Dual Interaction Network with Cross-Image Attention for Medical Image Segmentation

Jeonghyun Noh; Jinsun Park; Wangsu Jeon

REVIEW 4 major objections 6 minor 65 references

Dual Interaction Network with Cross-Image Attention for Medical Image Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Dual cross-attention between original and enhanced medical images, plus a Sobel boundary loss, reaches 93.25 Dice on cardiac MRI and 85.49 on abdominal CT.

desk verdict Useful incremental engineering, but the ACDC SOTA claim is not established because baselines mix protocols and the final model was selected on the same test set. read the letter →

arxiv 2509.05953 v1 pith:KHIZDCYX submitted 2025-09-07 cs.CV

Jeonghyun Noh , Wangsu Jeon , Jinsun Park This is my paper

classification cs.CV

keywords medicalimagesegmentationdualinteractivefusioncross-attentionenhancementboundarylossSobelgradientACDCSynapse

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that image enhancement helps medical segmentation only when the enhanced image is allowed to correct and be corrected by the original, rather than replacing it or being naively merged with it. It proposes a dual interactive fusion module (DIFM) that runs bidirectional cross-attention between feature maps of original and fuzzy-enhanced versions of the same image and then refines the result with global spatial attention, repeated at four encoder depths. It also adds a multi-scale boundary loss that matches Sobel gradients of prediction and ground truth at three resolutions. If correct, the approach yields the best reported average Dice on the ACDC cardiac MRI benchmark (93.25) and strong results on the Synapse abdominal CT benchmark (85.49 Dice, 10.74 HD95), beating 2D baselines while staying behind the 3D nnFormer on Synapse. The underlying value is a general recipe: use enhancement for structural cues while preserving original-image diagnostics through explicit mutual attention.

What carries the argument

DIFM (Dual Interactive Fusion Module): a skip-connection block that merges low- and high-level features from each of two image-specific encoders; runs bidirectional cross-attention (each branch's query attends to the other branch's key/value, plus residual) followed by a GELU-gated feed-forward network; then applies global spatial attention via channel-mean sigmoid gating; the two refined branches are added. The second mechanism is the multi-scale boundary loss L_bnd, a Sobel-filter gradient extractor applied to prediction and one-hot ground truth at scales 1, 1/2, and 1/4 with L1 difference. Together they carry the argument that enhancement information can be exploited without discarding or

What would settle it

Re-run the Table 2 baselines under the paper's exact split, 224x224 resize, augmentation, batch size, and Eq. (10) loss; if SwinUNet reproduces 91.32 instead of the listed 90.00 and the margin over FCT384 shrinks below 0.23 Dice, the claimed SOTA is a protocol artifact.

Watch

Extended reading notes

Core claim

The central claim is that the right way to combine an original medical image with an enhanced copy is neither feature concatenation nor one-directional cross-attention, but a symmetric interaction in which both branches query each other. DIFM first concatenates low-level edge-type features with high-level semantic features inside each branch, then applies cross-attention in both directions with residual connections, feeds the results through a gated FFN, and gates the merged features with sigmoid-activated global spatial attention. Because the two directions are added, the final fused feature retains information selected by both images. The multi-scale boundary loss compares Sobel-gradient m

Load-bearing premise

The load-bearing premise is that the comparison numbers for prior methods in Tables 1 and 2 were obtained under the same evaluation protocol as the proposed method; the paper does not state that it re-ran them, and the same baseline appears with different scores in different tables.

Editorial extensions

If this is right

Using original and enhanced images together through DIFM beats both single-image baselines: 93.25 vs 92.88 (original only) and 92.66 (enhanced only) on ACDC.
DIFM is transferable: plugging it into SwinUNet and FCT raises their Dice by 0.61 and 0.47 percentage points, suggesting the module, not the specific backbone, drives part of the gain.
The Sobel-gradient multi-scale boundary loss improves every architecture tested (SwinUNet +0.28, MERIT +0.12, FCT +0.04, Ours +0.34), so it can be used as a drop-in loss term.
The method generalizes beyond CT and MRI: it reports 97.00 Dice on CVC-ClinicDB and 93.85 on ISIC18, ahead of the compared methods.
The number of fusion modules matters: four DIFMs give 93.25 versus 92.78 with three, so cross-image interaction at every scale contributes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The I+I experiment (two copies of the original image, 92.92 Dice) shows that part of the gain over single-image input comes from the dual-encoder attention machinery itself; only about 0.33 Dice points (93.25 vs 92.92) can be attributed specifically to enhancement-provided content, so a controlled study with matched parameter counts would separate these effects.
The bidirectional cross-attention design is not specific to fuzzy enhancement; it should transfer to any paired inputs with complementary structure, such as T1/T2 MRI, CT/PET, or multi-contrast images, and could be tested by swapping the enhancement branch for a second modality.
Since the boundary loss improved several unrelated baselines uniformly, it is a plausible plug-in for existing 3D segmentation pipelines; the volumetric extension the authors discuss could be tested directly without the full DIFM.
At 323M parameters and 1860 GFLOPs the pipeline is expensive; a lighter dual-encoder variant (shared encoder or distilled enhancement branch) would test whether the accuracy gain survives at clinically deployable compute.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the architecture is plausible and the ablations are coherent, but the SOTA claim on ACDC doesn't survive contact with the paper's own re-runs. Table 2 lists FCT384 at 93.02, but the authors' own FCT under their protocol in Table 5 is 91.36; SwinUNet is 90.00 in Table 2 and 91.32 in Table 5. So the comparison mixes literature-reported baselines with their own runs, and the paper never discloses which is which. Given a 0.23pp margin over FCT384, that protocol mismatch alone could explain the result. More importantly, the final configuration was selected on the same ACDC test set used for the headline: enhancement method (Table 4), DIFM count (Table 6, 1→4 gives 86.86→93.25), cross-attention direction (Table 7), and loss components (Tables 3 and 8) are all chosen by test Dice. There are no repeated seeds or error bars anywhere. So the superiority claim is conditional at best.

