Federated Learning-Driven Skin Disease Diagnosis: Ensuring Data Privacy in Distributed Healthcare Systems

Abstract

Skin diseases such as melanoma, eczema, acne, psoriasis and basal cell carcinoma are common around the world that may lead to severe complications if diagnoses at last stage. Deep learning, in particular Convolutional Neural Networks (CNNs), has demonstrated promise for automated skin disease classification from dermatoscopic images. However, legacy central training involves pooling of images in a server which raises the related concerns of privacy, security and compliance (ex: HIPAA or GDPR). To that end, this thesis explores Federated Learning (FL) as a privacy-preserving solution for multi-class skin disease classification. A mixed skin disease dataset consisting of dermatology images collected from real hospital sources and publicly available online repositories is used in this study. A CNN-based classifier is developed and trained under two different settings: centralized training and federated training. The dataset is partitioned into two simulated healthcare clients representing hospitals with non-IID distributions: Client 1 contains melanoma and acne cases, while Client 2 contains psoriasis, eczema, and basal cell carcinoma. Standard preprocessing operations such as conversion, resizing, and data expansion are performed to help the model generalize better. To aggregate data from different nodes while avoiding end-to-end transmission of raw patient data, FLWR utilizes the Federated Average(FedAvg) method of updating models. Model quality is assessed using accuracy, loss, precision, recall, F1-score, and confusion matrices. Comparative experiments show that Federated Learning is capable of achieving results that are competitive with Centralized Learning, while still retaining full data localization and improved privacy protection. The study also points out practical challenges such as being non-IID, overhead of communication and convergence stability in federated settings. Overall, this study lends empirical support to the practical feasibility of Federated Learning in melanoma imaging. It also demonstrates how FLWR-based federated systems can enable privacy-conserving collaboration across the health institutions in skin-disease diagnosis--an approach where promising possibilities lie ahead for real-life medical AI (which needs to be put into practice).