用於光學相干斷層掃描之基於深度學習和聯邦學習框架之視網膜層分割技術

Abstract

在本研究中，我們提出了一種輕量級模型FPENet(α)，以FPENet為基底，用於處理專為邊緣設備設計的 OCT 影像中視網膜層分割。視網膜層分割是眼科診斷的重要工具，但其在資源有限的邊緣設備上應用時存在計算成本和精度之間的瓶頸。FedLion(α)在使用 HCMS資料集、NR206資料集及OCT5K資料集進行訓練和測試時，實現了高精度和高效率。該模型經過最佳化，實現了精度和計算成本之間的平衡。FPENet(α)可以有效地捕捉不同尺度的特徵，同時大幅降低計算成本，非常適合部署在如Raspberry Pi等資源有限的邊緣設備上，其輕量化設計使其在計算資源和內存容量方面具有顯著優勢。聯邦學習的部分我們以FedLion為基礎添加了L2正則化與學習率遞減，提出FedLion(α)，有效處理數據非獨立同分布的問題。數據顯示使用FPENet(α)與FedLion(α)進行聯邦學習，相較於原先只使用FPENet(α)，在HCMS資料集平均DICE係數提升了0.7%，在NR206資料集提升了3.75%，在OCT5K資料集提升了9.1%。In this study, we propose a lightweight model, FPENet(α), based on FPENet, designed for retinal layer segmentation in OCT images specifically for edge devices. Retinal layer segmentation is a crucial tool for ophthalmic diagnosis, but it faces the challenge of balancing computational cost and accuracy when applied to resource-constrained edge devices. FedLion(α) achieves high accuracy and efficiency when trained and tested on the HCMS dataset, NR206 dataset, and OCT5K dataset. This model has been optimized to balance precision and computational cost.FPENet(α) can effectively capture multi-scale features while significantly reducing computational cost, making it highly suitable for deployment on resource-limited edge devices such as the Raspberry Pi. Its lightweight design provides significant advantages in terms of computational resources and memory capacity. In the federated learning part, we enhanced FedLion by incorporating L2 regularization and learning rate decay, creating FedLion(α), which effectively addresses the issue of non-IID data.Data shows that using FPENet(α) and FedLion(α) for federated learning, compared to using FPENet(α) alone, increases the average DICE coefficient by 0.7% on the HCMS dataset, by 3.75% on the NR206 dataset, and by 9.1% on the OCT5K dataset.