基於深度學習之動態超音波影像的電腦輔助正中神經即時偵測方法

醫學超音波檢查是現今非常普及且有效的診斷方式。肌肉骨骼超音波因為涉及身體的部位眾多，解剖結構複雜，且組織紋理多變，造成超音波影像辨識難度相當高。在有限的範例圖譜以及需要大量練習才能精熟的情況下，導致初學者面臨學習門檻過高的困境。因此有必要發展一套電腦輔助自動化超音波影像即時標註系統，將超音波影像中不同器官組織自動標示出來，即時呈現在螢幕上，以協助醫師檢查與診斷。本論文選擇正中神經作為主要偵測目標，以及使用目前性能強大的人工智慧深度學習來實現物體偵測技術。由於需要演算法可達到即時偵測的速度，本論文使用YOLOv3作為物體偵測演算法。為確保實驗的公平性與客觀性，實驗過程中使用不同的資料集分割方式與多重交叉驗證。為增加對特定應用領域的偵測效果，在研究方法中加入對影像額外的前處理與後處理。經實驗結果證實，後處理可有效增加偵測流暢度以及提升準確率。根據實驗數據顯示，本研究方法在平均情況下可以同時達到90%以上的準確率以及80以上的影格率。在使用解析度來調整速度與準確率之間的權衡下，準確率最高可達94.19%以及影格率最高可達87.6幀。本論文證實了使用深度學習應用於超音波物體偵測，可同時達到高準確率與即時的偵測速度。

關鍵字

超音波影像；正中神經；影像辨識；深度學習；物體偵測；即時系統；卷積神經網路

並列摘要

Medical ultrasound examination is a very popular and effective diagnostic method nowadays. Musculoskeletal ultrasound involves many parts of the body, complex anatomical structure, and variable tissue texture, which makes it very difficult to identify ultrasound images. With limited sample pictures and a lot of practice to be proficient, beginners face the dilemma of too high learning threshold. Therefore, it is necessary to develop a computer-aided real-time ultrasound image automatic annotation system to automatically label different organs and tissues in the ultrasound image and present it on the screen in real time to assist physicians in examination and diagnosis. This thesis chooses the median nerve as the main detection target, and uses the current powerful artificial intelligence deep learning to realize the object detection technology. Due to the need for the algorithm to achieve real-time detection speed, this thesis uses YOLOv3 as the object detection algorithm. To ensure the fairness and objectivity of the experiment, different dataset partitioning methods and multiple cross-validation were used during the experiment. In order to increase the detection effect for the specific application field, additional preprocessing and postprocessing of the image are added to the research methods. The experimental results have confirmed that postprocessing can effectively increase detection fluency and improve accuracy. According to experimental data, this research method can achieve an accuracy of more than 90% and a frame rate of more than 80 at the same time on average. Using resolution to adjust the tradeoff between speed and accuracy, the accuracy can reach up to 94.19% and the frame rate can reach up to 87.6 frames. This thesis proves that the use of deep learning in ultrasound object detection can achieve both high accuracy and real-time detection speed.

並列關鍵字

ultrasound image ； median nerve ； image recognition ； deep learning ； object detection ； real-time system ； convolutional neural network

參考文獻

[1] "Ultrasound Equipment Market Size, Share and Industry Analysis." Fortune Business Insights. https://www.fortunebusinessinsights.com/industry-reports/ultrasound-equipment-market-100515 (accessed Oct. 23, 2020).

Google Scholar

[2] "Handheld portable ultrasound machine." Butterfly Network. https://www.butterflynetwork.com/ (accessed Oct. 23, 2020).

Google Scholar

[3] T.-G. Wang, W.-S. Chen, Y.-C. Wang, C.-H. Wu, M.-Y. Hsiao, and K.-V. Chang, Musculoskeletal Ultrasound Examination. Taipei, Taiwan: Leader Book, 2014.

Google Scholar

[4] "Arteries and Nerves of Upper Limb: Anterior View Anatomy." pediagenosis.com. https://www.pediagenosis.com/2019/09/arteries-and-nerves-of-upper-limb.html (accessed Sept. 24, 2020).

Google Scholar

[5] S. Liu et al., "Deep Learning in Medical Ultrasound Analysis: A Review," Engineering, vol. 5, no. 2, pp. 261-275, Apr. 2019, doi: 10.1016/j.eng.2018.11.020.

Google Scholar

國際替代計量

基於深度學習之動態超音波影像的電腦輔助正中神經即時偵測方法

未授權

主題瀏覽