Leveraging Interpretable AI for Deciphering Signature Histopathologic Patterns.

Area of Science:

• Dermatopathology
• Artificial Intelligence
• Computational Pathology

Background:

• Leukocytoclastic vasculitis (LCV) and microvascular occlusion (MVO) are distinct histopathologic patterns in purpura diagnoses.
• Differentiating LCV and MVO is crucial for accurate dermatologic diagnosis.

Purpose of the Study:

• To evaluate attention-based AI models for enhanced diagnostic accuracy in differentiating LCV and MVO.
• To demonstrate the utility of explainable AI in dermatopathology for identifying subtle histopathologic patterns.

Main Methods:

• Comparison of two attention-based AI models: clustering-constrained-attention multiple-instance learning (CLAM) and attention multiple instance learning (MIL).
• Analysis of whole slide images from 69 LCV and MVO biopsies.
• Generation of attention-based heatmaps for visualizing diagnostic regions.

Main Results:

• The CLAM model achieved superior performance across all evaluated metrics compared to the attention MIL model.
• Attention heatmaps successfully identified critical diagnostic areas, including subtle microvascular occlusion in MVO cases.

Conclusions:

• Attention-based AI models can significantly improve diagnostic accuracy in differentiating LCV and MVO.
• Explainable AI, through heatmaps, provides valuable insights for pathologists, enhancing the identification of histopathologic patterns.