Deep Learning for pericardial fat extraction and evaluation on a population study

The pericardial fat represents a powerful (promising) index which has been seen to correlate with several cardiovascular events. We propose a novel approach to automatically measure it from CT scans and we seek to explore how it is distributed in a study population. We studied a population of 1528 patients where 47 (3%) showed a cardiac event. The fat segmentation model was based on a deep neural network to identify the heart and from it, the pericardial fat was extract by threshold. Statistical analysis was finally computed to stratify the population according the quantity of pericardial fat. The high segmentation quality was reported with a Dice index (92.5%) and a Pearson coefficient (0.990, p<0.001). Notably, normalized pericardial fat volume was significantly higher in patients with cardiac event (73.46{+/-}30.84 vs 60.06{+/-}25.38 cm^3/m^2, p=0.005) with a AUC of 63,44% for discrimination of cardiac event. The proposed approach reached accurate and fast performance for the evaluation of the pericardial fat making it reliable for a population analysis.