Non-ischemic endocardial scar geometric remodeling toward topological machine learning.

Scar tissues have been important factors in determining the progression of myocardial diseases and the development of adverse cardiac failure outcomes. Accurate segmentation of the scar tissues can be helpful to the clinicians for risk prediction and better evaluation of cardiovascular diseases. Our goal is to apply topology data analysis toward machine learning algorithms to confirm the geometry of scar tissue, in addition to gaining better visualization and quantification of the scar tissue present. We have introduced architecture for integrating geometry in the form of topology toward machine learning. Morphological image processing was carried out to define the regions of the endocardial wall. We implemented convolutional neural networks on delayed enhancement cardiac computed tomography images for the recognition of scar tissue. Segmented two-dimensional images were stacked up to build the geometry of the scar area for visualization purposes. Mathematical calculations were executed for the validation of the scar tissue in addition to performing morphological image processing and marking the scar tissue present on the endocardial wall of the left ventricular. We applied convolutional neural network over convolution and pooling the layers with small sizes; we achieved 89.23% accuracy, 91.11% sensitivity, and 87.75% specificity, and found the dissimilarity distance between the normal endocardial tissue distances to be 9.37. This new concept in this study contributes toward a better understanding of scar structure and transmural variation of the endocardial wall of the left ventricular.