Automated Segmentation of Clot Components from Martius Scarlet Blue Stained Digital Histology using Deep Learning

Ischemic stroke (IS) remains a leading cause of morbidity and mortality, affecting approximately 795,000 people annually in the US, with $87 \%$ of these strokes being ischemic. Mechanical thrombectomy, a transformative treatment for IS, also provides an invaluable source of clot tissue for research, enabling insights into clot composition, structure, and potential recurrence risk. Histological analysis, particularly with Martius Scarlet Blue (MSB) staining, is the current standard for analyzing these tissues, allowing for clear differentiation of platelets and fibrin. However, the lack of fully automated analysis tools for MSB-stained slides limits consistency and scalability in clot analysis. In this study, we propose an automated approach to segment clot components on MSB-stained slides using Deeplabv3+ with a ResNet18 backbone. Using a heterogeneous dataset from STRIP AI, containing samples from stroke patients across 11 centers, we trained and evaluated the model on patch-and image-level segmentations, achieving a mean intersection-over-union (IoU) of $0.352 \pm 0.140$ and $0.450 \pm 0.114$, respectively, in the validation set. Misclassification trends included RBCs and platelets frequently identified as fibrin ($31.4 \%$ and $16.2 \%$) and platelets or blue artifacts labeled as background or WBCs. Qualitative analysis corroborated these patterns, revealing misclassifications at RBC-fibrin interfaces and misinterpretations of artifacts as WBCs, often due to staining inconsistencies. These findings emphasize the need for improved stain homogenization and expanded datasets to enhance model performance. This automated method offers a consistent, rapid, and scalable approach for analyzing clot tissue, potentially advancing biological insights and improving IS treatment strategies.