{"id":"9c54eb83-bd22-4781-a8c5-0714d4c0d058","arxiv_id":"2505.19733","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A semi-supervised framework that decomposes T1 and FA MRI features into shared and sequence-specific parts, then uses reliable pseudo-labels from unlabeled data, improves visual pathway segmentation accuracy.","lead":"This paper proposes a semi-supervised deep learning method that separates shared and unique information from two MRI sequences to better outline the brain's visual pathway. It uses unlabeled scans to reduce the need for manual labels, and the authors report gains over existing methods on three datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Statistical significance claims are unsupported: paired t-tests on n=2 test sets (MDM, multimodality) cannot yield p<0.05, and the 'Monte Carlo simulations' used to augment them are not described or validated.","rationale":"The reader's weakest_assumption focuses on the CFD decomposition ansatz in Eqs. (1)-(2). While this is a legitimate concern about the method's interpretability and whether the learned 'unique' features are truly unique, it is not the most load-bearing threat to the central claim. The central claim is empirical: the proposed framework 'achieves significant enhancements in segmentation performance' (Section IV.A). The evidence for this claim consists of DSC comparisons across three datasets, with statistical significance asserted via paired t-tests. On the MDM and multimodality brain MRI datasets, the test set has only 2 subjects. A paired t-test with n=2 has one degree of freedom and cannot yield p<0.05 for differences of the reported magnitude unless the paired differences are perfectly consistent; the paper's two-tailed p-values with asterisks thus require a non-standard procedure. The paper states it runs 'two Monte Carlo simulations' before computing the t-test, but the description is absent, so the significance claim is not verifiable. The reader's rationale does note the small test sets, so we partially agree, but the reader's formal weakest_assumption is the decomposition, which we consider secondary. Our concrete test—recomputing p-values on the actual 2 subjects without augmentation—would determine whether the statistical claim holds. If it does not, the paper's conclusion of superiority reduces to small point estimates, and the verdict should remain conditional pending proper statistical validation or a larger test set. We therefore keep the reader's CONDITIONAL verdict (UNCHANGED) while shifting the primary concern to statistical validity.","tokens_in":18731,"tokens_out":6798,"duration_ms":69864,"concrete_test":"Recompute the paired t-tests in Fig. 7b and Fig. 12 using only the actual test subjects (n=2) without the unspecified Monte Carlo augmentation; if any p<0.05 remains, report it. Also request the per-subject DSC values for all methods on the MDM and multimodality test sets to verify whether the claimed significance survives a legitimate permutation or bootstrap procedure on the real 2 subjects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of 'significant enhancements' (Section IV.A) is supported by paired t-tests on test sets of 10 (HCP), 2 (MDM), and 2 (multimodality) subjects. For the MDM and multimodality datasets, n=2 gives one degree of freedom; the critical t for a two-tailed p<0.05 is 12.71, so the reported DSC differences of 0.09 (Ours 0.70 vs UG-MCL 0.61, Table III) cannot be significant unless the within-pair variance is essentially zero. The paper acknowledges the tiny test set and states 'we run two Monte Carlo simulations before computing the paired t-test' (Fig. 7b, Fig. 12 note), but does not describe what is simulated, what null distribution is used, or how the p-values are derived. This is not a standard method for increasing statistical power and effectively manufactures significance from two subjects. In addition, hyperparameters (α, β, M, thres) are tuned by evaluating performance on the same test sets (Section IV.B), so the reported gains may reflect test-set overfitting. Without per-subject scores and a valid inference procedure, the observed DSC margins are point estimates on two subjects and cannot support the paper's significance claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semi-supervised framework for delineating the visual pathway (VP) from multi-parametric MRI (T1-weighted and FA images). The method has two main components: a correlation-constrained feature decomposition (CFD) module, which decomposes each sequence into unique and non-unique features using multi-parametric convolutional sparse coding, and a consistency-based sample enhancement (CSE) module, which selects reliable pseudo-labels from unlabeled data within a mean-teacher framework. The framework is evaluated on the HCP dataset, an MDM dataset, and a multimodality brain MRI dataset. The authors report DSC improvements over six semi-supervised baselines and