{"id":"67fced79-6e37-49ae-b175-05f8ddd3ccf6","arxiv_id":"2501.03838","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LM-Net reports state-of-the-art medical segmentation results on Kvasir-SEG, LGG, and breast ultrasound datasets with 5.4M parameters, using combined multi-branch convolutions, local window attention, and global feature-pyramid attention.","lead":"LM-Net is a lightweight U-shaped network that combines parallel multi-scale convolutions with local and global transformer branches for medical image segmentation. On three public datasets it reports higher Dice scores than several prior models while using roughly 5.4 million parameters and 4.66 GFLOPs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LGG evaluation uses image-level splits across only 110 patients, so the reported 95.48% mDice may reflect patient leakage; a patient-disjoint resplit is needed before the three-dataset SOTA claim can stand.","rationale":"The reader's weakest-assumption analysis identifies exactly the load-bearing concern: Section 4.2 describes an image-level 0.8:0.1:0.1 split with no mention of patient grouping, while the LGG dataset comprises 3,929 images from only 110 patients. Because adjacent MRI slices from one patient are highly correlated, random image-level splitting can leak the same patient's data into both training and test sets, inflating the reported metrics. This matters because the paper's headline contribution is 'state-of-the-art results on three datasets with different modalities'; if the LGG result is invalidated, the claim reduces to two datasets. The issue is an evaluation-protocol flaw rather than an internal inconsistency: the structural re-parameterization math in Section 3.2 is sound, the ablations are systematic, and the released GitHub code supports reproducibility. A patient-disjoint resplit is a straightforward correction, and Table 2 should be recomputed with modern LGG baselines under the same protocol. Because the paper's architecture itself may be sound and the defect is fixable, the appropriate verdict remains conditional rather than accept or reject. The reader's verdict already reflects this, so no adjustment is needed.","tokens_in":19192,"tokens_out":3267,"duration_ms":31881,"concrete_test":"Obtain the patient identifiers for the LGG dataset and perform a patient-stratified split, holding out whole patients rather than individual slices (e.g., 88/11/11 patients), while keeping the same preprocessing, augmentation, optimizer, loss, and 200-epoch schedule. Retrain LM-Net and the strongest baselines from Table 2 under this split. If LM-Net's mDice drops substantially, or if another baseline matches or exceeds it, the LGG state-of-the-art claim is not supported under a clinically meaningful evaluation protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 reports that 'each dataset is split into training, validation, and test subsets with a ratio of 0.8:0.1:0.1' with no patient-level grouping. The LGG dataset contains 3,929 images from only 110 patients, so a random image-level split will almost certainly place slices from the same patient in both training and test. Under such leakage, the model can memorize patient-specific intensity patterns and scanner characteristics, inflating the reported 95.48% mDice and 91.7% mIoU in Table 2. This is load-bearing because the abstract's central claim is state-of-the-art results on three datasets; LGG is one of the three supporting pillars. The Kvasir-SEG and BUS comparisons are not similarly affected by this issue, but the three-dataset claim fails if the LGG result cannot be reproduced under a patient-disjoint split.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LM-Net, a lightweight U-shaped architecture for medical image segmentation that combines a re-parameterizable multi-branch convolutional module with two transformer-based modules: LFT for local window self-attention and GFT for global self-attention over a feature pyramid. The authors claim state-of-the-art segmentation results on three public datasets (Kvasir-SEG, LGG, and breast ultrasound) while using only 4.66 GFLOPs and 5.4M parameters. The core technical idea is standard structural re-parameterization (Eqs. 3–9) to merge multiple convolutional branches into one inference-time convolution, and a local/global feature fusion scheme. Experiments compare against several U-Net variants and transformer hybrids, with ablations on the multi-branch kernels, GFT, and LFT.","tokens_in":19402,"tokens_out":3057,"duration_ms":28964,"significance":"If the reported results hold, LM-Net would be a practically useful lightweight medical segmentation model, especially for embedded computer-aided diagnosis. The manuscript's strengths include a correct and clearly explained re-parameterization identity, reproducibility-favorable details (training schedule, loss, augmentation), and relatively extensive ablation studies on the proposed modules. The efficiency claim (4.66 GFLOPs, 5.4M parameters) is specific and testable. However, the central 'state-of-the-art on three datasets' claim is currently under-supported because the LGG evaluation uses an image-level split with high risk of patient leakage, the Kvasir state-of-the-art comparison lacks error bars, and the LGG/BUSI baselines omit