[1] Mingyang Li, Xinyi Xu, Zuozhen Cao, Ruike Chen, Ruoke Zhao, Zhiyong Zhao, Xixi Dang, Kenichi Oishi, and Dan Wu. Multi-modal multi-resolution atlas of the human neonatal cerebral cortex based on microstructural similarity. NeuroImage, 272:120071, 2023.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.