Modified U-Net framework for left ventricle and myocardium segmentation and visual interpretation using Grad-CAM

This article introduces an advanced segmentation framework for the Left Ventricle (LV) and myocardium in Cardiac Magnetic Resonance Imaging (CMRI) based on a modified U-Net architecture and the Adam optimization algorithm. LV segmentation is essential in cardiac image analysis, supporting diagnosis and treating cardiovascular diseases. This paper presents a modified U-Net framework for the segmentation of the left ventricle and myocardium from cardiac MRI. To interpret the model decisions and validation, this work employed Gradient-weighted Class Activation Mapping (Grad-CAM) to provide qualitative evidence and enable a clearer understanding of which input regions have the greatest impact on the model predictions, achieving a Dice score of 97 % for CMRI ACDC 2017 Dataset. The modified U-Net in conjunction with Adam optimization has significantly improved segmentation accuracy and efficiency. This study evaluates the model performance in segmenting LV and myocardium during the end-systole (ES) and end-diastole (ED) phases in CMRI. The model was trained over several epochs, and Key measures to assess overlap similarity, spatial alignment, boundary proximity, and pixel classification accuracy were used to measure its effectiveness. Results revealed consistent improvements across all metrics with increasing epochs, highlighting the model's robustness. The system was evaluated on three publicly available datasets. Among them, the Sunnybrook Cardiac Dataset yielded the highest accuracy of 98 % in left ventricle (LV) and myocardium segmentation. This improved performance contributes to more precise treatment of cardiac conditions.