Explainable Fuzzy Learning for Cardiac Segmentation in 3D Short-Axis MRI

Accurate segmentation in short-axis cardiac MRI is crucial for extracting pa-rameters to diagnose cardiovascular diseases. However, deep learning models often use Max or Average pooling to reduce feature map size, risking loss of important information. In addition, most current research focuses mainly on the segmentation accuracy of machine learning models, while lacking inter-pretation of the segmentation results- a critical factor for clinical applica-tions. In this study, we propose a fuzzy deep learning model called 3DMFL-Net, which combines Explainable AI (XAI) techniques to segment cardiac structures including the Left Ventricle, Myocardium, and Right Ventricle from 3D short axis MRI images, while also providing interpretation of the segmentation results. The model adopts a 3D fuzzy pooling method to re-place traditional pooling methods. The 3D fuzzy pooling leverages fuzzy logic along with a Gaussian membership function to identify important features within the pooling window, allowing reduction of the feature map size while preserving essential information. To visualize and interpret the model’s deci-sions, we apply XAI with the 3D Grad-CAM method to generate heatmaps that highlight the important image regions involved in the segmentation process of the proposed model. The proposed model is evaluated on the M&Ms-2020 dataset, which includes multiple 3D cardiac MRI scans across various disease groups and is provided by different device vendors. The re-sults show that the Dice coefficient reaches 84.6% for the left ventricle, 76.6% for the myocardium, and 79.5% for the right ventricle. The proposed method suits intelligent healthcare deployment and helps clinicians trust segmentation outcomes.