Medical Image Object Detection Algorithm for Privacy-Preserving Federated Learning

Federated learning technology is a new type of multi-agency collaborative training model paradigm,which is widely used in many fields,among which model parameter privacy protection is a key issue.Aiming at the task of comprehensive lesion detection in CT images,privacy protected federated learning algorithm is proposed.A loosely coupled client-server architecture is deployed firstly.Then an improved RetinaNet detector is used on each client-side,and contextual convolution and backward attention mechanisms are introduced.Finally,federated training process is finished.Each client uses a local update strategy,adopts an adaptive training epoch,and adds the restriction term to the local objective function.The server uses an adaptive gradi-ent clipping strategy and a Gaussian noise differential privacy algorithm to update the global model parameters.Ablation analysis on the DeepLesion dataset demonstrates the importance of each part of the algorithm.Com-parative experimental results show that the improved RetinaNet detector effectively improves the detection performance of multi-scale lesions.Compared to the centralized data training model paradigm,the perform-ance of the model obtained by federated learning is slightly lower(with mAP of 75.33%and 72.80%respec-tively),but the training time is shortened by nearly 38%.The proposed algorithm effectively achieves a good trade-off between privacy protection,communication efficiency,and model performance.