What the paper does well: DIFM is a modular, clearly described fusion unit that improves SwinUNet by 0.61pp and FCT by 0.47pp in Table 5. Bidirectional cross-attention beats either unidirectional variant (Table 7), and the multi-scale Sobel boundary loss gives small but consistent gains across SwinUNet, MERIT, FCT, and the proposed model (Table 8). The writing is straightforward, the ablations are structured, and code is provided. The Synapse result (85.49, second to nnFormer) is honestly reported, and the supplementary experiments on ClinicDB and ISIC18 suggest the module generalizes.

Soft spots beyond the SOTA claim: the model is very heavy (323M parameters, 1860 GFLOPs) and the gains are modest relative to that cost. The 'global spatial attention' is a channel-mean sigmoid gate, and the boundary loss is a Sobel variant of Kervadec et al.; the novelty is in the combination, not the pieces. Still, the paper acknowledges the computational limitation and positions the work as high-accuracy rather than efficient.

Bottom line: this is a competent incremental engineering contribution. If the authors re-run all baselines under one protocol, report multiple seeds with error bars, and move model selection off the test set, the ACDC result would be credible. As it stands, I'd treat the method as promising but unproven at the claimed SOTA level. It deserves a serious referee, but the referee should ask for those experiments.

Referee Report

4 major / 6 minor

Summary. The paper proposes a dual-branch segmentation network for medical images. One branch processes the original input and the other a fuzzy-enhanced version; a Dual Interactive Fusion Module (DIFM) merges the two feature streams using feature merging, bidirectional cross-attention, and global spatial attention at multiple resolutions. Training adds a multi-scale Sobel-gradient boundary loss to CE and Dice losses. Experiments are reported on ACDC (93.25 Dice, claimed state of the art), Synapse (85.49 Dice, called competitive), and in the supplement on CVC-ClinicDB and ISIC18. Ablations cover enhancement choice, fusion strategy, number of DIFMs, cross-attention direction, loss terms, and decoder structure.

Significance. If the empirical claims held, the paper would offer a modular and reusable fusion mechanism for combining original and enhanced images, plus a simple boundary loss that could transfer to other segmenters. Strengths of the paper are its public code, the modular insertion of DIFM into SwinUNet and FCT, and the breadth of ablations on four public datasets. The significance is moderate because the central SOTA claim is currently benchmark-dependent and the evidence as reported is not rigorously controlled: baseline numbers are inconsistently sourced, the final model is selected on the test set, and no variance estimates are given for margins as small as 0.23 pp.

major comments (4)

