Learning Left Main Bifurcation Shape Features With An Autoencoder

Geometric characteristics of the coronary arteries have been suggested as potential markers for disease risk. However, evaluation of such characteristics rely on judgement by human experts, and are thus variable and may lack sophistication. Here we apply recent advances in 3D deep learning to automatically obtain shape representation of the Left Main Bifurcation (LMB) of the coronary artery. We train a Variational Auto-Encoder based on the FoldingNet architecture to encode LMB shape features in a 450-dimension feature vector. The geometric features of patient-specific LMBs can then be manipulated by modifying, combining or interpolating the feature vectors before decoding. We also show that these vectors, on average, perform better than hand-crafted features in predicting measures of adverse blood flow (oscillating shear index or 'OSI', relative residence time 'RRT' and time averaged wall shear stress 'TAWSS') with a 'R^2' goodness of fit value of 84.1% compared to 79.7%. These learned representations can also be used in other downstream predictive modelling tasks where an encoded version of a LMB is needed.