Multimodal Melanoma Detection with Federated Learning
Bless Lord Y. Agbley, Jianping Li, Amin Ul Haq, Edem Kwedzo Bankas, Sultan Ahmad, Isaac Osei Agyemang, Delanyo Kwame Bensah Kulevome, Waldiodio David Ndiaye, Bernard Cobbinah, Shoistamo Latipova
Abstract
Melanoma disease analysis is increasingly approached using statistical machine learning techniques, including deep learning. These techniques require large sizes of datasets. However, health institutions are inhibited from sharing their patients' data due to concerns regarding the privacy of subjects. This paper presents a methodology that utilizes Federated Learning (FL) in ensuring the preservation of subjects' privacy during training. We fused two modalities: skin lesion images and their corresponding clinical data. The performance of the global federated model was compared with the results of a Centralized Learning (CL) scenario. The FL model is on-par with the CL model with only 0.39% and 0.73% higher F1-Score and Accuracy performances, respectively, obtained by the CL model. Through extended fine-tuning, the performance difference could be further minimized. Moreover, the FL model was 3.27% more sensitive than the CL model, hence correctly classified more positives than the CL model. Our model also obtained competitive performance when compared with other models from literature. The results indicate the capability of federated learning in effectively learning high predictive models while ensuring no training data is shared among the participating clients.