Pathology-Based Ischemic Stroke Etiology Classification via Clot Composition Guided Multiple Instance Learning

Mara Pleasure, Ekaterina Redekop, Jennifer S Polson, Haoyue Zhang, Naoki Kaneko, William Speier, Corey W Arnold; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2023, pp. 2674-2683

Accurate identification of the etiology of acute ischemic stroke is crucial to prevent secondary strokes. With the growing utilization of endovascular thrombectomy as a treatment for acute ischemic stroke, there has been an increased interest in analyzing the removed clot tissue, opening possibilities for developing histology-based automated diagnostic methods for clot etiology prediction. In this paper, we propose an automated pipeline for Pathology-Based Ischemic Stroke Etiology Classification via Clot Composition Guided Multiple Instance Learning, that leverages the heterogeneity of the main clot components, specifically red blood cells (RBCs) and fibrin, to predict clot etiology solely using digital pathology data. We combine a publicly available dataset from 11 different medical Centers with a private dataset from the University of California, Los Angeles (UCLA). We train the model using center-wise leave-one-out cross-validation to create a model that can generalize to all 12 Centers. Additionally, we compare three different self-supervised methods for embedding the histology data from whole slide images (WSIs) for a downstream task and found that combining two feature sets results in the best performance for our dataset. Our solution resulted in 0.762+-0.141 AUC and 0.869+-0.139 PRAUC for differentiation between large-artery atherosclerosis and cardioembolic clot etiology. These results hint at the potential to construct a generalizable model for clot etiology prediction. Such a model could assist in treatment planning, thereby helping reduce the likelihood of recurrent strokes.