REVIEW 3 major objections 5 minor 29 references
CENet: Context Enhancement Network for Medical Image Segmentation
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read The paper's central claim is that CENet, a medical image segmentation network with two new context-enhancement modules, achieves the highest average Dice among compared state-of-the-art models on Synapse, ACDC, PH2, and HAM10000.
desk verdict CENet is a plausible modular segmentation network with a real PH2 ablation gain, but the SOTA claims rest on uncontrolled baseline comparisons and need a matched-protocol re-evaluation before they can be taken at face value. 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 load-bearing machinery is the DSEB-CFAM pair. The DSEB computes a two-scale edge residual $F_{\mathrm{edge}} = |F_{u1} - F_{u2}|$, where $F_{u1}$ and $F_{u2}$ are upsampled versions of the feature map downsampled at scales 0.75 and 0.5, and adds $\lambda F_{\mathrm{edge}}$ to the original features; the result then passes through Differential Attention, which splits queries and keys into two groups and subtracts the two attention maps to cut attention noise. The CFAM first recalibrates channels with a descriptor built from average, max, and standard-deviation pooling, then splits the feature map into channel groups processed by dilated depth-wise convolutions at rates 3, 5, and 8 plus global average pooling, followed by a weighted non-local block and a spatial recalibration MLP. This combination is what the paper credits for boundary detail, small-organ detection, and reduced feature redundancy.
What would settle it
Train CENet and the strongest compared model on Synapse from scratch under identical settings across multiple random seeds, and compare the seed-to-seed spread of average Dice; if the 0.29% gap falls within one standard deviation of that spread, the headline claim of state-of-the-art superiority is not supported by the evidence.
Extended reading notes
Core claim
The central claim is that CENet outperforms the compared state-of-the-art models on multi-organ CT, cardiac MRI, and dermoscopic segmentation by replacing plain skip connections and decoder stages with two context-aware components. The DSEB amplifies edge information through a two-scale downsampling/upsampling difference and then applies differential attention to suppress redundant context; the CFAM recalibrates channel statistics, aggregates multi-scale features with dilated depth-wise convolutions, denoises through a weighted non-local block, and spatially recalibrates. The reported results put CENet first in average Dice on Synapse (85.04), ACDC (92.18), PH2 (95.04), and HAM10000 (94.71), and the PH2 ablation attributes the gain to the components: Dice rises from 94.08 with neither module to 95.04 with both active.
Load-bearing premise
The load-bearing premise is that baseline numbers copied from earlier papers are directly comparable to CENet's own training runs, so that a 0.29% average-Dice lead on Synapse is a real effect rather than random seed noise.
Editorial extensions
If this is right
- On Synapse, the reported average Dice of 85.04 and HD95 of 8.84 imply that CENet's boundary error is far smaller than the next-best HD95 of 13.29, so the main measured gain is contour accuracy rather than bulk overlap.
- On ACDC, average Dice of 92.18 with right-ventricle and myocardium scores of 90.90 and 89.63 would make CENet the strongest of the compared models on every cardiac structure, not just on average.
- On PH2 and HAM10000, Dice scores of 95.04 and 94.71 would mean the boundary-enhancement design transfers from radiology to dermoscopy without retuning.
- The ablation table shows each added component (FEA, DiffAtt, wNLB, CCU) raises PH2 Dice in the reported runs, so the benefits are attributed to the modules sequentially rather than to any single one.
Reading between the lines
- Beyond the reported benchmarks, the DSEB's two-scale edge residual is effectively a learned edge-detection injection; the same recipe could be tested in other U-shaped architectures by inserting only that residual into their skip connections.
- The unusually large HD95 improvement relative to Dice on Synapse suggests that, if the numbers hold, contour-oriented clinical tasks such as radiotherapy planning and surgical margin assessment would benefit more than overlap-based evaluations show.
