Prediction of paravalvular leak post transcatheter aortic valve replacement using a convolutional neural network

Para-valvular leak (PVL) post transcatheter aortic valve replacement (TAVR) is a major complication of the TAVR procedure. Although many risk factors of PVL has have been identified, there is a lack of comprehensive and accurate way to predict PVL. Convolutional neural network has been extensively used in medical image classification, because of its extraordinary capability of feature extraction of visual cues. In this study we proposed a convolutional neural network framework to predict PVL based on pre-procedural CT images, in which we aimed to learn image features that were associated with the outcomes of the interventional procedures, such as the image cues related to the native tissue compositions and the tissue compliances to pressure load. We tested our model on the dataset of 168 patients' CT images, and performed a four-fold cross validation. Training and testing data were strictly separated. The performance of the proposed model achieved an accuracy of 79% using four-fold cross validation.