{"id":"ecc8ab69-ea4a-4106-aaec-f5d4b7365457","arxiv_id":"2507.19590","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"T-MPEDNet reports Dice scores of 97.6/89.1 on LiTS and 98.3/83.3 on 3DIRCADb for liver and tumor segmentation, surpassing twelve compared baselines.","lead":"A new AI system called T-MPEDNet automatically outlines livers and tumors in CT scans, combining several known deep-learning building blocks in one network. It reports slightly higher accuracy than twelve earlier methods on two public benchmark datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Superiority claim depends on an uncontrolled evaluation: Table 2 splits LiTS/3DIRCADb by slices, not patients, so same-volume slices can appear in train and test; Table 3 baselines come from unknown protocols. The 0.4–1.2 DSC margins are not attributable to T-MPEDNet.","rationale":"I read the paper's central claim as empirical superiority over twelve SOTA methods, not merely architectural novelty. The load-bearing condition is therefore a valid, controlled comparison. The Pith reader identified the evaluation setup—random split and unspecified baseline training protocols—as the weakest assumption, and I agree. My stress-test sharpens this into a concrete failure mode: Table 2 reports slice counts, indicating a slice-level rather than patient-level split, which creates same-patient leakage between training and test. Baseline scores from original papers are not controlled for preprocessing or split, so the reported margins of 0.4–1.2 DSC points cannot be assigned to T-MPEDNet's modules. The paper does contain real supporting material: a full module-by-module architecture description, an ablation study, and qualitative comparisons, and I do not see an internal inconsistency that independently invalidates the architecture. But none of that establishes the headline comparison. The proposed test—patient-level split plus matched retraining of the closest baselines—would settle the concern directly. Because the current manuscript does not provide the evidence needed to verify the central claim, I would move from CONDITIONAL to UNVERDICTED pending the requested patient-level split and code or retrained comparison.","tokens_in":22331,"tokens_out":6845,"duration_ms":75935,"concrete_test":"Ask the authors to release code and a patient-level split: 80/10/10 of the 131 LiTS volumes and 80/10/10 of the 20 3DIRCADb patients, ensuring no same-patient slices appear across sets, and recompute Table 3 under exactly that split and identical preprocessing for T-MPEDNet and at least the closest baselines (TD-Net, DefED-Net, HFRU-Net, RMAU-Net). If LiTS tumor DSC falls below 88.1 or liver DSC falls below 97.1, the stated superiority is an artifact of the slice-level split; if the margins survive on a patient-level holdout with retrained baselines, the concern is resolved.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The central claim is empirical superiority over twelve SOTA methods. For that claim to hold, the test split must be clean and the comparison must be controlled. Section 4.1.1 says the authors 'randomly split CT images into 80% for training, 10% for validation, and 10% for independent testing,' and Table 2 reports slice counts: LiTS has 54,065 training slices, 5,001 validation, and 4,573 test; 3DIRCADb has 2,258/282/282. If this is a slice-level split, as the table strongly suggests, then slices from the same patient volume can appear in both training and test. Adjacent axial CT slices of one liver are highly correlated, so a 2D network can effectively leak per-patient anatomy and peri-lesional context into test evaluation, inflating DSC. No patient-level or case-holdout split is stated. Separately, Table 3's baseline numbers are cited from original papers, but no evidence shows those numbers were produced under the same reformation steps (HU windowing, CLAHE, z-score normalization), the same 256×256 resampling, the same slice sampling, or the same train/validation/test split. Without a common protocol, reported margins of 0.4–1.2 DSC points—within the range of split-to-split variation—cannot be attributed to the proposed architecture. The module descriptions and ablation are coherent, but the headline 'surpassing all twelve methods' is currently unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes T-MPEDNet, a 2D progressive