{"id":"fd6b1b2d-70d0-447e-9049-c2ba3447d83d","arxiv_id":"2506.17712","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PDC-Net, combining directional strip convolutions, a memory module, and a mixture-of-experts decoder, is reported to improve pelvic radiation injury segmentation by 2 to 4 Dice points over baselines on a private 344-patient MRI dataset.","lead":"This paper reports a new deep learning network, PDC-Net, that automatically outlines pelvic radiation injury on MRI scans. The authors say it slightly beats six existing segmentation methods on their own private dataset, but release no code or data to check the claim.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run comparisons without error bars or significance tests leave the 2–3 point DSC advantage over EMCAD within the range of run-to-run noise; the SOTA claim is not yet established.","rationale":"The reader's verdict is CONDITIONAL and its weakest assumption is that the reported 2–3 point DSC differences are not within run-to-run variability. I agree with that assessment. My stress-test pass found no mathematical contradiction in the proposed modules: MDA is a plausible extension of strip convolutions, MGC resembles memory-augmented context modeling, and AFD is a MoE-style decoder. The load-bearing weakness is empirical. The paper's own text states there is no public dataset (Section 3.1), making independent verification impossible unless data and code are released. The comparison tables contain only single-run numbers, with no indication of how many times experiments were repeated, no variance estimates, and no significance testing. In medical segmentation, differences of 2–3 DSC points are commonly within seed-to-seed noise, especially on class-imbalanced private data. The ablation tables have the same limitation for each individual module contribution. I also noted small presentation issues, such as the text in Section 3.3 attributing the MGC improvement to AFD and the 'MPA' typo in Section 2.1, but these do not change the core concern. Because the central claim is a superiority claim supported solely by unreproducible single-run numbers, the paper should not be accepted as-is; however, the architecture is coherent and the concern is addressable by adding seeds, error bars, significance tests, and code or data release. That matches the reader's CONDITIONAL verdict, so no change is needed.","tokens_in":7196,"tokens_out":4280,"duration_ms":43484,"concrete_test":"Run the full Table 1 comparison at least five times per method with different random seeds on the same training/test case split, and report mean ± std for DSC, MCC, and HD. Then apply a paired test across the 70 test cases (e.g., Wilcoxon signed-rank or paired t-test) between PDC-Net and EMCAD, and likewise for the Table 2 ablation variants. If the DSC gap narrows to less than one point or the paired test yields p > 0.05, the state-of-the-art claim in Section 1 is not supported. Additionally, release the trained models, code, and test-set predictions so the reported numbers can be independently reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that PDC-Net achieves state-of-the-art performance (Section 1) rests entirely on Table 1, where PDC-Net exceeds the strongest baseline EMCAD by 2.52 DSC points on the negative class (48.79 vs 46.27) and 2.82 points on the positive class (49.12 vs 46.30), and on Table 2, where each proposed module adds 1–3 points. Every reported number is a single run: no multiple seeds, no error bars, no paired significance test, and no cross-validation are reported in Section 3.2. Baseline hyperparameters and training details are also unspecified. Because the dataset is private and no code or test-set predictions are released, the 2–3 point gaps are exactly the magnitude of typical seed-to-seed variability for medical segmentation networks, so the superiority claim could be an artifact of initialization, data ordering, or training noise. The ablations share this weakness: the 1.28–2.88 point contributions in Table 2 are also single-run and could lie within noise, which would undermine the claim that MDA, MGC, and AFD each contribute measurable gains. The issue is not that the architecture is internally inconsistent; it is that the empirical evidence is currently insufficient to distinguish the reported differences from run-to-run variability.