more recent segmentation methods. These issues are fixable but require additional experiments or a toned-down claim.","major_comments":[{"comment":"The LGG dataset contains 3,929 images from only 110 patients, and the split described in Section 4.2 is an image-level 0.8:0.1:0.1 split with no patient-level grouping. This almost certainly places slices from the same patient in both training and test, which can inflate the reported 95.48% mDice via patient-specific memorization. Because the abstract's central claim is state-of-the-art results on three datasets, this is load-bearing. Please re-run the LGG experiments with a patient-disjoint split (e.g., split by patient ID) and report whether the numbers in Table 2 change.","section":"Section 4.2 and Table 2"},{"comment":"The Kvasir-SEG state-of-the-art comparison reports only a single run for FCBFormer, SSFormer-L, and ESFPNet-L, while LM-Net is reported with standard deviations from Table 1. The 0.24-point mDice margin over FCBFormer (94.09 vs. 93.85) may be within run-to-run noise. Please provide error bars for all methods (e.g., multiple seeds or a paired significance test) before claiming to surpass prior state of the art.","section":"Table 5"},{"comment":"The baseline set for LGG and breast ultrasound consists mostly of older architectures (U-Net variants, TransUNet, Swin-Unet, FCN, DeepLabv3+). More recent and directly competitive segmentation methods, such as nnU-Net, PraNet, or other lightweight transformer-CNN hybrids, are missing. Without such baselines, the 'state-of-the-art' claim on these two datasets is not well supported. Either add more recent baselines or explicitly limit the claim to the compared methods.","section":"Tables 2 and 3"}],"minor_comments":[{"comment":"There is a typo: 'Specifcally' should be 'Specifically'.","section":"Section 4.2"},{"comment":"The summation in Eq. (7) implicitly requires that the four branch kernels (3×1, 1×3, 3×3, 5×5) be zero-padded to a common spatial size before element-wise addition. Please state this alignment explicitly, as it is an essential implementation detail for the re-parameterization to be valid.","section":"Equations (7)-(9)"},{"comment":"The per-class results for the ultrasound dataset are presented in a very cramped single-row format that is hard to read. Please reformat into separate columns or a multi-row structure for clarity.","section":"Table 4"},{"comment":"The sentence 'the second and third stages contain of three layers' contains a grammatical error; it should be 'consist of three layers'.","section":"Section 3.3"},{"comment":"The text says that LM-Net 'achieves competitive results with FCBFormer and SSFormer' in one place but the abstract and introduction claim 'state-of-the-art results surpassing previous methods'. Please make the strength of the claim consistent throughout the manuscript.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the LGG split leakage, which is readily fixable but essential before the three-dataset SOTA claim can be trusted. The manuscript otherwise has a sound technical core and reasonably detailed ablations. If the authors re-run LGG with a patient-disjoint split, add error bars to Table 5, and either include recent baselines or temper the SOTA wording, the paper could be acceptable. The citation pattern includes many self-citations of the corresponding author's group, but they appear in related work rather than as circular support for the main contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a clean architecture paper. LM-Net is a U-Net with three known ingredients: a RepVGG-style multi-branch block, a UperNet-like global feature pyramid transformer (GFT), and a local window attention transformer (LFT). None of these is new, but the combination is coherent, the re-parameterization equations (Eqs. 3-9) are correct, and the Kvasir ablations are properly controlled: multi-branch variants, GFT vs ASPP, single-scale vs multi-scale, LFT skip variants. The efficiency numbers (4.66 GFLOPs, 5.4M parameters) are a genuine selling point, and the code is on GitHub. Credit where due: the experimental design on Kvasir is fairly thorough, and the paper itself admits that random initialization may be inferior, which is a sign of honest reporting.\n\nThe soft spots are real but fixable. The \"state-of-the-art on three datasets\" claim is not fully supported. The LGG dataset has 3,929 images from only 110 patients, yet Section 4.2 splits 0.8/0.1/0.1 at image level with no patient grouping. Same-patient slices will almost certainly appear in both training and test, so the reported 95.48% mDice is likely inflated. The margins over Unet++ and Deeplabv3+ on LGG are 0.05 and 0.02 mDice, so a patient-disjoint resplit could easily flip the ranking. Also, the LGG and BUS comparisons omit modern baselines like nnU-Net and UNETR; Swin-Unet is ancient history now. The Kvasir Table 5 comparison reports no error bars for prior methods, and the 0.24 mDice gain over FCBFormer is within the paper's own standard deviation (0.11), so the SOTA claim there needs a significance test. Minor: Table 4 has messy formatting but the numbers are legible.