encoder-decoder for liver and tumor segmentation in CT volumes. The architecture combines Adaptive Feature Extraction (AdaFEx) stages, a Compressive Channel Recalibration (CCR) module, a Transformer-inspired Dynamic Contextual Attention (DCA) module, a Multi-Scale Atrous Spatial (MSAS) module, and a Morphological Boundary Refinement (MBR) postprocessing step. The authors report Dice scores of 97.6/89.1 for liver/tumor on a self-selected LiTS split and 98.3/83.3 on 3DIRCADb, and on this basis claim superiority over twelve prior methods. A five-row ablation on LiTS attributes performance gains to the individual components.","tokens_in":22694,"tokens_out":5601,"duration_ms":68224,"significance":"If the empirical claims are supported, the paper would provide a competitive 2D architecture for an important clinical segmentation task, with a clear module breakdown and ablations. The architectural components are described in enough detail to be reimplemented, and the use of two public datasets is appropriate. The main value would lie in the reported benchmark results. However, the significance is currently limited by the uncontrolled evaluation protocol: the data split appears to be at slice level rather than patient level, the comparison baselines are taken from disparate published protocols, and no uncertainty quantification is provided. These issues directly affect the central claim of outperforming all twelve compared methods.","major_comments":[{"comment":"The data split is described in Section 4.1.1 as randomly splitting CT images into 80% training, 10% validation, and 10% testing, and Table 2 reports slice counts (54,065/5,001/4,573 for LiTS and 2,258/282/282 for 3DIRCADb). This indicates a slice-level split rather than a patient-level or volume-level split. Adjacent axial slices from the same CT volume are highly correlated, so slices of the same liver and tumor can appear in both training and test sets, inflating the reported Dice scores. A patient-level split with case-level holdout, together with per-case metrics, is needed before the reported superiority can be accepted.","section":"Section 4.1.1, Table 2"},{"comment":"The comparison with twelve state-of-the-art methods relies on Dice scores taken from the original papers, with no evidence that those methods were run under the same preprocessing (HU windowing, CLAHE, z-score normalization), the same 256x256 resampling, the same slice selection, or the same train/validation/test partition. The reported margins over the second-best methods are only 0.4 to 1.2 DSC points, which is within the range of split-to-split and protocol-to-protocol variation. Without a controlled comparison under a common protocol, the claim that T-MPEDNet surpasses all twelve methods is not supported.","section":"Table 3, Section 4.2.1"},{"comment":"No error bars, standard deviations, number of independent runs, or significance tests are reported. The central comparison is based on single Dice values, and the smallest claimed margins are around 0.4 DSC points. At minimum, the authors should report mean and standard deviation over several runs with different seeds, and ideally paired statistical tests for the differences against the strongest baselines.","section":"Section 4.2.1, Tables 3 and 4"},{"comment":"The ablation interpretation in the text does not match the table. The sentence 'Removing the DCA module ... declines the liver DSC by 2.2% and tumor DSC by 5.2%' compares the full model with the row that removes DCA, MSAS, and MBR simultaneously, not DCA alone. Similarly, the reported 1.7% and 2.7% drops for removing MSAS are not the differences between consecutive rows. One-at-a-time ablations, with a single module removed while all others remain, are necessary to attribute performance changes to individual components.","section":"Section 4.2.2, Table 4"},{"comment":"The Morphological Boundary Refinement module is described as producing a boundary mask BMask by subtracting an eroded mask from the input mask, but the manuscript never specifies how BMask is combined with the network's output to obtain the final segmentation. Without an explicit final-mask equation, the reported 0.2% and 0.5% Dice improvements from MBR in Table 4 cannot be reproduced or verified. Please state the exact postprocessing rule and clarify whether the reported scores include this