Deep learning-based whole heart segmentation from CT images

Whole Heart Segmentation (WHS) aims to extract individual heart substructure and it is a critical step for the diagnosis, treatment planning, and evaluation of various cardiac disease. In this study, we examine the potential of Deep Learning (DL) neural networks to segment eight heart substructures from Computed Tomography (CT) scans. The four heart chambers, myocardium, aorta, pulmonary artery and left atrial appendage were manually annotated manually in 211 cardiac CTA exams and inspected by clinical experts. Those exams were used to train a multi-class 3D DL segmentation model. We investigated the impact of different network architectures, including UNet and its variants, CE-UNet, CE-A-Unet, with different input patch sizes and resolution. A test dataset comprising of 51 fully annotated exams was used to evaluate the model performance. Our findings indicate that, compared to UNet, neither CE-UNet or CE-A-Unet show superior performance. Moreover, the model with larger physical input patch size with coarser pixel resolution tends to achieve higher performance. The averaged dice score across all substructures was 0.91, which exceeds the current state-of-the-art.