{"id":"7d3f5393-607a-440e-b2b0-adff47411989","arxiv_id":"2502.00042","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LSU-Net combines depthwise separable convolutions, a spatial shift block, and adaptive multi-level loss to achieve strong organ segmentation with 1.08M parameters.","lead":"This paper introduces LSU-Net, a small U-shaped neural network that segments organs in medical scans using about one million parameters. It reports better accuracy than larger standard models on two public medical imaging datasets, suggesting lightweight AI can fit into clinics with limited computing power.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Baseline training protocol and statistical reliability are underspecified, making the claimed superiority margins unverifiable.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: the comparison protocol for baselines is unspecified and no statistical confidence is reported. The central claim is empirical, and its strength depends on the margins being reproducible and significant; the margin on MSD Colon is particularly small. I considered whether there is a deeper internal inconsistency, such as the unusual FLOPs increase in the 'w/o tokenized shift block' ablation, but this can be explained by the block's downsampling behavior and does not undermine the architecture internally. The lack of transformer baselines is a scope weakness, but the more fundamental issue is the unverifiable training protocol. This is an addressable reproducibility problem rather than a fatal flaw, so the conditional verdict remains appropriate. My additional recommendations—code release, multi-seed runs, and explicit baseline protocol—would strengthen the paper to a full acceptance, but they do not change the current conditional status. I agree with the reader on the weakest assumption and recommend no change to the verdict.","tokens_in":7005,"tokens_out":4347,"duration_ms":40267,"concrete_test":"Train the five comparison baselines (UNet, UNet++, Attention-UNet, Res-UNet, UNeXt) using the exact hyperparameters reported for LSU-Net (Adam, initial LR 1e-3, CosineAnnealingLR to 1e-5, 100 epochs, batch size 16 for UWMGI and 10 for MSD Colon, same resizing and normalization) on the official data splits, with at least 5 random seeds each. Report mean ± std of mIoU and DSC. Then check whether the reported LSU-Net margins (86.04 vs 84.02 mIoU on UWMGI; 72.79 vs 72.39 on MSD Colon) exceed the combined standard error. If the gaps fall within one standard deviation, the superiority claim is not statistically supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—LSU-Net outperforms most state-of-the-art segmentation architectures—rests entirely on the comparison in Tables 1 and 2. The manuscript never states whether all baselines (UNet, UNet++, Attention-UNet, Res-UNet, UNeXt) were trained with the same data splits, preprocessing, optimizer settings, or number of epochs. The Implementation Details section describes only LSU-Net's training, and no code is released. On MSD Colon, the margin over UNeXt is just 0.40 mIoU (72.79 vs 72.39), which is within typical run-to-run variance for deep segmentation models, especially without error bars or multiple seeds. The paper also cites Swin-UNet and TransUNet in the introduction but omits them from the comparison, so the 'most state-of-the-art' claim is broader than the evidence. These omissions are not fatal to the architecture, but they make the headline empirical claim unverifiable as reported. The ablations do show internal consistency, but they do not validate the comparative claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces LSU-Net, a lightweight U-shaped segmentation network that combines Light Conv Blocks (standard and depthwise separable convolutions) in shallow stages with Tokenized Shift Blocks (spatial shifts plus depthwise separable convolutions) in deeper stages, and trains with a multi-scale deep supervision loss whose per-level weights are learned via an automatic weighted loss scheme. The method is evaluated on the UWMGI and MSD Colon datasets against five baselines, reporting 86.04 mIoU / 89.43 DSC on UWMGI and 72.79 mIoU / 84.05 DSC on MSD Colon with 1.08 M parameters. The paper claims that LSU-Net outperforms most state-of-the-art segmentation architectures while being substantially lighter.","tokens_in":7119,"tokens_out":4680,"duration_ms":42902,"significance":"If the comparative results are reliable, LSU-Net offers a useful efficiency-accuracy trade-off for resource-constrained clinical settings: the architecture is simple, the design choices are clearly motivated by parameter reduction, and the module equations and Figure 1 provide a reasonably complete specification. The ablation study examines each proposed component, which is a strength. The main weakness is not the architecture itself but the empirical evidence supporting the headline claim: the baseline training protocol is not described, all metrics are single-run point estimates, and the closest competitor margin on MSD Colon is small. These issues are fixable with additional experiments and reporting, so the contribution is worth considering after a major revision.","major_comments":[{"comment":"The implementation details paragraph specifies only LSU-Net's