基于MRI的人工智能模型对高级别胶质瘤和脑转移瘤分类诊断的效能及验证OA
Diagnostic performance and validation of an MRI-based artificial intelligence model for differentiating high-grade glioma from brain metastasis
目的 构建基于脑部CE-T1WI和T2 FLAIR的人工智能(AI)模型,验证并评估其对脑高级别胶质瘤和转移瘤的诊断效能及临床应用价值.方法 回顾性收集有手术病理结果的272例脑肿瘤病人的脑MRI影像资料,其中高级别胶质瘤143例,脑转移瘤129例.将4名不同年资[低年资医生2名(阅片年限2、3年)、中年资医生2名(阅片年限5、8年)]放射科医生随机均分为2组,试验组2名医生在AI辅助下诊断,对照组2名医生单独诊断.2组医生对全部病人的MRI影像进行2轮交叉阅片,中间间隔3周的洗脱期.以病理结果作为诊断金标准,应用DBMH分析法比较2组医生诊断效能,计算受试者操作特征曲线下面积(AUC)、敏感度、特异度和准确度,并对比AI模型、中低年资医生亚组的诊断效能.结果 AI模型分类诊断的AUC为0.975(95%CI:0.956～0.993),敏感度为97.20%、特异度为97.67%,准确度为97.43%.试验组、对照组对脑高级别胶质瘤和转移瘤分类诊断的AUC值分别为0.934(95%CI:0.909～0.958)、0.707(95%CI:0.645～0.768),敏感度分别为98.08%、83.22%,特异度分别为69.19%、52.13%,准确度分别为84.38%、68.47%.低、中年资医生试验组诊断效能均高于对照组.结论 基于神经网络构建的AI模型在脑高级别胶质瘤和转移瘤的分类诊断中具有良好的效能,可提升放射科医生的诊断准确率,为临床诊治决策提供参考依据.
Objective To develop an artificial intelligence(AI)model based on contrast-enhanced T1-weighted(CE-T1WI)and T2 FLAIR MRI,and to validate and evaluate its diagnostic performance and clinical value in differentiating high-grade glioma(HGG)from brain metastasis.Methods A total of 272 patients with brain tumors confirmed by surgical pathology were retrospectively enrolled,including 143 cases of HGG and 129 cases of brain metastasis.Four radiologists with different levels of experience[two junior(2-3 years)and two mid-level(5-8 years)]were randomly assigned to two groups,AI-assisted test group with 2 radiologists+AI,and a non-AI control group with 2 radiologists only.All radiologists independently interpreted the MRI images of all patients in two rounds using a crossover design,including an AI-assisted test group and a non-AI control group,with a 3-week washout period between readings.Using pathology as the reference standard,diagnostic performance was compared using the DBMH method.The area under the receiver operating characteristic curve(AUC),sensitivity,specificity,and accuracy were calculated,and the performance of the AI model was compared with that of junior and mid-level radiologists.Results The AI model achieved a significantly higher AUC of 0.975(95%CI:0.956-0.993)for classification,with a sensitivity of 97.20%,specificity of 97.67%,and accuracy of 97.43%.For differentiating HGG from brain metastasis,the AUCs of the test group and control group were 0.934(95%CI:0.909-0.958)and 0.707(95%CI:0.645-0.768),respectively.Sensitivity was 98.08%versus 83.22%,specificity was 69.19%versus 52.13%,and accuracy was 84.38%versus 68.47%,respectively.Diagnostic performance in the AI-assisted test group was superior to that in the control group for both junior and mid-level radiologists.Conclusion The neural network-based AI model demonstrates excellent diagnostic performance in differentiating high-grade glioma from brain metastasis.It can improve the diagnostic accuracy of radiologists and provide valuable support for clinical diagnostic and therapeutic decision-making.
郑慧;赵倩茹;方慧;吕锟;张丹妮;曹鑫;鲍奕仿;耿道颖
东南大学附属徐州市中心医院影像科,徐州 221009复旦大学生物医学工程与技术创新学院复旦大学附属华山医院放射科复旦大学附属华山医院放射科复旦大学附属华山医院放射科复旦大学附属华山医院放射科复旦大学附属华山医院放射科复旦大学附属华山医院放射科
医药卫生
脑转移瘤高级别胶质瘤磁共振成像人工智能
Brain malignant tumorHigh grade gliomaMagnetic resonance imagingArtificial intelligence
《国际医学放射学杂志》 2026 (1)
39-43,89,6
国家自然科学基金项目(82372048)上海市科学技术委员会项目(22TS1400900,23S31904100,24SF1904201)
评论