The palmoplantar pustular psoriasis (PPP) disease produces numerous pustules and brown spots on the hands and feet of patients. Treatment requires a precise and continuous severity assessment of the disease. Dermatologists typically perform this task by studying the pustules visually: counting them, assessing their sizes, textures, locations and immediate surroundings. This process is a time intensive task, prone to inevitable inter-individual variations, which is an issue for the determination of treatment efficacy.
In this work we developed a deep-learning system to automatically evaluate the surface and quantity of pustules and brown spots from PPP patient photos. We trained a segmentation model based on the U-Net architecture using 121 PPP patients pictures. Each pictures was divided in smaller patches over which we applied several data augmentation techniques such as rotation, translation, cropping.
The performance of the system were evaluated over 30 new unseen patients pictures divided in 819 patches. We compared the model predictions with the dermatologists’ annotations and computed an intraclass correlation of 0.97 (CI95 0.97-0.98) for counts and 0.93 (CI95 0.92-0.94) for surface.
The model was able to successfully quantify pustules and brown spots lesions comparably with dermatologists. The produced metrics could be used in existing scoring systems such as PPPASI making them simpler to evaluate and more precise. In practice, the model could be deployed via smartphone apps enabling lesions evaluation by physician extenders before patients visit their dermatologists. The model could be integrated in both clinical pathways as well as prospective clinical trials to evaluate novel drugs for PPP.

ESDRLB003 - Automated Severity Grading of Palmoplantar Pustular Psoriasis

50th Annual Meeting of the European Society for Dermatological Research. VIRTUAL MEETING