Federated Learning for Collaborative Dental Diagnosis Across Hospitals Without Compromising Privacy
Annotatsiya
This paper examines how Federated Learning (FL) can be used to conduct dental diagnosis in many hospitals without threatening the privacy of patients. The approach is based on a Cross-Silo FL design where involved hospitals receive training on a common deep learning framework locally on their individual dental images (e.g., panoramic X-rays, CBCT). The model does not transfer raw data, but instead the update of the data is consolidated at a central server and sensitive data about the patient is kept confidential. In order to solve the problem of data heterogeneity and unbalances among hospitals, the research uses Federated Proximal (FedProx) a technique that becomes used to stabilize learning and enhances the convergence of the models when the data distributions are uneven. The experiment is conducted with the help of the TensorFlow Federated (TFF) framework of the model creation and coordination, which promotes efficient distributed learning and safe aggregation. The findings indicate that FL as implemented in dental diagnostics could facilitate cross-hospital training and keep privacy intact and the accuracy of diagnosis increased.