Ko Chi-Lun, Lin Shau-Syuan, Huang Cheng-Wen, Chang Yu-Hui, Ko Kuan-Yin, Cheng Mei-Fang, Wang Shan-Ying, Chen Chung-Ming, Wu Yen-Wen
Department of Biomedical Engineering, National Taiwan University, Taipei, Taiwan.
Department of Nuclear Medicine, National Taiwan University Hospital, Taipei, Taiwan.
Eur J Nucl Med Mol Imaging. 2023 Jan;50(2):376-386. doi: 10.1007/s00259-022-05953-z. Epub 2022 Sep 14.
Deep learning (DL) models have been shown to outperform total perfusion deficit (TPD) quantification in predicting obstructive coronary artery disease (CAD) from myocardial perfusion imaging (MPI). However, previously published methods have depended on polar maps, required manual correction, and normal database. In this study, we propose a polar map-free 3D DL algorithm to predict obstructive disease.
We included 1861 subjects who underwent MPI using cadmium-zinc-telluride camera and subsequent coronary angiography. The subjects were divided into parameterization and external validation groups. We implemented a fully automatic algorithm to segment myocardium, perform registration, and apply normalization. We further flattened the image based on spherical coordinate system transformation. The proposed model consisted of a component to predict patent arteries and a component to predict disease in each vessel. The model was cross-validated in the parameterization group, and then further tested using the external validation group. The performance was assessed by area under receiver operating characteristic curves (AUCs) and compared with TPD.
Our algorithm preprocessed all images accurately as confirmed by visual inspection. In patient-based analysis, the AUC of the proposed model was significantly higher than that for stress-TPD (0.84 vs 0.76, p < 0.01). In vessel-based analysis, the proposed model also outperformed regional stress-TPD (AUC = 0.80 vs 0.72, p < 0.01). The addition of quantitative images did not improve the performance.
Our proposed polar map-free 3D DL algorithm to predict obstructive CAD from MPI outperformed TPD and did not require manual correction or a normal database.
深度学习（DL）模型已被证明在从心肌灌注成像（MPI）预测阻塞性冠状动脉疾病（CAD）方面优于总灌注缺损（TPD）量化。然而，先前发表的方法依赖于极坐标图，需要手动校正和正常数据库。在本研究中，我们提出一种无极坐标图的三维DL算法来预测阻塞性疾病。
我们纳入了1861例接受碲化镉锌相机MPI检查及随后冠状动脉造影的受试者。受试者被分为参数化组和外部验证组。我们实施了一种全自动算法来分割心肌、进行配准和应用归一化。我们进一步基于球坐标系变换对图像进行扁平化处理。所提出的模型由一个预测通畅动脉的组件和一个预测每个血管疾病的组件组成。该模型在参数化组中进行交叉验证，然后使用外部验证组进行进一步测试。通过受试者操作特征曲线下面积（AUC）评估性能，并与TPD进行比较。
经目视检查证实，我们的算法准确地预处理了所有图像。在基于患者的分析中，所提出模型的AUC显著高于应激性TPD（0.84对0.76，p<0.01）。在基于血管的分析中，所提出的模型也优于局部应激性TPD（AUC=0.80对0.72，p<0.01）。添加定量图像并未提高性能。
我们提出的用于从MPI预测阻塞性CAD的无极坐标图三维DL算法优于TPD，且不需要手动校正或正常数据库。
