基于3D ResUNet的颅内动脉瘤自动测量模型的检测效能分析

Efficiency of 3D ResUNet-based automatic measurement model of intracranial aneurysm diameter

目的评价基于3D ResUNet的深度学习(DL)颅内动脉瘤自动测量模型对颅内不同长径动脉瘤的检测效能。方法回顾性分析老年颅内动脉瘤患者156例,根据CT血管成像(CTA)检查颅内动脉瘤长径分为小长径组(长径<5 mm)69例和大长径组(长径≥5 mm)87例,采用数字减影血管造影术(DSA)检查患者的头颅影像资料,由1位高年资影像医师独立盲法阅片,且由另1位高年资医师审核结果。采用DL模型3D ResUNet卷积神经网络实现载瘤血管定位、动脉瘤检出和瘤体长径测量。以DSA为标准,评估DL模型和影像医师检查的准确性、特异性、敏感性、检出率及测量性能。结果DSA、DL模型与影像医师分别对小长径组和大长径组瘤体长径测量值比较,无统计学差异[3.82(3.97,4.01)mm vs 3.76(3.74,3.94)mm vs 3.87(3.72,4.01)mm,8.45(7.82,9.76)mm vs 9.04(8.93,9.97)mm vs 9.12(8.07,10.16)mm,P>0.05]。以DSA检查为标准,影像医师检查瘤体长径的准确性、敏感性、特异性均略高于DL模型(P>0.05)。在小长径组中,DL模型与DSA检查,DL模型与高年资影像医师间的组内相关系数分别为0.705(95%CI:0.683~0.714)、0.929(95%CI:0.876~0.960);在大长径组中,DL模型与DSA检查,DL模型与高年资影像医师间的组内相关系数分别为0.817(95%CI:0.804~0.857)、0.940(95%CI:0.894~0.966)。结论基于DL自动测量模型在颅内动脉瘤的检出敏感性和长径测量效能上接近高年资影像医师的判别水平,可辅助临床医师进行颅内动脉瘤载瘤血管节段检出和瘤体长径测量。

Objective To assess the efficiency of 3D ResUNet deep learning(DL)-based automatic measurement model of intracranial aneurysm diameter in detecting intracranial aneurysms with a different diameter.Methods One hundred and fifty-six elderly intracranial aneurysm patients were divided into intracranial aneurysm diameter<5 mm group(n=69)and intracranial aneurysm diameter≥5 mm group(n=87).Their DSA data were read by a senior imaging technician independently and checked by another senior imaging technician.The located intracranial aneurysms were detected and their diameters were measured according to the 3D ResUNet DL convolutional neural network.The accuracy,specificity,sensitivity and efficiency of 3D ResUNet DLbased automatic measurement model of intracranial aneurysm diameter in detecting intracranial aneurysms were assessed according to their DSA parameters.Results No significant difference was detected in intracranial aneurysm diameter measured by DSA,3D ResUNet DL-based automatic measurement model and imaging technicians between the two groups[3.82(3.97,4.01)mm vs 3.76(3.74,3.94)mm vs 3.87(3.72,4.01)mm,8.45(7.82,9.76)mm vs 9.04(8.93,9.97)mm vs 9.12(8.07,10.16)mm,P>0.05]and in accuracy,sensitivity,specificity measured by 3D ResUNet DL-based automatic measurement model and DSA between the two groups(P>0.05).The correlation coefficient of intracranial aneurysm diameter detected by 3D ResUNet DL-based automatic measurement model and senior imaging technicians was 0.705(95%CI:0.683-0.714)and 0.929(95%CI:0.876-0.960)respectively in intracranial aneurysm diameter<5 mm group and that detected by 3D ResUNet DL-based automatic measurement model and senior imaging technicians was 0.817(95%CI:0.804-0.857)and 0.940(95%CI:0.894-0.966)respectively in intracranial aneurysm diameter≥5 mm group.Conclusion The sensitivity and efficiency of 3D ResUNet DL-based automatic measurement model and senior imaging technicians are similar in detecting intracranial aneurysms and can thus assist the clinicians in detecting intracranial aneurysms and measuring their diameters.