基于人工智能深度学习头颈部计算机断层血管造影后处理对狭窄评估的可行性研究

Feasibility study of the post process of CTA on head and neck based on AI deep learning on stenosis assessment

目的:探讨基于人工智能(AI)深度学习头颈部计算机断层血管造影(CTA)后处理对狭窄评估的可行性。方法:选取2022年1月至2023年6月北京中医医院顺义医院行头颈部CTA的108例患者,依据诊断方法的不同将其分为AI组和人工组,每组54例。采用双源CT进行检查,数据分别传输至深睿Dr.Wise头颈CTA智能辅助系统和Siemens Syngo View后处理工作站。图像评价采用5分制,观察比较两组头颈部动脉CTA图像后处理耗时和诊断耗时,以及头颈部动脉CTA图像质量、头颈部动脉狭窄程度诊断结果。结果:AI组图像后处理耗时和诊断耗时分别为(4.09±1.09)min和(3.22±1.23)min,与人工组比较差异均有统计学意义(t=52.315、24.509,P<0.05);两名医师对头颈部动脉各分支图像评分的组内相关系数(ICC)为0.996,AI组颈总动脉评分为(4.77±0.12)分,颈内动脉(4.56±0.13)分,椎动脉(4.55±0.16)分,大脑中动脉(4.78±0.16)分;人工组颈总动脉评分为(3.02±0.12)分,颈内动脉(3.02±0.12)分,椎动脉(3.02±1.14)分,大脑中动脉(3.11±1.09)分;两组头颈部动脉各主要分支图像评分比较,AI组均高于人工组,差异均有统计学意义(t=107.165、94.590、13.812、15.753,P<0.05)。以数字减影血管造影(DSA)结果为“金标准”,纳入的所有患者44段存在狭窄的动脉进行比较,DSA显示AI组轻度狭窄12例,中度狭窄12例,重度狭窄10例;人工组轻度狭窄12例,中度狭窄8例,重度狭窄7例,AI组诊断与“金标准”有较好一致性(Kappa=0.898,P<0.05)。狭窄符合率AI组均高于人工组,AI诊断头颈部动脉狭窄程度的灵敏度(96.80%)和特异度(66.52%)均高于人工组(68.36%、14.75%),特异度差异显著。结论:AI模式头颈部动脉CTA图像后处理耗时及诊断耗时及图像质量均优于人工模式,两者诊断效能一致性良好。

Objective:To explore the feasibility of the post process of computed tomography angiography(CTA)on head and neck based on artificial intelligence(AI)deep learning on stenosis assessment.Methods:A total of 108 patients who underwent CTA on head and neck at Shunyi Hospital of Beijing Traditional Chinese Medicine Hospital from January 2022 to June 2023 were selected,and they were divided into an AI group(54 cases)and a manual group(54 cases)based on different diagnostic methods.Dual source CT was used to conduct examination,and data was transmitted to“Shenrui Dr.Wise Head and Neck CTA Intelligent Assistance System”and“Siemens Syngo View Post Process Workstation”.The image evaluation was conducted on a 5-point scale.The time-consuming of post process and diagnosis of CTA images of the artery of head and neck between two groups were observed and compared,and the quality of CTA images of the artery of head and neck,and the diagnostic results of the degree of the artery stenosis of head and neck also were observed and compared between two groups.Results:The differences of the time-consuming of post process of images and the time-consuming of diagnosis between two groups were all statistically significant(t=52.315,24.509,P<0.05),respectively.The intra-group correlation coefficient(ICC)of each branch of the artery of head and neck was 0.996 between two physicians.In AI group,the score of common carotid artery was(4.77±0.12),and the score of internal carotid artery was(4.56±0.13),and the score of vertebral artery was(4.55±0.16),and the score of middle cerebral artery was(4.78±0.16).In manual group,the score of common carotid artery was(3.02±0.12),and the score of internal carotid artery was(3.02±0.12),and the score of vertebral artery was(3.02±1.14),and the score of middle cerebral artery was(3.11±1.09).The differences of the image scores of each main branch of the artery of head and neck between two groups were significant(t=107.165,94.590,13.812,15.753,P<0.05),respectively,and scores of AI group all higher than these of manual group.The result of digital subtraction angiography(DSA)was used as the"gold standard".A total of 44 stenosis arteries were included for comparison,and the DSA results indicated that there were 12 cases with mild stenosis,12 cases with moderate stenosis and 10 cases with severe stenosis in AI group,and there were 12 cases with mild stenosis,8 cases with moderate stenosis and 7 cases with severe stenosis in manual group.There was a favorable diagnostic consistency in AI group(Kappa=0.898,P<0.05).The compliance rate of each stenosis of AI group was higher than that of manual group.The sensitivity(96.80%)and specificity(66.52%)of AI group were all higher than those(68.36 and 14.75%)of manual group,and the difference of specificity was significant.Conclusion:The time-consuming of post process,the time-consuming of diagnosis and the quality of CTA image of the artery of head and neck in AI mode are better than those in manual mode,and the consistency of diagnostic efficacy between two modes is favorable.