Abstract 12220: Deep Learning Segmentation and Quantification of Epicardial Adipose Tissue in CT Calcium Score Images

Introduction: Studies have linked the epicardial adipose tissue (EAT) volume with the risk of adverse cardiovascular (CV) events, suggesting a role for clinical decision making. Manual quantification of EAT is time-consuming, requires specialized training, and is prone to human error, thus limiting its routine use in clinical practice. Hypothesis: Deep-learning approach can facilitate EAT quantification on archival coronary artery calcium scoring (CAC) scans. Methods: We developed a deep learning semantic segmentation method using the DeepLab-v3 plus network and applied it to gated, non-contrast cardiac CT calcium score scans. Analysts manually identified the enclosing pericardial sac using interactive visualization software. Our algorithm was trained to segment the image region enclosed by the pericardial sac. To keep the sac-increasing curvature, we segmented the heart into two halves. The lower half was analyzed from bottom-to-middle, while the slices in the upper half were analyzed from top-to-middle. Each 2D slice was associated with two consecutive slices, giving a three-slice image slab. A window/level of 350/40 Hounsfield unit (HU) was applied to axial slices providing a uniform HU-attention-window. A median filter with kernel size 3mm was applied to reduce noise. After deep learning semantic segmentation of the enclosing region, we determined EAT using the standard CT fat window [-190, -30] HU. Results: Using 89 CT scans (50 training/39 testing), our algorithm showed excellent results compared to ground-truth manual labeling. The total average Dice score was (88.52%±3.35) with a high correlation against manually derived EAT (R=98.52%, p-value Conclusions: Our developed end-to-end deep learning segmentation approach, using HU-attention-window and slab-of-slices gives accurate EAT volume segmentation. Our results provide an important avenue for automated analysis of EAT and its relationship with CV events in large data sets.