近來，隨著深度神經網路技術的快速發展，其在醫療影像領域的應用也日益增多，其中肺結節分割模型訓練就是其中之一，但礙於醫療影像牽涉到個人隱私、合法性，無法彼此共享交流，偏鄉地區醫院的數據量較小，可能導致模型性能在訓練時無法達到最佳化，因此在這樣的前提下採用聯邦學習架構結合本地模型做訓練，會是最適合的選擇。 聯邦學習是一種新穎的機器學習方法，可以達到實現分散式學習的同時，也維護資料安全性。聯邦學習訓練中，將由伺服器端發送初始化模型給各參與聯邦的客戶端做本地訓練，且各個客戶端使用獨立的本地數據，彼此不共享隱私數據，僅藉由回傳模型訓練權重至伺服器端聚合，更新後的模型權重再回傳給客戶端做訓練，使模型能學習不同數據的多樣性，來提高整體的性能及可靠性。 本篇論文採用Flower作為模擬環境，並假設兩間不同地理位置的醫院，彼此數據分佈不均，藉由聯邦學習架構所帶來的數據多樣性，來優化最終分割的準確度。 ;Recent advancements in deep neural network technologies have significantly increased their applications in medical imaging. Nonetheless, the sensitive nature of medical data and legal constraints prevent data sharing, particularly in rural areas where hospitals have limited data availability. This limitation can hinder the optimization of model training. Under these circumstances, federated learning provides an optimal solution by enabling local model training without data exchange, thereby maintaining data privacy. Federated learning is a novel machine learning method that facilitates distributed learning while maintaining data security. In this process, a server sends an initial model to federated clients for local training. Each client uses their independent data without sharing private information. They then return their model parameters to the server for aggregation. The updated parameters are redistributed to the clients for further training, enabling the model to learn from diverse data, thus enhancing overall performance and reliability. The paper adopts Flower as the simulation environment and assumes two hospitals in different geographical locations, with unevenly distributed data between them. By leveraging the data diversity brought by the federated learning framework, the aim is to optimize the final segmentation accuracy .
