冠狀動脈光學同調斷層掃描影像深度學習斑塊診斷模型之建構

Abstract

冠狀動脈疾病(CAD)為目前世界上的主要死亡原因之一，其成因多為動脈粥狀硬化於血管管壁內部堆積，導致血管管徑縮小，影響血流量，當血管狹窄逐漸嚴重，可能導致心肌缺氧甚至心肌梗塞。冠狀動脈介入治療(PCI)為目前普遍之治療方法，PCI時常會使用氣球擴張術搭配血管支架治療血管狹窄，將血管狹窄處撐開並維持其形狀，使管徑擴大讓足夠血流通過，但若血管上有嚴重的鈣化斑塊，因鈣化斑塊為硬性斑塊，會使支架不容易擴張到預期大小，因此於血管支架放置前得到鈣化斑塊的分布與嚴重程度為相當重要的資訊。目前常使用冠狀動脈內光學同調斷層掃描影像(OCT)對血管壁的組成成分進行分析。 由於在一段血管中便會產生上百張OCT影像，導致心臟科醫師在短時間完整判讀之困難，因此本研究提出了基於深度學習的自動化鈣化斑塊分割演算法，以輔助醫師迅速進行判斷。 本研究於建構相關方法時，依據OCT影像的特性進行設計，模型於分割鈣化斑塊時除了影像中的二維資訊之外，同時也會得到相鄰幾張影像提供的資訊，使得模型能夠較精確的分割整個鈣化斑塊。 本研究於鈣化斑塊分割的結果上，鈣化區段為單位之Dice係數於5-Fold交叉驗證中達到0.7688 ± 0.1332，於額外測試集達到0.8102 ± 0.1005；以影像為單位之Dice係數於5-Fold交叉驗證中達到0.7624 ± 0.1970，於額外測試集達到0.7851 ± 0.2031。同時以額外測試集的鈣化斑塊分割結果，計算了針對OCT影像的鈣化分數，鈣化分數的準確度(Accuracy)達到了92%，能夠良好預測鈣化對支架放置的影響。

Coronary artery disease (CAD) is one of the main causes of death in the world. Most of its causes are the accumulation of atherosclerosis inside the vessel wall, resulting in the reduction of vessel diameter and affecting blood flow. When the blood vessel narrows gradually, it may lead to myocardial hypoxia or even myocardial infarction. Coronary interventional therapy (PCI) is a common treatment method at present. In PCI, percutaneous transluminal coronary angioplasty (PTCA) and vascular stent are often used to stretch the narrowed part of the blood vessel and maintain its shape, so that the diameter of the lumen can be enlarged to allow sufficient blood flow to pass through. If there are severe calcified plaques on the vascular stent, which will make it difficult for the stent to expand to the expected size. Because the calcified plaques are hard plaques. Therefore, it is very important to obtain the distribution and severity of calcified plaques before stent placement. At present, intracoronary optical coherence tomography (OCT) is often used to analyze the composition of the vessel wall. Since hundreds of OCT images are generated in a segment of blood vessels, it is difficult for cardiologists to fully interpret them in a short time. Therefore, this study proposes an automatic calcified plaque segmentation algorithm based on deep learning to assist physicians in making judgments. When constructing the relevant methods, this study was designed according to the characteristics of OCT images. When the model segmented calcified plaques, in addition to the two-dimensional information in the images, the information provided by several adjacent images was also obtained, so that the model could more accurate segmentation of the entire calcified plaque. In the results of calcified plaque segmentation in this study, the Dice coefficient of segment as a unit, it reached 0.7688 ± 0.1332 in 5-Fold cross-validation, and reached 0.8102 ± 0.1005 in the additional test set. The Dice coefficient of image as a unit, it reached 0.7624 ± 0.1970 in 5-Fold cross-validation, and reached 0.7851 ± 0.2031 in the additional test set. In this study, the calcified plaque segmentation results of the additional test set were used to calculate the calcium score for OCT images, and the accuracy of the calcium score reached 92%, which can well predict the effect on stent placement.