基于人工智能的冠状动脉CT血管成像检测阻塞性冠状动脉狭窄效能的研究

目的 以有创冠状动脉造影（ICA）为参考标准,探讨人工智能(AI)辅助的冠状动脉CT血管成像（CCTA）诊断阻塞性冠状动脉狭窄的效能。 方法 回顾性收集行CCTA检查并于3个月内行ICA检查的50例疑患冠状动脉疾病（CAD）的病人,男34例,女16例,平均年龄（61.8±8.5）岁。AI软件、不同年资医师（低/中/高年资）及AI+不同年资医师分别对入组病人CCTA影像进行后处理并解读。将ICA和CCTA上冠状动脉管腔狭窄≥50%定义为阻塞性冠状动脉狭窄。采用Agatston积分法测量病人的钙化积分值,并将病人分为低钙化组（钙化积分<100）和高钙化组（钙化积分≥100）。采用独立样本t检验对AI、医师及AI+医师的图像后处理和解读时间进行两两比较。以ICA为参考标准,分析AI在不同研究水平和高/低钙化组的诊断价值,并比较AI、不同年资医师和AI+不同年资医师的诊断敏感度、特异度、阳性预测值、阴性预测值、准确度及受试者操作特征（ROC）曲线下面积（AUC）。采用Pearson卡方检验或Fisher精确概率检验比较组间差异,采用DeLong检验比较AUC。 结果 50例病人共分析195支血管424个节段。AI和AI+医师的平均后处理和解读时间均低于单独医师诊断的时间（均P<0.05）,AI的时间较低/中/高年资医师分别减少了80%、76.8%和75%;AI+低/中/高年资医师较单独医师分别减少了67%、64%、57.9%。在病人、血管及节段水平,AI诊断阻塞性冠状动脉狭窄的敏感度分别为93.7%、83.1%、67.7%,特异度为50.0%、89.0%、91.0%,准确度为92%、86.7%、85.6%,阳性预测值为97.8%、83.1%、69.8%,阴性预测值为25%、89.0%、90.2%,AUC为0.87、0.89、0.83;在血管及节段水平,AI对低钙化组的特异度高于高钙化组（均P<0.05）。在血管水平,AI诊断的AUC值均低于中/高年资医师（均P<0.05）;其余研究水平,AI与其他不同年资医师诊断的AUC值差异均无统计学意义（均P>0.05）。3种研究水平下,AI+低/中/高年资医师诊断的AUC值与单独不同年资医师诊断的AUC值差异均无统计学意义（均P>0.05）。 结论 AI辅助的CCTA诊断阻塞性冠状动脉狭窄具有较好的诊断效能,且明显缩短后处理时间,可能成为临床医师诊断阻塞性冠状动脉狭窄的有效辅助工具。

Diagnostic performance of artificial intelligence based coronary CT angiography in detecting obstructive coronary artery disease

Objective To investigate the diagnostic performance of artificial intelligence (AI) based coronary CT angiography (CCTA) in detecting obstructive coronary artery stenosis with invasive coronary angiography (ICA) as reference standard. Methods A retrospective analysis was performed on 50 patients with suspected coronary artery disease (CAD) who underwent CCTA examination and ICA examination within 3-month interval, including 34 males and 16 females, with an average age of (61.8±8.5) years. AI software, cardiovascular radiologists with different experiences (low/intermediate/high experience), and AI+radiologists with different experiences independently performed image post-processing and interpretation of CCTA. Luminal stenosis ≥ 50% was defined as obstructive coronary artery stenosis in both ICA and CCTA. The calcium scores on a per-patient were measured by Agatston integral method, they were divided into two groups: low calcification group (calcium scores<100) and high calcification group (calcium scores≥100). Independent sample t test was used to compare the difference in post-processing time between AI and radiologists, and between AI+radiologists and radiologists. The diagnostic sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and area under the reciever operating characteristic curve(AUC) of AI were calculated for low/high calcification groups and low/intermediate/high experienced doctors with ICA as reference standard. The diagnostic performances were further compared among the AI, radiologists, and AI+radiologists. The Pearson chi-square test, Fisher’s exact test, and DeLong test were used to compare the AUC differences in performances between groups when appropriate. Results A total of 195 vessels and 424 segments were analyzed in 50 patients. The average post-processing and interpretation times of AI and AI+ radiologists were shorter than that of independent radiologists (all P<0.05). The average post-processing and interpretation times of AI was reduced by 80%, 76.8%, and 75% compared with low/intermediate/high experienced radiologists, respectively. The times of AI+low/ intermediate/high experienced radiologists were reduced by 67%, 64%, and 57.9% compared with independent low/intermediate/high experienced radiologists, respectively. On a per-patient, per-vessel and per-segment basis, with ICA as reference method, the AI overall diagnostic sensitivities for detecting obstructive coronary artery stenosis were 93.7%, 83.1%, 67.7%, the specificities were 50.0%, 89.0%, 91.0%, the accuracies were 92%, 86.7%, 85.6%, the positive predictive values were 97.8%, 83.1%, 69.8%, the negative predictive values were 25%, 89.0%, 90.2%, with the corresponding AUCs of 0.87, 0.89, 0.83, respectively. On a per-vessel and per-segment basis, the specificity of the low calcification group was higher than that of the high calcification group (all P<0.05). On a per-vessel level, the AUC of AI was lower than that of intermediate/high experienced radiologists (both P<0.05). On other study levels, there was no significant difference in AUCs between AI and other different experienced radiologists (all P>0.05). Comparison of AUCs between AI+low/intermediate/high experienced radiologists and low/intermediate/high radiologists alone, there were no significant differences(all P<0.05). Conclusions AI based CCTA has a good diagnostic performance in detecting coronary stenosis with shorter postprocessing time, which can become an effective tool in helping clinicians to detecting coronary stenosis.