Data-Driven Ischemic Stroke Clot Phenotyping from Whole-Slide Histopathology Images

Mechanical thrombectomy (MT) is a predominant treatment modality for acute ischemic stroke. In vitro MT experiments are critical for testing the efficacy of thrombectomy devices, and are traditionally performed on artificial clots fabricated from animal blood. These clots are generally categorized as red blood cell (RBC)-rich or fibrin-platelet aggregate (FP)-rich. Although clinical studies have shown that clots are more complex in structure, with implications for device testing, clot histological architecture has not been quantified. We hypothesized that computational image analysis can be used to quantify clot structure and compute more informative phenotypes defined by diversity in histological patterns rather than percent composition. Brightfield whole-slide images (WSIs) of H&E stained clots from n=68 patients were acquired. Digital image processing techniques were applied to compartmentalize (segment) clot WSIs into RBC-FP regions as well as engineer 204 image features (textural and geometric) from RBC and FP compartments. Unsupervised learning was used to identify five computational clot phenotypes from engineered features. While three phenotypes were distinguishable based on composition alone - RBC-rich, FP-rich, and FP-rich with small RBC regions - two "mixed"phenotypes required more intensive analysis of computed features. More specifically, phenotypes 3 and 5 were distinguishable based on the regional distribution of RBCs and FP - larger, focal RBC regions were observed in cluster 3, whereas smaller, diffuse RBC regions amidst contiguous FP regions were observed in cluster 5. These observations were corroborated by quantitative analyses performed on the features. Quantification of diverse histological patterns in digital clot pathology paves the way for future research investigating how clots of different phenotypes are related to procedural and cognitive outcomes, and how they can be mimicked in in vitro MT testbeds.