{"id":"b1f2a5ac-3a70-4c8f-80b2-404ce664cbe1","arxiv_id":"2506.10325","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SWDL-Net improves semi-supervised intracranial hemorrhage segmentation by learning from differences between a Laplacian pyramid upsampler and a convolutional upsampler, reaching 89.3% Dice with 2% labeled data.","lead":"This paper introduces SWDL-Net, a semi-supervised deep learning method for 3D brain hemorrhage segmentation from CT scans, which combines two different upsampling routes and learns from their differences to work with very few labeled examples. A generalist reader might care because it shows a way to achieve strong segmentation accuracy when expert annotations are scarce, a common bottleneck in medical imaging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set hyperparameter selection in Sec. IV-C inflates the reported SOTA margins; no validation split is described in Sec. III-F.","rationale":"The reader's weakest_assumption targets the usefulness of the difference signal (Eq. 2). That is a reasonable architectural concern, and Table V gives only one ablation as direct evidence. However, the most load-bearing requirement for the paper's central claim is not that the difference signal works in isolation; it is that the comparison showing SOTA is valid. The paper provides no validation split, so the hyperparameter searches in Sec. IV-C are effectively model selection on the test set. Selection bias of this kind can easily produce margins of 1-2 Dice points, which is exactly the size of the reported improvements (SWDL over LeFeD: 1.79; DL over no-DL: 1.17). The Wilcoxon tests are computed on the same optimized test set, so they cannot be used as evidence of significance. A correct validation protocol is a standard, concrete fix; if the margins survive, the claim stands. I therefore keep the CONDITIONAL verdict rather than escalating, but the condition should explicitly require re-evaluation with a held-out validation split and corrected Table VI.","tokens_in":15302,"tokens_out":4600,"duration_ms":56530,"concrete_test":"Re-run the PBGTH 2%-label experiment with a three-way, patient-level split (e.g., 70% train / 10% validation / 20% test), tuning mu, T, xi, and the baselines' hyperparameters only on the validation fold, then compute all metrics on the untouched test fold. If SWDL's Dice advantage over LeFeD falls below the reported 1.79 points or the Wilcoxon p-values lose significance, the SOTA claim is not supported by the current experimental design.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-F describes an 80/20 train/test split with no validation set, and Section IV-C selects the hyperparameters mu=1.5, T=3, and xi=1e-3 by maximizing performance on that same 20% test split (Tables II-IV). The reported headline Dice of 89.32% is therefore the best point in a small hyperparameter grid evaluated on the test cases, not an unbiased estimate of generalization. The baselines (UAMT, URPC, LeFeD) are not given this selection advantage, so the claimed margins (e.g., +1.79 Dice over LeFeD) and the Wilcoxon significance tests in Fig. 3 are optimistically biased. This is an evaluation-validity problem, not an attack on the architecture: the difference-learning mechanism itself has some support from Table V (SWDL vs SWDL-DL: 89.32 vs 88.15). But the central SOTA claim depends on a fair comparison, and the current protocol cannot support it. A second, independent data-entry issue appears in Table VI (supervised VNet ASD CI identical to HD95 CI), which further reduces confidence in the reported BHSD numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SWDL-Net, a semi-supervised 3D intracranial hemorrhage segmentation framework combining a CNN encoder, a deep convolutional decoder, and a deep Laplacian pyramid upsampling decoder. The key mechanism is stratum-wise difference learning: feature differences between the two decoders are fed back into the encoder across iterations (Eq. 2). The method is evaluated on a private 271-case PBGTH dataset with 2% labeled data and on the public BHSD 2024 dataset with 5% labeled data, reporting Dice scores of 89.32% and 51.81%, respectively, and claiming state-of-the-art performance over UAMT, URPC, and LeFeD.","tokens_in":15540,"tokens_out":2822,"duration_ms":35284,"significance":"If the reported improvements are unbiased, the paper makes a useful contribution: the dual-decoder difference-learning mechanism is well motivated, the Laplacian pyramid decoder is a principled way to sharpen boundaries, and the authors state that code is released. The component ablation (Table V) provides evidence that difference learning and deep supervision each contribute to performance. However, the evaluation protocol has a load-bearing flaw: hyperparameters are selected on the test split, which inflates the reported margins and undermines the central SOTA claim. The significance is