{"id":"6a5c8fbb-3b02-45c8-96af-6a38103826d9","arxiv_id":"2505.16283","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The framework adds CutMix augmentation and an entropy-variance uncertainty score to prototype consistency learning, reporting SOTA on three medical datasets, but the paper reprints the authors' own BIBM 2024 publication (reference [23]).","lead":"EPCL-JUDA combines CutMix-augmented labeled data, a joint uncertainty score, and prototype consistency learning to improve semi-supervised medical image segmentation. The authors report state-of-the-art Dice scores on left atrium, pancreas, and aortic dissection datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unlabeled prototype equations (8)-(9) use a ground-truth indicator 1[y_p=c] on data with no labels; literal reading leaks labels, intended reading leaves the pseudo-label mask undefined, so the SOTA claim is not supported as written.","rationale":"The reader's weakest assumption (JUQ reliability) is a real limitation but not the single most load-bearing one: the ablation in Table IV shows that replacing JUQ with entropy costs only 0.36 Dice, so the SOTA claim does not rest on JUQ alone. By contrast, the unlabeled prototype pooling in Eqs. (8)-(9) is the core mechanism for transferring unlabeled information; if it is leaky, the entire semi-supervised evaluation is invalid, and if it is merely a typo, the paper fails to specify the actual mask. The problem definition explicitly lacks labels for D_u, so the symbol y in those equations is undefined for unlabeled samples. This is an internal inconsistency, not a disagreement with consensus, and it directly undermines the central SOTA claim. The concrete check of re-deriving the prototypes from the defined pseudo-label will settle it. I recommend REJECT for the current version, with resubmission possible after correcting the equations, re-running experiments, and clarifying the masking procedure; the duplicated baseline numbers in Table II (e.g., UA-MT 77.26 at both 10% and 20% labeled) should also be checked, as they suggest the comparisons may not be reliable.","tokens_in":16159,"tokens_out":7992,"duration_ms":63686,"concrete_test":"Independently re-derive Eqs. (8)-(9) using only quantities defined for unlabeled data: replace 1[y_p=c] with either (i) argmax(hat_pl^p_{u,a}) or (ii) the raw probability hat_pl^p_{u,a}(c). If neither yields a well-defined prototype (e.g., division by zero, or no correspondence with the ablation pattern in Table IV), then the equations as written are not the implemented method; if the literal ground-truth version is needed to reproduce Table I, label leakage is confirmed. Also spot-check whether the baseline rows repeated identically across 10%/20% in Table II (UA-MT, SASSNet, DTC) match the cited original papers.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is empirical SOTA performance, so the most load-bearing condition is that the unlabeled prototype construction in Section III-D is valid and uses only unlabeled information. That condition fails in Eqs. (8)-(9). In Section III-A, labels y_a are defined only for the labeled subset D_l; the unlabeled subset D_u contains only images x_a. Yet Eqs. (8) and (9) compute unlabeled prototypes with the hard mask 1[y_a^p = c], the same ground-truth indicator used in Eq. (7) for labeled data. If these equations are taken literally, unlabeled prototypes are built from ground-truth labels, which leaks label information into the 'semi-supervised' pipeline and would trivially inflate the reported Dice/Jaccard numbers. If the authors intended the pseudo-label hat_pl^p_{u,a} instead, the paper never specifies how a probability vector becomes the mask 1[...] (argmax? threshold? soft weighting?), and the denominator sum_p 1[y_p=c] can be zero for absent classes. Either way, the method is not reproducible from the manuscript, and the SOTA superiority in Tables I-III cannot be verified. This is more fundamental than the JUQ formula concern: even a perfect JUQ would not fix the undefined or leaky masking in the prototype pooling step.