Two-stage deep learning framework for myocardial infarction segmentation in late-gadolinium enhancement magnetic resonance imaging images

Accurate segmentation of myocardial structures and infarct regions in late-gadolinium enhancement magnetic resonance imaging (LGE-MRI) is essential for diagnosing ischemic heart disease (IHD). However, traditional and single-stage deep learning (DL) methods struggle with small or low-contrast regions such as myocardial scars. This study proposes a two-stage DL framework to address these limitations. Stage 1 segments the LV cavity using DeepLabv3+ (ResNet50), and Stage 2 segments the myocardium and scar using DeepLabv3+ (Xception). The framework was developed through four phases: baseline evaluation, loss and optimizer exploration, two-stage pipeline integration, and final validation with post-processing. Both models were trained using Dice loss and Adam optimizer. Final testing showed high segmentation performance for the LV cavity (Dice = 0.947) and myocardium (Dice = 0.7351). Scar segmentation remained challenging (Dice = 0.0556) due to small size and low contrast. Nonetheless, the modular design enhanced anatomical accuracy and reduced inter-class misclassification, demonstrating its potential for clinical cardiac image analysis.