[Tables 2, 5, and 8] The comparison protocol for the headline SOTA claim is not defined. The same methods appear with different ACDC Dice values across tables: SwinUNet is 90.00 in Table 2 but 91.32 in the authors' re-runs in Table 5; FCT is 92.84 (FCT224) in Table 2 but 91.36 in Table 5; MERIT is 92.32 in Table 2 but 91.30 in Table 8. Table 5 states it was run under Sec. 4.2, but Table 2 gives no source or protocol statement. The claimed 0.23 pp advantage over FCT384 is therefore not established under a single training protocol. If Table 2 is literature-reported, the comparison is uncontrolled; if Table 5/8 are the controlled re-runs, then Table 2 is internally inconsistent and the correct margin over FCT is 1.89 pp against 91.36 rather than 0.23 pp against 93.02.
[Sec. 4.4 and Supplementary Sec. 7.1, 7.2, 7.5] The final configuration is selected on the ACDC test set. Enhancement method (Table 4), number of DIFMs (Table 6: 1 to 4 modules gives 86.86 to 93.25), cross-attention direction (Table 7), loss terms (Tables 3/8), and decoder (Table 11) are all evaluated by test Dice. The paper reports 10 validation patients in Sec. 4.1 but no ablation uses this validation set for model selection. With the closest baseline margin at 0.23 pp and no repeated-seed variance, test-set selection alone could plausibly explain the reported ordering. The authors should either perform ablations on a validation split and report the final test result only for the chosen configuration, or explicitly justify and statistically support test-set-based selection.
[Sec. 3.1.3 and Eq. (2)] The architecture as written is not reproducible. Section 3.1.3 states that the dual cross-attention is window-based ('computing attention independently within each window'), but Eq. (2) defines a global softmax attention over Q K^T / sqrt(d), with no window size, window partition, or shift strategy anywhere in the paper or figure. If windowing is in fact used, Eq. (2) is incomplete and the window configuration is a free parameter that must be reported; if windowing is not used, the statement in Sec. 3.1.3 is contradicted by the formulation. Either way, the core module cannot be implemented from the text as it stands.
[Sec. 4.3.1 and Tables 1-2] All comparison claims are based on single training runs. The reported advantages are small: 0.23 pp over FCT384 on ACDC and 0.59 pp over MERIT on Synapse. ACDC has 20 test patients; Synapse has 12 test scans. In this regime, run-to-run seed variability and patient subsampling can easily exceed these margins. Report at least three seeds with mean and standard deviation, or a paired statistical test, for the main comparisons and for the key ablations. Without this, the quantitative claim that the method 'consistently outperforms' baselines is not supported by the data presented.

minor comments (6)

[Table 3] The checkmark format is very hard to read. It is not clear which rows correspond to baseline concatenation, feature merging, cross-attention, global spatial attention, and which rows include L_CE and L_Dice. Please rewrite the table with explicit 'on/off' entries or named configurations.
[Sec. 3.1.3 and Eq. (6)] The text says the model learns channel-wise weights, but Eq. (6) is a fixed, parameter-free gate: sigmoid of the channel mean. Please revise the wording to describe what the operation actually does.
[Sec. 4.5 vs Supplementary Sec. 11] The MLP decoder parameter count is given as 25.27M in the main text and 25.67M in the supplement. Please reconcile.
[Tables 1 and 2] There are typographical inconsistencies in method names: 'TrasnsCASCADE' in Tables 1-2 versus 'TransCASCADE' elsewhere, and 'TrasnsCASCADE' is not the name used in the cited reference. Please correct.
[Supplementary Sec. 6] The heading contains 'limitation,s' instead of 'limitations'. Minor proofreading issue.
[Abstract/Conclusion] The phrase 'SOTA performance on the ACDC dataset' is repeated in the abstract and conclusion. Given the protocol concerns above, it would be safer to state that the method achieves 'competitive results' until the baseline protocol is clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmark results, not derivations that reduce to their own inputs.

full rationale

The paper is an empirical segmentation method paper, not a derivation chain: DIFM is defined by architectural equations (1)-(6), the loss by Eqs. (7)-(10), and the enhancement by the externally cited fuzzy method [36] with explicit formulas in the supplementary (Eqs. (11)-(12)). No fitted parameter is renamed as a prediction, and no stated result is equivalent by construction to an input. The author self-citations ([3]-[5], [37]) are contextual or supporting, and the use of fuzzy enhancement is additionally validated by the paper's own ablations in Table 4, so the self-citation is not load-bearing. The selection of the number of DIFM modules, enhancement type, and loss components on the ACDC test set is a methodological validity concern (possible selection bias) rather than circularity under the definitions used here: it is not an equation-level reduction and does not make the central benchmark claim true by construction. The paper is self-contained against external benchmarks, so no circular step is identified.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on standard deep-learning machinery plus domain assumptions about enhancement complementarity and benchmark comparability. The paper introduces no mathematical derivations, but it contains several hand-chosen hyperparameters, the most consequential being the enhancement method, the number of DIFMs, the loss weights, and the boundary-loss scales. None are accompanied by sensitivity analysis.

free parameters (5)

Loss weights alpha and beta = 0.3, 0.7 (from MERIT [41])
Weights in Eq. (10) are set without sensitivity analysis and directly modulate the total loss.
Number of DIFM modules = 4
Supplementary Table 6 shows accuracy increasing with the count; 4 was chosen because it produced the best test Dice.
Enhancement method = Fuzzy enhancement [36]
Selected among HE, CLAHE, and Zero-DCE based on ACDC Dice in Table 4; no held-out validation protocol is described.
Boundary loss downsampling scales = i in {1, 2, 4}
The scales in Eq. (9) are chosen without ablation of alternative scale sets.
Input resolution = 224x224
Fixed for all experiments; some prior methods such as FCT384 use a different resolution, complicating direct comparison.