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes EPCL-JUDA, a semi-supervised 3D medical image segmentation method built on a Mean-Teacher architecture. Labeled images are augmented with CutMix and concatenated with originals to produce labeled prototypes by masked average pooling. Original and augmented unlabeled data are processed separately, with a joint uncertainty quantification (JUQ) combining normalized prediction variance and entropy to weight pseudo-labels. Labeled and unlabeled prototypes are fused into global prototypes, and prototype-to-feature cosine similarity maps are supervised with labels and weighted pseudo-labels through consistency losses. A small prototype network is introduced to lower GPU memory relative to using decoder features. The paper reports Dice, Jaccard, 95HD, and ASD on Left Atrium, Pancreas-NIH, and Type B Aortic Dissection at 10% and 20% labeled ratios, claiming state-of-the-art results, plus ablations of JUQ, data augmentation, the prototype network, and unlabeled-data combination patterns.","tokens_in":16409,"tokens_out":6523,"duration_ms":50018,"significance":"The claimed contribution is incremental but potentially useful: augmenting the prototype pool with CutMix and weighting pseudo-labels by a product of variance- and entropy-based reliability are reasonable design choices, and the memory reduction from the prototype network is a concrete engineering benefit. The reported Dice improvements over strong baselines (e.g., 92.24 vs. 91.69 for UPCoL on LA at 20%) are consistent across three datasets and four metrics, and the ablations isolate the main components. However, the contributions as written are not verifiable: the unlabeled prototype equations appear to use ground-truth labels, the self-citation in [23] describes the identical method, and the quantitative comparisons lack error bars or significance tests. These issues must be resolved before the empirical superiority claim can be accepted.","major_comments":[{"comment":"In Section III-A, labels y_a are defined only for the labeled subset D_l, while D_u contains only images. Nevertheless, Eqs. (8) and (9) compute the unlabeled prototypes p_c^{u1} and p_c^{u2} using the ground-truth indicator 1[y_a^p = c], the same mask as the labeled prototype in Eq. (7). Taken literally, this leaks label information into the semi-supervised pipeline and would invalidate the reported results; if the intended mask is the reliable pseudo-label \\hat{pl}^p_{u,a} from Eq. (6), the paper never specifies how a probability vector is converted to the binary mask 1[\\cdot] (argmax, threshold, or soft weighting), and the denominator \\sum_p 1[y_a^p = c] can be zero for classes absent from a sample. The unlabeled prototype construction must be rewritten with an explicit pseudo-label mask, a defined threshold/soft-weighting rule, and a non-zero denominator handling, so that the SOTA claims in Tables I-III can be verified without label leakage.","section":"Section III-D, Eqs. (8)-(9)"},{"comment":"Section II-B and reference [23] present EPCL-JUDA [23] as prior work (\"Building upon these concepts, EPCL-JUDA [23] combines...\"), yet [23] has the identical title, the same authors, and reports the same core method. At the same time, the paper's contributions in Section I introduce EPCL-JUDA as a new proposal. The manuscript needs a clear statement of what this version adds over the BIBM 2024 paper (e.g., extended experiments, new ablations, memory analysis), or the novelty claim must be revised; otherwise the contribution is self-referential and cannot be evaluated.","section":"Section II-B and References"},{"comment":"Section III-C, Eq. (6) is not a valid reliability weighting as typeset. It reads \\hat{pl}^p_{u,a} = (1/(HWD))(1 - JUQ \\sum_p JUQ) \\otimes pl^p_{u,a}; the term \"1 - JUQ \\cdot \\sum_p JUQ\" is dimensionally mixed and can be negative or unbounded, and the factor 1/(HWD) is unexplained. This pseudo-label formula is the input to the unlabeled prototypes (Eqs. (8)-(9)) and the consistency losses (Eqs. (12)-(13)), so it must be corrected to the intended normalization (e.g., (1 - JUQ/\\sum_p JUQ)) and its range stated.","section":"Section III-C, Eq. (6)"},{"comment":"Section IV-B states that 5-fold cross-validation is used for LA and Pancreas-NIH and 3-fold for TBAD, yet Tables I-III report only single point estimates for every metric. Without means, standard deviations, or significance tests, the claimed superiority over previous state of the art (e.g., Table I: EPCL-JUQ 92.24 Dice vs. UPCoL 91.69) may be within fold-to-fold variability. Please report mean \\pm std over folds and, where possible, paired significance tests.","section":"Section IV-B and Tables I-III"},{"comment":"Section III-D, Eq. (10) defines p_c^u = \\lambda_1 p_c^l + \\lambda_2 p_c^m, but the surrounding text says that original and augmented unlabeled prototypes are fused to form p_c^u, and p_c^m