- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper presents a way to represent brain functional connectivity without first dividing the brain into fixed atlas regions. It learns a continuous map that assigns a functional embedding to each cortical location. A shared population map captures common patterns, while constrained individual adjustments capture differences between people. The method then adjusts the locations, sizes, and shapes of regions to find connections associated with age, sex, or behavioral scores. Experiments on HCP-D, PNC, and HBN show stronger reported held-out associations than selected region pairs from Schaefer-100 and Schaefer-1000 atlases.
Critical strengths
-
The continuous representation allows brain regions to change location, size, and shape for different questions, while capturing both shared population patterns and individual differences.
-
Experiments on three datasets show stronger reported held-out associations than selected Schaefer atlas connections. Repeat-scan experiments and ablation studies help assess reliability and explain the effects of model choices.
Critical weaknesses
- Comparisons are limited to fixed Schaefer atlases. Ward and k-means also learn regions from data rather than relying on predefined atlas boundaries. Comparisons with these and supervised parcellation methods are needed to establish the benefit of the continuous representation.
- The best region configuration is reported based on held-out performance, which may overestimate generalization. A separate validation set or nested cross-validation would provide a stronger evaluation.
Questions for authors
Q1 How does the method compare with Ward, k-means, and supervised parcellation methods, which also learn regions from data rather than using predefined atlas boundaries?
Q2 How could this method be extended to clinical datasets such as ADHD-200, ABIDE, and ADNI, particularly when the available sample is small? Would a cohort-matched population field, stronger regularization, or simpler region geometry be needed?
Q3 How are the initial radius r_0 and contour complexity K selected without using held-out results? Please clarify the validation procedure, split seeds, and number of region-optimization steps.
Q4 The proposed method receives geometry optimization and a sweep over radius and contour complexity, while the Schaefer baselines only select the best parcel pair. How do the authors account for this difference in tuning effort? Would the gains remain under comparable tuning budgets?
5 comments
Response to Reviewer ubRR
The best region configuration is reported based on held-out performance. A separate validation set or nested cross-validation would provide a stronger evaluation. How are the initial radius and contour complexity selected without using held-out results? Would the gains remain under comparable tuning budgets?
We apologize for the ambiguity in our original wording. While Table 3 reported the best-performing region configuration based on the held-out results in Table 9, we did so primarily because of the page limit, to avoid displaying the full grid in the main text. More importantly, the choice among configurations turns out to matter little: in Table 9, the median held-out of every configuration of our method is higher than that of the Schaefer baselines. Following the reviewer's suggestion, we now report both alternatives the reviewer suggests:
-
A configuration fixed in advance: , , with center, radius and contour all fitted.
-
Nested cross-validation: within each training split, and are chosen by 5-fold cross-validation on training subjects; the model is refit on the full training split and scored once on the held-out split.
Both are in the revised Table 3 above. For both settings, regions are optimized with Adam (learning rate 0.02) for 450 steps: 150 at each of three mask temperatures, , and . Performance is reported over 12 random train–test splits using random seed 101.
Regarding the tuning budget, all hyperparameter selection is performed exclusively on training subjects, and all methods are evaluated on the same held-out splits. Under this evaluation protocol, our method with nested cross-validation achieves a higher median held-out than the tuned atlas baselines in all six settings. More importantly, even with a configuration fixed in advance, without any hyperparameter tuning, our method outperforms every baseline in five of the six settings. In the remaining setting (PNC sex), its performance is nearly identical to that of the best baseline (15.9 versus 16.0 for k-means). These results demonstrate that our improvements are robust to configuration selection and cannot be explained by a larger hyperparameter tuning budget.
How could this method be extended to clinical datasets such as ADHD-200, ABIDE, and ADNI, particularly when the available sample is small
For applications in clinical datasets such as ADHD-200, the population field need not be fitted on the clinical cohort. It can come from a large normative dataset, matched in age, with the clinical subjects' individual fields fitted against it. Our PNC and HBN results in Table 3 already work this way: both use the population field trained on HCP-D, only the individual fields are fitted to their own subjects.
When scans are short or noisy, the residual penalty can be increased to pull each individual field toward the population field. Table 2 shows this trade-off between in-sample fit and test–retest reliability on HCP-D subjects. Because the best depends on scan length and quality, it can be re-chosen for new small datasets.
For region geometry in small samples, we recommend a fixed, relatively large radius and a low contour order . This limits the number of fitted parameters and reduces the risk of overfitting.
More broadly, our results already show that the framework transfers across developmental cohorts that differ in population, scanner, scan length and phenotypes, using a population field learned on a single cohort. Extending these experiments to a wider range of clinical datasets and tasks is the next step we plan.
To Reviewer ubRR
Thank you again for your constructive feedback. In our response, we have clarified the configurations of all methods, explained how our framework can be applied to clinical cohorts, and added new results with nested cross-validation (revised Table 3).
As the discussion period is coming to an end, we would greatly appreciate it if you could let us know whether our responses have addressed your concerns. We would be glad to answer any further questions or provide additional clarification.
Response to Reviewer ubRR
We thank the reviewer for the constructive review and detailed questions. We appreciate that the reviewer recognized the flexibility of our regions in location, size and shape, and valued our experiments across three datasets, including repeat-scan and ablation analyses. The main concerns can be summarized as: (1) comparisons limited to fixed Schaefer atlases, (2) selection on held-out data and unequal tuning budgets, and (3) extension to small clinical datasets. We address each point in detail below:
Comparisons with these and supervised parcellation methods are needed to establish the benefit of the continuous representation.
We added three suggested dataset-specific region-learning baselines to the updated Table 3 below:
Revised Table 3: held-out t2, median [IQR] over 12 splits (35% of subjects held out). All choices are made on training subjects by 5-fold cross-validation within each training split: the number of parcels for Schaefer, Ward and k-means (100 or 1000), the number of parcels and penalty λ for SBP (100 or 1000, λ∈0,0.1,0.3,1,3), and r0∈4,8,12∘ and K∈2,4,6 for Ours, nested. Ours, fixed: r0=8∘, K=4. Bold: best per row.
In all these methods:
Each subject's individual field is pooled within each parcel, and the connection between two parcels is the cosine of their pooled fields. Each hemisphere is divided into K/2 parcels, as in the Schaefer atlases, and every parcel pair is a candidate. After regressing out the covariates (coefficients fitted on training subjects), the pair with the largest training ∣t∣ is selected and scored once on held-out subjects.
We used nested cross-validation, the number of parcels K∈100,1000 (and λ for SBP) is chosen by mean validation t2 over 5 inner folds of each training split; the parcel pair is re-selected inside every inner fold.
Results indicate our continuously optimized regions still achieve stronger associations on held-out subjects. Nevertheless, we would emphasize that brain parcellation methods serve a different purpose. Our framework focuses on localization: identifying where in the brain FC is most strongly associated with a phenotype. This is distinct from the goal of partitioning the entire brain into discrete parcels. Therefore, when interpreting these results, we would like to remind the reviewer that our central scientific question is whether continuous, gradient-based optimization enabled by our INR can identify region boundaries that yield stronger brain–phenotype associations. This represents a fundamentally new perspective that has not been systematically explored in previous work.