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PDC-Net, a deep learning architecture for segmenting pelvic radiation injury (PRI) from T2-weighted MRI. The method combines three novel modules: a Multi-Direction Aggregation (MDA) module using strip convolutions in four directions, a Memory-Guided Context (MGC) module that maintains dataset-level memory to reduce background interference, and an Adaptive Fusion Decoder (AFD) based on a Mixture-of-Experts framework. The authors introduce the first large-scale private PRI dataset (344 cases, 8049 images) and report quantitative comparisons against six existing segmentation methods, along with ablation studies. The central claim is that PDC-Net achieves state-of-the-art performance on PRI segmentation.","tokens_in":7444,"tokens_out":3858,"duration_ms":37507,"significance":"If the reported improvements are reproducible, this paper would be a valuable contribution as the first AI-based method specifically designed for PRI segmentation, and the dataset itself is a significant resource for the community. The architectural ideas are well motivated by the strip-like and circular morphology of the target structures, and the paper compares against a reasonable set of recent baselines. The strengths are the clear problem framing, the large in-house dataset, and the modular design that allows each component to be ablated. However, the empirical evidence for the central claim is currently weak because all reported numbers come from single training runs with no error bars, no significance tests, and no external validation, leaving the reported 2–3 point DSC advantages within the range of run-to-run variability.","major_comments":[{"comment":"The claim that PDC-Net achieves state-of-the-art performance rests on single-run DSC differences of 2.52 points (negative class) and 2.82 points (positive class) over the strongest baseline EMCAD. No error bars, repeated seeds, or significance tests are reported. On a private dataset with no code release, these gaps are within the range of seed-to-seed variability commonly observed for medical segmentation networks, so the superiority claim is not yet established. Please report results over multiple seeds with paired significance tests (e.g., Wilcoxon signed-rank or paired bootstrap) and provide the variance.","section":"Section 3.2, Table 1"},{"comment":"The ablation study is subject to the same single-run limitation. The claimed contributions of each module—MDA (2.61 points on the negative class vs. PConv), MGC (1.28 points vs. EFA), and AFD (2.88 points vs. U-shape)—are each based on one run, with no confidence intervals. Since these differences are comparable to typical seed-to-seed variability, the conclusion that each proposed module contributes a measurable gain is not supported. Please add repeated runs and statistical tests, or at least report per-seed results and the mean±std for each configuration.","section":"Section 3.2, Table 2"},{"comment":"The comparison protocol is underspecified. The paper does not state whether all baselines were trained with the same hyperparameters, preprocessing, loss functions, and data splits, nor how the 3D method UNETR++ was adapted to the 2D slice-level task. Without this information, the observed improvements over baselines could reflect differences in training setup rather than architectural merit. Please document the exact training protocol for every baseline, including any modifications made for fair comparison.","section":"Section 3.1, Compared Methods"},{"comment":"The evaluation uses a private in-house dataset, and no code or test-set predictions are released. While private clinical data are acceptable, the lack of any external verification, combined with the absence of error bars, makes the headline result impossible to validate. Please consider releasing pseudonymized features, test-set predictions, or a public benchmark subset, and at minimum provide a detailed description of the data-split procedure, annotation protocol, and any quality control steps.","section":"Section 3.1, Datasets"}],"minor_comments":[{"comment":"In the text following Eq. (2), 'the final output of the MPA module' should be 'the final output of the MDA module'.","section":"Section 2.1, Eq. (2)"},{"comment":"In the paragraph for the MGC ablation, the sentence 'our AFD shows an improvement of 1.28%' should refer to the MGC module, not AFD.","section":"Section 3.3, MGC ablation"},{"comment":"The dimensions in Eqs. (3)–(4) are not fully defined: after max and average pooling, what is the spatial size of Avg(f_patch) and Max(f_patch)? The symbol S in the memory bank is used without explicit definition, and the shapes in Eq. (4) appear inconsistent (element-wise operations between a FC output and the pooled features). Please clarify the tensor shapes for each step.","section":"Section 2.2, Eqs. (3)–(4)"},{"comment":"The text says 'we apply adaptive pooling' but Eq. (7) uses AvgPool. Please specify the pooling type and the output dimension before the FC layer.","section":"Section 2.3, Eq. (7)"},{"comment":"The ablation tables report only DSC. Since Table 1 includes MCC, ACC, and HD, reporting at least MCC or HD in the ablations would provide a more complete picture of the contribution of each module.