- An untested consequence of the ablation is that the wNLB and CCU contribute most of the measured gain only after FEA and DiffAtt are present, so a follow-up could determine whether the modules are complementary or whether one subsumes the other.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CENet, a U-shaped medical image segmentation network built on a PVTv2-b2 encoder with two new components: the Dual Selective Enhancement Block (DSEB), which replaces plain skip connections and combines a Feature Edge Amplifier with differential attention, and the Context Feature Attention Module (CFAM), a Transformer-style decoder block with channel calibration, multi-scale dilated depth-wise convolutions, a weighted non-local block, and a spatial recalibration MLP. The method is evaluated on Synapse (CT multi-organ), ACDC (cardiac MRI), PH2, and HAM10000 (dermoscopic) datasets, reporting the highest average Dice scores among the compared methods in each case. The authors claim that CENet outperforms state-of-the-art methods in multi-organ segmentation and boundary preservation, and they provide qualitative visualizations, an ablation study on PH2, and a public code repository.
Significance. If the reported results hold, CENet would be a practically interesting addition to the medical segmentation literature: it combines a relatively light PVTv2-b2 encoder (33.39M parameters) with a decoder that attempts to address both boundary detail and multi-scale context, and the authors release the code, which supports reproducibility. The design ideas, particularly the differential-attention skip connection and the multi-scale decoder with explicit noise suppression, are plausible and follow recent trends. However, the significance of the empirical claim is currently limited by the evaluation protocol: baseline numbers are taken from earlier publications without matched training conditions, no statistical significance or variance information is reported, and the only ablation that isolates the proposed components is performed on a single small dataset. The paper is therefore a reasonable technical contribution whose central 'SOTA' claim requires substantially stronger evidence.
major comments (3)
- [Section 3.2, Tables 1 and 2] The central claim that CENet 'significantly outperforms' state-of-the-art methods is not supported by the reported evidence. All baseline scores are quoted from prior papers that do not use the same training protocol as CENet; in particular, CENet is trained with BDoU loss [24], while the cited baselines (e.g., TransUNet, Swin-UNet, MSA2Net) are not reported to use this loss. A boundary-aware loss can independently improve boundary metrics, so the observed gains cannot be attributed solely to DSEB and CFAM. Furthermore, no error bars, confidence intervals, or significance tests are given. On Synapse the margin over MSA2Net is only 0.29 average DSC (85.04 vs. 84.75), and CENet is worse on several individual organs (Gallbladder 68.29 vs. 74.35, Stomach 81.68 vs. 84.03); this margin is comparable to typical seed-to-seed variation in medical segmentation. To justify the abstract's claim, the authors should either run all baselines under identical training conditions (including loss and input resolution) or provide paired statistical tests, ideally with multiple runs.
- [Table 3 and Section 3.2] The ablation study used to attribute gains to DSEB and CFAM is conducted only on PH2. The abstract and introduction claim improvements in multi-organ segmentation and boundary preservation across radiology and dermoscopic datasets, but no ablation is reported on Synapse, ACDC, or HAM10000. Without ablations on those datasets, the paper does not directly show that the two proposed modules are responsible for the average gains reported in Tables 1 and 2. Please add component ablations (at least with and without DSEB and CFAM) on the other three datasets, or explicitly limit the claim to the PH2 setting.
- [Eq. (4) and Section 2.2] There is a direct contradiction in the definition of the MCA. The text states that the channel splits are processed with dilation rates 'e.g., 3, 5, and 8', while Eq. (4) defines d_k ∈ {6, 8, 12}. The actual dilation rates used in the experiments must be stated unambiguously. More generally, the architectural description is too underspecified for reproduction: Eq. (1) uses scales s1=0.75, s2=0.5, s0=1.0 without defining the downsampling and upsampling operators D and U (e.g., whether s is a scale factor or a target resolution), and the channel split described as 'less than 10 percent of all the channels' is informal. Please clarify these definitions, as they are load-bearing for implementing the proposed modules.
minor comments (5)
- [Eq. (3)] The input to the CCU is denoted X in the text but F in Eq. (3); please make the notation consistent throughout Section 2.2.
