BONE METASTASIS CLASSIFICATION BASED ON FEDERATED LEARNING FOR SPECT WHOLE-BODY BONE SCANNING

Early detection of bone metastasis is important, given its severe impact on patients’ quality of life. A dataset containing 977 annotated whole-body bone scan (WBS) images from diverse sources was constructed. An innovative framework combining federated learning (FL) based on the residual convolutional neural network was adopted, and the rectified linear unit (ReLU) in ResNet was replaced with LeakyReLU to optimize the network. The experiments were carried out from two aspects: FL and residual network. Compared with various models, the results show that the FedAvg model has excellent performance while protecting privacy, and the improved ResNet algorithm has an accuracy exceeding 95% in the bone metastasis classification task. The experiment indicates that this framework can enhance the efficiency and accuracy of bone metastasis diagnosis, protect privacy, and lay a foundation for clinical applications.