Pan-BCS: An Enhanced Panoptic Biventricular 3D Cardiac Assistive Model Integrating Feature Pyramid Networks and Parallel Semantic Segmentation

Abstract

Cardiac image segmentation, which is used to assess heart function in cardiac magnetic resonance imaging (CMRI), is a crucial step in the early diagnosis of cardiovascular disease. An improved panoptic biventricular 3D cardiac MRI segmentation (Pan-BCS) is proposed to handle the various biventricular shapes during the different cardiac cycles by comprising the Feature Pyramid Network (FPN) as a backbone for the feature extraction to extract multiscale features. Additionally, Pan-BCS offers a parallel semantic segmentation using ResUNet and instance segmentation using Mask-RCNN to enhance the biventricular segmentation. Promising results are obtained when training and testing the proposed model using the MyoPs and Automatic Cardiac Diagnosis Challenge (ACDC 2017) dataset. Pan-BCS demonstrates superior performance in biventricular segmentation compared to the state-of-the-art. Dice symmetry coefficient (DSC) performance gap for Pan-BCS's left ventricle (LV), myocardium (Myo), and right ventricle (RV) segmentation on ACDC 2017 are 3.82%, 1.26%, and 1.13%, respectively. In addition, Pan-BCS segments the LV, Myo, and RV imply a difference in the average performance of 2.59%, 5.49%, and 6.44%, respectively.