therefore conditional on fixing the evaluation-validity issues.","major_comments":[{"comment":"No validation split is described; Sec. III-F specifies an 80/20 train/test split only, while Sec. IV-C selects mu=1.5, T=3, and xi=1e-3 by maximizing performance on that same 20% test split (Tables II-IV). The headline Dice of 89.32% is thus the best point of a small hyperparameter grid evaluated on the test cases, not an unbiased estimate. Baselines are not given the same selection advantage, so the claimed margins and the Wilcoxon significance tests in Fig. 3 are optimistically biased. The authors must introduce a validation split (or nested cross-validation), tune on validation only, and then report test performance for all methods under the same protocol.","section":"III-F and IV-C"},{"comment":"The supervised VNet row reports ASD = 7.00 with 95% CI [14.32, 26.42], which is identical to the HD95 CI and impossible for an ASD mean below the CI lower bound; this is a clear data-entry error. The UAMT Dice CI [41.92, 95.90] is also implausibly wide relative to the mean 48.70. Because the BHSD results are used to support the generalizability claim, these numbers must be corrected and the statistical summaries recomputed.","section":"Table VI"},{"comment":"The paper describes difference learning as the core mechanism (Eq. 2, encoder feedback of inter-decoder feature differences), but the unlabeled loss in Eq. (8) is only an MSE consistency between the two decoder outputs. It is not clear how the difference signal in Eq. (2) is trained or whether the gradient path through the feedback loop is active during optimization. Without this clarification, the ablation SWDL vs. SWDL-DL (Table V) is difficult to interpret: the improvement could come from the extra consistency loss or from the difference feedback, rather than from the claimed mechanism.","section":"II-C"},{"comment":"The comparison set contains only three SSL baselines (UAMT, URPC, LeFeD). Given the claim of state-of-the-art performance, the evaluation should include at least one or two recent strong SSL segmentation methods, such as MC-Net+, UniMatch, or comparable 3D methods cited in the introduction. The current comparison is too narrow to support the 'current state-of-the-art' wording.","section":"IV-B"}],"minor_comments":[{"comment":"In the URPC row, the Accuracy CI lower bound is 84.27 while the mean is 94.95; this appears to be a typo and should be corrected.","section":"Table I"},{"comment":"The deep supervision loss uses both L and S for the number of strata; the notation should be made consistent (S is defined in the text, but Eq. (7) sums over L).","section":"II-C, Eq. (7)"},{"comment":"The symbol f_theta_DL is mentioned in the paragraph before Eq. (8) but is not defined; it appears to refer to the DelPU decoder and should be denoted consistently.","section":"II-C"},{"comment":"Reference [42] is cited for UA-MT but is titled 'Unsupervised domain adaptation based on source-guided discrepancy'; the correct UA-MT reference is [20], or the citation should be fixed.","section":"V (References)"},{"comment":"The abstract states that 'Code and data have been released', but the PBGTH dataset is a private clinical dataset; please clarify exactly what is released (code, trained models, or only the public benchmark usage).","section":"Abstract"},{"comment":"The stratum-adaptive depth rule in Eq. (6) gives D=2 for the input size 16x64x64 (max dim 64), but the text does not state the actual D used in the experiments; please specify the effective pyramid depth for all reported results.","section":"II-B"}],"recommendation":"major_revision","confidential_remarks":"The central architectural idea is interesting and the code release is valuable, but the evaluation-validity problem (test-set hyperparameter selection) is serious enough that the reported SOTA margins cannot be trusted as-is. The data-entry errors in Table VI also need correction. I would support reconsideration after the authors re-run the experiments with a proper validation split and fix the reporting issues. The paper is within scope for the journal, but the novelty is moderate and the experimental support needs to be strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. What's actually new: SWDL-Net is a dual-decoder 3D ICH segmentation net where one decoder uses deep Laplacian pyramid upsampling, the other uses transposed convolutions, and the stratum-wise difference between the two decoder feature maps is fed back into the encoder. That combination isn't in the cited literature. The paper also ships code and data, includes the public BHSD benchmark, and the ablations give the difference-learning mechanism real support: removing DL drops Dice from 89.32 to 88.15 on the private set. The authors are upfront about limitations (inter-slice context, subtype generalizability), which I appreciate.