REVIEW 3 major objections 5 minor 24 references
LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images
T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read A 1.08M-parameter U-Net variant reports higher mIoU than most larger baselines on two organ-segmentation benchmarks.
desk verdict A sensible lightweight U-Net combination whose empirical superiority claim is not yet verifiable because baselines, error bars, and code are missing. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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.
Reading between the lines
- 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.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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 (3)
- [Section 3, Implementation Details and Tables 1-2] 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.
- [Tables 1-4] 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 1 and Tables 1-2] 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.
minor comments (5)
- [Title and Section 2] The title and abstract contain spacing artifacts ("AUTOMA TIC", "SEGMENTA TION"), and Section 2 contains a typo in "shif trefers" just before Eq. (8).
- [Tables 3 and 4] 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 2, Tokenized Shift Block] 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.
- [Eqs. (6)-(7)] 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.
- [Figure 3 caption] 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.
Circularity Check
No circular reasoning found: LSU-Net is an empirical architecture comparison against external benchmarks with cited building blocks.
full rationale
The paper's central claim is that LSU-Net outperforms several segmentation baselines on UWMGI and MSD Colon. This claim is established by direct experiments on public datasets, not by a derivation from its own assumptions. The core components, Light Conv Block and Tokenized Shift Block, are assembled from standard operations and previously published blocks, which are cited rather than derived in this paper. The dynamic loss weights in Eq. (11) are learned during training via the cited AWL formulation; they are not fitted to the reported test metrics, so no quantity called a prediction is identical to an input by construction. The ablations in Tables 3 and 4 test each module's contribution empirically. There are no self-citations that carry the argument, no uniqueness theorem imported from the authors, and no ansatz smuggled in via citation: the Spatial Shift Block and Automatic Weighted Loss are external prior work used as components. The main reviewer concern, that baseline training protocols and error bars are underspecified, is a reproducibility and statistical-validity issue, not circularity. Therefore the derivation chain is self-contained with respect to circularity.
Assumptions & free parameters
assumptions (4)
- domain assumption UWMGI and MSD Colon datasets are representative benchmarks for abdominal organ segmentation.
- domain assumption Baseline models were trained and evaluated under comparable conditions.
- domain assumption Deep supervision via downsampled ground-truth masks at each level is a valid training signal.
- domain assumption The AWL loss formula from [16] is applicable to multi-scale segmentation losses.
Cite this review
Pith. "Pith review of LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images." pith.science (2026). https://pith.science/paper/UVLCWAAA
@misc{pith2026250200042,
author = {Pith},
title = {Pith review of: LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images},
year = {2026},
howpublished = {\url{https://pith.science/paper/UVLCWAAA}},
note = {Machine review of arXiv:2502.00042}
}
read the original abstract
UNet and its variants have widespread applications in medical image segmentation. However, the substantial number of parameters and computational complexity of these models make them less suitable for use in clinical settings with limited computational resources. To address this limitation, we propose a novel Lightweight Shift U-Net (LSU-Net). We integrate the Light Conv Block and the Tokenized Shift Block in a lightweight manner, combining them with a dynamic weight multi-loss design for efficient dynamic weight allocation. The Light Conv Block effectively captures features with a low parameter count by combining standard convolutions with depthwise separable convolutions. The Tokenized Shift Block optimizes feature representation by shifting and capturing deep features through a combination of the Spatial Shift Block and depthwise separable convolutions. Dynamic adjustment of the loss weights at each layer approaches the optimal solution and enhances training stability. We validated LSU-Net on the UWMGI and MSD Colon datasets, and experimental results demonstrate that LSU-Net outperforms most state-of-the-art segmentation architectures.
Reference graph
Works this paper leans on
-
[1]
INTRODUCTION Medical image segmentation is crucial for extracting struc- tural details to aid diagnosis and treatment. Deep learning has significantly advanced the field with diverse network ar- chitectures, especially the U-shaped encoder-decoder design starting with UNet [1]. Subsequent models, such as UNet++ [2], Attention U-Net [3], UNet3+ [4], Rollin...
-
[2]
LSU-Net: Lightweight Automatic Organs Segmentation Network For Medical Images
METHOD This section provides detailed information about LSU-Net, including its overall network architecture, core components, and loss functions. The core components discussed include the Light Conv Block and the Tokenized Shift Block. The Overall Architecture. We propose LSU-Net, a lightwei- ght network designed for abdominal organ segmentation through e...
work page Pith review arXiv 2025
-
[3]
First, we outline some implementation details
EXPERIMENTS In this section, we validate our proposed network architecture on the UWMGI [13] and MSD Colon [14] datasets. First, we outline some implementation details. Then, we compare LSU-Net with other state-of-the-art architectures on the two datasets and analyze the experimental results. Implementation Details. We develop LSU-Net using Py- torch [17]...
-
[4]
CONCLUSION In this paper, We propose LSU-Net, a lightweight abdominal organ segmentation network that uses feature space shift and multi-scale loss to enhance the recognition of organ regions, while employing lightweight convolutions to reduce model parameters. We conduct extensive experiments on multiple abdominal datasets, demonstrating that the propose...