comparable or better performance than fully supervised methods in low-label regimes.","tokens_in":19030,"tokens_out":3228,"duration_ms":37153,"significance":"If the reported results are reliable, the work would be a useful contribution to annotation-efficient VP delineation and to the broader literature on semi-supervised multi-modal medical image segmentation. The CFD idea of explicitly separating shared and unique sequence features is interesting and the use of a consistency-based pseudo-label selection mechanism is well motivated. The paper also includes comparisons with several recent SSL methods, ablations of the main components, and a model-complexity analysis. However, the current evidence for the central claim of 'significant enhancements' is weak because the statistical analysis is invalid for the smallest test sets, hyperparameters appear to be tuned on the same test sets, and an important internal inconsistency in the reported hyperparameters is not resolved.","major_comments":[{"comment":"The statistical significance claim for the MDM and multimodality datasets is not supported by the reported procedure. Both test sets contain only 2 subjects, so a paired t-test has 1 degree of freedom and requires |t| > 12.71 for two-tailed p < 0.05. The reported DSC differences (e.g., 0.70 vs 0.61 in Table III) cannot produce such a p-value unless the within-pair variance is essentially zero. The statement that 'we run two Monte Carlo simulations before computing the paired t-test' (Fig. 7b and Fig. 12 notes) is not accompanied by any description of what is simulated, what null distribution is used, or how the p-values are derived. Without per-subject scores and a valid inference procedure, the observed margins are point estimates on two subjects and cannot support the paper's significance claim.","section":"IV.A.2 / Fig. 7b"},{"comment":"The hyperparameters (M, thres, alpha, beta) appear to be selected by evaluating performance on the same test sets used to report the main results. Section IV.B describes varying M and thres while reporting DSC, HD95, and ASD on what appears to be the test set, and Fig. 8 shows alpha/beta selection using DSC and ASD on the MDM dataset. No validation split is described anywhere. This creates a risk of test-set overfitting and makes the reported gains over baselines difficult to interpret. The authors should either introduce a separate validation set or use nested cross-validation for hyperparameter selection.","section":"IV.B / Tables IV-V / Fig. 8"},{"comment":"There is an internal inconsistency in the reported hyperparameters. Section III.B.2 states 'We set alpha = 10, beta = 1, M = 3, and threshold value thres = 0.05 as optimal values', but Section IV.B.1 states 'To obtain the optimal values of M, we fix alpha = 1, beta = 1, and thres = 0.05'. Likewise, Tables IV and V, which study M and thres, appear to use alpha = 1, while the main results and the component ablation (Table VI) use alpha = 10. The authors need to clarify which hyperparameter values were used for each reported result, since this affects the validity of the ablation conclusions.","section":"IV.B.1 vs III.B.2"},{"comment":"The central assumption of the CFD module is that each MRI sequence can be decomposed exactly into a non-unique component plus the convolution of learned unique filters and sparse feature maps. This assumption is never validated: the reconstruction residual of Eqs. (1)-(2) is not reported, and the t-SNE visualization in Fig. 1 is qualitative. If the decomposition is not faithful, the claimed benefit of separating unique and non-unique features could be an artifact of the decomposition loss rather than a genuine property of the data. The authors should report the reconstruction error of the CFD module and ideally compare the learned unique features against an alternative decomposition or a simple concatenation baseline with the same network.","section":"II.A / Eqs. (1)-(2)"}],"minor_comments":[{"comment":"The abstract says 'two public datasets, and one in-house Multi-Shell Diffusion MRI (MDM) dataset', but Section III.A states 'Two datasets were used in this study, including the Human Connectome Project (HCP) dataset, the Multi-shell Diffusion MRI (MDM) dataset, and the multimodal brain MRI dataset'. This is a three-dataset evaluation, and the MDM dataset appears to be a public traveling-subject dataset (reference [23]) rather than in-house. Please correct the wording.","section":"III.A / Abstract"},{"comment":"The decomposition loss in Eq. (7) is written as PCC(f_T1, f_FA)^2 / (epsilon + PCC(c_T1, c_FA)). Minimizing this loss reduces the unique-feature correlation and, through the denominator, also rewards a larger non-unique correlation, but the sign conventions and the role of epsilon are not explained. Please clarify the intended optimization direction and why a negative PCC in the denominator is not problematic.","section":"II.A / Eq. (7)"},{"comment":"The