- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
THEME addresses brain tumour segmentation when 1,2 or 3 of the four MRI sequences (FLAIR, T1, T1ce, T2) are missing. Instead of synthesising missing images or completing a shared latent code, each modality expert produces a spatial feature map and three 64d summaries z_{m,t}, one per region task (ET, TC/ET, WT/TC), via query-based pooling biased by region masks (GT in training, detached expert predictions at inference). A cross-attention module predicts the summaries of missing modalities, with heads evaluated in order.
A global router fuses observed spatial maps; a task router fuses observed and completed summaries per task; each summary drives a spatial gate and residual in a shared decoder whose outputs are summed into ET/TC/WT logits.
Training uses two nested availability subsets per subject, with Dice+BCE on both views, an L1+cosine reconstruction loss against detached full-input summaries, ranking losses on outputs and evidence, and gate regularizers.
Evaluation covers all 15 non-empty modality subsets on BraTS 2018/2020 with 5 seeds, cumulative and leave-one-out ablations, an efficiency audit, and a Prostate158 study.
Critical strengths
THEME predicts a small summary per missing sequence and tumor region, supervised by what the full-input expert would produce. Spatial detail comes only from acquired sequences, so the model never relies on synthesized voxels. Indexing by region fits the problem, since a missing sequence affects ET and WT differently. The added cost is small (0.74M parameters, 40.7 GFLOPs).
Moreover the authors contribute with an informative evaluation design. All 15 subsets, two releases, five seeds, separate WT/TC/ET scores and a severe-missing summary. Seed and subset variation are kept apart, and the per-subset results show that which sequence is available matters, not only how many.
Critical weaknesses
W1: The paper does not test its main idea directly. A test of this kind would replace the three per-task summaries with one summary per modality and keep everything else the same. In Table 2, the last row only adds the training losses, some of which follow SimMLM [11], and it still gains +3.57 WT and +3.55 TC on BraTS 2018. So it is not clear how much of the gain comes from the architecture and how much from the training.
Table 2 names the module added at each step, but it does not describe the Baseline itself, how missing summaries are handled before EC is added, or which training objective the intermediate rows use. Without this, the gains of each step are hard to interpret.
Some ablations are missing. First, a missing slot is filled with a fixed mix of the mean of the observed summaries and the predicted summary, so it is the mean plus a down-weighted learned correction. Results with the mean alone are not reported, so the value of the completion module is unclear, and the mixing weights are not justified. Second, summaries are pooled with ground-truth masks in training but predicted masks at test time. Testing with ground-truth masks would show how much this costs. Third, the error of the predicted summaries is never reported, for example as a function of |A|.
The paper contrasts its approach with image synthesis, but the only synthesis-based methods (Hyper-GAE, MouseGAN++) appear in the appendix. A two-stage baseline, which synthesizes the missing sequences with a many-to-many model (e.g., a conditional diffusion model) and then segments the completed volume with a strong segmenter, would directly test whether completing compact summaries is better than completing images.
The low similarities in Table 5 are expected, since each summary is pooled over a different region with its own query. They do not show that per-task summaries help segmentation.
W2: Some per-subset values are hard to interpret under the subset definitions in Table S1. In Table S11 (BraTS 2020 ET), THEME obtains about 80 on subsets without T1ce (T1-only, T1+T2, FLAIR+T1+T2), close to its full-input score, but 57.0 to 62.7 on five of the eight subsets that include T1ce. Similar cases appear in BraTS 2018 TC subset 11 and BraTS 2020 WT FLAIR-only. Across baselines, the single-sequence subset with the highest ET also falls in different columns. If the rows follow different subset orderings, the per-subset comparisons and the analysis in Section 4.4 cannot be read as reported.
The paper does not state whether the baselines were re-trained or taken from earlier work. The BraTS 2018 per-subset values of the baselines shared with AdaMM [29] appear to match those reported in the AdaMM work, which used a different split and training budget. This would conflict with the statement in 4.1 that all methods share the same split and evaluation, and it is unclear how the five-seed standard deviations in Table 1 were obtained for these methods.
Eleven of the 18 baselines appear only in the appendix, and some score above main-table entries (e.g., SMUNet's ET of 58.07). TMFormer, SMML, UniMRSeg, D3Seg, SIUM, and Konwer et al. are cited but not evaluated on BraTS. DC-Seg and M3AE are evaluated only on Prostate158, and PASSION and LS3M are not discussed. A modality-dropout U-Net is also missing.
Only Dice is reported, with no paired test, and the TC gain is within seed variation. The data splits are not stated, and it is unclear whether the 128^3 center crop used for validation is also applied at test time.
Note: The evaluation relies on BraTS 2018 and 2020, which are small by current standards and partly overlap, since the 2020 release includes most 2018 cases. Recent releases provide substantially larger cohorts (e.g., over 1200 pre-treatment training cases in BraTS 2023/2024) and have adopted lesion-wise Dice and HD95 as standard metrics. In addition, gains on pre-treatment glioma segmentation with complete inputs have become small in recent challenges, and the community has moved toward settings such as post-treatment glioma, where segmentation is harder and clinically more relevant. Evaluating on a recent release with lesion-wise metrics, and ideally on a post-treatment cohort, would strengthen the relevance of the contribution.
Questions for authors
- Do all rows of the per-subset tables follow the subset order of Table S1? If not, please provide corrected tables.
- Which baselines were re-trained under your protocol, and which were taken from earlier work? For the latter, what do the standard deviations in Table 1 measure?
- For every method, the BraTS 2020 ET values differ from the BraTS 2018 values by a nearly constant offset, and the across-subset standard deviations are the same. How were the BraTS 2020 results produced?
- What does the Table 2 Baseline contain? Can you report the ablations listed in W1, in particular a single shared summary per modality?
- What is the training cost per epoch compared with the baselines?