assumptions (4)

domain assumption Fuzzy-enhanced images retain diagnosis-relevant structure and provide complementary features for segmentation.
The dual-input design depends on this; Table 4 gives empirical support, but the paper provides no guarantee across modalities or enhancement types.
domain assumption ACDC and Synapse Dice and HD95 values are comparable across the reported baseline methods.
Tables 1 and 2 mix literature-reported numbers with the authors' pipeline; baseline protocol compatibility is not documented.
ad hoc to paper The window-based cross-attention implementation uses an unstated window configuration that preserves enough global context.
Section 3.1.3 introduces window-based attention without defining window size or partitioning, even though Eq. (2) is written as global attention.
standard math Softmax attention, LayerNorm, GELU, and Sobel filtering behave as standard differentiable components.
These are standard tools imported from prior literature and used without proof in Eqs. (2), (5), and (15).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dual Interaction Network with Cross-Image Attention for Medical Image Segmentation." pith.science (2026). https://pith.science/paper/KHIZDCYX

@misc{pith2026250905953,
  author       = {Pith},
  title        = {Pith review of: Dual Interaction Network with Cross-Image Attention for Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KHIZDCYX}},
  note         = {Machine review of arXiv:2509.05953}
}

read the original abstract

Medical image segmentation is a crucial method for assisting professionals in diagnosing various diseases through medical imaging. However, various factors such as noise, blurriness, and low contrast often hinder the accurate diagnosis of diseases. While numerous image enhancement techniques can mitigate these issues, they may also alter crucial information needed for accurate diagnosis in the original image. Conventional image fusion strategies, such as feature concatenation can address this challenge. However, they struggle to fully leverage the advantages of both original and enhanced images while suppressing the side effects of the enhancements. To overcome the problem, we propose a dual interactive fusion module (DIFM) that effectively exploits mutual complementary information from the original and enhanced images. DIFM employs cross-attention bidirectionally to simultaneously attend to corresponding spatial information across different images, subsequently refining the complementary features via global spatial attention. This interaction leverages low- to high-level features implicitly associated with diverse structural attributes like edges, blobs, and object shapes, resulting in enhanced features that embody important spatial characteristics. In addition, we introduce a multi-scale boundary loss based on gradient extraction to improve segmentation accuracy at object boundaries. Experimental results on the ACDC and Synapse datasets demonstrate the superiority of the proposed method quantitatively and qualitatively. Code available at: https://github.com/JJeong-Gari/DIN

Figures

Figures reproduced from arXiv: 2509.05953 by the authors.

**Figure 1.** Medical image segmentation results with various input configurations: (a) Original image only, (b) Enhanced image only, (c) Proposed, and (d) Ground truth. The proposed method shows more accurate prediction compared to singleimage outputs (a) and (b) because it benefits from both the original and enhanced images effectively. enhancement. As a result, the segmentation model suffers from performance degradation ( [P… view at source ↗

**Figure 2.** The overall architecture of the proposed network. Fuzzy image enhancement [36] is used to improve the quality of the original image. The ConvNext-base [22] encoder extracts image features, while the DIFM exploits the interaction between original and enhanced feature maps utilizing an attention mechanism to fuse them. The decoder is an MLP decoder [23], which generates a segmentation mask. around object boundaries. D… view at source ↗

**Figure 3.** Visual comparisons of segmentation results on the (a) Synapse and (b) ACDC datasets. Yellow boxes highlight regions in which our method excels at segmentation. approach demonstrated superior performance for the Myo and LV, with Dice scores of 91.08% and 96.50%. By effectively extracting and integrating features from original and enhanced images using DIFM, our method achieves significant performance improvements co… view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Visual comparisons of enhancement results (a) Original image and (b) Fuzzy image enhancement [36] image. Eq. (12), emphasizing the organ’s attributes, including shape and intensity. Consequently, previously subtle features in Fig. 4a, become more pronounced in Fig. 4b.…

**Figure 5.** Figure 5: Decoder [23] architecture. in size and channel dimensions, to uniform configuration. Specifically, it transforms all feature maps to a channel dimension of 768 and a feature map size of 224×224, using bilinear interpolation for resizing. The concatenation-based fusion …

**Figure 6.** Figure 6: Additional visual comparison of segmentation results on the ACDC dataset. (a) TransUNet [10], (b) SwinUNet [11], (c) MERIT [41], (d) FCT [12], (e) only 𝐼, (f) only 𝐸, (g) Ours, and (h) GT, respectively. Yellow boxes highlight regions in which our method excels at segme…

**Figure 7.** Figure 7: Additional visual comparison of segmentation results on the Synapse dataset. (a) TransUNet [10], (b) SwinUNet [11], (c) MERIT [41], (d) FCT [12], (e) only 𝐼, (f) only 𝐸, (g) Ours, and (h) GT, respectively. Yellow boxes highlight regions in which our method excels at se…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 57 canonical work pages

[1]