\n\nNone of this is a load-bearing flaw in the architecture itself. The risk is in the empirical claim, not the math. For a reader working on efficient medical segmentation, the GFT/LFT design and the multi-branch re-param block are worth studying. The paper deserves a serious referee: someone should ask for a patient-disjoint LGG split, modern baselines on LGG and BUS, and error bars or significance tests on the Kvasir comparison. If those hold up, it is a useful contribution. I would not cite it as SOTA until then, but I would send it to review.","headline":"A competent lightweight segmentation architecture with honest ablations, but the three-dataset SOTA claim is overreached by an image-level LGG split and thin baselines.","tokens_in":19919,"tokens_out":2180,"would_cite":false,"duration_ms":22112,"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":"LM-Net claims that a 5.4M-parameter hybrid CNN-transformer network can match or beat much heavier medical segmentation models on three public benchmarks.","keywords":["Medical image segmentation","Multi-scale feature representation","Local Feature Transformer","Global Feature Transformer","Lightweight network","Structural re-parameterization","U-shaped encoder-decoder","Hybrid CNN-Transformer"],"falsifier":"Run a patient-stratified cross-validation on LGG—all slices of each patient confined to a single fold—and compare the resulting mDice with the reported 95.48%; a large drop would show the headline result depends on image-level splitting.","tokens_in":19005,"feed_emoji":"🩻","tokens_out":10507,"duration_ms":89070,"temperature":0.7,"pith_summary":"LM-Net is a U-shaped medical image segmentation network designed to be accurate while staying small enough for clinical computer-aided diagnosis. The paper's central claim is that a deliberate two-level use of multi-scale features—parallel convolutions of different kernel sizes within each stage, plus feature-pyramid transformers across stages—lets one model beat heavier baselines on polyp, low-grade glioma, and breast ultrasound segmentation. The Local Feature Transformer (LFT) handles fine boundary texture with windowed self-attention; the Global Feature Transformer (GFT) captures whole-lesion context with global self-attention. On the three test sets the model reports 94.09%, 95.48%, and 90.96% mean Dice, respectively, at 4.66G FLOPs and 5.4M parameters, which the paper offers as evidence that a small hybrid CNN-transformer can match or exceed much larger models. The paper's argument is that local and global cues, plus same-level and cross-level scales, are complementary and together remove the usual over-segmentation, under-segmentation, and blurry-boundary failure modes.","feed_headline":"5.4M-parameter network tops three medical segmentation tests","feed_subtitle":"Local and global multi-scale transformers lift Dice scores on polyps, gliomas, and breast ultrasound at 4.66G FLOPs.","key_machinery":"The load-bearing mechanism is the multi-branch module and the convolution algebra behind structural re-parameterization. Because convolution is homogeneous and additive ($F \\otimes (\\lambda K) = \\lambda (F \\otimes K)$ and $F \\otimes K_1 + F \\otimes K_2 = F \\otimes (K_1 + K_2)$), the four parallel branches—each a convolution plus batch normalization—can be folded into one ordinary convolution at inference, so training gets multi-scale receptive fields while inference pays only single-convolution cost. On top of that, GFT feeds a five-level feature pyramid into a global self-attention transformer, and LFT feeds adjacent-stage local pyramids into windowed self-attention; together they give the decoder both global context and fine boundary detail. The feature-pyramid-plus-transformer design is what lets the network use multi-scale information across levels, not just within them.","core_discovery":"The central claim the paper argues for is that the persistent failure modes of medical segmentation—over-segmentation, under-segmentation, and blurry boundaries—can be addressed by combining multi-scale representation at two levels of analysis, and that this can be done in a lightweight network. Within each encoder stage, a multi-branch module runs four depth-wise convolutions with kernels 3×1, 1×3, 3×3, and 5×5 in parallel, then merges them at inference through structural re-parameterization. Across stages, GFT builds a global feature pyramid from all encoder outputs and feeds it to global self-attention, while LFT builds local pyramids from adjacent stages and feeds them to windowed self-attention. The decoder then fuses local texture and global semantics layer by layer. The paper reports the consequence as state-of-the-art mean Dice of 94.09% on Kvasir-SEG, 95.48% on LGG, and 90.96% on breast ultrasound, using only 5.4M parameters and 4.66G FLOPs, and argues the gain comes not from extra capacity but from local-global and same-level/cross-level complementarity.","pith_inferences":["The paper does not test a patient-stratified split on LGG; if one were run, the reported 95.48% mDice might shrink when slices of the same tumor are kept out of training.","Because LFT and GFT are described as plug-and-play, attaching them to other U-shaped encoders would show whether the gain comes from the transformers or from the specific multi-branch backbone.","The same convolution-additivity identity could be applied to other branch configurations, so the