由於超音波能窺視體內,所以其是一項有價值的儀器能進行診斷。此外,肌肉骨骼超音波是最早進入照護端的領域且已被廣泛運用。因為其他成像工具無法在門診被臨床醫師使用,以至於肌肉骨骼超音波對於臨床醫師來說是一項可行的替代方案。然而,許多身體部位有著複雜的解剖結構和可變的組織紋理,使得初學者難以找到適合的導師,亦難從只有少量圖片的現有教科書中來學習。我們實驗室成員的前作中已經應用了基於深度學習的演算法來檢測肘關節周圍的正中神經,但仍有改進的空間。因此,在此論文,我們使用YOLOv5演算法來實現肌肉骨骼超音波精準且即時的檢測來減緩上述的問題。實驗結果表明,YOLOv5在交叉驗證中mAP達到80%以上,在測試集中mAP達到89%,這意味著和前作使用的YOLOv3相比,在交叉驗證中mAP進步了2%,而測試集中進步了7%。除此之外,YOLOv5的FPS也達到了即時的基準。
Ultrasound is a valuable instrument to make a medical diagnosis since it can peer inside the body. Moreover, musculoskeletal ultrasound is one of the earliest to enter the point-of-care field and has been widely used. As other imaging tools cannot be operated by the clinicians in the out-patient clinic, musculoskeletal ultrasound is a practicable alternative for the clinicians. However, because many body parts come with complex anatomical structures and changeable tissue texture, it is hard for beginners to find suitable tutors and to learn from existing textbooks with few pictures. Previous work by our laborious member has applied a deep learning-based algorithm to detect median nerve around an elbow joint, but it leaves much to be desired. Therefore, in this thesis, we use YOLOv5 algorithm to achieve accurate and real-time detection for musculoskeletal ultrasound image detection to alleviate the problem as mentioned above. In contrast to the algorithm the previous work applied, YOLOv3, the experimental results have shown that, in mAP50, YOLOv5 achieves over 80% for cross-validation, improving by over 2%, and 89% for testing, improving by 7%. Furthermore, The FPS of YOLOv5 also achieves the real-time benchmark.