重建层厚(1/2/3mm)对人工智能检测肺结节效能的影响

Effect of Reconstruction Thickness(1/2/3 mm)on Effectiveness of Artificial Intelligence in Detecting Pulmonary Nodules

目的比较人工智能(Artificial Intelligence,AI)辅助诊断系统对胸部CT不同层厚图像的肺结节检测敏感度、准确性、假阳性率,评估适合AI检测的最佳层厚。方法采用西门子二代双源CT前瞻性采集190例胸部CT并分别进行1/2/3 mm层厚骨算法重建。应用基于深度学习的计算机辅助诊断系统对三组图像分别进行结节自动检测。将AI自动检测结果与金标准对比,分别计算并比较各组图像人工智能检测的敏感度和假阳性率。结果1、2、3 mm层厚图像自动检出肺结节分别为1403、853、1077个,正确结节数分别为1103、607、401个,敏感度分别为0.833±0.195、0.473±0.258、0.301±0.239,假阳性率分别为1.58/CT、1.29/CT、3.56/CT。结论2 mm层厚胸部CT图像在检测>4 mm和亚实性结节时敏感度并不弱于1 mm层厚图像,而假阳性率更是远低于1 mm图像。综合评估,1 mm层厚的整体检测效能,优于其它层厚。

Objective To compare the sensitivity,accuracy and false positive rate of artificial intelligence(AI)auxiliary diagnosis system in detecting pulmonary nodules in different slice thickness images of chest CT,and evaluate the optimal slice thickness suitable for AI detection.Methods A total of 190 chest CT cases were prospectively collected by Siemens dual source CT and reconstructed with 1/2/3 mm thick bone algorithm.The computer-aided diagnosis system based on deep learning was used to automatically detect nodules in three groups of images.The AI automatic detection results were compared with the gold standard,and the sensitivity and false positive rate of each group were calculated and compared.Results The number of lung nodules automatically detected by the 1 mm,2 mm,3 mm layer thickness image were 1403,853,1077,and the correct number of nodules were 1103,607,401,the sensitivity was 0.833±0.195,0.473±0.258,0.301±0.239,and the false positive rate was 1.58/CT,1.29/CT,3.56/CT,respectively.Conclusion The sensitivity of chest CT images with slice thickness of 2 mm in detecting subsolid nodules with size larger than 4 mm was not weaker than that of chest CT images with slice thickness of 1 mm,and the false positive rate was lower than that of chest CT images with slice thickness of 1 mm.Comprehensive evaluation,the overall detection efficiency of 1 mm layer thickness is better than other layer thicknesses.