text states that on the MDM dataset the method showed 'comparable performance over FS-1, FS-2, and FS-3', but Table III reports only FS-1. FS-2 and FS-3 appear in Fig. 5 but not in Table III; please include all three fully supervised baselines in the corresponding table for reproducibility.","section":"IV.A.3 / Table III"},{"comment":"The HD95 value for FS-1 in Table II is reported as 1.25 ± 0.00, which is suspicious for a standard deviation and suggests either a rounding artifact or a measurement issue. Please double-check this entry and report the appropriate precision.","section":"Table II / FS-1 row"},{"comment":"Section IV.B.3 refers to a hyperparameter lambda_1 that is 'defined as a time-dependent Gaussian function', but this term does not appear in Eq. (14), which uses alpha, delta, and beta. Please align the notation between the text and the objective function.","section":"IV.B.3 / Equation (14)"}],"recommendation":"major_revision","confidential_remarks":"The central method is plausible and the topic is relevant, but the paper's main evidence base needs substantial rework before it can be accepted. In particular, the statistical significance claim on n=2 test sets should be replaced with per-subject results and a valid inference procedure, and the hyperparameter tuning protocol needs to be clarified to exclude test-set overfitting. I also note that the ground-truth generation method (reference [25]) and the FS-1 baseline (reference [32]) share authors with the current paper; this is not circular, but the overlap should be explicitly disclosed so that the comparison is interpreted correctly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is a reasonable new assembly: they take Deng et al.'s disentangled convolutional sparse coding, add a Pearson-correlation loss to separate unique from non-unique features across T1 and FA, and wrap it in a mean-teacher framework with a consistency-based pseudo-label filter. For visual pathway delineation with scarce labels, that combination is genuinely new, and the ablations suggest both components contribute. The reported DSC gains over UG-MCL and other SSL baselines are consistent across three datasets, and the qualitative figures show visibly better delineation of the optic nerve and tract. I also give them credit for testing on an in-house MDM dataset and comparing against six SSL and three fully supervised methods.\n\nThe paper's load-bearing statistical claims do not hold up, however. The paired t-tests on the MDM and multimodality datasets have n=2; with one degree of freedom, p<0.05 is impossible unless the within-pair variance is essentially zero. The paper's own note says they ran 'two Monte Carlo simulations' before the t-test, but there is no description of what was simulated or how the null distribution was constructed. That is not a standard fix; it reads as a way to manufacture significance. Even on HCP (n=10), the per-subject scores are not reported, so the p-values in Fig. 7a are hard to verify.\n\nThe experimental design has a second soft spot: hyperparameters (alpha, beta, M, threshold) appear to be tuned on the same test sets. The ablation in Section IV.B fixes alpha=1 for the M sweep, but the loss-weight ablation and Table VI use alpha=10. That internal inconsistency suggests the reported optimal values are not stable, and without a validation split the gains could partly reflect test-set overfitting.\n\nA more technical concern: equations (1)-(2) assume exact decomposition of each image into non-unique features plus a convolution of unique filters and sparse maps. The reconstruction residual is never quantified, and the filters are learned jointly, so the decomposition could be fitting the loss rather than capturing true shared and unique structure. The t-SNE visualization is suggestive, but not proof.\n\nBottom line: the method is plausible and the direction is worth pursuing, but the current paper overstates its statistical support. I would not cite it as-is. It deserves peer review, though, because the core idea is sound and the weaknesses are fixable with a proper validation split, per-subject results, a described resampling procedure, and ideally code and data release. A serious referee should ask for those before acceptance.","headline":"Plausible new combination of disentangled sparse coding and mean-teacher consistency for visual pathway segmentation, but the statistical evidence is too thin to support the significance claims as written.","tokens_in":731,"tokens_out":1087,"would_cite":false,"duration_ms":26585,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A semi-supervised framework that splits T1 and FA MRI scans into shared and sequence-specific features delineates the visual pathway more accurately than seven existing methods, reaching a Dice score of 0.85 on HCP with only 16 labeled…","keywords":["visual pathway delineation","multi-parametric MRI","semi-supervised learning","feature