Medical support platform for melanoma analysis and detection based on federated learning

Abstract

Advances in computer science and medicine have led to the emergence of artificial intelligence as a key tool in the medical and scientific fields. Its application in the diagnosis and treatment of diseases, such as cancer, has proven to be fundamental in improving early detection and saving lives. This article presents a proposal based on Deep Learning to develop a model capable of detecting melanomas in the skin from clinical images. The aim is to provide doctors with a tool to support early identification of this type of cancer, considering additional factors such as sun exposure and the patient's skin tone. To optimize diagnostic accuracy and prevent data silos, a collaborative learning technique called Federated Learning is implemented. The FL framework employs a weighted averaging algorithm for model aggregation, allowing locally trained models to contribute to a continuously improving global model without sharing patient data. Experiments show that the proposed federated model achieved an accuracy of 89.1% and a ROC AUC of 0.9251, demonstrating performance comparable to centralized training while preserving privacy. In addition, a web application is presented to manage and process the information efficiently, making it easier for doctors to consult and analyze the results.