S. Jang, H. Song, Y. J. Shin, J. Kim, J. Kim, K. W. Lee, S. S. Lee, W. Lee, S. Lee, K. H. Lee, Deep learning–based automatic detection algorithmforreducingoverlookedlungcancersonchestradiographs, Radiology 296 (3) (2020) 652–661. J. Noh et al.:Preprint submitted to ElsevierPage 7 of 15 Dual Interaction Network

work page 2020
[2]

S.Lee,J.Park,J.Park,Crossformer:Cross-guidedattentionformulti- modal object detection, Pattern Recognition Letters 179 (2024) 144– 150

work page 2024
[3]

J. Kim, J. Noh, M. Jeong, W. Lee, Y. Park, J. Park, Adnet: Non-local affinity distillation network for lightweight depth completion with guidancefrommissinglidarpoints,IEEERobot.andAutomat.Letters (2024)

work page 2024
[4]

S. Ha, Y. Kim, J. Park, Interdimensional knowledge transfer for se- manticsegmentationonlidarpointclouds,IEEERobot.andAutomat. Letters (2024)

work page 2024
[5]

H. Son, J. Noh, S. Jeon, C. Wang, W.-K. Jeong, Mc-inr: Effi- cient encoding of multivariate scientific simulation data using meta- learning and clustered implicit neural representations, arXiv preprint arXiv:2507.02494 (2025)

work page arXiv 2025
[6]

Ronneberger, P

O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks forbiomedicalimagesegmentation,in:Int.Conf.Med.Img.Comput. and Comput. Assist. Interv., 2015, pp. 234–241

work page 2015
[7]

Z.Zhou,M.M.RahmanSiddiquee,N.Tajbakhsh,J.Liang,Unet++: A nested u-net architecture for medical image segmentation, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv., 2018, pp. 3– 11

work page 2018
[8]

Huang, L

H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y. Iwamoto, X. Han, Y.-W. Chen, J. Wu, Unet 3+: A full-scale connected unet for medical image segmentation, in: Int. Conf. Acoust., Speech and Signal Pro- cess., 2020, pp. 1055–1059

work page 2020

Show all 65 references

[9]

methods 18 (2) (2021) 203–211

F.Isensee,P.F.Jaeger,S.A.Kohl,J.Petersen,K.H.Maier-Hein,nnu- net: a self-configuring method for deep learning-based biomedical image segmentation, Nat. methods 18 (2) (2021) 203–211

2021
[10]

Zhou, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306 (2021)

J.Chen,Y.Lu,Q.Yu,X.Luo,E.Adeli,Y.Wang,L.Lu,A.L.Yuille, Y. Zhou, Transunet: Transformers make strong encoders for medical image segmentation, arXiv preprint arXiv:2102.04306 (2021)

2021 arXiv
[11]

H. Cao, Y. Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, M. Wang, Swin-unet: Unet-like pure transformer for medical image segmenta- tion, in: Eur. Conf. Comput. Vis., 2022, pp. 205–218

2022
[12]

Tragakis, C

A. Tragakis, C. Kaul, R. Murray-Smith, D. Husmeier, The fully convolutional transformer for medical image segmentation, in: Wint. Conf. Applicat. Comput. Vis., 2023, pp. 3660–3669

2023
[13]

Gupta, S

A. Gupta, S. Gupta, R. Katarya, Instacovnet-19: A deep learning classificationmodelforthedetectionofcovid-19patientsusingchest x-ray, Appl. Soft Comput. 99 (2021) 106859

2021
[14]

T.Iqball,M.A.Wani,Covid-19andpneumoniadetectionusingdeep weighted ensemble model, in: Int’l Conf. Comput. Sustain. Glob. Dev., 2022, pp. 337–340

2022
[15]

Saifullah, R

S. Saifullah, R. Dreżewski, Advanced medical image segmentation enhancement:Aparticle-swarm-optimization-basedhistogramequal- ization approach, Appl. Sci. 14 (2) (2024) 923

2024
[16]

J. Dolz, K. Gopinath, J. Yuan, H. Lombaert, C. Desrosiers, I. B. Ayed, Hyperdense-net: a hyper-densely connected cnn for multi- modalimagesegmentation,IEEETrans.Med.Imaging38(5)(2018) 1116–1126

2018
[17]

Zhang, N

Y. Zhang, N. He, J. Yang, Y. Li, D. Wei, Y. Huang, Y. Zhang, Z. He, Y. Zheng, mmformer: Multimodal medical transformer for incompletemultimodallearningofbraintumorsegmentation,in:Int. Conf.Med.Img.Comput.andComput.Assist.Interv.,2022,pp.107– 117

2022
[18]

J. Zhu, M. Ge, Z. Chang, W. Dong, Crcnet: Global-local context and multi-modality cross attention for polyp segmentation, Biomed. Signal Process. Control 83 (2023) 104593

2023
[19]

Z. Xing, L. Yu, L. Wan, T. Han, L. Zhu, Nestedformer: Nested modality-aware transformer for brain tumor segmentation, in: Int. Conf.Med.Img.Comput.andComput.Assist.Interv.,2022,pp.140– 150

