Federated learning (FL) provides a decentralized method for model training and effectively allows for adherence to the principles of data privacy, which are crucial in applications related to healthcare. This work looks into the case of FL application to dermatological condition classification using a comprehensive image dataset that encompasses nine different skin diseases. The first approach was based on the Flower framework, which uses standard aggregation-FedAvg. However, it came with a number of problems, mainly around communication and issues associated with non-Independent and Identically Distributed data. In that respect, it was decided to move to a custom, manually simulated approach using Convolutional Neural Network (CNN)s. This alternative approach includes a carefully designed CNN over different client datasets, with synchronized rounds of training and equal local epochs. This manually built FL model resulted in comparatively consistent validation performance with a good generalization ability. Moreover, the class-wise accuracy was state-of-the-art, especially for classes such as Melanoma, Nevus, and Birthmark. This model, therefore, can be put to use in Non-Independent and Identically Distributed (non-IID), is easy to deploy and requires no complex infrastructure, and scales well in privacy-sensitive industries. In this project, manual FL highlights an applicable and robust alternative in deploying Artificial Intelligence (AI) for real-world, medical-related applications.
Krishnan et al. (Mon,) studied this question.