Scholarlink

Deep learning–based medical image analysis is driving digital transformation in biomedical engineering by enabling automatic assessment of cardiac function from MRI data. However, the black-box nature of modern segmentation networks remains a major barrier to their adoption in clinical workflows. In this study, we propose Fuzzy Attention 3D U-Net (FA-3DUNet), a fuzzy-attention–enhanced 3D U-Net architecture for multi-class short-axis cardiac MRI segmentation, targeting three anatomical structures: the left ventricle (LV), right ventricle (RV), and myocardium (MYO). The proposed fuzzy attention module improves spatial focus in ambiguous boundary regions and enhances segmentation robustness under multi-center and multi-vendor conditions, supporting practical deployment in real-world medical imaging systems. To improve model transparency, we further introduce a hybrid explainability framework that integrates multi-layer Grad-CAM analysis with intrinsic fuzzy attention maps. Unlike conventional approaches that focus solely on the final output layer, the proposed layer-wise strategy enables visualization of feature evolution across encoder and decoder stages, providing insight into how 3D networks progressively refine structural and semantic representations. In addition, we analyze model attention quantitatively by aggregating Grad-CAM and attention responses across a set of representative cases from the M&Ms-2 dataset, allowing qualitative slice-wise inspection to be supported by reproducible statistical characterization of attentional behavior. Experimental results on the multi-vendor M&Ms-2 benchmark demonstrate that FA-3DUNet achieves a mean Dice score of 88.19%, outperforming the original Attention U-Net while maintaining consistent performance across diverse cardiac conditions. Overall, the proposed method offers both methodological advancement and practical value in explainable cardiac MRI segmentation, contributing to trustworthy AI-assisted diagnostics in biomedical engineering and digital healthcare applications.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explaining Fuzzy Attention in 3D U-Net for Multi-class Cardiac MRI Segmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Thi-My-Nga Nguyen,
  • Anh-Cang Phan

摘要

Deep learning–based medical image analysis is driving digital transformation in biomedical engineering by enabling automatic assessment of cardiac function from MRI data. However, the black-box nature of modern segmentation networks remains a major barrier to their adoption in clinical workflows. In this study, we propose Fuzzy Attention 3D U-Net (FA-3DUNet), a fuzzy-attention–enhanced 3D U-Net architecture for multi-class short-axis cardiac MRI segmentation, targeting three anatomical structures: the left ventricle (LV), right ventricle (RV), and myocardium (MYO). The proposed fuzzy attention module improves spatial focus in ambiguous boundary regions and enhances segmentation robustness under multi-center and multi-vendor conditions, supporting practical deployment in real-world medical imaging systems. To improve model transparency, we further introduce a hybrid explainability framework that integrates multi-layer Grad-CAM analysis with intrinsic fuzzy attention maps. Unlike conventional approaches that focus solely on the final output layer, the proposed layer-wise strategy enables visualization of feature evolution across encoder and decoder stages, providing insight into how 3D networks progressively refine structural and semantic representations. In addition, we analyze model attention quantitatively by aggregating Grad-CAM and attention responses across a set of representative cases from the M&Ms-2 dataset, allowing qualitative slice-wise inspection to be supported by reproducible statistical characterization of attentional behavior. Experimental results on the multi-vendor M&Ms-2 benchmark demonstrate that FA-3DUNet achieves a mean Dice score of 88.19%, outperforming the original Attention U-Net while maintaining consistent performance across diverse cardiac conditions. Overall, the proposed method offers both methodological advancement and practical value in explainable cardiac MRI segmentation, contributing to trustworthy AI-assisted diagnostics in biomedical engineering and digital healthcare applications.