Deep Learning Based Volumetric Segmentation of Heart Ventricles for Assessment of Cardiac Disease Using MRI

Diagnosis of cardiovascular diseases through cardiac MRI imaging plays a crucial role. Manual evaluation is time consuming and prone to errors. With the help of deep learning, a lot of traction has been developed for cardiac imaging diagnosis. In this study, we present a fully automated pipeline for the segmentation of left ventricle, right ventricle, myocardium, and classification of cardiovascular diseases into five classes using the cardiac MRI scans from the ACDC dataset. We adopted Segnet architecture for segmenta-tion and made a comparative analysis using 2D and 3D approach. Best results were obtained using 2D approach with dice scores of 0.877(RV), 0.877(MYO), 0.937(LV) on the test set. We later on use the segmentation outputs to extract quantitative features to develop a robust classifier that gave us an overall accuracy of 85% on the test set and 0.81,0.89 scores of precisions and recall. Our proposed approach is computationally efficient and can be used for making critical decisions during diagnosis.