Mycosis fungoides: a double look at 10x and 20x helps AI distinguish this cutaneous lymphoma from benign dermatoses

Out of 6,267 images from 463 patients, a set of convolutional networks achieved 89% sensitivity for screening, and a model based on clinical data achieved 96.6% accuracy for the stage.

Mycosis fungoides is a rare form of cutaneous T-cell lymphoma, often confused in its early stages with inflammatory dermatoses that are not serious, which delays diagnosis. A team on arXiv offers a diagnostic framework that analyzes histopathology slides at two magnifications at the same time: 20x for cell detail, 10x for the organization of the tissue around it.

Two tools for two steps

The first part combines two convolutional neural networks in late fusion, one per magnification. Trained and evaluated on an expanded set of 6,267 images (4,306 of mycosis fungoides, 1,961 of other conditions) from 463 patients, it achieves 83.58% accuracy and 89.13% sensitivity in separating the disease from other dermatoses. The authors note that the best results come from a priority given to cellular detail, placed in its context.

The second part, a random forest fed with 16 clinical variables, is used to specify the stage of confirmed cases (plaques or tumors): 96.6% accuracy and 93.8% sensitivity. The authors see this as a basis for a decision support tool in dermatology, which remains to be validated under clinical conditions.