基於聯邦式學習框架的血管內超音波影像斑塊分割之深度學習技術

Abstract

深度學習技術已經廣泛地被應用在諸多醫學影像分析上，相較於傳統的醫學影像分析高度仰賴專業醫師及放射師的經驗判斷，深度學習的方法可以提供更為穩定及精確的判斷，因此，深度學習模型可以在臨床治療時有效的協助醫生進行決策。 本研究使用影像分割技術應用於血管內超音波影像 (IVUS) 的斑塊分割，以U-Net模型為基礎，設計一個兩階段的IVUS分割模型用以標示出影像中外彈性膜、流明區域以及斑塊的位置，骰子相似係數分別為0.88、0.87和0.70，分割結果與專業醫師具有一致性。模型提供準確與即時的預測結果，是手術過程中不可或缺的有效分割工具。 為了建立一個泛化能力高的模型，通常會需要搜集更多的資料來訓練模型，但在實務上，基於醫病隱私等緣故，跨機構間的交換資料是有困難的。本研究設計了一套聯邦式學習框架，使各機構可以在不互相交換資料的情況下，共同在分散式的資料集上訓練模型。為了提升參與者投入合作的意願，本研究同時提出了一套公平的商業模式作為激勵機制。與最先進的演算法相比，所提出的演算法在運算和通訊成本、安全性及公平性等面向展現了優越性。透過建構一個周全的聯邦式學習框架，希冀增加本研究應用於實務問題的可能性。 本研究的主要貢獻歸納如下: (1) 提出一套高效能的斑塊分割系統，能夠準確且即時的辨識 IVUS 影像中斑塊的位置。 (2) 設計一個周全的聯邦式學習框架，相較於傳統演算法在運算及通訊成本、安全性及公平性等面向具有優勢。

Deep learning technologies have been widely used in medical image analysis. Compared with traditional methods that rely heavily on the experience of professional physicians and radiologists, deep learning methods can provide more stable and accurate judgments. Therefore, deep learning models can effectively assist doctors in making decisions during clinical treatment. This thesis adopted the image segmentation technique for the task of plaque segmentation from intravascular ultrasound (IVUS) images. Based on the U-Net model, a two-stage IVUS segmentation model was designed to annotate the external elastic membrane (EEM), lumen area, and plaque burden in IVUS images with a dice similarity coefficient of 0.88, 0.87, and 0.70, respectively. The segmentation results showed close agreement with human experts. The proposed model provides precise and real-time segmentation masks and is an efficient segmentation tool essential during surgery. In general, collecting large volumes of training data can make the model has a better generalization capacity. However, exchanging data across institutions would be challenging in practice for patient privacy and other reasons. In this study, a federated learning framework is proposed. Institutions can collaboratively train models on distributed data sets without any data exchange. In order to enhance the willingness of participants to invest, this study also proposed a fair business model as an incentive mechanism. Compared with the state-of-the-art algorithm, the proposed algorithm exhibited superiority in computational and communication costs, security, and fairness. It is expected to put this study into practice by fulfilling a comprehensive federated learning framework. The main contributions of this study are summarized as follows: (1) Propose a high-performance plaque burden segmentation system that can accurately and instantly identify the location of plaque burden in IVUS images. (2) Develop a comprehensive federated learning framework that outperforms the traditional algorithm regarding computational and communication costs, security, and fairness.