- [Tables 1 and 3] Table 1 reports CENet FLOPs as 12.76 G, while Table 3 gives 12.84 G for the full configuration with the same parameters; please clarify whether the difference is due to input resolution, counting method, or a typo.
- [Figure 4] The text adjacent to Figure 4 contains garbled '/uni...' strings that appear to be a rendering artifact; these should be removed or replaced with the intended caption material.
- [Section 3.2 and Figure 3] The 'boundary detail preservation' claim is supported mainly by qualitative visualizations; consider reporting quantitative boundary-aware metrics (for example, HD95 on Synapse is already given, but no boundary metric is reported for ACDC or the dermoscopic datasets) to substantiate the claim.
- [References] Several baseline numbers come from papers by the same research group (e.g., [2], [13], [19]); please state explicitly whether those numbers were produced under the same evaluation protocol as the other baselines, since comparability is essential to the claims.
Circularity Check
No circular derivation: CENet's claims are empirical and the reported gain is not constructed from its own definitions or fitted parameters.
full rationale
The paper's central claim is empirical: CENet's DSEB and CFAM modules are evaluated on public datasets and compared with published baselines. Nothing in the method section defines a target metric in terms of a fitted parameter, and no reported Dice score is obtained by rearranging baseline numbers. The DSEB's FEA and DiffAtt, and the CFAM's CCU, MCA, wNLB, and SRM, are architectural components whose outputs are measured after training; the equations describe the forward computation, not a derivation of the benchmark numbers. The comparison tables quote prior results, including several from the same research group (HiFormer [9], MSA2Net [13], D-LKA [2]), but these are independently published baselines and CENet's own scores are produced by a separate training run; quoting them does not make the compared quantity equal to the input by construction. The ablation on PH2 reports a small gain from adding DSEB and CFAM, but this is a controlled experiment on one dataset, not a fitted parameter renamed as a prediction. The main weakness is statistical comparability: baseline numbers come from different training protocols and no significance tests are reported, so the 0.29% Synapse margin over MSA2Net may be within run-to-run variance. That is a correctness and fairness concern, not circularity. No specific reduction of a predicted quantity to its own input is exhibited, so the circularity score is minimal.
Assumptions & free parameters
free parameters (4)
- DSEB scales (s1, s2, s0) =
0.75, 0.5, 1.0
- MCA dilation rates =
3, 5, 8 in text; 6, 8, 12 in Eq. (4)
- CFAM channel split allocation =
C1=C2=C3, C4 less than 10 percent of channels, exact value unspecified
- FEA weighting lambda =
learned vector in R^d
assumptions (6)
- domain assumption ImageNet-pretrained PvTv2 features transfer to medical images.
- domain assumption The BDoU loss is beneficial for all four datasets.
- domain assumption Baseline numbers from cited papers are reliable and comparable to the paper's own runs.
- domain assumption Standard optimization converges to representative performance.
- ad hoc to paper The difference of two resampled feature maps isolates useful edge details.
- ad hoc to paper Differential attention noise subtraction transfers from NLP to visual skip connections.
