Automated and Semi-Automated Frameworks for Left Ventricle Volume and Ejection Fraction Computation from CMR Images

Cardiovascular diseases (CVDs) remain a leading cause of mortality worldwide. Cardiovascular Magnetic Resonance (CMR) imaging is widely recognized as the gold standard for the non-invasive assessment of cardiac structure and function. Among the key clinical biomarkers, left ventricular volume and ejection fraction are routinely estimated from CMR images to evaluate cardiac performance. This paper presents both fully automated and semi-automated frameworks for LVEF estimation at end-systole and end-diastole, with the primary distinction lying in their segmentation methodologies. The proposed pipeline consists of image enhancement, heart localization using Fourier transform techniques, segmentation, and volume estimation. In the semi-automated system, segmentation is achieved through a learning-free region growing algorithm, augmented by global thresholding and an optimized seed point selection strategy. The automated approach employs a deep convolutional neural network based on the U-Net architecture. Both systems incorporate inter-slice spatial continuity to enhance segmentation consistency across short-axis slices of the left ventricle. Experimental evaluation was conducted on a dataset comprising 4,400 short-axis CMR images. The proposed methods achieved Continuous Ranked Probability Score (CRPS) values of 0.053 and 0.015 for the automated and semi-automated systems, respectively. The results demonstrate the effectiveness and robustness of the proposed approaches, particularly the learning-free method, in handling anatomical and imaging variability. These findings underscore the potential of the proposed systems to support reliable and efficient LVEF quantification, thereby contributing to improved clinical workflows and decision-making in cardiovascular diagnostics.