-
心血管疾病严重威胁着全球人类的健康,据报道,2017年全球范围内约1700万人病死于心血管疾病,而其中有3/4在发展中国家[1]。我国心血管疾病患者目前有2.9亿,因此对于心血管疾病的早诊断和防治是刻不容缓的[2]。随着影像技术的迅猛发展,冠状动脉CT血管造影(coronary computed tomography angiography,CCTA)已经成为冠状动脉疾病危险因素的临床评估标准,并且已应用于冠状动脉支架植入患者的随访中,以评估植入支架的通畅性及判断是否有相关并发症[3]。但是,CCTA对心血管疾病的诊断严重依赖于人工图像后处理质量以及诊断医师的经验,尤其我国的医疗资源严重紧缺,对所有心血管疾病患者均行CCTA临床评估是难以实现的。
人工智能(artificial intelligence,AI)通俗来讲就是模仿人类思维,是具有“认知”功能的机器,其通过“学习”来“解决问题”。AI作为21世纪三大尖端技术之一,在众多领域均获得了广泛的应用,并取得了丰硕的成果。AI在医疗领域可用于计算患者给药剂量、肿瘤药物的选择、高危患者的监测,甚至手术的实施,且CT图像报告解读更是AI的优势所在[4]。将AI应用于CCTA图像后处理以及诊断报告的书写,势必能够缓解我国紧张的医疗资源,但是冠状动脉AI对于CCTA图像处理、诊断报告的准确性以及是否能够代替临床医师,目前尚不明确。本研究从冠状动脉AI与人工后处理、诊断报告等方面分别进行对比,以此评估冠状动脉AI在CCTA的应用前景。
-
在64例患者中,5例患者AI冠状动脉重建失败,余59例患者重建图像合格,其中男性36例、女性23例,年龄(62.75±14.32)岁。
冠状动脉AI图像后处理耗时约3 min,而人工图像后处理时间约为20~30 min,AI后处理耗时约为人工耗时的10%,并且冠状动脉AI后处理的合格率为92.2%(59/64)。如图1所示,冠状动脉AI图像后处理能够对各支血管进行自动命名。AI后处理(图1中D~F)与人工后处理(图1中A~C)相比,冠状动脉图像分支更多、更长、管壁更光滑、细节小分支显示更全面;在冠状动脉拉直图像中,发现冠状动脉AI处理的图像(图1中H)较人工处理图像(图1中G)的血管更加清晰,并且能够自动识别冠状动脉狭窄。此外,本研究还发现冠状动脉AI也能够自动识别植入的冠状动脉支架,这都暗示着冠状动脉AI在图像后处理中的应用价值。
图 1 CT增强扫描后冠状动脉AI与人工处理图像的对比
Figure 1. The comparison between artificial intelligence post-processing image and artificial post-processing image of coronary artery after CT enhanced scan
虽然冠状动脉AI在图像后处理上表现优异,但本研究同时发现,冠状动脉AI图像后处理有不足之处。如表1所示,冠状动脉AI图像后处理的合格率为92.2%(59/64),原始图像质量为1分及2分的病例共有5例,AI后处理全部为不合格;而图像质量达到3分时,AI图像后处理合格率达到100%。AI后处理图像(图2中B)与人工后处理图像(图2中A)相比,AI分割遗漏导致右冠状动脉血管截断、命名错误,无法全面显示血管,其后处理失败的原因之一为钙化斑块周围伪影导致管腔被掩盖(图2中C),而人工后处理可以通过工作站手动添加使截断血管“再生”。这也就体现出冠状动脉AI图像后处理严重依赖原始图像质量的弊端。
项目 图像质量等级(分) 5 4 3 2 1 患者(例) 45 10 4 3 2 患者比例(%) 70.31 15.62 6.25 4.69 3.12 人工智能后处理的合格率(%) 100 100 100 0 0 注:表中,CT:计算机体层摄影术 表 1 64例患者CT增强扫描后冠状动脉的原始图像质量 评分及人工智能后处理的合格率
Table 1. The original image quality score of 64 patients and the corresponding qualified rate of coronary artificial intelligence post-processing after CT enhanced scan
-
冠状动脉AI图像的诊断报告在图像重建后即可完成(<1 min),而人工出具一份冠状动脉的诊断报告通常需要耗时15 min左右。与AI后处理一样,AI在报告的诊断上同样高效。如表2所示,冠状动脉AI斑块检出的灵敏度达93.3%,与人工报告的灵敏度(92.0%)相当。本研究中,人工报告对斑块的检出不存在假阳性结果,其特异度达100%,而冠状动脉AI对其检测的特异度为93.8%(表3)。3支分支血管中,冠状动脉AI对左前降支病变的特异度最低(88.0%)。
斑块