step.","section":"Section 3.2.3, Eq. (12), Fig. 8"}],"minor_comments":[{"comment":"The summation in Eq. (4) runs to H×C, while the text states the spatial dimensions are H×W; the summation index j is also undefined. Please correct the upper limit and define the index.","section":"Eq. (4)"},{"comment":"Eq. (6) writes f_j^CCR = a_k ⊙ z_k, mixing indices j and k. This should presumably be f_k^CCR = a_k ⊙ z_k for each channel k.","section":"Eq. (6)"},{"comment":"The notation σ^Sigmoid and α^ReLU is unconventional and hard to read; standard mathematical notation for the sigmoid and ReLU functions should be used.","section":"Eq. (5)"},{"comment":"Eq. (11) uses plus signs between the four atrous-convolution terms, although the text says the multi-scale features are concatenated. Please use a concatenation symbol or clarify the operation.","section":"Eq. (11)"},{"comment":"There are typographical issues in the column headers, including 'V oxel spacing' with a space, and the Z-axis spacing range is written as '01.6 - 4.05 mm' instead of 1.6-4.05 mm.","section":"Table 2"},{"comment":"The abstract appears twice at the beginning of the document, and the arXiv header says 'Preprint submitted to Expert Systems with Applications' while the highlight page says the paper is accepted in Biomedical Signal Processing and Control. Please align these statements and remove the duplication.","section":"Title page and Abstract"}],"recommendation":"major_revision","confidential_remarks":"The central issue is that the headline empirical claim is not yet supported by a controlled evaluation. The slice-level split and the use of non-reproduced baseline numbers are the two points that most need to be addressed. If the authors can provide a patient-level split, a common protocol for at least the strongest baselines, and uncertainty estimates, the paper may become acceptable. The architecture itself is coherent and the ablation, once corrected, could be informative."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: T-MPEDNet is a competent engineering paper. The architecture is a new combination of parts that are each well known, the ablation is honestly done, but the headline claim — that it surpasses twelve state-of-the-art methods — is not supported by the evaluation as reported.\n\nWhat is actually new: the specific assembly of a progressive encoder-decoder, channel recalibration on skip connections (CCR), dynamic-context attention with dynamic convolutions (DCA), an atrous multiscale module (MSAS), and morphological boundary refinement (MBR). The related work cites each component properly, and the paper does not pretend otherwise; the contribution is the combination plus the empirical study. Credit where due: the ablation in Table 4 is complete and internally consistent. CCR is the workhorse (removing it costs 4.1/7.7 DSC on liver/tumor), MBR is modest (0.2/0.5), and they report that honestly. The limitations paragraph in the Discussion — 2D training ignores inter-slice context, and downsampling to 256×256 loses resolution — is candid and lands.\n\nSoft spots, in order. First, the split. Section 4.1.1 says the authors \"randomly split CT images\" into 80/10/10, and Table 2 reports slice counts per fold. That reads like slice-level splitting, not patient-level. For a 2D network, adjacent slices from one volume are near-duplicates; if they land in both train and test, per-case anatomy leaks and Dice inflates. The fix is a patient- or case-level holdout, reported as volumes per fold. Second, the SOTA numbers in Table 3 are taken from the original papers, with no evidence those models used the same preprocessing (HU windowing, CLAHE, 256×256 resampling, z-score normalization) or the same split. The claimed margins are 0.4–1.2 DSC points, which is exactly the size of split-to-split variation. As reported, the superiority claim cannot be attributed to the architecture. Third, there are no error bars or significance tests, and no code is provided. Fourth, minor equation sloppiness — indices in Eq. 4 and Eq. 6, and '+' in Eq. 11 where the figure shows concatenation. Cosmetic, but a referee should ask for a clean pass. The MBR mechanism is also under-explained; its measured contribution is small, so the opacity there is a minor concern, not a fatal one.