is never defined. Additionally, \\lambda_1 and \\lambda_2 are described as fusion coefficients for the two unlabeled prototypes, which does not match the equation. This equation must be corrected to show how p_c^{u1} and p_c^{u2} are combined before the global fusion in Eq. (11).","section":"Section III-D, Eq. (10)"}],"minor_comments":[{"comment":"The acronym is inconsistent throughout: EPCL-JUDA in the title and abstract, EPCL-JUQ in the result tables and discussion, and EPCL-JUA in Section V and Fig. 3; standardize one name.","section":"Throughout"},{"comment":"Table IV lists the setting as \"Labeled 8(20%) Unlabled 0\", which contradicts Table I's 16(20%) labeled / 64(80%) unlabeled for LA; correct the scan counts and the \"Unlabled\" typo.","section":"Table IV"},{"comment":"Table II repeats the same values for UA-MT, SASSNet, and DTC under both 10% and 20% labeled ratios (e.g., UA-MT 77.26/63.82/11.90/3.06), which looks like a copy-paste error; the 20% rows should be actual measurements.","section":"Table II"},{"comment":"Fig. 3's caption contains garbled path strings instead of a readable description of the reliability maps.","section":"Fig. 3"},{"comment":"The code is only promised (\"will be released soon\") and no repository link is given; since the equations are not fully reproducible as written, providing the code would materially help verification.","section":"Abstract and Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"The main concern for the editor is novelty disclosure: reference [23] is the same paper by the same authors, so this arXiv manuscript appears to be a conference-to-journal extension without an explicit novelty statement. I would ask the editor to verify the relationship to the BIBM 2024 publication and whether the experimental tables in this version are new or extended. Also, the label-leakage notation issue in Eqs. (8)-(9) should be resolved before the paper is considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is the same paper as the authors' BIBM 2024 conference paper (ref [23], identical title and authors), and the text admits as much by citing EPCL-JUDA [23] in the related work. So the novelty claim collapses; there is no new contribution here beyond the earlier conference version.\n\nWhat is actually useful: the prototype network idea. Running decoder features through three 3D convs before masked average pooling cuts memory from ~18 GB to ~4 GB in their ablation, and it doesn't hurt accuracy. If that result is real, it is a transferable engineering trick worth borrowing. The broader recipe of CutMix-augmented labeled data plus entropy/variance uncertainty for pseudo-labels is a reasonable extension of UPCoL and BCP, and the reported Dice numbers are consistently a bit higher than baselines.\n\nThe soft spots are real. First, the duplication. Submitting the same work to arXiv after a conference publication is fine only if you disclose it; here they cite their own prior paper as if it were someone else's building block, which is misleading. Second, and worse, the unlabeled prototype equations (8) and (9) use 1[y_a^p=c], the ground-truth mask, on data that by definition has no labels. If taken literally, that leaks supervision and would trivially inflate the Dice/Jaccard numbers. If the authors meant the pseudo-label mask, they never say how a probability vector becomes that mask (argmax? threshold? soft?). So the method is unreproducible as written, and the SOTA claim is not supported. This is load-bearing, not a typo in a footnote. The JUQ formula (Eq. 5) is also ad hoc, with no derivation or validation beyond one qualitative figure. There are no error bars or significance tests despite cross-validation, and no code, so the empirical gains are unverifiable.\n\nWho gets value from this: people working on prototype-based semi-supervised segmentation might want to read the memory-reduction ablation and possibly steal that idea. But they should not cite this version as a clean source, and they should wait until the authors release a corrected manuscript with the mask defined and the prior publication disclosed. A serious editor should desk reject this as a duplicate submission; it is not a new paper. If the authors want a journal version, they'd need substantial new experiments, a real derivation of JUQ, and a fix to Equations (8)-(9).","headline":"A reprint of the authors' own BIBM 2024 paper with an undefined label-leaking mask in the unlabeled prototype equations; the memory trick is the