own training (Adam, initial learning rate 0.001, CosineAnnealingLR, 100 epochs, batch sizes 16 for UWMGI and 10 for MSD Colon). It does not state whether the five baseline networks were retrained under the same protocol or whether their numbers were taken from the literature. Data splits, preprocessing (including resizing and normalization), optimizer settings, and training duration can change mIoU by several points; without this information the reported margins (e.g., 86.04 vs 84.02 on UWMGI) cannot be attributed to the architecture. Please specify the baseline training protocol for both datasets or, if the numbers are cited from prior work, give the exact sources and verify that the evaluation settings are comparable.","section":"Section 3, Implementation Details and Tables 1-2"},{"comment":"All quantitative results are reported as single-run point estimates, with no standard deviations, confidence intervals, or number of seeds. In Table 2, LSU-Net's mIoU advantage over UNeXt is only 0.40 percentage points (72.79 vs 72.39) and the DSC advantage is 0.21 points (84.05 vs 83.84), which is within the run-to-run variability commonly observed in deep segmentation training. The same issue affects the ablation comparisons in Tables 3 and 4, where several differences are below 0.5 points (e.g., 72.51 vs 72.79 in Table 4). Please repeat experiments with at least three seeds, report mean and standard deviation, and use a paired test where appropriate, especially for the comparisons against UNeXt and the w/o MDL ablation.","section":"Tables 1-4"},{"comment":"The introduction names Swin-UNet and TransUNet as state-of-the-art Transformer-based segmentation models, but neither appears in the comparison tables. The headline claim that LSU-Net \"outperforms most state-of-the-art segmentation architectures\" is therefore broader than the evaluated set, which contains only convolutional baselines plus UNeXt. Either add Transformer-based baselines (or report their published scores on these datasets under matching protocols) or rephrase the claim to refer specifically to the compared lightweight convolutional models.","section":"Section 1 and Tables 1-2"}],"minor_comments":[{"comment":"The title and abstract contain spacing artifacts (\"AUTOMA TIC\", \"SEGMENTA TION\"), and Section 2 contains a typo in \"shif trefers\" just before Eq. (8).","section":"Title and Section 2"},{"comment":"The table headers have missing spaces (e.g., \"Params(M)GFLOPs\" and \"w/o tokenized shift block0.26\"), which makes the tables harder to read.","section":"Tables 3 and 4"},{"comment":"The text says the shifted feature parts are \"fused through a partitioned attention mechanism,\" but Eq. (8) only shows a residual addition and no attention operation is defined in the equations; please either specify the attention step or reword the description.","section":"Section 2, Tokenized Shift Block"},{"comment":"The shift equations use slice notation that presupposes the channel dimension is divisible by 4 and the spatial dimensions are larger than 2; stating these constraints explicitly would improve implementation clarity.","section":"Eqs. (6)-(7)"},{"comment":"The caption says \"origin image\" instead of \"original image,\" and the row labels would be clearer if each row's dataset name appeared directly on the figure rather than only in the caption.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a medical imaging venue and the architectural contribution is plausible, but the empirical validation needs substantial strengthening before acceptance. I do not see grounds to question novelty disclosure, since the building blocks are drawn from cited prior work; the main concern is completeness of the comparative evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a competent lightweight U-Net variant, but the headline claim is not yet verified. The paper combines known blocks—depthwise separable convolutions, the Spatial Shift Block from S2-MLPv2, and Liebel and Körner's adaptive weighted loss—into a U-Net scaffold. The combination is new, and the ablations show each module helps on both datasets. The architecture is clearly described, and the parameter count (1.08M) is genuinely low.\n\nThe soft spot is the comparison. The Implementation Details section describes only LSU-Net's own training; it never states whether UNet, UNet++, Attention-UNet, Res-UNet, and UNeXt were trained with the same splits, preprocessing, optimizer, or epochs. All tables report single runs without standard deviations. On MSD Colon, the margin over UNeXt is 0.40 mIoU, which is within normal run-to-run variance. The paper also cites Swin-UNet and TransUNet in the introduction but omits them from the experiments, so 'most state-of-the-art' is broader than what was actually compared. No code is released, so the numbers cannot be checked.\n\nAll of this is fixable, but it is load-bearing: the central claim is that LSU-Net beats larger models in both accuracy and efficiency. The efficiency claim is solid; the accuracy claim is plausible but unproven as reported. The ablations are internally consistent and the loss-weighting mechanism is described well enough to reproduce.