2022
[20]

Zheng, H

J. Zheng, H. Liu, Y. Feng, J. Xu, L. Zhao, Casf-net: Cross-attention and cross-scale fusion network for medical image segmentation, Comput. Methods Programs Biomed. 229 (2023) 107307

2023
[21]

Y. Lin, Y. Chen, K.-T. Cheng, H. Chen, Few shot medical image segmentation with cross attention transformer, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv., 2023, pp. 233–243

2023
[22]

Z. Liu, H. Mao, C.-Y. Wu, C. Feichtenhofer, T. Darrell, S. Xie, A convnet for the 2020s, in: IEEE Conf. Comput. Vis. Pattern Recog., 2022, pp. 11976–11986

2022
[23]

E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, P. Luo, Segformer: Simple and efficient design for semantic segmentation with transformers, Adv. Neural Inform. Process. Syst. 34 (2021) 12077–12090

2021
[24]

Bernard, A

O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P.-A. Heng, I. Cetin, K. Lekadir, O. Camara, M. A. G. Ballester, et al., Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?, IEEE Trans. Med. Imaging ...

2018
[25]

Landman, Z

B. Landman, Z. Xu, J. Igelsias, M. Styner, T. Langerak, A. Klein, Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv. Worksh., 2015

2015
[26]

Dosovitskiy, L

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T.Unterthiner,M.Dehghani,M.Minderer,G.Heigold,S.Gelly,etal., An image is worth 16x16 words: Transformers for image recognition at scale, arXiv preprint arXiv:2010.11929 (2020)

2010 arXiv
[27]

Z.Liu,Y.Lin,Y.Cao,H.Hu,Y.Wei,Z.Zhang,S.Lin,B.Guo,Swin transformer: Hierarchical vision transformer using shifted windows, in: Int. Conf. Comput. Vis., 2021, pp. 10012–10022

2021
[28]

Huang, Z

X. Huang, Z. Deng, D. Li, X. Yuan, Y. Fu, Missformer: An effective transformer for 2d medical image segmentation, IEEE Trans. Med. Imaging 42 (5) (2022) 1484–1494

2022
[29]

M. M. Rahman, M. Munir, R. Marculescu, Emcad: Efficient multi- scale convolutional attention decoding for medical image segmenta- tion, in: IEEE Conf. Comput. Vis. Pattern Recog., 2024, pp. 11769– 11779

2024
[30]

T. Chen, C. Wang, Z. Chen, Y. Lei, H. Shan, Hidiff: hybrid diffusion frameworkformedicalimagesegmentation,IEEETrans.Med.Imag- ing (2024)

2024
[31]

J. Ruan, J. Li, S. Xiang, Vm-unet: Vision mamba unet for medical image segmentation, arXiv preprint arXiv:2402.02491 (2024)

2024 arXiv
[32]

Neural Inform

Y.Liu,Y.Tian,Y.Zhao,H.Yu,L.Xie,Y.Wang,Q.Ye,J.Jiao,Y.Liu, Vmamba: Visual state space model, Adv. Neural Inform. Process. Syst. 37 (2024) 103031–103063

2024
[33]

Zhang, J

Y. Zhang, J. Yang, J. Tian, Z. Shi, C. Zhong, Y. Zhang, Z. He, Modality-aware mutual learning for multi-modal medical image seg- mentation, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv., 2021, pp. 589–599

2021
[34]

C.-F.R.Chen,Q.Fan,R.Panda,Crossvit:Cross-attentionmulti-scale vision transformer for image classification, in: Int. Conf. Comput. Vis., 2021, pp. 357–366

2021
[35]

X. Li, S. Ma, J. Xu, J. Tang, S. He, F. Guo, Transiam: Aggregating multi-modal visual features with locality for medical image segmen- tation, Expert Syst. Appl. 237 (2024) 121574

2024
[36]

Patrascu, Color image enhancement using the lrgb coordinates in thecontextofsupportfuzzification,arXivpreprintarXiv:1502.04499 (2015)

V. Patrascu, Color image enhancement using the lrgb coordinates in thecontextofsupportfuzzification,arXivpreprintarXiv:1502.04499 (2015)

2015 arXiv
[37]

Noh, W.-S

J.-H. Noh, W.-S. Jeon, S.-Y. Rhee, Object recognition methodology for detecting multiple diseases in musculoskeletal and chest images of companion animals, International Journal of Fuzzy Logic and Intelligent Systems 25 (2) (2025) 125–135

2025
[38]

Y.Wang,Y.Zhang,F.Hou,Y.Liu,J.Tian,C.Zhong,Y.Zhang,Z.He, Modality-pairinglearningforbraintumorsegmentation,in:Int.Conf. Med. Img. Comput. and Comput. Assist. Interv., 2021, pp. 230–240

2021
[39]