位置人工 人工智能 真阳性(个) 假阴性(个) 灵敏度(%) 真阳性(个) 假阴性(个) 灵敏度(%) 左前降支 37 4 90.2 40 1 97.6 回旋支 11 1 91.7 11 1 91.7 右冠状动脉 21 1 95.4 19 3 86.4 总计 69 6 92.0 70 5 93.3 表 2 人工智能与人工对冠状动脉斑块检出的灵敏度
Table 2. Sensitivity of artificial intelligence and the artificial to detection of coronary plaque
斑块位置 真阴性(个) 假阳性(个) 特异度(%) 左前降支 22 3 88.0 回旋支 45 2 95.7 右冠状动脉 38 2 95.0 总计 105 7 93.8 表 3 人工智能对冠状动脉斑块检出的特异度
Table 3. Specificity of artificial intelligence for detection of coronary plaque
本研究结果发现,合格的后处理图像并不代表正确的AI诊断报告,如图3中A、B所示,虽然冠状动脉AI将心肌桥成功处理,但是诊断报告并未提及。从图3中C、D发现,冠状动脉AI将分布在左前降支较细管腔内的造影剂误报为钙化斑块,其他的还有管壁毛糙造成的非钙化斑块的误报。
人工智能在冠状动脉CT血管成像后处理和诊断报告的初步评估
Artificial intelligence in coronary CT angiography post-processing and preliminary evaluation of diagnostic reports
-
摘要:
目的 探讨人工智能(AI)在冠状动脉CT血管造影(CCTA)的图像后处理和诊断报告中的应用价值。 方法 选取重庆医科大学附属第三医院于2019年4月至7月就诊的64例疑似冠心病患者,其中男性40例、女性24例,年龄(62.16±14.13)岁。所有患者均行CCTA扫描,按照李克特量表评分标准对原始图像质量进行评分,分别进行人工和AI图像后处理,比较二者的用时及合格率、诊断报告的用时及对冠状动脉斑块的诊断效能。 结果 CCTA扫描后,冠状动脉AI图像后处理的时间约3 min,合格率为92.2%(59/64);人工后处理的时间为20~30 min。与人工处理相比,冠状动脉AI后处理的图像中冠状动脉管壁更光滑、小分支显示更全面、血管对比更清晰,并且能自动识别冠状动脉狭窄。冠状动脉AI图像的诊断报告在图像重建后即可完成(<1 min),而人工的诊断报告需15 min左右才能完成。冠状动脉AI与人工对冠状动脉斑块检出的灵敏度几乎一致,分别为93.3%和92.0%;人工诊断报告对斑块检出的特异度达100%,而AI的特异度为93.8%。 结论 冠状动脉AI在图像后处理速度、图像质量及报告诊断的效率方面具有一定优势,有望成为CCTA分析的有效辅助工具。 -
关键词:
- 人工智能 /
- 冠状血管 /
- 计算机体层摄影血管造影术 /
- 图像处理,计算机辅助 /
- 影像诊断
Abstract:Objective To explore the value of coronary artificial intelligence (AI) in the post-processing and diagnosis of coronary CT angiography (CCTA). Methods Sixty-four patients with suspected coronary heart disease who were admitted to Third Affiliated Hospital of Chongqing Medical University from April to July 2019, including 40 males and 24 females, aged (62.16±14.13) years, were randomly selected. All patients underwent coronary CT angiography. The original image quality was scored in accordance with the Likert scoring standard, and artificial and AI image post-processing were carried out. The time, qualified rate, time of the diagnosis report, and diagnostic efficiency of the two were compared. Results The post-processing time of AI images of the coronary arteries was about 3 min, and the time of artificial post-processing was 20−30 min after