Federated Learning for Masked Psoriasis Severity Classification

Psoriasis is a chronic skin disease that has various appearances and severity depending on the patient, and it requires continuous observation of the disease during several months of treatment. It is difficult to track changes in psoriasis severity using a patient's personal device owing to data security issues. Recently, convolutional neural networks (CNN) and federated learning (FL) approaches for data security have shown remarkable performance in vision tasks on medical images. However, in a client environment, disease images acquired from personal devices are unconstrained, and data loss can occur because of various environmental and physical noises. We used masking modeling to overcome data deformation and damage. In addition, we propose a masked attention model to improve the severity classification performance by extracting discriminative severity features from the masked image. As a result, when the masking ratio was set to 0.5, the severity classification of the FL-based masked attention model yielded the best classification performance, with an F1-score of 0.88. Psoriasis severity classification using the proposed method ensured data security and was robustly performed even during data deformation and damage.