Multi-Task and Federated Learning for Breast and Lung Cancer Screening and Diagnosis: A Survey and Future Research Directions.

backgroundBreast cancer (BrC) and lung cancer (LuC) are two forms of aggressive cancer that affect both men and women worldwide. Recently, multitask learning (MTL) and federated learning (FL) techniques have proven to be efficient in increasing the robustness of deep learning (DL)-based models by performing multiple tasks simultaneously and preserving the confidentiality of medical data.

methodsThis paper presents a survey of MTL and FL methods for BrC and LuC screening and diagnosis using the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) methodology. Comprehensive tables were created to highlight the performances of both MTL models and FL environments.

resultsThe main challenges identified were the lack of hybrid MTL models that combine hard and soft sharing, heterogeneous imaging data, and edge FL systems.

conclusionsFL environments obtain competitive performance compared with centralized MTL models, highlighting their potential to preserve medical data confidentiality without compromising performance. Future research directions could include MTL-based models incorporated in FL environments, hybrid MTL models that combine both hard- and soft-sharing parameter methods, and the use of blockchain techniques to increase the security of FL environments.