內視鏡超音波影片中標誌性血管的分類與分割

Abstract

在過去的一個世紀中，超音波成像已確立其作為關鍵診斷影像技術的地位。在近四十年內，內視鏡超音波（Endoscopic Ultrasound, EUS）進一步擴展了超音波成像的能力與應用彈性。然而，在執行 EUS 時，一項主要挑戰在於於探頭移動的情況下，如何快速且容易地定位可靠的解剖學標誌。本研究將深度學習應用於 EUS 探頭偏離標準化位置時，在 EUS 影片中偵測標誌性血管的問題。我們提出的方法能提升單一輸出分割模型的中心線準確度，使其表現可媲美多任務分割與中心線模型，並具有以下貢獻：第一，我們提出 C-Dice，作為centerline Dice 損失函數的延伸；第二，我們建立一個新的 EUS 影片資料集，包含標誌性血管的分類與分割標註；第三，我們探討在 EUS 分割訓練中，將 EUS 分類作為預適應步驟所帶來的效益。結果代表著在創建用於 EUS 的動態術中位置估計系統方面邁出了有希望的一步。

Over the past century, ultrasound imaging has cemented itself as a key diagnostic imaging modality. In the past 40 years, endoscopic ultrasound (EUS) has further expanded the power and versatility of ultrasound imaging. A key challenge in performing EUS is quickly and easily locating reliable anatomical landmarks, especially when the probe is moving. This work applies deep learning to the challenge of finding landmark blood vessels in EUS videos when the EUS probe is away from standardized positions. Our proposed method improves the centerline accuracy of a single-output segmentation model to those comparable to a multi-task segmentation and centerline model, making the following contributions: First, we propose C-Dice, an expansion on the centerline Dice loss function. Second, we introduce a novel dataset of EUS videos with classification and segmentation labels for landmark blood vessels. Third, we investigate the benefits of EUS classification as a pre-adaptation step in EUS segmentation training. These results represent a promising step forward towards creating a dynamic, intra-operative position estimation system for EUS.