\n\nThe soft spot is evaluation validity. Section III-F describes an 80/20 train/test split with no validation set, and Section IV-C picks mu, T, and xi by maximizing performance on that same 20% test split. The headline 89.32% Dice is the best point in a small grid filtered through the test set, not an unbiased generalization estimate. Baselines UAMT, URPC, LeFeD don't get this selection advantage, so the claimed margins and the Wilcoxon stars in Fig. 3 are optimistically biased. This is a protocol problem, not a pointer to a bad architecture — but it means the SOTA claim as written can't be trusted. There's also a clear data-entry error in Table VI: the supervised VNet ASD row shows 7.00 with the same CI as the HD95 row ([14.32, 26.42]), which is impossible. Copy-paste slip, but it doesn't build confidence.\n\nDoes it deserve a serious referee? Yes, with heavy revision. The architecture is novel, the code and data exist, and the core mechanism has direct ablation support. The fix is straightforward: rerun with a validation split or nested CV, retune the baselines under the same protocol, and correct the tables. That's a major-revision path, not a desk reject. I'd bring it to the reading group as a cautionary tale about test-set-based hyperparameter selection, but I wouldn't cite the current numbers until they survive a cleaner evaluation.","headline":"New architecture and ablations are real, but test-set hyperparameter tuning and a CI copy-paste error undermine the SOTA claim.","tokens_in":16062,"tokens_out":3803,"would_cite":false,"duration_ms":41372,"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 dual-decoder network segments brain hemorrhages with just 2% labeled CT scans, reaching 89.32% Dice.","keywords":["intracranial hemorrhage segmentation","semi-supervised learning","Laplacian pyramid","deep convolutional upsampling","difference learning","3D CT","dual-decoder","low-label regime"],"falsifier":"A concrete test would be to replace the difference term in Equation 2 with random Gaussian noise of the same magnitude (or with a fixed non-informative constant) and measure whether Dice on the PBGTH 2% setting drops significantly; if it does not, the specific difference signal is not responsible for the reported gain.","tokens_in":15119,"feed_emoji":"🧠","tokens_out":1478,"duration_ms":17744,"temperature":0.7,"pith_summary":"The paper proposes SWDL-Net, a semi-supervised framework for 3D intracranial hemorrhage (ICH) segmentation that works with very few labeled CT volumes. The method trains two decoders—one using deep transposed convolutions and one using Laplacian pyramid upsampling—and feeds the stratum-wise difference between their features back into the encoder. The authors claim this difference learning lets the model use abundant unlabeled data effectively, outperforming existing semi-supervised methods (UAMT, URPC, LeFeD) on a 271-case clinical dataset with only 2% labels and on the public BHSD benchmark with 5% labels. If correct, the approach offers a practical path to accurate ICH segmentation in settings where expert annotations are scarce.","feed_headline":"Dual-decoder network segments brain bleeds with 2% labeled CT scans","feed_subtitle":"SWDL-Net feeds decoder-feature differences back into the encoder, hitting 89.32% Dice on a 271-case hemorrhage dataset.","key_machinery":"The key mechanism is 'stratum-wise difference learning' (SWDL) combined with a dual-decoder architecture: a deep convolutional (DC) decoder and a Deep Laplacian Pyramid Upsampling (DelPU) decoder. At each iteration, the difference $\\Delta^{(s,p)}$ between the two decoders' feature maps at stratum $s$ is computed and added into the encoder's next-iteration feature computation via $y^s_{E,p} = f^s_{\\theta_E}(y^{s-1}_{E,p} + \\xi \\Delta^{(s-1,p-1)})$. The DelPU decoder builds a Laplacian pyramid with Gaussian downsampling, inter-level differential extraction, and weighted reconstruction governed by an edge-sharpening hyperparameter $\\mu$. A stratum-adaptive rule chooses pyramid depth based on input dimension. The unsupervised loss is MSE between the two decoders' outputs, while the supervised loss combines Dice, cross-entropy, and deep supervision on the DC path.","core_discovery":"SWDL-Net establishes that a hybrid dual-decoder architecture with stratum-wise difference learning can achieve near-fully-supervised performance in low-label regimes. With only 2% labeled data on the private PBGTH dataset, it attains Dice of 89.32% versus 92.48% for fully supervised VNet, i.e., 96.6% of the fully supervised score, and improves HD95 by 41.3–66.9% over the compared SSL baselines. On the public BHSD dataset, with 5% labeled data, it reaches 51.81% Dice, out-performing LeFeD by 1.11 points. The central claim is that feeding the difference between the DC decoder and DelPU decoder features back into the encoder (Equation 2) provides a supervisory signal that leverages the Laplacian pyramid's edge-sharpening and