\n\nThis paper is for readers working on efficient medical segmentation who want a sensible lightweight U-Net with standard tricks. It does not break new ground, and the empirical superiority claim needs more support. But it is not a waste of referee time. I would send it out, with the expectation that the authors add error bars, specify the baseline training protocol, release code, and include at least one transformer baseline.","headline":"A sensible lightweight U-Net combination whose empirical superiority claim is not yet verifiable because baselines, error bars, and code are missing.","tokens_in":7683,"tokens_out":2547,"would_cite":false,"duration_ms":22778,"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 1.08M-parameter U-Net variant reports higher mIoU than most larger baselines on two organ-segmentation benchmarks.","keywords":["lightweight medical image segmentation","U-Net","depthwise separable convolution","spatial shift","deep supervision","multi-scale loss","abdominal organ segmentation","encoder-decoder network"],"falsifier":"Retrain UNet, UNet++, Attention-UNet, Res-UNet, and UNeXt on the UWMGI and MSD Colon datasets using the exact splits, preprocessing, optimizer, scheduler, epochs, and seeds used for LSU-Net, and report per-run variance; if the mIoU margins (86.04 versus 84.02 on UWMGI, 72.79 versus 72.39 on MSD Colon) fall within run-to-run noise, the central claim of outperformance is not established.","tokens_in":6769,"feed_emoji":"🩻","tokens_out":9947,"duration_ms":82275,"temperature":0.7,"pith_summary":"LSU-Net is a lightweight encoder-decoder network for segmenting organs in medical images, built as a U-Net in which shallow stages use Light Conv Blocks and deep stages use Tokenized Shift Blocks. The design goal is to keep parameter count and computation low enough for clinical use while retaining segmentation accuracy. The paper reports that the 1.08M-parameter model reaches 86.04% mIoU on the UWMGI dataset and 72.79% mIoU on the MSD Colon dataset, beating most of the larger U-Net variants it is compared with, and that a dynamically weighted multi-scale loss adds a consistent gain in ablation tests. A sympathetic reader would care because it claims the usual accuracy-versus-efficiency tradeoff can be shifted: a fraction of the parameters of standard U-Nets may suffice for these tasks.","feed_headline":"1.08M-parameter U-Net beats larger rivals on organ scans","feed_subtitle":"Spatial shift and lightweight convolutions reach 86% mIoU on abdominal organs with a fraction of the compute","key_machinery":"The network keeps the classic encoder-decoder skeleton and assigns different modules to different depths: Light Conv Blocks in the shallow stages capture local features with few parameters, while Tokenized Shift Blocks in the deep stages capture regional dependencies. A Light Conv Block is a residual stack of a 3x3 convolution, batch normalization and GELU, a depthwise separable convolution, group normalization, and 1x1 convolutions, with a 1x1 skip connection in the encoder. A Tokenized Shift Block starts with a Spatial Shift Block that expands the feature map along the channel dimension, shifts three channel groups in different spatial directions, and fuses them, then applies depthwise separable convolution, group normalization, and 1x1 convolutions; when downsampling, an Overlap Patch Embedding is prepended. The Multi-scale Deep Loss (MDL) applies Automatic Weighted Loss (AWL), combining per-level losses as $\\sum_i \\frac{1}{2\\sigma_i^2}L_i + \\sum_i \\ln(1+\\sigma_i^2)$, so the per-level weights are learned during training and the logarithmic term keeps them bounded.","core_discovery":"On its own terms, the paper's central claim is that a U-shaped segmentation network assembled from Light Conv Blocks and Tokenized Shift Blocks, trained with a Multi-scale Deep Loss under Automatic Weighted Loss, outperforms most existing segmentation architectures on the UWMGI and MSD Colon datasets while using only 1.08M parameters and about 1.10 GFLOPs. The reported numbers are 86.04% mIoU and 89.43% DSC on UWMGI, ahead of UNet, UNet++, Attention-UNet, Res-UNet, and UNeXt, and 72.79% mIoU and 84.05% DSC on MSD Colon, ahead of the same baselines, with the narrowest margin over the similarly lightweight UNeXt (72.39% mIoU). The ablations attribute the gain to all three components, with the multi-scale loss contributing the clearest improvement on the multi-class UWMGI dataset.","pith_inferences":["Beyond the paper: if the comparison protocol is exactly what it appears to be, the reported margins suggest that parameter count is not the binding constraint on these two benchmarks, so even lighter or faster variants may be achievable.","Implicit extension: the Automatic Weighted Loss combination of per-scale losses is architecture-agnostic and could be applied to other multi-scale or deeply supervised segmentation networks.","Editorial inference: the paper gives no run-to-run variance for any model, so a multi-seed reproducibility study would reveal how much of the reported superiority is a stable architecture effect versus training luck.","Testable extension: replacing the Tokenized Shift Block's hand-designed shift directions with learned shifts, or varying the number of shifted channel groups, could show whether the exact shift pattern is what matters or simply the channel-wise receptive-field