Hendrycks, K

D. Hendrycks, K. Gimpel, Gaussian error linear units (gelus), arXiv preprint arXiv:1606.08415 (2016)

2016 arXiv
[40]

Kervadec, J

H. Kervadec, J. Bouchtiba, C. Desrosiers, E. Granger, J. Dolz, I. B. Ayed, Boundary loss for highly unbalanced segmentation, in: Int. Conf. Med. Imaging with deep Learn., 2019, pp. 285–296

2019
[41]

M.M.Rahman,R.Marculescu,Multi-scalehierarchicalvisiontrans- former with cascaded attention decoding for medical image segmen- tation, arXiv preprint arXiv:2303.16892 (2023)

2023 arXiv
[42]

H.Wang,S.Xie,L.Lin,Y.Iwamoto,X.-H.Han,Y.-W.Chen,R.Tong, Mixed transformer u-net for medical image segmentation, in: Int. J. Noh et al.:Preprint submitted to ElsevierPage 8 of 15 Dual Interaction Network Conf. Acoust., Speech and Signal Process., 2022, pp. 2390–2394

2022
[43]

C. You, R. Zhao, F. Liu, S. Dong, S. Chinchali, U. Topcu, L. Staib, J. Duncan, Class-aware adversarial transformers for medical image segmentation, Adv. Neural Inform. Process. Syst. 35 (2022) 29582– 29596

2022
[44]

M. M. Rahman, R. Marculescu, Medical image segmentation via cascadedattentiondecoding,in:Wint.Conf.Applicat.Comput.Vis., 2023, pp. 6222–6231

2023
[45]

H.-Y. Zhou, J. Guo, Y. Zhang, L. Yu, L. Wang, Y. Yu, nnformer: Interleaved transformer for volumetric segmentation, arXiv preprint arXiv:2109.03201 (2021)

2021 arXiv
[46]

C. Li, C. Guo, C. C. Loy, Learning to enhance low-light image via zero-reference deep curve estimation, IEEE Trans. Pattern Anal. Mach. Intell. 44 (8) (2021) 4225–4238

2021
[47]

Bernal, F

J. Bernal, F. J. Sánchez, G. Fernández-Esparrach, D. Gil, C. Ro- dríguez, F. Vilariño, Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians, Com- puterized medical imaging and graphics 43 (2015) 99–111

2015
[48]

Codella, V

N. Codella, V. Rotemberg, P. Tschandl, M. E. Celebi, S. Dusza, D. Gutman, B. Helba, A. Kalloo, K. Liopyris, M. Marchetti, et al., Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the international skin imaging collaboration (isic), arXiv preprint arXi...

2018 arXiv
[49]

Oktay, J

O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Mi- sawa, K. Mori, S. McDonagh, N. Y. Hammerla, B. Kainz, et al., Attention u-net: Learning where to look for the pancreas, arXiv preprint arXiv:1804.03999 (2018)

2018 arXiv
[50]

L.-C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, A. L. Yuille, Deeplab:Semanticimagesegmentationwithdeepconvolutionalnets, atrous convolution, and fully connected crfs, IEEE Trans. Pattern Anal. Mach. Intell. 40 (4) (2017) 834–848

2017
[51]

Conf.Med.Img.Comput.andComput.Assist.Interv.,2020,pp.263– 273

D.-P.Fan,G.-P.Ji,T.Zhou,G.Chen,H.Fu,J.Shen,L.Shao,Pranet: Parallel reverse attention network for polyp segmentation, in: Int. Conf.Med.Img.Comput.andComput.Assist.Interv.,2020,pp.263– 273

2020
[52]

A. Lou, S. Guan, H. Ko, M. H. Loew, Caranet: context axial reverse attention network for segmentation of small medical objects, in: MedicalImaging2022:ImageProcessing,Vol.12032,2022,pp.81– 92

2022
[53]

T. Kim, H. Lee, D. Kim, Uacanet: Uncertainty augmented context attention for polyp segmentation, in: ACM Int. Conf. Multimedia, 2021, pp. 2167–2175

2021
[54]

J.Wang,Q.Huang,F.Tang,J.Meng,J.Su,S.Song,Stepwisefeature fusion: Local guides global, in: International conference on medical imagecomputingandcomputer-assistedintervention,2022,pp.110– 120

2022
[55]

B. Dong, W. Wang, D.-P. Fan, J. Li, H. Fu, L. Shao, Polyp-pvt: Polypsegmentationwithpyramidvisiontransformers,arXivpreprint arXiv:2108.06932 (2021)

2021 arXiv
[56]

Y.Zhang,H.Liu,Q.Hu,Transfuse:Fusingtransformersandcnnsfor medical image segmentation, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv., 2021, pp. 14–24

2021
[57]

J. M. J. Valanarasu, V. M. Patel, Unext: Mlp-based rapid medical image segmentation network, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv., 2022, pp. 23–33

2022
[58]

Perumal, T

S. Perumal, T. Velmurugan, Preprocessing by contrast enhancement techniques for medical images, Int. J. Pure Appl. Math. 118 (18) (2018) 3681–3688. J. Noh et al.:Preprint submitted to ElsevierPage 9 of 15 Dual Interaction Network Supplementary: Dual Interaction Network with Cr...

