- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes STH-Brain, a self-supervised foundation model for resting-state fMRI time series. Whereas most prior work builds a functional connectivity (FC) matrix from the whole scan, or from sliding windows, STH-Brain operates directly on ROI time series and introduces two modules: Physical-Time Rotary Encoding (PTRE), which maps elapsed time in seconds to rotary phases, and Spatial-Aware Encoding (SAE), which injects atlas coordinates into within-network spatial attention. These are combined with a hierarchy that retains ROI-level states alongside network-level tokens, trained with dual-level latent prediction and visibility-based routing. Across seven disorder-classification tasks, STH-Brain reports the best results over 13 baselines, and it also reports the best restuls for age prediction across seven disease groups, plus gains on an external SCZ/HC cohort and a zero-shot MDD→BP transfer. Interpretability maps are provided.
Critical strengths
S1. Performance is strong across a broad set of benchmarks, and the evaluation covers a sufficient number of tasks and baselines (13 models, 7 classification tasks, age prediction, external and zero-shot evaluation).
S2. The zero-shot transfer setting (MDD/HC → BP/HC without target labels) is an interesting and underexplored evaluation for fMRI foundation models.
S3. The idea of a hierarchy that retains both ROI-level and network-level representations, rather than collapsing ROIs into network tokens, is appealing.
Critical weaknesses
W1. (Minor) Citation formatting is incorrect throughout: in-text citations appear as “Li et al. (2021)” where a parenthetical citation is intended. Please use \citep / \citet appropriately.
W2. All downstream experiments use a single fixed subject-level split (7:1:2), with the 10 seeds varying only training randomness. It is well known that performance on neuro-disorder benchmarks varies substantially across splits, so the reported standard deviations understate the true variance, and the paired t-tests across seeds do not establish robustness to data partitioning. In addition, several datasets (ABIDE-I/II, ADHD-200, REST-MDD) are multi-site, but the paper does not describe how sites are handled in the partitions (site-stratified, or possibly site-leaking across train/test). Without repeated splits and site-aware partitioning details, I am not fully convinced by the current results.
W3. The main text depends heavily on the appendix. For example, the loss-weight sensitivity results discussed at L453–L457 are not shown in the main paper at all (only Figs. 5–8 in Appendix H.5); if results are discussed in the main text, the corresponding figure should appear there. Similarly, the local temporal encoder that underlies PTRE (L167–171) is only described in Appendix A, including key design choices such as the 8-s blocks and 1-s query grid. This makes the method difficult to understand from the main text alone.
W4. The contribution of PTRE is unclear. (a) The model already interpolates every scan onto a fixed 1-s query grid in 8-s blocks (Appendix A.2), which by itself removes TR heterogeneity; it is not clear what PTRE adds over simply resampling all scans to a common TR. (b) Once blocks are fixed-length and non-overlapping, the block-center timestamp is 8b + 4, so PTRE’s phase is a linear function of the block index and the learnable multiplier can absorb the scale. Unless blocks are missing, PTRE appears nearly equivalent to standard RoPE over block indices. (c) The ablation “w/o PTRE” does not specify what replaces it (block-index RoPE, sinusoidal PE, or no position encoding), so it cannot distinguish “physical time vs. frame index” from “rotary vs. no position encoding.” A comparison against block-index RoPE is needed to support the central claim.
W5. The motivation for spherical coordinates in SAE over Euclidean coordinates (or pairwise Euclidean distance) is not explained, and the paper itself notes that the encoding does not reflect Euclidean distance (Appendix B.1). An ablation comparing spherical rotary encoding against a Euclidean-coordinate or distance-based variant is needed.
W6. The four loss terms are not ablated. The paper reports sensitivity to the loss weights (Appendix H.5), but a sweep over weights does not show whether each term is necessary; a leave-one-out ablation would clarify the contribution of each component of the pretraining objective, which is a core part of the claimed contribution.
Questions for authors
Q1. The CC200 atlas (200 ROIs) is coarser than parcellations used by some baselines (e.g., Brain-JEPA uses 450 ROIs). Did the authors try finer parcellations with STH-Brain? Figure 3 shows CC200 outperforming Power264; if the gap persists at higher resolution, what is the plausible explanation, given that the method is designed to preserve ROI-level states?
Q2. All comparisons are with full fine-tuning. Since the goal of pretraining is a generalizable representation space, did the authors evaluate STH-Brain and the baselines under linear probing (or frozen-encoder) protocols, as done in Brain-Semantoks? This would help separate representation quality from fine-tuning capacity.