Cite this review
Pith. "Pith review of CENet: Context Enhancement Network for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/ZQO5FMB4
@misc{pith2026250518423,
author = {Pith},
title = {Pith review of: CENet: Context Enhancement Network for Medical Image Segmentation},
year = {2026},
howpublished = {\url{https://pith.science/paper/ZQO5FMB4}},
note = {Machine review of arXiv:2505.18423}
}
read the original abstract
Medical image segmentation, particularly in multi-domain scenarios, requires precise preservation of anatomical structures across diverse representations. While deep learning has advanced this field, existing models often struggle with accurate boundary representation, variability in organ morphology, and information loss during downsampling, limiting their accuracy and robustness. To address these challenges, we propose the Context Enhancement Network (CENet), a novel segmentation framework featuring two key innovations. First, the Dual Selective Enhancement Block (DSEB) integrated into skip connections enhances boundary details and improves the detection of smaller organs in a context-aware manner. Second, the Context Feature Attention Module (CFAM) in the decoder employs a multi-scale design to maintain spatial integrity, reduce feature redundancy, and mitigate overly enhanced representations. Extensive evaluations on both radiology and dermoscopic datasets demonstrate that CENet outperforms state-of-the-art (SOTA) methods in multi-organ segmentation and boundary detail preservation, offering a robust and accurate solution for complex medical image analysis tasks. The code is publicly available at https://github.com/xmindflow/cenet.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[24]
Sun, F., Luo, Z., Li, S.: Boundary difference over union loss for medical image segmentation (2023)
work page 2023
-
[1]
IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)
Azad, R., Aghdam, E.K., Rauland, A., Jia, Y., Avval, A.H., Bozorgpour, A., Karimijafarbigloo, S., Cohen, J.P., Adeli, E., Merhof, D.: Medical image segmen- tation review: The success of u-net. IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)
work page 2024
-
[2]
Azad, R., Niggemeier, L., Huttemann, M., Kazerouni, A., Aghdam, E.K., Velichko, Y., Bagci, U., Merhof, D.: Beyond self-attention: Deformable large kernel attention for medical image segmentation (2023)
work page 2023
-
[3]
Bernard, O., Lalande, A., Zotti, C., Cervenansky, F., Yang, X., Heng, P.A., Cetin, I., Lekadir, K., Camara, O., Ballester, M.A.G., et al.: Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE transactions on medical imaging37(11), 2514–2525 (2018)
2018
-
[4]
In: Proceedings of the European Conference on Computer Vision Workshops(ECCVW) (2022)
Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin-unet: Unet-like pure transformer for medical image segmentation. In: Proceedings of the European Conference on Computer Vision Workshops(ECCVW) (2022)
work page 2022
-
[5]
arXiv preprint arXiv:2102.04306 (2021)
Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021)
arXiv 2021
-
[6]
In: Proceedings of the European conference on computer vision (ECCV)
Chen, L.C., Zhu, Y., Papandreou, G., Schroff, F., Adam, H.: Encoder-decoder with atrous separable convolution for semantic image segmentation. In: Proceedings of the European conference on computer vision (ECCV). pp. 801–818 (2018)
2018
-
[7]
arXiv preprint arXiv:2010.11929 (2020)
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020)
arXiv 2020
Show all 29 references
-
[8]
Computational visual media9(4), 733–752 (2023)
Guo, M.H., Lu, C.Z., Liu, Z.N., Cheng, M.M., Hu, S.M.: Visual attention network. Computational visual media9(4), 733–752 (2023)
2023
-
[9]
In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision
Heidari, M., Kazerouni, A., Soltany, M., Azad, R., Aghdam, E.K., Cohen-Adad, J., Merhof, D.: Hiformer: Hierarchical multi-scale representations using transform- ers for medical image segmentation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vi...
2023
-
[10]
IEEE Transactions on Medical Imaging pp
Huang, X., Deng, Z., Li, D., Yuan, X., Fu, Y.: Missformer: An effective transformer for 2d medical image segmentation. IEEE Transactions on Medical Imaging pp. 1–1 (2022).https://doi.org/10.1109/TMI.2022.3230943
2022
-
[11]
IEEE Transactions on Medical Imaging (2022)