the convolution path's detail precision.","pith_inferences":["A natural testable extension is to apply the same stratum-wise difference feedback to other multi-decoder consistency frameworks (e.g., different upsampling kernels or transformer-based decoders) to see whether the benefit is specific to Laplacian-vs-convolutional diversity or generic to any two sufficiently different decoders.","The paper does not isolate how much of the gain comes from the difference-feedback loop versus simply using two decoders with different losses; a controlled experiment with feature concatenation instead of subtraction would clarify the mechanism.","Since the method shows a larger relative gain on the private PBGTH dataset (lesion-consistent basal ganglia/thalamic hemorrhages) than on the more varied BHSD, one might infer that the difference signal is most useful when the label distribution is relatively homogeneous.","The reported Dice improvement on BHSD is modest (51.81 vs 50.70 for LeFeD) and the confidence intervals overlap; a reader should interpret the 'state-of-the-art' claim on BHSD with caution."],"forward_implications":["If the results hold, semi-supervised ICH segmentation can be performed accurately with only a handful of labeled scans (2–5%), greatly reducing annotation cost in emergency neuroimaging.","The difference-learning mechanism between Laplacian-pyramid and convolutional upsampling decoders suggests that other anisotropic or edge-critical medical segmentation tasks (e.g., vessel or tumor boundary delineation) could benefit from the same dual-decoder discrepancy signal.","The reported Dice of 89.32% on PBGTH and 51.81% on BHSD with so few labels indicates that the method could serve as a strong baseline for future SSL work on volumetric CT segmentation.","Because inference requires only the encoder and primary decoder, the method could be deployed in clinical workflows without the computational overhead of the pyramid module."],"supporting_citations":[{"why":"UAMT is the uncertainty-aware teacher-student baseline that SWDL compares against on both datasets.","marker":"[20]"},{"why":"URPC is the uncertainty-rectified pyramid consistency baseline, a key SSL comparator.","marker":"[14]"},{"why":"LeFeD is the consistency-guided differential decoding method that SWDL directly extends and beats.","marker":"[38]"},{"why":"VNet is the fully supervised baseline whose performance SWDL is measured against.","marker":"[39]"},{"why":"BHSD is the public benchmark dataset used for external validation.","marker":"[27]"},{"why":"U-Net is the architectural ancestor for the encoder-decoder design.","marker":"[2]"}],"fun_headline_variants":["2% labeled CTs achieve 96.6% of fully supervised ICH Dice","Dual-decoder difference learning segments brain bleeds from 2% labels","Laplacian pyramid + deep conv difference nets near-full Dice at 2% labels","SWDL-Net: segment ICH with 2% labeled scans via stratified difference learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the difference between the two decoders' features is a meaningful, learnable signal for unlabeled data, rather than noise that simply adds variance to the encoder's features.","fun_headline_variants_meta":{"raw":{"variants":["2% labeled CTs achieve 96.6% of fully supervised ICH Dice","Dual-decoder difference learning segments brain bleeds from 2% labels","Laplacian pyramid + deep conv difference nets near-full Dice at 2% labels","SWDL-Net: segment ICH with 2% labeled scans via stratified difference learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000944,"raw_usage":{"total_tokens":4051,"prompt_tokens":982,"completion_tokens":3069,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2979}},"tokens_in":598,"tokens_out":3069,"duration_ms":23943,"temperature":1.0,"reasoning_tokens":2979,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:28:59.216501+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test would be to replace the difference term in Equation 2 with random Gaussian noise of the same magnitude (or with a fixed non-informative constant) and measure whether Dice on the PBGTH 2% setting drops significantly; if it does not, the specific difference signal is not responsible for the reported gain.","supporting_citations":[{"cited_title":"Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmenta- tion,","cited_arxiv_id":null,"evidence_quote":"UAMT is the uncertainty-aware teacher-student baseline that SWDL compares against on both datasets."},{"cited_title":"Consistency-guided differential decoding for enhancing semi- supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"LeFeD is the consistency-guided differential decoding method that SWDL directly extends and beats."},{"cited_title":"Bhsd: A 3d multi-class brain hemorrhage segmentation dataset,","cited_arxiv_id":null,"evidence_quote":"BHSD is the public benchmark dataset used for external validation."}],"review_version":1}