On Applications of Deep Learning to Improve Blood Clot Permeability Predictions

Abstract

Despite blood clot heterogeneity being linked to treatment outcomes for stroke, the current stroke standard-of-care is unable to leverage clot heterogeneity as a variable for supporting treatment decisions. In this work, we couple computational fluid dynamics with machine learning to generate quantitative estimates of clot permeability—a key microstructural parameter—for a wide range of clot representations. Specifically, we train (i) multilayer perceptrons based on extracted radiomic features from simulated images and (ii) convolutional neural networks trained directly on these images. These models show that clot permeability can be predicted from simulated images and associated radiomic features, paving the way for more customized, patient-specific treatments.