深度卷積網絡超音波檢測系統：多部位檢測與臨床應用

Abstract

超音波作為即時且非侵入式的診斷方法，在臨床上眾多科別中被廣泛利用，然而超音波影像的判讀並不容易。在以往，實習醫生需要經過大量的練習與學習來獲取經驗，才能準確的判讀超音波影像。 為此，我們團隊在先前開發了基於YOLO家族模型深度神經網絡的超音波檢測系統，分別在前臂及肩膀單一視角的數據集上做的訓練，同時透過半監督式學習與幀間關係，減少實際訓練資料的人工標記數量與檢測顯示結果的改善。 在本研究中，我們提出了在更新的YOLOv8模型上應用的檢測系統，延續半監督式學習並稍做改善，同時將訓練集擴大到前臂與肩膀的多個視角，並透過腳本運行先檢測身體部位後再進行個別部位的細部檢測，建立起多部位、多視角檢測系統，最後更完成可視化的使用者介面針對臨床上的使用需求做進一步的改善。 根據研究結果，YOLOv8模型能夠有效的分辨前臂與肩膀的超音波影像，在個別部位的模型裡，前臂與肩膀的多視角模型都得到良好的結果，進行半監督式學習後效能也有所提升。檢測系統實際使用時，在訓練集上能夠反應出與訓練結果相符的成像，甚至因為幀間關係的引入影像上能有更好的召回率。 本研究初步建立起多部位、多視角的超音波檢測系統，雖然僅有前臂與肩膀作為數據集，但作為一個方法學，在更多的訓練資料加入後，便能夠如法炮製的進行模型的訓練，並從而使此系統變得更加完整。

Ultrasound is utilized wildly across numerous clinical disciplines as a real-time and non-invasive diagnostic method. However, interpreting ultrasound images is not easy. Traditionally, medical interns had to undergo extensive practice and learn to gain experience in accurately interpreting ultrasound images. To tackle this challenge, our team developed an ultrasound image annotation system utilizing the YOLO model family, specifically trained on forearm and shoulder datasets captured from a single perspective. Through semi-supervised learning and the “Relation of Frames,” we reduce the manual labels needed for training data and improve detection outcomes. In this study, I propose a detection system applied to the YOLOv8 model, continuing with semi-supervised learning with some improvements. The training set was expanded to include multi-views of the forearm and shoulder. The system runs scripts to detect body parts and then performs detailed inspections of individual parts, establishing a multi-site, multi-view detection system. Finally, a user interface was developed based on clinical usage needs for further improvements. According to the research findings, the YOLOv8 model can effectively differentiate forearm and shoulder ultrasound images. The models for both body parts in multi-views showed good results individually, improving performance after semi-supervised learning. In practical use, the detection system's performance on the training set reflects the results, with the introduction of the “Relation of Frames” leading to better recall rates in the imaging. This study initially establishes a multi-site, multi-view ultrasound detection system. Although it only includes the forearm and shoulder as datasets, as a methodology, with more training data added, the model training can be replicated, making the system more comprehensive.