expansion."],"forward_implications":["A U-Net with 1.08M parameters and roughly 1.10 GFLOPs can match or beat several much larger baselines on both a multi-class and a single-class organ-segmentation benchmark.","The Tokenized Shift Block shows that spatial-shift operations can serve as a low-parameter alternative to heavier attention or large MLP blocks in deep encoder-decoder stages.","Multi-scale deep supervision with automatically learned per-layer weights improves mIoU and DSC in both reported ablations.","The ablation results imply that both lightweight convolution and tokenized shift modules contribute to the final accuracy, not just one of them.","At 1.08M parameters, the architecture is small enough to consider for clinical settings with limited computational resources, which is the stated motivation of the paper."],"supporting_citations":[{"why":"It defines the U-Net encoder-decoder architecture that LSU-Net extends and serves as the main baseline it must beat.","marker":"[1]"},{"why":"It supplies the nested UNet++ baseline used for comparison.","marker":"[2]"},{"why":"It supplies the attention-gated Attention-UNet baseline used for comparison.","marker":"[3]"},{"why":"It is the closest lightweight baseline; UNeXt's reported scores are the immediate bar for LSU-Net.","marker":"[6]"},{"why":"It is the UWMGI dataset, the multi-class abdominal segmentation benchmark where the largest margins are reported.","marker":"[13]"},{"why":"It is the MSD Colon dataset, the single-class benchmark used to test transferability.","marker":"[14]"},{"why":"It provides the Spatial Shift Block that the Tokenized Shift Block is built around.","marker":"[15]"},{"why":"It provides the Automatic Weighted Loss formula that combines the per-level losses in Multi-scale Deep Loss.","marker":"[16]"},{"why":"It provides the Res-UNet baseline used for comparison.","marker":"[18]"}],"fun_headline_variants":["Lightweight U-Net with 1.08M params tops larger organ segmentation models","Slim 1.08M-parameter U-Net outsegments heavyweights on organ scans","1.08M-param U-Net: light enough for clinics, sharp on organs","Shift-based U-Net cuts parameters, still tops organ mIoU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that every baseline network was trained under exactly the same protocol as LSU-Net (same data splits, preprocessing, optimizer, learning-rate schedule, and epochs); the paper does not describe the baseline training procedure and reports no error bars, so if that assumption fails the reported margins could shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight U-Net with 1.08M params tops larger organ segmentation models","Slim 1.08M-parameter U-Net outsegments heavyweights on organ scans","1.08M-param U-Net: light enough for clinics, sharp on organs","Shift-based U-Net cuts parameters, still tops organ mIoU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000504,"raw_usage":{"total_tokens":2447,"prompt_tokens":921,"completion_tokens":1526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":1434}},"tokens_in":537,"tokens_out":1526,"duration_ms":10864,"temperature":1.0,"reasoning_tokens":1434,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:43:58.439935+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain UNet, UNet++, Attention-UNet, Res-UNet, and UNeXt on the UWMGI and MSD Colon datasets using the exact splits, preprocessing, optimizer, scheduler, epochs, and seeds used for LSU-Net, and report per-run variance; if the mIoU margins (86.04 versus 84.02 on UWMGI, 72.79 versus 72.39 on MSD Colon) fall within run-to-run noise, the central claim of outperformance is not established.","supporting_citations":[{"cited_title":"Transbts: Multimodal brain tumor segmentation using transformer,","cited_arxiv_id":null,"evidence_quote":"It is the MSD Colon dataset, the single-class benchmark used to test transferability."},{"cited_title":"Medical transformer: Gated axial-attention for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"It is the UWMGI dataset, the multi-class abdominal segmentation benchmark where the largest margins are reported."},{"cited_title":"Deep learning has significantly advanced the field with diverse network ar- chitectures, especially the U-shaped encoder-decoder design starting with UNet [1]","cited_arxiv_id":null,"evidence_quote":"It defines the U-Net encoder-decoder architecture that LSU-Net extends and serves as the main baseline it must beat."},{"cited_title":"LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images","cited_arxiv_id":"2502.00042","evidence_quote":"It supplies the nested UNet++ baseline used for comparison."},{"cited_title":"First, we outline some implementation details","cited_arxiv_id":null,"evidence_quote":"It supplies the attention-gated Attention-UNet baseline used for comparison."},{"cited_title":"Unet++: A nested u- net architecture for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"It is the closest lightweight baseline; UNeXt's reported scores are the immediate bar for LSU-Net."},{"cited_title":"Axial- deeplab: Stand-alone axial-attention for panoptic seg- mentation,","cited_arxiv_id":null,"evidence_quote":"It provides the Automatic Weighted Loss formula that combines the per-level losses in Multi-scale Deep Loss."}],"review_version":1}