CCTA. The qualified rate of AI post-processing of the coronary arteries was 92.2% (59/64). Compared with manual processing, the AI images of the coronary arteries after processing were smoother, had more small branches and clearer vessel contrast, and can automatically identify coronary artery stenosis. The diagnosis report of coronary artery AI images was completed immediately after image reconstruction (< 1 min), whereas the artificial diagnosis report was about 15 min. The sensitivity of AI plaque in the coronary artery was almost the same as that of artificial detection (i.e., 93.3% and 92.0%, respectively). The specificity of the artificial diagnosis report was 100% and that of AI was 93.8%. Conclusion Coronary AI has certain advantages in image post-processing speed, image quality and efficiency of reporting diagnosis, and is expected to be an effective auxiliary tool for CCTA analysis. -
表 1 64例患者CT增强扫描后冠状动脉的原始图像质量 评分及人工智能后处理的合格率
Table 1. The original image quality score of 64 patients and the corresponding qualified rate of coronary artificial intelligence post-processing after CT enhanced scan
项目 图像质量等级(分) 5 4 3 2 1 患者(例) 45 10 4 3 2 患者比例(%) 70.31 15.62 6.25 4.69 3.12 人工智能后处理的合格率(%) 100 100 100 0 0 注:表中,CT:计算机体层摄影术 表 2 人工智能与人工对冠状动脉斑块检出的灵敏度
Table 2. Sensitivity of artificial intelligence and the artificial to detection of coronary plaque
斑块
位置人工 人工智能 真阳性(个) 假阴性(个) 灵敏度(%) 真阳性(个) 假阴性(个) 灵敏度(%) 左前降支 37 4 90.2 40 1 97.6 回旋支 11 1 91.7 11 1 91.7 右冠状动脉 21 1 95.4 19 3 86.4 总计 69 6 92.0 70 5 93.3 表 3 人工智能对冠状动脉斑块检出的特异度
Table 3. Specificity of artificial intelligence for detection of coronary plaque
斑块位置 真阴性(个) 假阳性(个) 特异度(%) 左前降支 22 3 88.0 回旋支 45 2 95.7 右冠状动脉 38 2 95.0 总计 105 7 93.8 -
[1] GBD 2017 Causes of Death Collaborators. Global, regional, and national age-sex-specific mortality for 282 causes of death in 195 countries and territories, 1980-2017: a systematic analysis for the Global Burden of Disease Study 2017[J]. Lancet, 2018, 392(10159): 1736−1788. DOI: 10.1016/S0140−6736(18)32203−7. [2] 胡盛寿, 高润霖, 刘力生, 等. 《中国心血管病报告2018》概要[J]. 中国循环杂志, 2019, 34(3): 209−220. DOI: 10.3969/j.issn.1000−3614.2019.03.001.
