用於光學同調斷層掃描之基於深度學習和聯邦學習框架之視網膜積液分割技術

Abstract

在眼科領域，光學相干斷層掃描（OCT）是檢測眼病的關鍵技術。偏鄉資源有限僅能使用輕量化設備，但其計算能力不足，難以支撐較為大型模型的訓練，以及數據缺乏和隱私問題阻礙醫院數據共享。首先針對輕量化設備，基於LEDNet設計了高效的LEDNet(α)模型，通過調整通道、添加Shuffle Attention模塊和Group Normalization。使用成本低廉的樹莓派5進行訓練，適合偏鄉需求，為解決隱私問題，引入聯邦學習，通過上傳本地模型參數聚合全局模型，避免資料直接上傳。本研究提出Krum(α)算法，在客戶端損失函數中添加近端項並考慮模型自適應性，改善淘汰機制，改進基於歐氏距離淘汰惡意模型的Krum算法。最後實驗結果顯示，在AROI、DUKE、UMN和RETOUCH數據集上，AROI積液類別提高了3.4%，DUKE提高了5.9%，UMN提高了2.4%，RETOUCH提高了1.4%。In the field of ophthalmology, Optical Coherence Tomography (OCT) is a key technology for detecting eye diseases. In resource-limited rural areas, only lightweight devices can be used, but it’s computational power is insufficient to support the training of large models. Additionally, data scarcity and privacy issues hinder data sharing between hospitals. First, for lightweight devices, an efficient LEDNet(α) model was designed based on LEDNet by adjusting channels, adding Shuffle Attention modules, and using Group Normalization. Training was conducted using the low-cost Raspberry Pi 5, suitable for rural needs. To solve privacy issues, federated learning was introduced, aggregating a global model by uploading local model parameters, avoiding direct data upload. This study proposes the Krum(α) algorithm, improving the Krum algorithm based on Euclidean distance by adding a proximal term in the client's loss function and considering model adaptability, improving the elimination mechanism. Finally, experimental results showed that on the AROI, DUKE, UMN, and RETOUCH datasets, the AROI exudate category improved by 3.4%, DUKE by 5.9%, UMN by 2.4%, and RETOUCH by 1.4%.