[Deep learning reconstruction algorithm for coronary CT angiography in assessing obstructive coronary artery disease caused by calcified lesions: the clinical application value].

Objective: To investigate the image quality of coronary CT angiography (CCTA) subjected to deep learning-based reconstruction algorithm (DLR) method and its diagnostic performance for stenosis caused by coronary calcified lesions. Methods: We enrolled 33 consecutive patients with known or suspected coronary artery disease (CAD) who underwent CCTA and subsequently invasive coronary angiography (ICA) within 1 month in the department of radiology, Peking Union Medical College Hospital between February 2020 and February 2021. Among them, there are 26 males and 7 females, age range from 45 to 86 (61.9±9.0) years. The CCTA images were reconstructed with DLR and hybrid iterative reconstruction (HIR). Image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) were calculated on the aorta root, left main artery, proximal left anterior descending, left circumflex, and right coronary artery of the CCTA images and were used to evaluate the objective image quality (IQ). Subjective IQ score was graded using Likert four-point scale (1 for excellent and 4 for poor). The diagnostic performance of obstructive coronary artery disease caused by calcified lesions on CCTA subjected to DLR and HIR methods were evaluated using ICA as the reference standard. Results: A total of 123 lesions in 33 patients were included in the analysis. Image noise of DLR image was significantly lower than that on HIR image(defined as the standard deviation of the attenuation values in the aortic root: 18.12±3.66 vs 24.19±5.71, P<0.001), CNR and SNR of DLR image in the aortic root were higher (CNR:43.83±23.73 vs 26.38±9.69, P<0.001,SNR:26.66±7.83 vs 21.23±8.65, P<0.001). Subjective scores of DLR was better than HIR image (1.12±0.41 vs 1.46±0.60,P<0.001). The sensitivity, specificity and accuracy of DLR and HIR images for diagnosing obstructive coronary artery disease caused by calcified lesions were 100.0%, 77.4%, 78.9% and 100.0%, 63.5%, 65.9%%, respectively. The number of false positive cases on DLR image decreased by 38% compared with HIR. Conclusions: Artificial intelligence based DLR can significantly reduce the image noise and improve the image quality of CCTA. DLR helps to improve the diagnostic performance of CCTA in assessing obstructive coronary artery disease caused by calcified lesions, which may have good clinical application value.目的： 探讨基于深度学习重建算法（DLR）的冠状动脉CT血管成像（CCTA）图像质量和对钙化病变所致冠状动脉狭窄的诊断价值。 方法： 前瞻性纳入2020年2月至2021年2月北京协和医院放射科确诊或拟诊冠心病的33例患者，其中男26例，女7例，年龄45~86（61.9±9.0）岁。所有患者接受CCTA检查并于1个月内进行有创冠状动脉造影（ICA）检查。采用DLR和混合迭代重建算法（HIR）重建CCTA图像。分别在主动脉根部、左主干开口、左前降支近段、左回旋支近段及右冠状动脉近段选取不同的感兴趣区测量两种图像的噪声、信噪比（SNR）、对比噪声比（CNR），并以Likert 4级评分法进行图像质量主观评分（1分，优秀；4分，不能诊断）。以ICA为金标准，计算基于DLR和HIR的CCTA诊断钙化斑块所致冠状动脉血流梗阻性病变的诊断效能。 结果： 共33例患者的123处病变纳入分析。DLR图像的噪声低于HIR图像（定义为主动脉根部CT值的标准差：18.12±3.66比24.19±5.71，P<0.001），CNR和SNR均高于HIR图像（主动脉根部CNR：43.83±23.73比26.38±9.69，P<0.001，SNR：26.66±7.83比21.23±8.65，P<0.001），主观评分优于HIR图像（1.12±0.41比1.46±0.60，P<0.001）。DLR与HIR对于诊断钙化病变所致冠状动脉血流梗阻性病变的灵敏度、特异度和准确度分别为100.0%、77.4%、78.9%和100.0%、63.5%、65.9%。与HIR相比，DLR图像上CCTA的假阳性病例减少38%。 结论： 基于人工智能的DLR重建算法能够显著降低CCTA图像噪声并提高图像质量。DLR有助于提高CCTA对钙化斑块所致冠状动脉血流梗阻性病变的诊断效能，具有良好的临床应用价值。.