MSA‐MaxNet: Multi‐Scale Attention Enhanced Multi‐Axis Vision Transformer Network for Medical Image Segmentation DOI Creative Commons
Wei Wu, Junfeng Huang, Mingxuan Zhang

et al.

Journal of Cellular and Molecular Medicine, Journal Year: 2024, Volume and Issue: 28(24)

Published: Dec. 1, 2024

ABSTRACT Convolutional neural networks (CNNs) are well established in handling local features visual tasks; yet, they falter managing complex spatial relationships and long‐range dependencies that crucial for medical image segmentation, particularly identifying pathological changes. While vision transformer (ViT) excels addressing dependencies, their ability to leverage remains inadequate. Recent ViT variants have merged CNNs improve feature representation segmentation outcomes, yet challenges with limited receptive fields precise persist. In this work, we propose MSA‐MaxNet. Specifically, our model utilises an encoder–decoder structure, using MaxViT blocks apply multi‐axis self‐attention (Max‐SA) as the encoder global extraction. To restore map's resolution during upsampling operations, a symmetric block–based decoder layers employed. address mismatches skip connections of UNet architecture, introduce convolutional block attention module (CBAM). Furthermore, design multi‐scale (MCBAM) based on CBAM, which enhance refine connection. We evaluate performance MSA‐MaxNet three publicly available imaging datasets, including Synapse multi‐organ ACDC cardiac analysis Kvasir‐SEG gastrointestinal polyp detection. Notably, achieves state‐of‐the‐art (SOTA) Dice scores 85.59% 95.26% respectively, 40.28 M parameters. Additionally, two smaller versions meet demands various scenarios. summary, work provides robust framework diverse tasks, offering potential applications early cancer detection, cardiovascular disease diagnosis comprehensive organ‐level assessments.

Language: Английский

MPKU‐Net: A U‐Shaped Medical Image Segmentation Network Based on MLP and KAN DOI

Peng Chen,

Huihui Wang,

Qin Jin

et al.

International Journal of Imaging Systems and Technology, Journal Year: 2025, Volume and Issue: 35(3)

Published: May 1, 2025

ABSTRACT The UNET architecture has been widely adopted for image segmentation across various domains, owing to its efficient and powerful performance in recent years. Its application enhancement medical primarily involve convolutional neural network (CNN) Transformer. However, both methods have fundamental limitations. CNN struggle capture global features, which greatly reduces the computational complexity but compromises effectiveness. Transformers excel at capturing features demand substantial parameters computations fail effectively extract local features. To address these challenges, we propose a U‐shaped model, MPKU‐NET, integrates multilayer perception (MLP) with Knowledge‐Aware Networks (KAN) architecture, aiming characteristics coordinated manner. MPKU‐NET flexible rolling Flip operation that, along MLP Network (KAN), creates WE‐MPK modules thorough learning of effectiveness is proven by extensive testing on BUSI, CVC, GlaS datasets. results demonstrate that MPKU‐Net consistently outperforms several used networks, including U‐KAN, Rolling‐U‐net, U‐Net ++, terms model accuracy, highlighting as scalable solution segmentation. code uploaded: https://github.com/cp668688/MPKU‐Net .

Language: Английский

Citations

0
Attention-enhanced Separable Residual with Dilation Net for Medical Image Segmentation DOI
Leyi Xiao, Yang Liu, Chaodong Fan

et al.

Neurocomputing, Journal Year: 2025, Volume and Issue: unknown, P. 130434 - 130434

Published: May 1, 2025

Language: Английский

Citations

0
Adapting Classification Neural Network Architectures for Medical Image Segmentation Using Explainable AI DOI Creative Commons
Artūrs Ņikuļins, Edgars Edelmers, Kaspars Sudars

et al.

Journal of Imaging, Journal Year: 2025, Volume and Issue: 11(2), P. 55 - 55

Published: Feb. 13, 2025

Segmentation neural networks are widely used in medical imaging to identify anomalies that may impact patient health. Despite their effectiveness, these face significant challenges, including the need for extensive annotated data, time-consuming manual segmentation processes and restricted data access due privacy concerns. In contrast, classification networks, similar capture essential parameters identifying objects during training. This paper leverages this characteristic, combined with explainable artificial intelligence (XAI) techniques, address challenges of segmentation. By adapting tasks, proposed approach reduces dependency on To demonstrate concept, Medical Decathlon 'Brain Tumours' dataset was utilised. A ResNet network trained, XAI tools were applied generate segmentation-like outputs. Our findings reveal GuidedBackprop is among most efficient effective methods, producing heatmaps closely resemble masks by accurately highlighting entirety target object.

Language: Английский

Citations

0
MSA‐MaxNet: Multi‐Scale Attention Enhanced Multi‐Axis Vision Transformer Network for Medical Image Segmentation DOI Creative Commons
Wei Wu, Junfeng Huang, Mingxuan Zhang

et al.

Journal of Cellular and Molecular Medicine, Journal Year: 2024, Volume and Issue: 28(24)

Published: Dec. 1, 2024

ABSTRACT Convolutional neural networks (CNNs) are well established in handling local features visual tasks; yet, they falter managing complex spatial relationships and long‐range dependencies that crucial for medical image segmentation, particularly identifying pathological changes. While vision transformer (ViT) excels addressing dependencies, their ability to leverage remains inadequate. Recent ViT variants have merged CNNs improve feature representation segmentation outcomes, yet challenges with limited receptive fields precise persist. In this work, we propose MSA‐MaxNet. Specifically, our model utilises an encoder–decoder structure, using MaxViT blocks apply multi‐axis self‐attention (Max‐SA) as the encoder global extraction. To restore map's resolution during upsampling operations, a symmetric block–based decoder layers employed. address mismatches skip connections of UNet architecture, introduce convolutional block attention module (CBAM). Furthermore, design multi‐scale (MCBAM) based on CBAM, which enhance refine connection. We evaluate performance MSA‐MaxNet three publicly available imaging datasets, including Synapse multi‐organ ACDC cardiac analysis Kvasir‐SEG gastrointestinal polyp detection. Notably, achieves state‐of‐the‐art (SOTA) Dice scores 85.59% 95.26% respectively, 40.28 M parameters. Additionally, two smaller versions meet demands various scenarios. summary, work provides robust framework diverse tasks, offering potential applications early cancer detection, cardiovascular disease diagnosis comprehensive organ‐level assessments.

Language: Английский

Citations

0