Self-supervised AI system for differentiating mycosis fungoides and benign inflammatory dermatoses

Area of Science:

• Dermatology
• Artificial Intelligence
• Computational Pathology

Background:

• Mycosis fungoides (MF) is the most common cutaneous T-cell lymphoma.
• Early MF diagnosis is challenging due to overlapping features with benign inflammatory dermatoses (BIDs).
• Misclassification leads to delayed treatment for MF patients.

Purpose of the Study:

• To evaluate a self-supervised AI system for supporting MF diagnosis.
• To assess AI's ability to differentiate MF from common inflammatory mimics.
• To integrate histopathology and clinical data for improved diagnostic accuracy.

Main Methods:

• Developed a self-supervised multimodal AI system.
• Integrated whole-slide histopathology and clinical variables.
• Validated the system on training, internal, and external patient cohorts.
• Conducted a reader study with dermatopathologists assessing AI assistance.

Main Results:

• The multimodal AI model achieved macro-AUCs > 0.85.
• AI assistance improved diagnostic accuracy (macro-balanced accuracy) for both junior and senior dermatopathologists.
• AI improved sensitivity and specificity across diagnostic categories.
• Interpretability analyses showed AI-generated heatmaps aligned with histopathologic features.

Conclusions:

• A multimodal, self-supervised AI system can support dermatopathologists.
• The AI provides interpretable, probability-based guidance for MF classification.
• This technology aids in differentiating MF from its mimics, potentially improving patient outcomes.