- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes FReMCI, which combines low-frequency amplitude sharing (PLAS), reliability-guided region mixing (ReSRI), and foreground memory injection (MCI) for semi-supervised ultrasound segmentation. The authors report the highest average Dice on five tasks from four dataset sources at three labeling ratios with eighteen baselines. They also provide component ablations and parameter sensitivity analysis.
Critical strengths
- Promising results on semi-supervised segmentation. Tab. 1 to Tab. 3 report the highest average Dice at different labeling ratios. Tab. 4 also shows gains from different module combinations.
- Simple inference. The additional modules are used only during training and do not add steps at inference.
Critical weaknesses
-
The experiments do not show reliable advantages over the baselines. Actual baseline settings and training costs are not fully reported, making it difficult to assess the trade-off between performance and training cost across methods. The experimental discussion mainly repeats table scores without analyzing trends or outliers. In particular, it does not explain why Dice on UDIAT drops when the labeled fraction increases from 3/10 to 1/2. Missing setup details, such as data splits, and the encrypted supplement make the codes difficult to check and reduce confidence in reproducibility. These gaps leave it unclear whether the gains are stable and worth the extra training cost.
-
The method description and ablations do not fully support the main design claims. The motivation sounds reasonable, but the specific claims in the Introduction need clearer support from prior work or direct experiments. Appendix D claims independence from memory, but Eq. 7 and Alg. 1 use a student prediction affected by memory to compute reliability. The figures also contain conflicting or missing connections. Fig. 1 does not clearly label the PLAS and ReSRI modules or show which operations belong to each module. In Fig. 2, "Top-k extract" is connected directly to the input image, while Section 3.3 states that the top-1 and top-2 values come from the teacher probabilities. Table 4 shows gains from some module combinations, but all variants share the reliability estimator and PLAS is only tested with ReSRI. These tests do not fully explain why the specific design choices are needed or whether the modules address the problems described.
Questions for authors
-
Supplement and data splits. Please provide accessible supplementary code and explain the construction of the Aneil/Aryan subsets, data splits, and labeled-sample selection. Are patients or image sequences kept separate across splits where relevant? Are test sets fixed across labeling ratios, and is each smaller labeled set included in the larger one?
-
Baseline results and training cost. Were the baseline results rerun for this study or taken from previous papers? If rerun, did all methods use the same data splits, labeled samples, preprocessing, and evaluation procedure? If taken from previous papers, please identify those results and explain whether their settings match this study. Please also report the actual backbones, pretraining, training budgets, runtime, and peak memory for FReMCI and the strongest baselines. Are the gains still present under comparable training budgets?
-
Results across labeling ratios. The average Dice advantage over the strongest baseline becomes smaller as the labeled fraction increases. Is this trend consistent across repeated runs? Please provide a comparison with a strong baseline at a higher labeling ratio, such as 3/4, and a 1/1-label fully supervised reference under comparable settings. How does the trade-off between labeling effort, training cost, and segmentation performance change across these ratios? Please also explain the UDIAT performance drop from 3/10 to 1/2 labels.
-
Method details and evidence. Appendix D states that pseudo-label computation is independent of memory, but Eq. 7 and Alg. 1 use a student prediction affected by memory to compute reliability. Please explain this difference and make the text, figures, and code consistent. What studies or experiments support the problems described in the Introduction and show that the modules solve them?
I thank the authors for the response and additional experiments.
However, some concerns are still not resolved. Splits are image-level without patient or sequence separation. Training cost and budget-matched comparisons cover one baseline on two datasets, and their gains lie within the reported deviations. Table 4 is still a single run, and the new reliability controls differ by less than one deviation. The protocol details appear only in the response and not in the revised paper.
In addition, the supplement does not seem to contain code for the new experiments.
In conclusion, I maintain my rating.