Hu SS, Gao RL, Liu LS, et al. Summary of the 2018 Report on Cardiovascular Diseases in China[J]. Chin Circ J, 2019, 34(3): 209−220. DOI: 10.3969/j.issn.1000−3614.2019.03.001.[3] Sun ZH, Al Moudi M, Cao Y. CT angiography in the diagnosis of cardiovascular disease: a transformation in cardiovascular CT practice[J]. Quant Imaging Med Surg, 2014, 4(5): 376−396. DOI: 10.3978/j.issn.2223−4292.2014.10.02. [4] Kermany DS, Goldbaum M, Cai WJ, et al. Identifying Medical Diagnoses and Treatable Diseases by Image-Based Deep Learning[J]. Cell, 2018, 172(5): 1122−1131.e9. DOI: 10.1016/j.cell.2018.02.010. [5] Cury RC, Abbara S, Achenbach S, et al. CAD-RADSTM Coronary Artery Disease-Reporting and Data System. An expert consensus document of the Society of Cardiovascular Computed Tomography (SCCT), the American College of Radiology (ACR) and the North American Society for Cardiovascular Imaging (NASCI). Endorsed by the American College of Cardiology[J]. J Cardiovasc Comput Tomogr, 2016, 10(4): 269−281. DOI: 10.1016/j.jcct.2016.04.005. [6] Bahl M, Barzilay R, Yedidia AB, et al. High-Risk Breast Lesions: A Machine Learning Model to Predict Pathologic Upgrade and Reduce Unnecessary Surgical Excision[J]. Radiology, 2018, 286(3): 810−818. DOI: 10.1148/radiol.2017170549. [7] Rajpurkar P, Irvin J, Ball RL, et al. Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists[J/OL]. PLoS Med, 2018, 15(11): e1002686[2019-11-12]. https://journals.plos.org/plosmedicine/article?id=10.1371/journal.pmed.1002686. DOI: 10.1371/journal.pmed.1002686. [8] Obermeyer Z, Emanuel EJ. Predicting the Future- Big Data, Machine Learning, and Clinical Medicine[J]. N Engl J Med, 2016, 375(13): 1216−1219. DOI: 10.1056/NEJMp1606181. [9] Chockley K, Emanuel E. The End of Radiology? Three Threats to the Future Practice of Radiology[J]. J Am Coll Radiol, 2016, 13(12): 1415−1420. DOI: 10.1016/j.jacr.2016.07.010. [10] Jha S, Topol EJ. Adapting to Artificial Intelligence: Radiologists and Pathologists as Information Specialists[J]. JAMA, 2016, 316(22): 2353−2354. DOI: 10.1001/jama.2016.17438. [11] 黄增发, 王翔. 人工智能冠状动脉CT血管成像在冠心病诊断中的应用[J]. 放射学实践, 2018, 33(10): 1017−1021. DOI: 10.13609/j.cnki.1000−0313.2018.10.008.
Huang ZF, Wang X. Artificial intelligence-based coronary computed tomography angiography in the evaluation of coronary artery disease[J]. Radio Pract, 2018, 33(10): 1017−1021. DOI: 10.13609/j.cnki.1000−0313.2018.10.008.[12] Wolterink JM, Leiner T, de Vos BD, et al. Automatic coronary artery calcium scoring in cardiac CT angiography using paired convolutional neural networks[J]. Med Image Anal, 2016, 34: 123−136. DOI: 10.1016/j.media.2016.04.004. [13] Lessmann N, van Ginneken B, Zreik M, et al. Automatic Calcium Scoring in Low-Dose Chest CT Using Deep Neural Networks With Dilated Convolutions[J]. IEEE Trans Med Imaging, 2018, 37(2): 615−625. DOI: 10.1109/TMI.2017.2769839. [14] Slomka PJ, Betancur J, Liang JX, et al. Rationale and design of the REgistry of Fast Myocardial Perfusion Imaging with NExt generation SPECT (REFINE SPECT)[J/OL]. J Nucl Cardiol, 2018[2019-11-12]. https://link.springer.com/article/10.1007%2Fs12350-018-1326-4. DOI: 10.1007/s12350−018−1326−4. [15] Betancur J, Hu LH, Commandeur F, et al. Deep Learning Analysis of Upright-Supine High-Efficiency SPECT Myocardial Perfusion Imaging for Prediction of Obstructive Coronary Artery Disease: A Multicenter Study[J]. J Nucl Med, 2019, 60(5): 664−670. DOI: 10.2967/jnumed.118.213538. [16] Dey D, Slomka PJ, Leeson P, et al. Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art Review[J]. J Am Coll Cardiol, 2019, 73(11): 1317−1335. DOI: 10.1016/j.jacc.2018.12.054.