Abstract
Leukocytoclastic vasculitis (LCV) and microvascular occlusion (MVO) represent distinct histopathologic patterns underlying dermatologic diagnoses of purpura. This study explores the potential of interpretable artificial intelligence (AI) models to enhance diagnostic accuracy and provide insights in differentiating these conditions. We compared the performance of two interpretable AI models in analyzing whole slide images of LCV and MVO cases. The models were trained and evaluated using whole slide images from a cohort of 69 clinically and histopathologically confirmed biopsies. The best-performing model achieved an accuracy of 0.895 (95% CI: 0.889 - 0.902, p-value < 0.001). Attention-based heatmaps effectively highlighted key diagnostic regions for both conditions. For MVO, the model identified expected areas of vascular occlusion in the dermis and subcutaneous fat. Notably, an unexpected finding was the consistent highlighting of subtle areas of occlusion in the superficial papillary dermis adjacent to the epidermis, a feature not typically emphasized in traditional histopathologic evaluation of MVO. In LCV cases, heatmaps showed prominent highlighting in areas of neutrophilic inflammation and vessel damage, with higher attention weights compared to MVO cases. The model's ability to detect subtle features demonstrates its potential to augment human expertise in histopathologic analysis, providing diagnostic insights and uncovering novel histopathologic features. By offering a new perspective on tissue analysis, these models could enhance our understanding of disease processes and contribute to diagnostic criteria in dermatopathology.