Accuracy Assessment of a Deep-Learning Based Segmentation Tool Over Right Ventricle Short-Axis Slices

Abstract

The right ventricle (RV) is the largest heart cavity playing a vital role in the cardiac cycle and function. However, to assess its function in cardiac magnetic resonance imaging (CMRI), the segmentation is an essential but very challenging task due to the RV complex crescent shape, thin boundaries, as well as the high variability of images. Indeed, to overcome these issues, several approaches have been proposed. In this paper, we aim to highlight the impact of various short-axis slices on the segmentation process. For that sake, based on the Cvi42 advanced deep learning base software, we assess the effectiveness of the RV contour delineation process over an entire MRI cardiac short-axis slice. We consider for evaluating the accuracy, the Dice Score Metric (DSM), as well as some RV functional parameters such as the Ejection Fraction, the End-Systolic-Volume, and the End-Diastolic-Volume. Our experimental assessment is done using two CMRI exams of both healthy and diseased patients. We consider for evaluation the whole short-axis frames from base to apex and from end-systole to end-diastole. The evaluation results show that: (1) The segmentation accuracy is influenced by the slice level especially on the basal and apical slices, (2) A huge accuracy decreasing for the case of a patient having dysplasia.