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

Mycosis fungoides (MF), the most common cutaneous T-cell lymphoma, is challenging to diagnose in its early stages because clinical and histopathologic features often overlap with benign inflammatory dermatoses (BIDs), leading to misclassification and delayed treatment.

To evaluate whether a self-supervised AI system can support diagnostic decision-making among MF and common inflammatory mimics.

This retrospective two-center study included patients with confirmed diagnoses of MF or BIDs. A self-supervised multimodal system integrating whole-slide histopathology and routine clinical variables was developed to perform multiclass differential classification across MF and common BIDs. Cases were partitioned into training, internal validation, and independent external validation cohorts at the patient level. Clinical utility was assessed in a reader study in which dermatopathologists reviewed cases with and without system assistance.

Across 786 WSIs from 532 patients, the multimodal model achieved macro-AUCs greater than 0.85 in both validation sets, with macro-balanced accuracy of 0.837 (95% CI, 0.778-0.897; n=106) internally and 0.762 (95% CI, 0.724-0.802; n=260) externally, representing improved performance over the unimodal histopathology model. In the reader study, AI assistance was associated with improved macro-balanced accuracy among both junior (0.778 to 0.818) and senior (0.805 to 0.859) dermatopathologists, accompanied by consistent improvements in macro-averaged sensitivity and specificity across all diagnostic categories. Interpretability analyses generated heatmaps that were generally consistent with recognized histopathologic features of both MF and BIDs, aligning with dermatopathologist interpretation.

This multimodal, self-supervised system may support dermatopathologists by providing interpretable, probability-based guidance for the classification of MF and its common inflammatory mimics.