\n\nWho this is for: applied medical imaging researchers working on liver/tumor segmentation who want an architectural comparison point. It deserves a serious referee. Everything wrong here is fixable — state the split explicitly, run the baselines under one protocol or stop claiming superiority, add dispersion, release code. I would not lean on the reported numbers myself until that happens.","headline":"Coherent architecture and an honest ablation, but the 'beats all twelve' claim rests on an uncontrolled comparison: split may be slice-level and baselines are borrowed.","tokens_in":23206,"tokens_out":5687,"would_cite":false,"duration_ms":61064,"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":"T-MPEDNet claims state-of-the-art CT liver and tumor segmentation, with 97.6% and 89.1% Dice on LiTS and 98.3% and 83.3% on 3DIRCADb.","keywords":["liver tumor segmentation","computed tomography","progressive encoder-decoder","transformer attention","multi-scale atrous convolution","channel recalibration","morphological boundary refinement","Dice similarity coefficient"],"falsifier":"Run T-MPEDNet and all twelve compared methods on the same 80/10/10 split of LiTS and 3DIRCADb, with identical preprocessing (HU windowing, CLAHE, z-score normalization, and 256 by 256 resizing) and the same optimizer and augmentation, then compare Dice on the identical test slices; if the margins shrink below the reported 0.4 to 1.2 points or reverse sign, the central claim fails.","tokens_in":22160,"feed_emoji":"🩻","tokens_out":7295,"duration_ms":73905,"temperature":0.7,"pith_summary":"The paper aims to show that a new 2D neural network, T-MPEDNet, can segment the liver and liver tumors in CT scans more accurately than twelve published methods. It reports Dice similarity coefficients of 97.6% for liver and 89.1% for tumor on the LiTS benchmark, and 98.3% and 83.3% on 3DIRCADb, with the largest reported gains on tumors, the harder class. The architecture combines a progressive five-stage encoder-decoder with channel-wise feature recalibration, a transformer-inspired attention module, multi-scale dilated convolutions, and a morphological boundary-refinement step. If the results hold, the network would be an effective automated contouring tool for liver diagnosis and treatment planning.","feed_headline":"New network tops 12 methods in CT liver-tumor segmentation","feed_subtitle":"T-MPEDNet reports Dice of 97.6% for liver and 89.1% for tumor on LiTS.","key_machinery":"The load-bearing idea is a progressive encoder-decoder built from five Adaptive Feature Extraction (AdaFEx) stages, joined by skip connections that go through a Compressive Channel Recalibration (CCR) module before each decoder stage. A Dynamic Contextual Attention (DCA) module applies dynamic convolutions before multi-head self-attention to capture long-range spatial context, a Multi-Scale Atrous Spatial (MSAS) module uses parallel dilated convolutions with dilation rates 1, 4, 8, and 12 to span local and global scales, and a Morphological Boundary Refinement (MBR) post-processor erodes the output mask to isolate and sharpen boundary pixels.","core_discovery":"The paper's central claim is that T-MPEDNet surpasses all twelve compared methods on both public benchmarks. The network couples a five-stage progressive encoder-decoder whose skip connections pass through a compressive channel recalibration unit with a transformer-inspired dynamic contextual attention module and a multi-scale atrous spatial module, then erodes the predicted mask to sharpen boundaries. On LiTS it reports Dice similarity coefficients of 97.6% and 89.1% for liver and tumor; on 3DIRCADb it reports 98.3% and 83.3%. Ablations on LiTS attribute the largest single contributions to channel recalibration and dynamic attention.","pith_inferences":["A controlled re-run with all twelve baselines retrained under identical splits and preprocessing would test whether the reported 0.4 to 1.2 point margins are due to the architecture; the paper does not report such a re-run.","Because boundary refinement is a generic morphological operation, it may transfer to other low-contrast organ segmentation tasks, such as pancreas or kidney tumors, without architectural changes.","The largest gains come from modules that reweight channels and capture long-range context, so a similar combination could help small-lesion segmentation in other CT modalities; testing this would require