Huang, X., Deng, Z., Li, D., Yuan, X., Fu, Y.: Missformer: An effective transformer for 2d medical image segmentation. IEEE Transactions on Medical Imaging (2022)
2022
-
[12]
Drones7(5), 287 (2023) 10
Jamil, S., Jalil Piran, M., Kwon, O.J.: A comprehensive survey of transformers for computer vision. Drones7(5), 287 (2023) 10
2023
-
[13]
arXiv preprint arXiv:2407.21640 (2024)
Kolahi, S.G., Chaharsooghi, S.K., Khatibi, T., Bozorgpour, A., Azad, R., Heidari, M., Hacihaliloglu, I., Merhof, D.: Msa 2net: Multi-scale adaptive attention-guided network for medical image segmentation. arXiv preprint arXiv:2407.21640 (2024)
2024 arXiv
-
[14]
IEEE Journal of Biomedical and Health Informatics27(10), 4828–4839 (2023)
Kuang, H., Wang, Y., Liang, Y., Liu, J., Wang, J.: Bea-net: Body and edge aware network with multi-scale short-term concatenation for medical image segmentation. IEEE Journal of Biomedical and Health Informatics27(10), 4828–4839 (2023). https://doi.org/10.1109/JBHI.2023.3304662
2023
-
[15]
In: Proc
Landman, B., Xu, Z., Igelsias, J., Styner, M., Langerak, T., Klein, A.: Miccai multi- atlas labeling beyond the cranial vault–workshop and challenge. In: Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge. vol. 5, p. 12 (2015)
2015
-
[16]
In: Proceedings of the IEEE/CVF international conference on computer vision
Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin transformer: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 10012–10022 (2021)
2021
-
[17]
arXiv preprint arXiv:2305.00837 (2023)
Ma, Q., Mao, K., Wang, G., Xu, L., Zhao, Y.: Lcaunet: A skin lesion segmentation network with enhanced edge and body fusion. arXiv preprint arXiv:2305.00837 (2023)
2023 arXiv
-
[18]
arXiv preprint arXiv:1804.03999 (2018)
Oktay, O., Schlemper, J., Folgoc, L.L., Lee, M., Heinrich, M., Misawa, K., Mori, K., McDonagh, S., Hammerla, N.Y., Kainz, B., et al.: Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999 (2018)
2018 arXiv
-
[19]
In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
Rahman, M.M., Munir, M., Marculescu, R.: Emcad: Efficient multi-scale convolu- tional attention decoding for medical image segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11769– 11779 (2024)
2024
-
[20]
In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, Oc- tober 5-9, 2015, Proceedings, Part III 18
Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, Oc- tober 5-9, 2015, Proceedings, Part III 18....
2015
-
[21]
arXiv preprint arXiv:2402.02491 (2024)
Ruan, J., Xiang, S.: Vm-unet: Vision mamba unet for medical image segmentation. arXiv preprint arXiv:2402.02491 (2024)
2024 arXiv
-
[22]
Medical Image Analysis88, 102802 (2023)
Shamshad, F., Khan, S., Zamir, S.W., Khan, M.H., Hayat, M., Khan, F.S., Fu, H.: Transformers in medical imaging: A survey. Medical Image Analysis88, 102802 (2023)
2023
-
[23]
Sohn, K., Hao, Y., Lezama, J., Polania, L., Chang, H., Zhang, H., Essa, I., Jiang, L.: Visual prompt tuning for generative transfer learning (2022)
2022
-
[25]
In: Proceedings of the AAAI conference on artificial intelligence
Wang, H., Cao, P., Wang, J., Zaiane, O.R.: Uctransnet: rethinking the skip connec- tions in u-net from a channel-wise perspective with transformer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 2441–2449 (2022)
2022
-
[26]
In: ICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP)
Wang, H., Xie, S., Lin, L., Iwamoto, Y., Han, X.H., Chen, Y.W., Tong, R.: Mixed transformer u-net for medical image segmentation. In: ICASSP 2022-2022 IEEE international conference on acoustics, speech and signal processing (ICASSP). pp. 2390–2394. IEEE (2022)
2022
-
[27]
Computational Visual Media8(3), 415–424 (2022)
Wang, W., Xie, E., Li, X., Fan, D.P., Song, K., Liang, D., Lu, T., Luo, P., Shao, L.: Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media8(3), 415–424 (2022)
2022
-
[28]
In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition
Wang, X., Girshick, R., Gupta, A., He, K.: Non-local neural networks. In: Pro- ceedings of the IEEE conference on computer vision and pattern recognition. pp. 7794–7803 (2018) CENet: Context Enhancement Network 11
2018
-
[29]
arXiv preprint arXiv:2410.05258 (2024)
Ye, T., Dong, L., Xia, Y., Sun, Y., Zhu, Y., Huang, G., Wei, F.: Differential trans- former. arXiv preprint arXiv:2410.05258 (2024)
2024 arXiv
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.