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a medical image analysis venue, but the experimental evidence needs to be substantially strengthened before the SOTA claim can be accepted. The authors should be asked to provide error bars or per-seed results, and to document the baseline training protocol carefully. If the authors cannot access multiple seeds (e.g., due to compute constraints), the claims should be appropriately toned down to 'comparable performance' or 'promising results'. The private dataset is a concern but not disqualifying if the statistical analysis is made rigorous."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this paper attacks a real gap: nobody has published an automated segmentation method for pelvic radiation injury, and the authors built an in-house dataset of 344 cases / 8k T2-MRI slices with manual annotations for injured and non-injured pelvic structures. That alone is a useful contribution. Second, the paper's central 'state-of-the-art' claim is not yet backed by the evidence: every number in Tables 1 and 2 is a single run, with no error bars, no multiple seeds, no significance tests, no code, and no public data.\n\nThe architecture is a coherent combination of existing ideas, not a conceptual leap. Strip convolutions come from SegNext, extended here with diagonal directions; the memory bank is adapted from space-time memory networks; the decoder is an MoE variant with cross-stage feature mixing. The clinical motivation—strip-like muscles and circular hollow organs—makes sense, and the modules are straightforward to follow.\n\nThe soft spot is the evaluation. The 2.5–2.8 DSC points over EMCAD in Table 1 could easily be seed noise in medical segmentation, and the ablation differences (1.3–2.9 points) have the same problem. Baseline hyperparameters are not specified, and the memory bank's capacity and sliding-window size are not stated, so even with code, the method would be hard to reproduce exactly. The dataset is private, so external verification is impossible right now. I don't see any internal inconsistency or circular reasoning; the issue is purely that the evidence is too weak to support the superiority claim.\n\nWho's this for? People working on PRI or on segmentation of hollow/strip-like organs. It is a reasonable first attempt and a useful reference for the task, but I wouldn't treat its numbers as a benchmark yet.\n\nI would send this to peer review, not desk reject it. The task is new and the fix is concrete: report multiple seeds with variance, run paired significance tests, release code and data, specify the hyperparameters, and tone down the SOTA claim until the comparisons are statistically defensible.","headline":"A genuinely new segmentation task and dataset, with a coherent architecture and an unproven SOTA claim due to single-run evaluations.","tokens_in":7969,"tokens_out":3659,"would_cite":true,"duration_ms":36830,"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 pattern divide-and-conquer network for pelvic radiation injury segmentation achieves the best reported Dice and MCC on a 344-case MRI dataset, with every module contributing a measurable gain.","keywords":["pelvic radiation injury","MRI segmentation","deep learning","image segmentation","strip convolution","memory-guided context","mixture-of-experts","medical image analysis"],"falsifier":"Run PDC-Net, EMCAD, and the ablation variants under at least five random seeds on the same 274-case training split and the fixed 70-case test split; if the 95% confidence interval for the positive-class DSC difference between PDC-Net and EMCAD includes zero, the reported state-of-the-art claim is not supported.","tokens_in":6999,"feed_emoji":"🩻","tokens_out":16846,"duration_ms":142550,"temperature":0.7,"pith_summary":"Segmentation of pelvic radiation injury (PRI) from MRI is the kind of task that decides whether radiotherapy follow-up can be quantified: the injured structures are hollow organs and thin muscles that appear as circular or strip-like shapes, and injured and uninjured regions can sit in the same slice. The paper's central claim is that a pattern divide-and-conquer network, PDC-Net, fits these shapes and resolves the confusing