Toward an automatic detection of cardiac structures in short and long axis views

Abstract

Objective: This work aims to create an automatic detection process of cardiac structures in both short-axis and long-axis views. A workflow inspired by human thinking process, for better explainability. Methods: we began by separating the images into two classes: long axis and short axis, using a Residual Network model. Then, we used Particle Swarm Optimization for general segmentation. After segmentation, a characterization step based on shape descriptors calculated from bounding box and ANOVA for features selection were applied on the binary images to detect the location of each region of interest: lung, left and right ventricle in the short-axis view, the aorta, the left heart (left atrium and ventricle), and the right heart (right atrium and ventricle) in the long axis view. Results: we achieved a 90% accuracy on view separation. We have selected: Elongation, Compactness, Circularity, Type Factor, for short axis identification; and:Area, Centre of Mass Y, Moment of Inertia XY, Moment of Inertia YY, for long axis identification. Conclusion: a successful separation of long axis and short axis views allows for a better characterization and detection of segmented cardiac structures. After that, any method can be applied for segmentation, attribute selection, and classification. Significance: an attempt to introduce explainability into cardiac image segmentation, we tried to mimic the human workflow while computerizing each step. The process seems to be valid and added clarity and interpretability to the detection