Multi-approach technique implementation for segmentation of left ventricular myocardial infarction and quantification

Abstract

Abstract Accurate segmentation and optimal diagnosis are required for myocardial infarction (MI) because it is a life-threatening cardiovascular condition. Therefore, its quantification is necessary to facilitate accurate diagnosis and make beneficial decisions to prevent a heart attack. Advanced cutting-edge techniques, such as deep learning, promise more precise MI segmentation and potentially provide an automated segmentation of annotated cardiac MRI (CMRI) datasets. However, segmentation accuracy is strongly affected by domain shift, low contrast of infarcted tissue, and scarcity of annotated datasets. This study introduces an innovative framework as a hybrid deep-learning framework combining Cycle-Consistent Generative Adversarial Networks (CycleGAN), Discrete Wavelet Transform (DWT), and U-Net to achieve robust MI segmentation. CycleGAN is employed to generate synthetic infarcted images from healthy MRI scans, addressing data scarcity and domain variability. DWT is integrated into the segmentation pipeline to preserve high-frequency spatial details and enhance edge localization, which are crucial for accurate infarct boundary detection. Finally, a U-Net model is trained on a hybrid dataset comprising synthetic and real images with wavelet-based feature fusion. Experiments dataset demonstrate that the proposed method significantly outperforms baseline U-Net models, achieving Dice improvements of 4–8% for infarct segmentation and 2–4% for myocardium segmentation. The results highlight the effectiveness of combining CycleGAN-based data augmentation with multi-resolution wavelet features for improved cardiac tissue delineation.