-
[5]
U-net: Convolutional networks for biomedical im- age segmentation,
Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical im- age segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th in- ternational conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. Springer, 2015, pp. 234– 241
work page 2015
-
[6]
Unet++: A nested u- net architecture for medical image segmentation,
Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang, “Unet++: A nested u- net architecture for medical image segmentation,” 2018
work page 2018
-
[7]
Attention u-net: Learning where to look for the pancreas,
Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al., “Attention u-net: Learning where to look for the pancreas,” arXiv preprint arXiv:1804.03999 , 2018
arXiv 2018
-
[8]
Unet 3+: A full- scale connected unet for medical image segmentation,
Huimin Huang, Lanfen Lin, Ruofeng Tong, Hongjie Hu, Qiaowei Zhang, Yutaro Iwamoto, Xianhua Han, Yen-Wei Chen, and Jian Wu, “Unet 3+: A full- scale connected unet for medical image segmentation,” in ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2020, pp. 1055–1059
work page 2020
Show all 24 references
-
[9]
Rolling-unet: Revitalizing mlp’s ability to efficiently extract long-distance depen- dencies for medical image segmentation,
Yutong Liu, Haijiang Zhu, Mengting Liu, Huaiyuan Yu, Zihan Chen, and Jie Gao, “Rolling-unet: Revitalizing mlp’s ability to efficiently extract long-distance depen- dencies for medical image segmentation,” in Proceed- ings of the AAAI Conference on Artificial Intelligence , 202...
2024
-
[10]
Unext: Mlp-based rapid medical image segmentation network,
Jeya Maria Jose Valanarasu and Vishal M Patel, “Unext: Mlp-based rapid medical image segmentation network,” in International conference on medical image comput- ing and computer-assisted intervention. Springer, 2022, pp. 23–33
2022
-
[11]
Swin- unet: Unet-like pure transformer for medical image seg- mentation,
Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang, “Swin- unet: Unet-like pure transformer for medical image seg- mentation,” in European conference on computer vision. Springer, 2022, pp. 205–218
2022
-
[12]
Transunet: Transformers make strong en- coders for medical image segmentation,
Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou, “Transunet: Transformers make strong en- coders for medical image segmentation,” arXiv preprint arXiv:2102.04306, 2021
2021 arXiv
-
[13]
Medical transformer: Gated axial-attention for medical image segmentation,
Jeya Maria Jose Valanarasu, Poojan Oza, Ilker Haci- haliloglu, and Vishal M Patel, “Medical transformer: Gated axial-attention for medical image segmentation,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Con- ference, Strasbour...
2021
-
[14]
Transbts: Multimodal brain tumor segmentation using transformer,
Wenxuan Wang, Chen Chen, Meng Ding, Jiangyun Li, Hong Yu, and Sen Zha, “Transbts: Multimodal brain tumor segmentation using transformer,” 2021
2021
-
[15]
Axial attention in multidimensional transformers,
Jonathan Ho, Nal Kalchbrenner, Dirk Weissenborn, and Tim Salimans, “Axial attention in multidimensional transformers,” arXiv preprint arXiv:1912.12180, 2019
1912 arXiv
-
[16]
Axial- deeplab: Stand-alone axial-attention for panoptic seg- mentation,
Huiyu Wang, Yukun Zhu, Bradley Green, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen, “Axial- deeplab: Stand-alone axial-attention for panoptic seg- mentation,” in European conference on computer vision. Springer, 2020, pp. 108–126
2020
-
[17]
Uw-madison gi tract im- age segmentation,
happyharrycn, Maggie, Phil Culliton, Poonam Yadav, and Sangjune Laurence Lee, “Uw-madison gi tract im- age segmentation,” 2022
2022
-
[18]
The medical seg- mentation decathlon,
Michela Antonelli, Annika Reinke, Spyridon Bakas, Keyvan Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ron- neberger, Ronald M Summers, et al., “The medical seg- mentation decathlon,” Nature communications, vol. 13, no. 1, pp. 4128, 2022
2022
-
[19]
S 2-mlpv2: improved spatial-shift mlp architecture for vision,
Tan Yu, Xu Li, Yunfeng Cai, Mingming Sun, and Ping Li, “S 2-mlpv2: improved spatial-shift mlp architecture for vision,” arXiv preprint arXiv:2108.01072, 2021
2021 arXiv
-
[20]
Auxiliary tasks in multi-task learning,
Lukas Liebel and Marco K ¨orner, “Auxiliary tasks in multi-task learning,” arXiv preprint arXiv:1805.06334, 2018
2018 arXiv
-
[21]
Pytorch: An imperative style, high-performance deep learning library,
Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al., “Pytorch: An imperative style, high-performance deep learning library,” Advances in neural information processing systems, ...
2019
-
[22]
Road extraction by deep residual u-net,
Zhengxin Zhang, Qingjie Liu, and Yunhong Wang, “Road extraction by deep residual u-net,” IEEE Geo- science and Remote Sensing Letters , vol. 15, no. 5, pp. 749–753, 2018
2018
-
[23]
Adam: A method for stochastic optimization,
Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014
2014 arXiv
-
[24]
Sgdr: Stochastic gradient descent with warm restarts,
Ilya Loshchilov and Frank Hutter, “Sgdr: Stochastic gradient descent with warm restarts,” 2017
2017
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.