近年來，心血管疾病一直都是台灣前三大死亡原因，在全球更是最主要的死亡原因。如果病人的心臟鈣化指數高的話，就有極高度的心血管疾病的風險。對醫生來說，要從大量的心臟血管的CT醫學影像中標註出鈣化斑塊是一項頗耗費時間與精力的工作。因此，本論文希望能夠將醫學影像處理技術和深度學習做結合，來開發出一套心臟血管鈣化偵測系統，以便減輕醫師的工作量以及減少因疲勞而升高的誤判率。 本論文所開發出心臟血管鈣化偵測系統包含四個模組：（1）心臟血管擷取模組：利用影像處理技術，從心臟血管的CT醫學影像中擷取出冠狀動脈血管的部分、（2）鈣化斑塊偵測模組：使用YOLOv3模型並搭配影像處理技術，偵測並且標記出鈣化斑塊的發生位置、（3）鈣化斑塊血管分段定位模組：使用CNN模型判斷鈣化斑塊的位置是在血管的上中下的哪個位置、（4）編輯工具模組：可以搭配上述的三個模組，在讀取心臟血管的CT醫學影像後，自動生成一張心臟鈣化評估報告圖，透過此編輯工具，可讓醫師直接增修圖上的鈣化斑塊的位置與數量，藉此減輕醫師的工作量。 實驗結果顯示，本系統的心臟血管擷取模組可成功擷取90.32%的血管影像，鈣化斑塊偵測的結果，F1 score達到0.8，平均偵測到的鈣化斑塊正確率達到90%，鈣化斑塊血管位置準確率則有81%，整體準確率大於八成。因此，本系統具備一定程度之可用性。 ;In recent years, cardiovascular disease has been the top three causes of death in Taiwan, and it is the leading cause of death in the world. If the patient′s heart calcification index is high, there is a high risk of cardiovascular disease. For doctors, labeling calcified plaques from a large number of CT medical images of coronary arteries is a time-consuming and energetic task. Therefore, this thesis hopes to combine medical image processing technology with deep learning to develop a coronary artery calcification detection system to reduce the workload of physicians and reduce the false positive rate due to fatigue. This thesis developed a coronary artery calcification detection system, which consists of four modules: (1) the coronary artery capture module: using image processing technology to extract the vascular part of the coronary artery from a CT medical image of the coronary artery, (2) the calcified plaque detection module: using the YOLOv3 model with image processing technology to detect and mark the location of calcified plaques, (3) the calcified plaque vascular segmentation positioning module: using a CNN model to decide which position of the calcified plaque is in the upper, middle, and lower parts of the coronary artery, (4) the editing tool module: via the use of above-mentioned three modules, it can automatically generate a coronary artery calcification evaluation report after reading the CT medical image of the heart. The editing tool allows the physician to directly modify the position and the number of the calcified plaques on the report in order to reduce the workload of the physician. The experimental results are as follows. The coronary artery capture module can successfully segment the coronary artery from a CT medical image of the heart with 90.32% correct rate. The F1 score of the calcified plaque detection module reaches 0.8. The average classification accuracy and positioning accuracy rate of the calcified plaque classification and coronary artery positioning module is 90% and 81%, respectively. Overall accuracy of the coronary artery calcification detection system is greater than 80%. Therefore, the system has a certain degree of usability.