only real novelty.","tokens_in":17005,"tokens_out":3763,"would_cite":false,"duration_ms":29387,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that enriching class prototypes with CutMix-augmented data and JUQ-weighted pseudo-labels yields state-of-the-art semi-supervised medical image segmentation.","keywords":["semi-supervised learning","medical image segmentation","prototype learning","uncertainty quantification","data augmentation","CutMix","Mean Teacher","memory-efficient 3D segmentation"],"falsifier":"A controlled experiment on the same framework comparing JUQ against plain entropy as the uncertainty measure, evaluated on a dataset with added label noise or a domain shift: if entropy matches or beats JUQ there, the claimed benefit of the joint measure would be falsified. More directly, computing the correlation between JUQ weights and voxel-wise prediction correctness on a held-out set would test whether the reliability map ranks voxels correctly.","tokens_in":15900,"feed_emoji":"🩻","tokens_out":4411,"duration_ms":33597,"temperature":0.7,"pith_summary":"This paper tries to establish that semi-supervised medical image segmentation can be improved by enlarging the pool of samples used to build class prototypes, adding CutMix augmentation and a joint uncertainty measure to filter pseudo-labels. The method, EPCL-JUDA, builds on the Mean-Teacher framework and fuses prototypes from labeled, original unlabeled, and augmented unlabeled data into global prototypes. The authors report Dice 92.24, Jaccard 85.57, 95HD 4.63, and ASD 1.47 on the Left Atrium dataset at 20% labeled ratio, with leading numbers on Pancreas-NIH and Type B Aortic Dissection. They also claim a prototype network cuts GPU memory substantially, from 18.65 GB to 4.21 GB when generating features from the first decoder layer. If true, this would ease both the annotation burden and the hardware demands of 3D medical segmentation.","feed_headline":"Augmented prototypes lift semi-supervised segmentation","feed_subtitle":"Reports Dice 92.24 on the Left Atrium with 20% labels, and a 4x cut in prototype GPU memory.","key_machinery":"The central object is the global class prototype, formed by fusing labeled and unlabeled prototypes. Labeled prototypes are computed by masked average pooling on features from original plus CutMix-augmented labeled data; unlabeled prototypes are pooled separately from original and augmented unlabeled features, masked by pseudo-labels weighted by JUQ. JUQ multiplies normalized distributional uncertainty (exponential of negative variance) with normalized complement entropy to produce per-voxel reliability weights. A prototype network of three 3D convolutional layers compresses feature maps to the number of classes before upsampling, which is what reduces memory. Cosine similarity maps between global prototypes and student features then drive the consistency losses.","core_discovery":"The central claim is that EPCL-JUDA outperforms previous state-of-the-art semi-supervised segmentation methods by combining four ingredients: CutMix-augmented labeled data for prototype generation, joint uncertainty quantification (JUQ) to weight pseudo-labels, fusion of labeled and unlabeled prototypes into global prototypes, and a lightweight prototype network for memory-efficient feature compression. On the LA dataset at 20% labeled ratio, the method achieves Dice 92.24, Jaccard 85.57, 95HD 4.63, and ASD 1.47; on Pancreas-NIH and TBAD it reports similarly leading scores. The ablation studies attribute the largest gain to the data augmentation, with JUQ and the prototype network contributing smaller but consistent improvements.","pith_inferences":["The JUQ formula is hand-designed and unvalidated against ground-truth voxel correctness; its advantage over plain entropy may be dataset-dependent and should be tested under label noise or domain shift.","The memory savings from the prototype network could enable larger batch sizes or higher-resolution inputs in 3D segmentation, which might further improve accuracy beyond what the paper reports.","Since CutMix on labeled data enriches prototype diversity, other augmentation strategies that preserve label alignment, such as Mixup or controlled synthetic deformations, could be plugged into the same fusion framework.","Equation 6 applies a soft reliability weighting to pseudo-labels; a hard threshold variant might behave differently under distribution shift and is worth exploring."],"forward_implications":["If the reported