Multimodal CMR Image Myocardial Lesion Segmentation and Application Based on AI-Enabled System Design

In this study, an end-to-end deep learning model for cardiomyopathy segmentation in multimodal cardiac magnetic resonance (CMR) images is proposed. The model adopts a coarse to fine frame, using 2D U-Net for initial coarse segmentation of cardiac structures, and then for fine segmentation of cardio myocardial lesions through 3D nnU-Net. The model integrates multimodal data (LGE, T2w, bSSFP) and adopts a two-stage training process of 50% cross-validation to improve segmentation accuracy. Through post-processing technology, the model results are aligned with clinical data, supporting 3D myocardial volume visualization and 4D spatiotemporal analysis. Experiments on the MyoPS 2020 dataset show that the model shows high accuracy (Dice > 0.9) in myocardial structure segmentation and significantly improves the detection ability of lesions. In addition, the model is optimized for efficient training iteration and real-time reasoning capabilities, making it suitable for clinical applications. Future research will focus on small object detection and the improvement of cross-vendor generalization capabilities.