context by assigning three modules to three pattern classes—directional strip convolutions for local shape, a dataset-level memory bank for global context, and a mixture-of-experts decoder for adaptive fusion. Evaluated on the first large-scale PRI dataset (344 pelvic cases, 8,049 axial T2-weighted MRI images), PDC-Net reports the best Dice and Matthews correlation among six comparison methods, with positive-class DSC 49.12 versus 46.30 for the strongest baseline EMCAD. The ablations show each module adding a measurable gain over its direct alternative, consistent with the divide-and-conquer rationale.","feed_headline":"PDC-Net beats six baselines on pelvic radiation injury MRI","feed_subtitle":"Strip convolutions in four directions plus a dataset memory bank separate injured from uninjured pelvic tissue.","key_machinery":"The load-bearing machinery is the pattern divide-and-conquer composition: three modules, each responsible for one failure mode, assembled so the decoder can route among them. The Multi-Direction Aggregation (MDA) module is a bank of four strip convolutions in directions $[1,0]$, $[0,1]$, $[1,1]$, and $[-1,1]$ with kernel size 9, whose outputs are concatenated and element-wise multiplied by a max-pooled detail branch; it exists because square convolutions capture circular and strip-like organ shapes inefficiently. The Memory-Guided Context (MGC) module maintains a dataset-level memory bank $\\mathcal{M}\\in\\mathbb{R}^{1\\times1\\times S\\times K}$; pooled feature patches query it through softmax similarity weights, so the network accumulates and retrieves foreground-specific patterns across images and can suppress background. The Adaptive Fusion Decoder (AFD) is a mixture-of-experts decoder: features from four stages are patch-shuffled, passed through four depthwise-convolution experts with $1\\times n$ and $n\\times1$ kernels for $n\\in\\{3,5,7,9\\}$, and fused by sigmoid-gated per-expert weights. Together they implement the paper's claim that local shape, global context, and fusion should be handled by dedicated, separable mechanisms.","core_discovery":"PDC-Net is built on a PVT-v2-b2 encoder followed by three components. The Multi-Direction Aggregation (MDA) module splits the encoder features into four channel groups and applies strip convolutions in horizontal, vertical, left-diagonal, and right-diagonal directions with kernel size 9, then multiplies the concatenated result by a max-pooling detail branch, letting the network fit the circular and elongated shapes of the rectal wall, anal sphincter, puborectalis muscle, and bladder. The Memory-Guided Context (MGC) module slices the deep feature map into patches, pools each patch, and queries a memory bank $\\mathcal{M}\\in\\mathbb{R}^{1\\times1\\times S\\times K}$ with softmax-computed similarity weights, reinforcing foreground-related patterns and suppressing background. The Adaptive Fusion Decoder (AFD) mixes features from four stages through patch shuffle, applies four depthwise-convolution experts with $n\\times1$ and $1\\times n$ kernels for $n\\in\\{3,5,7,9\\}$, and combines expert predictions with sigmoid-gated weights. On the in-house dataset of 344 pelvic cases (8,049 T2-weighted axial images), the paper reports negative-class DSC 48.79 and positive-class DSC 49.12, compared with 46.27 and 46.30 for EMCAD, and shows in ablations that MDA, MGC, and AFD each outperform their direct replacements.","pith_inferences":["Inference: the same divide-and-conquer recipe—shape-specific strip convolutions, dataset-level memory, and expert-fusion decoding—should transfer to other thin or hollow anatomical structures such as vessels or nerves; running the modules on a public vessel or nerve segmentation benchmark would test this.","Inference: because the memory bank is trained on the training set's global patterns, its benefit may depend on scanner protocol; a multi-center MRI evaluation would show whether the bank needs re-initialization when the data distribution shifts.","Inference: the memory-bank design suggests an incremental-learning use the paper does not explore—freeze the encoder and update only the bank as new cases arrive—which could matter for clinical deployment where data accumulate over time."],"forward_implications":["Pelvic radiation injury segmentation now has a concrete automated baseline: 344 cases and 8,049 images, with negative/positive DSC of 48.79/49.12 and MCC of 50.31/50.98, against which later methods can be compared.","Four-direction strip convolutions, including diagonals, improve over two-direction strip convolutions, so thin and diagonal anatomical structures