new experiments.","Training on 2D slices at 256 by 256 resolution leaves inter-slice context unused, so a 3D or higher-resolution variant is a natural next step that could push tumor Dice further; the paper lists this direction as future work."],"forward_implications":["On LiTS, the full model reports 97.6% Dice for liver and 89.1% for tumor; on 3DIRCADb it reports 98.3% and 83.3%.","The ablation study shows that the compressive channel recalibration module is the largest single contributor, with its removal dropping liver Dice by 4.1 points and tumor Dice by 7.7 points on LiTS.","The dynamic contextual attention module contributes the second-largest margin, with its removal dropping liver Dice by 2.2 points and tumor Dice by 5.2 points.","The morphological boundary refinement adds a smaller margin (0.2 liver, 0.5 tumor Dice on LiTS) and works as a post-processing step, so it could be reused without retraining the network."],"supporting_citations":[{"why":"Supplies the LiTS benchmark and ground-truth labels on which the headline Dice scores are computed.","marker":"[19]"},{"why":"Provides the 3DIRCADb benchmark used to test generalization of liver and tumor segmentation.","marker":"[20]"},{"why":"U-Net is the architectural ancestor whose skip-connection design T-MPEDNet extends with recalibration.","marker":"[9]"},{"why":"TD-Net is the strongest attention-based baseline on LiTS liver and a required comparison.","marker":"[8]"},{"why":"HFRU-Net is the strongest baseline on 3DIRCADb liver and a required comparison.","marker":"[13]"},{"why":"RMAU-Net is the strongest baseline for 3DIRCADb tumor and a required comparison.","marker":"[14]"},{"why":"DefED-Net is the strongest baseline for LiTS tumor and a required comparison.","marker":"[41]"},{"why":"Supplies the self-attention formulation that the DCA module adapts with dynamic convolutions.","marker":"[17]"}],"fun_headline_variants":["T-MPEDNet outperforms 12 methods in liver-tumor CT segmentation","T-MPEDNet tops 12 baselines in CT liver-tumor segmentation","CT liver-tumor segmentation: T-MPEDNet beats 12 methods","T-MPEDNet hits 97.6% liver Dice, 89.1% tumor on LiTS","T-MPEDNet uses transformer attention for liver-tumor CT segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on the baseline methods being evaluated under the same data split and preprocessing as the new network; if they were not, the reported advantages could come from the test setup rather than from the architecture.","fun_headline_variants_meta":{"raw":{"variants":["T-MPEDNet outperforms 12 methods in liver-tumor CT segmentation","T-MPEDNet tops 12 baselines in CT liver-tumor segmentation","CT liver-tumor segmentation: T-MPEDNet beats 12 methods","T-MPEDNet hits 97.6% liver Dice, 89.1% tumor on LiTS","T-MPEDNet uses transformer attention for liver-tumor CT segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001814,"raw_usage":{"total_tokens":7176,"prompt_tokens":1022,"completion_tokens":6154,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":6044}},"tokens_in":638,"tokens_out":6154,"duration_ms":41712,"temperature":1.0,"reasoning_tokens":6044,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:14:39.477788+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run T-MPEDNet and all twelve compared methods on the same 80/10/10 split of LiTS and 3DIRCADb, with identical preprocessing (HU windowing, CLAHE, z-score normalization, and 256 by 256 resizing) and the same optimizer and augmentation, then compare Dice on the identical test slices; if the margins shrink below the reported 0.4 to 1.2 points or reverse sign, the central claim fails.","supporting_citations":[{"cited_title":"Soler, A","cited_arxiv_id":null,"evidence_quote":"Provides the 3DIRCADb benchmark used to test generalization of liver and tumor segmentation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TD-Net is the strongest attention-based baseline on LiTS liver and a required comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HFRU-Net is the strongest baseline on 3DIRCADb liver and a required comparison."},{"cited_title":"Jiang, J","cited_arxiv_id":null,"evidence_quote":"RMAU-Net is the strongest baseline for 3DIRCADb tumor and a required comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DefED-Net is the strongest baseline for LiTS tumor and a required comparison."}],"review_version":1}