A hybrid segmentation and classification CAD framework for automated myocardial infarction prediction from MRI images

Abstract

Early diagnosis of myocardial infarction (MI) is critical for preserving cardiac function and improving patient outcomes through timely intervention. This study proposes an annovaitive computer-aided diagnosis (CAD) system for the simultaneous segmentation and classification of MI using MRI images. The system is evaluated under two primary approaches: a serial approach, where segmentation is first applied to extract image patches for subsequent classification, and a parallel approach, where segmentation and classification are performed concurrently using full MRI images. The multi-class segmentation model identifies four key heart regions: left ventricular cavity (LV), normal myocardium (Myo), myocardial infarction (MI), and persistent microvascular obstruction (MVO). The classification stage employs three AI-based strategies: a single deep learning model, feature-based fusion of multiple AI models, and a hybrid ensemble model incorporating the Vision Transformer (ViT). Both segmentation and classification models are trained and validated on the EMIDEC MRI dataset using five-fold cross-validation. The adopted ResU-Net achieves high F1-scores for segmentation: 91.12% (LV), 88.39% (Myo), 80.08% (MI), and 68.01% (MVO). For classification, the hybrid CNN-ViT model in the parallel approach demonstrates superior performance, achieving 98.15% accuracy and a 98.63% F1-score. These findings highlight the potential of the proposed CAD system for real-world clinical applications, offering a robust tool to assist healthcare professionals in accurate MI diagnosis, improved treatment planning, and enhanced patient care.