decomposition","correlation-constrained","consistency-based sample enhancement","mean teacher","convolutional sparse coding"],"falsifier":"Quantify the reconstruction residual of Eqs. (1) and (2) on held-out T1 and FA images: if the residual is large relative to the signal, or if a plain two-stream U-Net receiving the same multi-parametric input achieves the same Dice score at the same label budgets, the central claim that the decomposition drives the improvement would be unsupported.","tokens_in":1738,"feed_emoji":"🧠","tokens_out":2918,"duration_ms":75779,"temperature":0.7,"pith_summary":"The paper tries to establish that the visual pathway, the brain structures that carry visual signals and are costly to label, can be delineated accurately from multi-parametric MRI using very few manual labels. The proposed framework combines a correlation-constrained feature decomposition (CFD) that splits each MRI sequence into unique and non-unique features with a consistency-based sample enhancement (CSE) module that harvests reliable pseudo-labels from unlabeled data. The authors claim this combination outperforms six semi-supervised baselines on HCP, MDM, and multimodal brain MRI datasets, and that with very scarce labels it matches or beats fully supervised models trained on far more annotations. If true, the method would make clinically usable visual pathway mapping feasible with as few as 1 to 16 labeled subjects per dataset.","feed_headline":"Visual pathway delineation hits 0.85 DSC with just 16 labeled scans","feed_subtitle":"Decomposing T1 and FA MRI into shared and unique features lets the model exploit unlabeled scans.","key_machinery":"The load-bearing machinery is the correlation-constrained feature decomposition (CFD), which models each MRI sequence $x$ as a sum of a non-unique component $c$ and a convolution of sequence-specific filters with sparse feature maps (Eqs. 1 and 2). The unique feature maps are predicted by a learned convolutional sparse coding network and serve as inputs to the final delineation branch, while the non-unique components align the two sequences during training. A decomposition loss $L_{dcp}$ drives the Pearson correlation between non-unique features up and between unique features down. The second moving part, consistency-based sample enhancement (CSE), uses a teacher model (an exponential moving average of the student) and $M$ noise-augmented passes; unlabeled samples whose prediction variance exceeds a threshold are excluded from the unsupervised consistency loss.","core_discovery":"The central claim is that decomposing the two MRI sequences, T1-weighted and fractional anisotropy, into unique and non-unique components, and feeding only the unique features into the final segmentation stage, lets a mean-teacher semi-supervised network delineate the visual pathway from very few labeled scans. The decomposition is enforced by a loss that maximizes the Pearson correlation of non-unique features across sequences and minimizes the correlation of unique features, while the CSE module discards unlabeled samples whose student and teacher predictions are inconsistent across multiple noise-augmented views. The paper reports a DSC of 0.85 on the HCP test set with 16 labels against 0.79 for the best semi-supervised baseline (UG-MCL), 0.81 with 8 labels, 0.74 on the MDM dataset with 3 labels, and up to a 10-point DSC improvement on the multimodal brain MRI dataset with 2 labeled subjects.","pith_inferences":["The decomposition ansatz of Eqs. (1) and (2) is never validated by measuring the reconstruction residual on held-out images, so a direct test would be to quantify that residual and check whether it correlates with segmentation accuracy.","The method's reliance on paired T1 and FA images may transfer to other multi-parametric MRI combinations, but the paper's own comparison shows T1+FA clearly beats T1+T2 or T2+FA, so sequence choice is load-bearing.","The CSE module selects consistent samples with a hard threshold of 0.05; a softer weighting by the inconsistency score could extract more signal from the unlabeled pool without risking overconfidence, which the paper does not explore.","If the CFD decomposition genuinely separates shared from sequence-specific structure, the unique features could serve downstream tasks beyond binary delineation, such as tractography initialization or quantifying structural changes in visual disorders."],"forward_implications":["With 16 labeled HCP subjects, the method reaches a Dice score of 0.85, a 6-point gain over UG-MCL, the best semi-supervised baseline.","With only 1 labeled MDM subject, the method reaches 0.70 DSC, a 9-point gain over UG-MCL and a 36-point gain over the supervised baseline.","In low-label regimes the method matches or exceeds fully supervised models: on MDM with 3 labels it reaches 0.74 DSC versus 0.72 for FS-1 trained on all 8 labeled subjects.","Ablations attribute a 3-point DSC gain on MDM to the CSE module and