2018
[59]

The details of these supplementary components are described in the following sections

Sup: Introduction In this supplementary material, we present additional ablation studies, image enhancement methods, the structure of the encoder and decoder, the gradient extractor of multi- scale boundary loss, limitation,s and further visual compar- ison. The details of the...
[60]

We demonstrated the efficacy of our method through a number of DIFMs

Additional Ablation Studies We have conducted additional ablation studies on the ACDC and Synapse datasets. We demonstrated the efficacy of our method through a number of DIFMs. Weevaluatethecross-attentionmethodanddemonstrate the superiority of our multi-scale boundary loss f...

1989
[61]

In thispaper,weusefuzzyimageenhancement[36],whichhas proveneffectiveinvariousworks[13,14],forenhancingthe originalimage.Thefuzzyimageenhancementutilizesfuzzy J

Image Enhancement Inmedicalimaging,preprocessingthroughenhancement techniques is necessary for effective noise handling [58]. In thispaper,weusefuzzyimageenhancement[36],whichhas proveneffectiveinvariousworks[13,14],forenhancingthe originalimage.Thefuzzyimageenhancementutilize...
[62]

Encoder and Decoder 9.1. Encoder In this paper, we employ ConvNext-base [22] as our en- codertoextractdiversestructuralattributesfromtheoriginal andenhancedimage.ConvNextisdesignedtocompetewith vision transformers, incorporating various architectural en- hancements,suchaseffic...
[63]

The gradient extractor employs the Sobel filter to quickly and accurately identify boundaries

Multi-Scale Boundary Loss Ours𝐿 𝑏𝑛𝑑 implements a gradient extractor to efficiently compute gradients such as boundaries and intensities be- tween adjacent pixels. The gradient extractor employs the Sobel filter to quickly and accurately identify boundaries. The Sobel filter, t...
[64]

Computational Cost The computational cost of the proposed method is sig- nificant. This study focuses on achieving high accuracy to support precise diagnostic assistance, which necessitated the use of two heavy encoders, ConvNext-base [22], re- sulting in a total of 323.31M pa...
[65]

Additional Visual Comparisons Todemonstratethesuperiorityofourmethod,additional qualitative comparison results for the ACDC and Synapse datasets are presented in Fig. 6, Fig. 7, respectively. No- tably, in the case of GB, the quantitative results in Fig. 7 show relatively lowe...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

S. Jang, H. Song, Y. J. Shin, J. Kim, J. Kim, K. W. Lee, S. S. Lee, W. Lee, S. Lee, K. H. Lee, Deep learning–based automatic detection algorithmforreducingoverlookedlungcancersonchestradiographs, Radiology 296 (3) (2020) 652–661. J. Noh et al.:Preprint submitted to ElsevierPage 7 of 15 Dual Interaction Network

work page 2020

[2] [2]

S.Lee,J.Park,J.Park,Crossformer:Cross-guidedattentionformulti- modal object detection, Pattern Recognition Letters 179 (2024) 144– 150

work page 2024

[3] [3]

J. Kim, J. Noh, M. Jeong, W. Lee, Y. Park, J. Park, Adnet: Non-local affinity distillation network for lightweight depth completion with guidancefrommissinglidarpoints,IEEERobot.andAutomat.Letters (2024)

work page 2024

[4] [4]

S. Ha, Y. Kim, J. Park, Interdimensional knowledge transfer for se- manticsegmentationonlidarpointclouds,IEEERobot.andAutomat. Letters (2024)

work page 2024

[5] [5]

H. Son, J. Noh, S. Jeon, C. Wang, W.-K. Jeong, Mc-inr: Effi- cient encoding of multivariate scientific simulation data using meta- learning and clustered implicit neural representations, arXiv preprint arXiv:2507.02494 (2025)

work page arXiv 2025

[6] [6]

Ronneberger, P

O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks forbiomedicalimagesegmentation,in:Int.Conf.Med.Img.Comput. and Comput. Assist. Interv., 2015, pp. 234–241

work page 2015

[7] [7]

Z.Zhou,M.M.RahmanSiddiquee,N.Tajbakhsh,J.Liang,Unet++: A nested u-net architecture for medical image segmentation, in: Int. Conf. Med. Img. Comput. and Comput. Assist. Interv., 2018, pp. 3– 11

work page 2018

[8] [8]

Huang, L

H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y. Iwamoto, X. Han, Y.-W. Chen, J. Wu, Unet 3+: A full-scale connected unet for medical image segmentation, in: Int. Conf. Acoust., Speech and Signal Pro- cess., 2020, pp. 1055–1059

work page 2020