人工智能技术对低年资医师在磁共振颅脑病变筛查中的辅助价值

The Auxiliary Value of Artificial Intelligence Technique for Young Radiologists in MRI Brain Lesion Screening

目的:探讨低年资医师结合人工智能(AI)辅助诊断系统在磁共振影像筛查颅脑病变中的价值。方法:回顾性分析复旦大学附属华山医院颅脑磁共振影像数据296例,其中正常数据94例,异常数据202例(肿瘤数据99例,非肿瘤数据103例)。由两位经验丰富的影像科医师和一名具有15年以上磁共振影像诊断经验的医师阅片并结合病理结果制定金标准。再选取两位低年资医师,分别对296例数据进行独立阅片,经过4周洗脱期后在AI辅助下重新阅片并记录阅片结果,统计并比较前后两次阅片中正常检出正确率,异常检出正确率,以及对肿瘤患者和非肿瘤患者的检出率。结果:对正常数据,两次阅片的正确率分别为79.79%和80.85%,差异不具有统计学意义(χ^(2)=0.0336,P=0.8544)。对异常数据,两次阅片的正确率分别为64.36%和80.20%,差异具有统计学意义(χ^(2)=12.6497,P=0.0003)。对肿瘤数据,两次阅片的检出率分别为66.67%和81.82%,差异具有统计学意义(χ^(2)=5.9423,P=0.0147)。对非肿瘤数据,两次阅片的检出率分别为62.14%和78.64%,差异具有统计学意义(χ^(2)=6.7308,P=0.0094)。结论:AI对低年资医师基于磁共振影像筛查颅脑病变有较好的辅助价值。

Purpose:To explore the auxiliary diagnostic value of artificial intelligence(AI)technique in MRI diagnosis of brain lesions for the young radiologists.Methods:A total of 296 cases of brain MRI imaging data from Huashan Hospital affiliated to Fudan University were retrospectively analyzed,including 94 cases of normal data,202 cases of abnormal data(99 cases of neoplasm data and 103 cases of non-neoplasm data).Two experienced radiologists and one radiologist with more than 15 years'experience in MRI diagnosis reviewed the images and established the gold standard in conjunction with the pathological results.Then,two young radiologists independently diagnosed the 296 cases,and the results were recorded.After four weeks of washout period,the doctors re-diagnosed with the assistance of AI.The accuracy of normal detection,abnormal detection,and the detection rates of neoplasm patients and non-neoplasm patients were analyzed and compared between the two reviewing.Results:For normal cases,the accuracy rates of independent diagnosis and AI-assisted diagnosis were 79.79% and 80.85%,respectively,with no statistically significant difference(χ^(2)=0.0336,P=0.8544).For abnormal cases,the accuracy rates were 64.36% and 80.20%,respectively,with statistically significant difference(χ^(2)=12.6497,P=0.0003).For neoplasm cases,the detection rates of independent diagnosis and AI-assisted diagnosis were 66.67% and 81.82%,respectively,and the difference was with statistical significance(χ^(2)=5.9423,P=0.0147).For nonneoplasm cases,the detection rates were 62.14%and 78.64%,respectively,also with statistically significant difference(χ^(2)=6.7308,P=0.0094).Conclusions:AI showed significant auxiliary value for young radiologists to screen for brain lesions based on MRI.