gains hold, prototype-based semi-supervised segmentation can benefit directly from data augmentation on the labeled branch, not just on the unlabeled branch.","The JUQ weighting scheme could be transferred to other pseudo-label pipelines as a generic per-voxel confidence filter.","The prototype network shows that generating prototypes from low-dimensional class-feature maps is a viable memory-saving alternative to using high-dimensional decoder features.","The strong results on the aortic dissection dataset suggest the method handles thin, complex anatomical structures such as vessel walls."],"supporting_citations":[{"why":"UPCoL is the prototype consistency baseline that EPCL-JUDA extends, and its decoder-feature prototype approach is the memory baseline the prototype network improves upon.","marker":"[26]"},{"why":"CutMix supplies the augmentation that creates enriched labeled samples for prototype generation, a core component of the method.","marker":"[53]"},{"why":"Mean Teacher provides the EMA teacher-student framework that underlies the consistency learning architecture.","marker":"[37]"},{"why":"Masked average pooling from PANet is the operation used to compute class prototypes from features.","marker":"[39]"},{"why":"FUSSNet fuses two sources of uncertainty and is a direct uncertainty-modeling baseline that JUQ builds on and compares against.","marker":"[46]"},{"why":"UA-MT introduces uncertainty-guided pseudo-label filtering in the Mean-Teacher setting, motivating the reliability weighting used in JUQ.","marker":"[52]"}],"fun_headline_variants":["Prototype uncertainty beats SOTA in semi-supervised segmentation","Uncertainty and augmentation cut memory 4x, boost segmentation","Uncertainty-aware prototypes cut GPU memory 4x, lift segmentation","EPCL-JUDA: fewer labels, better segmentation","Prototype consistency with uncertainty slashes GPU memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The joint uncertainty formula (JUQ) is assumed to rank voxel reliability correctly for both original and augmented unlabeled data, but the paper does not validate this ranking against ground truth; if the weights mis-rank voxels, the pseudo-labels and unlabeled prototypes built from them are corrupted.","fun_headline_variants_meta":{"raw":{"variants":["Prototype uncertainty beats SOTA in semi-supervised segmentation","Uncertainty and augmentation cut memory 4x, boost segmentation","Uncertainty-aware prototypes cut GPU memory 4x, lift segmentation","EPCL-JUDA: fewer labels, better segmentation","Prototype consistency with uncertainty slashes GPU memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001206,"raw_usage":{"total_tokens":4951,"prompt_tokens":908,"completion_tokens":4043,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":3959}},"tokens_in":524,"tokens_out":4043,"duration_ms":25227,"temperature":1.0,"reasoning_tokens":3959,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:04:18.842612+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment on the same framework comparing JUQ against plain entropy as the uncertainty measure, evaluated on a dataset with added label noise or a domain shift: if entropy matches or beats JUQ there, the claimed benefit of the joint measure would be falsified. More directly, computing the correlation between JUQ weights and voxel-wise prediction correctness on a held-out set would test whether the reliability map ranks voxels correctly.","supporting_citations":[{"cited_title":"In: International Conference on Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"UPCoL is the prototype consistency baseline that EPCL-JUDA extends, and its decoder-feature prototype approach is the memory baseline the prototype network improves upon."},{"cited_title":"IEEE transactions on medical imaging 41(3), 608– 620 (2021)","cited_arxiv_id":null,"evidence_quote":"Mean Teacher provides the EMA teacher-student framework that underlies the consistency learning architecture."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"Masked average pooling from PANet is the operation used to compute class prototypes from features."},{"cited_title":"In: Medical Image Computing and Computer Assisted Intervention– MICCAI 2019: 22nd International Conference, Shenzhen, China, Octo- ber 13–17, 2019, Proceedings, Part II 22","cited_arxiv_id":null,"evidence_quote":"UA-MT introduces uncertainty-guided pseudo-label filtering in the Mean-Teacher setting, motivating the reliability weighting used in JUQ."}],"review_version":1}