Deep-Learning-Based Myocardial Pathology Detection

Cardiovascular diseases are the top cause of death worldwide. Commonly, physicians screen suspected pathological patients with histological examinations and blood tests. Since these clinical parameters are frequently ambiguous, they are routinely extended by delayed-enhancement magnetic resonance imaging of the myocardium. We propose a method combining deep learning and classical machine learning to differentiate between pathological and normal cases. A convolutional neural network infers a segmentation of the left myocardium from a magnetic resonance image as a preliminary step. This segmentation is employed to determine radiomics-based features describing the morphology and texture of the myocardium. Subsequently, a multilayer perceptron deduces pathological cases from these radiomics features and clinical observations. The presented method demonstrates an accuracy of 0.96 and an F2-score of 0.98 on a nested cross-validation.