observed falloff with more branches is a finding about this model, not a general limit of re-parameterization.","A controlled input-resolution sweep could test whether the 5×5 branch matters at 256×256 or is an artifact of that choice."],"forward_implications":["If the reported numbers hold, a 5.4M-parameter model at 4.66G FLOPs can replace much heavier U-Net variants in computer-aided diagnosis pipelines where memory and latency matter.","The LFT/GFT pair implies that boundary quality, quantified by lower Hausdorff distance on Kvasir and ultrasound, does not require a larger or deeper network, only a decoder that receives both local textures and global semantics.","Because the multi-branch module re-parameterizes into a single convolution, the multi-scale benefit is nearly free at deployment, with inference cost staying close to that of one convolutional layer.","The reported faster convergence without ImageNet pre-training suggests the local-global hybrid can be trained from scratch on small medical datasets, lowering the barrier for new segmentation tasks.","State-of-the-art results across endoscopy, MRI, and ultrasound imply the architecture is not tuned to one imaging modality, making it a plausible general-purpose medical segmentation backbone."],"supporting_citations":[{"why":"This is the structural re-parameterization work that supplies the convolution homogeneity and additivity used to fuse the four branches into one convolution at inference.","marker":"[25]"},{"why":"This paper's feature-pyramid design is the template for GFT's global pyramid built from all encoder stages.","marker":"[51]"},{"why":"This local window self-attention mechanism gives LFT its local inductive bias and boundary-texture extraction with linear complexity.","marker":"[56]"},{"why":"The Squeeze-and-Excitation layer inside the multi-branch module recalibrates channel features.","marker":"[10]"},{"why":"Kvasir-SEG is the polyp dataset where LM-Net reports 94.09% mDice and compares against prior methods.","marker":"[26]"},{"why":"These references define the LGG low-grade glioma dataset used for the 95.48% mDice result.","marker":"[30, 31]"},{"why":"This is the breast ultrasound dataset with normal, benign, and malignant classes where LM-Net reports 90.96% mDice.","marker":"[32]"},{"why":"FCBFormer is the best prior Kvasir baseline, at 93.85% mDice, that LM-Net says it surpasses.","marker":"[29]"},{"why":"ESFPNet-L is the lightweight prior Kvasir baseline that LM-Net compares against and beats.","marker":"[27]"},{"why":"SSFormer-L is another Kvasir state-of-the-art baseline that LM-Net reports outperforming.","marker":"[28]"}],"fun_headline_variants":["LM-Net: 5.4M params, multi-scale fusion tops three medical sets","Lightweight multi-scale net with local-global transformers hits SOTA","Hybrid CNN-ViT: 5.4M parameters beat prior medical segmentation","Multi-scale local-global attention yields SOTA on 3 modalities"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 80/10/10 split of the LGG dataset, which is made per image rather than per patient, treats slices from the same tumor as independent, so training and test could share a patient's data.","fun_headline_variants_meta":{"raw":{"variants":["LM-Net: 5.4M params, multi-scale fusion tops three medical sets","Lightweight multi-scale net with local-global transformers hits SOTA","Hybrid CNN-ViT: 5.4M parameters beat prior medical segmentation","Multi-scale local-global attention yields SOTA on 3 modalities"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2206,"prompt_tokens":1076,"completion_tokens":1130,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":692,"completion_tokens_details":{"reasoning_tokens":1046}},"tokens_in":692,"tokens_out":1130,"duration_ms":9122,"temperature":1.0,"reasoning_tokens":1046,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:45:38.838348+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a patient-stratified cross-validation on LGG—all slices of each patient confined to a single fold—and compare the resulting mDice with the reported 95.48%; a large drop would show the headline result depends on image-level splitting.","supporting_citations":[{"cited_title":"Repvgg: Making vgg-style convnets great again","cited_arxiv_id":null,"evidence_quote":"This is the structural re-parameterization work that supplies the convolution homogeneity and additivity used to fuse the four branches into one convolution at inference."},{"cited_title":"Unified perceptual parsing for scene understanding","cited_arxiv_id":null,"evidence_quote":"This paper's feature-pyramid design is the template for GFT's global pyramid built from all encoder stages."},{"cited_title":"Kvasir-seg: A segmented polyp dataset","cited_arxiv_id":null,"evidence_quote":"Kvasir-SEG is the polyp dataset where LM-Net reports 94.09% mDice and compares against prior methods."},{"cited_title":"Fcn-transformer feature fusion for polyp segmentation","cited_arxiv_id":null,"evidence_quote":"FCBFormer is the best prior Kvasir baseline, at 93.85% mDice, that LM-Net says it surpasses."},{"cited_title":"Stepwise feature fusion: Local guides global","cited_arxiv_id":null,"evidence_quote":"SSFormer-L is another Kvasir state-of-the-art baseline that LM-Net reports outperforming."}],"review_version":1}