a 1-point gain on both datasets to the decomposition loss.","On the multimodal brain MRI dataset, the method improves DSC by up to 10 points over the six compared semi-supervised methods using only 2 labeled subjects."],"supporting_citations":[{"why":"Supplies the disentangled convolutional sparse coding approach that the CFD module adapts to MRI sequences.","marker":"[15]"},{"why":"Supplies the mean-teacher self-ensembling architecture that the CSE module builds on for semi-supervised learning.","marker":"[17]"},{"why":"Supplies the learned convolutional sparse coding (LCSC) solver used to compute the unique feature maps.","marker":"[19]"},{"why":"Provides the HCP dataset and its high-quality dMRI and T1-weighted acquisition protocol.","marker":"[22]"},{"why":"Provides the preprocessing and visual pathway ground-truth generation pipeline used across all datasets.","marker":"[25]"},{"why":"Supplies the U-shaped encoder-decoder delineation architecture and cross-modal supervision design that the segmentation branch is similar to.","marker":"[26]"},{"why":"Supplies the uncertainty-aware self-ensembling baseline (UA-MT) that the CSE module extends.","marker":"[27]"},{"why":"Supplies the UG-MCL baseline, the strongest semi-supervised competitor that the proposed method is compared against and outperforms.","marker":"[29]"}],"fun_headline_variants":["Semi-supervised VP delineation hits 0.85 DSC with 16 labels","Decomposing T1 and FA features improves visual pathway segmentation","16 labeled scans: 0.85 DSC for visual pathway delineation","Cross-sequence feature split: few labels, high VP accuracy","Mean-teacher framework exploits unlabeled MRI for VP delineation"],"cache_read_input_tokens":21632,"weakest_assumption_plain":"The framework assumes each MRI sequence is exactly a sum of a shared non-unique component and a convolution of learned unique filters with sparse maps; if that decomposition is not faithful, the claimed benefit of separating features could be an artifact of the loss rather than real structure in the data.","fun_headline_variants_meta":{"raw":{"variants":["Semi-supervised VP delineation hits 0.85 DSC with 16 labels","Decomposing T1 and FA features improves visual pathway segmentation","16 labeled scans: 0.85 DSC for visual pathway delineation","Cross-sequence feature split: few labels, high VP accuracy","Mean-teacher framework exploits unlabeled MRI for VP delineation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1572,"prompt_tokens":947,"completion_tokens":625,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":533}},"tokens_in":563,"tokens_out":625,"duration_ms":84554,"temperature":1.0,"reasoning_tokens":533,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:06:56.297082+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Quantify the reconstruction residual of Eqs. (1) and (2) on held-out T1 and FA images: if the residual is large relative to the signal, or if a plain two-stream U-Net receiving the same multi-parametric input achieves the same Dice score at the same label budgets, the central claim that the decomposition drives the improvement would be unsupported.","supporting_citations":[{"cited_title":"Spatial probabilistic distribution map-based two-channel 3D U-net for visual pathway segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the disentangled convolutional sparse coding approach that the CFD module adapts to MRI sequences."},{"cited_title":"Interpretable Multi- modal Image Registration Network Based on Disentangled Convolu- tional Sparse Coding,","cited_arxiv_id":null,"evidence_quote":"Supplies the mean-teacher self-ensembling architecture that the CSE module builds on for semi-supervised learning."},{"cited_title":"Semi-supervised strong-teacher consis- tency learning for few-shot cardiac mri image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the HCP dataset and its high-quality dMRI and T1-weighted acquisition protocol."},{"cited_title":"Reproducibility of multi-shell diffusion tractography on traveling subjects: a multicenter study prospective,","cited_arxiv_id":null,"evidence_quote":"Provides the preprocessing and visual pathway ground-truth generation pipeline used across all datasets."},{"cited_title":"Test– retest reproducibility of a multi-atlas automated segmentation tool on multimodality brain mri,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-shaped encoder-decoder delineation architecture and cross-modal supervision design that the segmentation branch is similar to."},{"cited_title":"CNTSeg: A multimodal deep-learning-based network for cranial nerves tract segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the uncertainty-aware self-ensembling baseline (UA-MT) that the CSE module extends."},{"cited_title":"Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the UG-MCL baseline, the strongest semi-supervised competitor that the proposed method is compared against and outperforms."}],"review_version":1}