Identification of Right Ventricle Based on Adaptive UNet Architecture from Cardiac MR Images

Abstract

The anatomical functionality of the heart can be described very well using cardiac MR images. In cardiac MR images, the role of right and left ventricle delineation is very important. The right ventricle cardiac image segmentation is more critical due to the crescent varying shape of the organ and its fuzzy borders.

This paper presents Adaptive UNet architecture with varying depths to identify suitability for our selected subjects. We introduce our contribution for building four kinds of UNet models based on varying hyper-parameters of the architecture and tested on eighteen patients’ data set selected images till end-systole. The UNet4 model has achieved accuracy up to 94%. Training any network for any irregular-shaped organ shows better accuracy in the deep learning method at the cost of time to train the model. But this proposed Adaptive UNet architecture takes less time to train and can be applicable for a limited data set and give predictions with promising results. Due to continuous blood flow, the difficulty in delineating RV can be overcome with this Adaptive UNet architecture as it is less effective to connected object classes.