are better served by directional kernels than by square kernels alone.","A dataset-level memory bank queried by pooled patches can separate positive from negative class patterns without dilated convolutions or self-attention, under the class imbalance this dataset exhibits.","An MoE decoder that mixes features from multiple stages by patch shuffle and expert-specific depthwise convolutions beats both a U-shaped decoder and a vanilla MoE decoder.","PDC-Net is presented as the first AI method for PRI segmentation, so this paper defines the task's evaluation setup (metrics and dataset split) for the field."],"supporting_citations":[{"why":"It supplies the PVT-v2-b2 encoder features that the MDA and MGC modules refine and that the AFD decoder consumes.","marker":"[13]"},{"why":"U-Net is the primary baseline in Table 1 and the U-shape decoder replaced by the AFD in the decoder ablation.","marker":"[10]"},{"why":"EMCAD is the strongest competing method; its DSC and MCC numbers set the threshold that the state-of-the-art claim must beat.","marker":"[9]"},{"why":"The space-time memory network is the origin of the memory-bank mechanism that MGC adapts from video segmentation to dataset-level context.","marker":"[7]"},{"why":"The mixture-of-experts framework and its medical-segmentation adaptation are the foundation that the AFD extends with patch shuffle and expert customization.","marker":"[5,8]"},{"why":"Embedding-free attention is one of the two direct alternatives that the MGC module is ablated against in Table 2b.","marker":"[17]"},{"why":"Multi-scale representation is the second direct alternative that the MGC module is ablated against in Table 2b.","marker":"[16]"},{"why":"SegNeXt's horizontal and vertical strip convolution is the design that MDA extends to diagonal directions.","marker":"[3]"}],"fun_headline_variants":["PDC-Net splits patterns to conquer pelvic radiation injury","Strip convolution directions and memory bank boost PRI segmentation","Mix-of-experts decoder lifts pelvic MRI injury accuracy","Four-direction strip convs unmask pelvic injury on MRI","PDC-Net: divide-and-conquer for clearer pelvic MRI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that PDC-Net is superior rests on the assumption that its 2-3 percentage-point Dice advantages over the strongest baseline are real differences rather than run-to-run noise, because the paper reports a single run per method without error bars or significance tests.","fun_headline_variants_meta":{"raw":{"variants":["PDC-Net splits patterns to conquer pelvic radiation injury","Strip convolution directions and memory bank boost PRI segmentation","Mix-of-experts decoder lifts pelvic MRI injury accuracy","Four-direction strip convs unmask pelvic injury on MRI","PDC-Net: divide-and-conquer for clearer pelvic MRI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1645,"prompt_tokens":1123,"completion_tokens":522,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":739,"completion_tokens_details":{"reasoning_tokens":442}},"tokens_in":739,"tokens_out":522,"duration_ms":5545,"temperature":1.0,"reasoning_tokens":442,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:02:17.783328+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run PDC-Net, EMCAD, and the ablation variants under at least five random seeds on the same 274-case training split and the fixed 70-case test split; if the 95% confidence interval for the positive-class DSC difference between PDC-Net and EMCAD includes zero, the reported state-of-the-art claim is not supported.","supporting_citations":[{"cited_title":"In: Medical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"U-Net is the primary baseline in Table 1 and the U-shape decoder replaced by the AFD in the decoder ablation."},{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"The space-time memory network is the origin of the memory-bank mechanism that MGC adapts from video segmentation to dataset-level context."},{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Embedding-free attention is one of the two direct alternatives that the MGC module is ablated against in Table 2b."},{"cited_title":"In: International Conference on Learning Repre- sentations (2024)","cited_arxiv_id":null,"evidence_quote":"Multi-scale representation is the second direct alternative that the MGC module is ablated against in Table 2b."},{"cited_title":"Advances in Neural Information Processing Systems35, 1140–1156 (2022)","cited_arxiv_id":null,"evidence_quote":"SegNeXt's horizontal and vertical strip convolution is the design that MDA extends to diagonal directions."}],"review_version":2}