應用YOLO深度學習方法即時偵測X光冠狀動脈狹窄

Real-time Detection of Stenosis on X-Ray Coronary Angiography Using YOLO Approach

李祖銘(Tzu-Ming Lee)

指導教授 ： 陳泰賓

義守大學/智慧科技學院/資訊工程學系/碩士(2022年)

摘要

冠狀動脈疾病(Coronary Artery Disease, CAD)連續二十年位居世界十大死因排名第一，臨床中已有多項影像學檢查工具及功能性影像可得知患者是否罹患冠狀動脈疾病，受限於各非侵入性造影與檢查工具都存在一定之偽陰性及偽陽性，面對冠狀動脈疾病伴隨之高風險死亡率，臨床將藉由介入性心導管冠狀動脈造影術作為最終之診斷方式。
本文使用YOLO(You Only Look Once)深度學習物件偵測之方法於X光心臟血管攝影(X-ray Coronary Angiography, XCA)中，即時偵測冠狀動脈狹窄並加以外框標示位置。本研究資料共收集120筆冠狀動脈血管攝影動態影像，經手動轉換為2,708張已標記冠狀動脈狹窄處座標之靜態影像。以隨機亂數分配方式拆分影像資料為訓練集1,624張(60%)、測試集542張(20%)及驗證集542張(20%)，影像分析工具使用MATLAB Computer Vision Toolbox建立YOLO v2狹窄偵測模型，並以ResNet-50與MobileNet V2作為特徵擷取網路並比較各種不同降採樣大小特徵圖之效能及影像即時辨識速度。將經由圖形處理器(Graphics Processing Unit, GPU)訓練完成之ResNet-50與MobileNet V2模型於驗證集影像進行偵測，最佳之模型由ResNet-50網路搭配當中activation_40_relu之特徵層與Batch size 16所得出，平均精確度(mean Average Precision, mAP)為97.9%、平均交聯集面積(Average IoU)為78%、精確度(Precision)為98.7%、召回率(Recall)為99.1%及F1-Score 98.9%；然而，最佳影像辨識速度32 FPS (Frames Per Second)由MobileNet V2網路模型搭配當中之block_6_depthwise_relu特徵層與Batch size 8所得出，且獲得94.2%之平均精確度(mAP)、平均交聯集面積(Average IoU)76.2%、精確度(Precision)為96.2%、召回率(Recall)為98.5%及F1-Score 97.4%。
經一系列研究結果表明，使用YOLO物件偵測演算法為X光冠狀動脈造影狹窄之偵測帶來了高準確度與影像即時辨識之速度，在測試集影像下所有模型平均精確度均在84%以上，並提供最高32 FPS之影像辨識速度，藉此全自動化冠狀動脈血管攝影狹窄即時偵測方法將能替智慧醫療奠定良好之基礎。

關鍵字

心導管 ； 狹窄偵測 ； 卷積神經網路 ； 物件偵測演算法

並列摘要

Coronary Artery Disease (CAD) is the first of the leading causes of death worldwide. X-ray Coronary Angiography (XCA) is one of the most utilized methods to assess CAD and is still considered the gold standard in clinical. This chapter discusses the deep learning method of YOLO (You Only Look Once) in X-ray Coronary Angiography, which can detect coronary artery stenosis in real-time and mark it with an outline.
A total of 120 coronary angiograms were collected in this study, which were manually converted into 2,708 static images with the coordinates of the coronary stenosis marked. The image data is divided into 1,624 (60%) training (60%), 542 (20%) testing, and 542 (20%) validation datasets by random number allocation. The image analysis tool uses MATLAB Computer Vision Toolbox to establish YOLO stenosis detection. The model was tested ResNet-50 and MobileNet V2 were used as feature extraction networks to compare the difference of feature layers with different down-sampling sizes and the image recognition speed. The ResNet-50 and MobileNet V2 coronary artery stenosis detection models trained on the NVIDIA RTX 3080 Ti Graphics Processing Unit (GPU) are used for validation dataset detection.
The evaluation of detection accuracy included mean average precision (mAP), average intersection over union (IoU), precision, recall, and F1-Score were 97.9%, 78%, 98.7%, 99.1%, and 98.9% provided by “activation_40_relu” of the ResNet-50 network. Based on the feature layer and Batch size 16. However, the best image recognition speed of 32 FPS is obtained from the “block_6_depthwise_relu” feature layer and Batch size 8 in the MobileNet V2 network model. The mAP, average IoU, precision, recall, and F1-Score were 94.2%, 76.2%, 96.2%, 98.5%, and 97.4%.
In this automatic stenosis detection method, YOLO brings high accuracy and speed of real-time image identification for XCA stenosis detection. All of the mAP were greater than 84% under the validation dataset. The speed of image recognition was close to 32 FPS. In the future, more effective cases should be involved to obtain a robust and accurate stenosis detection model.

並列關鍵字

Cardiac Catheterization ； Stenosis Detection ； YOLO v2 ； Convolutional Neural Network ； Object Detection Algorithm

參考文獻

[1] Rizvi A, Han D, Danad I, Ó Hartaigh B, Lee JH, Gransar H, Stuijfzand WJ, Roudsari HM, Park MW, Szymonifka J, Chang HJ, Jones EC, Knaapen P, Lin FY, Min JK, Peña JM. Diagnostic Performance of Hybrid Cardiac Imaging Methods for Assessment of Obstructive Coronary Artery Disease Compared With Stand-Alone Coronary Computed Tomography Angiography: A Meta-Analysis. JACC Cardiovasc Imaging. 2018;11(4):589-599.

[2] Çimen S, Gooya A, Grass M, Frangi AF. Reconstruction of coronary arteries from X-ray angiography: A review. Med Image Anal. 2016;32:46-68.

[3] Weerakkody, Y., Luong, D. ACC/AHA classification of coronary lesions. Reference article, Radiopaedia.org.

[4] Molenaar MA, Selder JL, Nicolas J, Claessen BE, Mehran R, Bescós JO, Schuuring MJ, Bouma BJ, Verouden NJ, Chamuleau SAJ. Current State and Future Perspectives of Artificial Intelligence for Automated Coronary Angiography Imaging Analysis in Patients with Ischemic Heart Disease. Curr Cardiol Rep. 2022;24(4):365-376.

[5] Joseph R, Santosh D, Ross G, Ali F. You Only Look Once: Unified, Real-Time Object Detection. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016;779-788.