AI辅助系统评估双层探测器光谱CT虚拟平扫肺结节检出的价值探究

Evaluation of Pulmonary Nodule Detection by AI Diagnosis System in Dual⁃Layer Detector Spectral CT Virtual Non⁃Contrast Images

目的探讨人工智能(AI)辅助诊断系统在胸部光谱CT虚拟平扫中评估肺结节检出的可行性。方法回顾性搜集55例于光谱CT行胸部平扫及动静脉增强扫描患者的图像资料,包括67个高危结节。将常规平扫(TNC)、重组的动脉期虚拟平扫(VNC⁃A)和静脉期虚拟平扫(VNC⁃V)图像导入AI运算,统计AI检出的肺结节数目、直径以及识别高危结节的形态学特征。经放射科医师鉴别结节的真假阳性,并采用盲法对高危结节形态学特征进行肉眼评估。计算并比较AI在3组图像中检出结节的敏感度、阳性预测值及假阳性率,采用Bland⁃Altman散点图分别比较VNC⁃A、VNC⁃V与TNC中AI识别的共同结节长径的平均差异。同时评估基于AI识别和放射科医师主观评价的肺部高危结节形态学特征的一致性。此外,测量并比较3组图像中胸主动脉、肺动脉干、肩胛肌和胸壁脂肪的CT值、背景噪声、信噪比,以客观评估图像质量。记录各期扫描的辐射剂量。结果3组图像中AI检出肺结节的敏感度差异无统计学意义(P=0.345),而阳性预测值及假阳性率差异有统计学意义(P=0.007、0.002),VNC⁃A的阳性预测值明显更低,假阳性率更高。基于AI或放射科医师评估VNC⁃A、VNC⁃V中高危结节的形态学特征方面,与TNC相当,而以放射科医师识别结果作为金标准时,除AI识别的血管集束征数目明显少于放射科医师(P<0.001),其他形态特征的AI识别性能与放射科医师无统计学差异。Bland⁃Altman散点图分析结果显示,TNC和VNC⁃A、VNC⁃V中共同病灶长径的平均差异分别为0.151 mm、0.057 mm。在图像质量的客观评价中,除脂肪外,VNC图像在评估胸部不同组织展现出与TNC相近的CT值(P值均>0.05),具有较低背景噪声,更高的信噪比。使用VNC图像替代TNC可使患者所受总有效辐射剂量减少31.65%。结论基于AI辅助诊断系统联合光谱CT虚拟平扫技术,更推荐VNC⁃V,其在显著降低辐射剂量的同时,可以保证高质量图像,提供优异的结节检测性能,并基本还原形态学特征。

Objective To explore the feasibility of Artificial Intelligence(AI)assisted diagnostic system in evaluating pulmonary nodules detection in spectral CT virtual non⁃contrast images of the chest.Methods Image data were retrospectively collected from 55 patients who underwent chest conventional plain scan(TNC),and arterial and venous phase enhancement scans,including 67 high⁃risk nodules.Images from the TNC,reconstructed virtual non⁃contrast in the arterial phase(VNC⁃A)and the venous phase(VNC⁃V)were imported into AI for calculations,documenting the number of pulmonary nodules detected by AI,diameters,and the identification morphological features of high⁃risk nodules.Radiologists differentiated true and false positives of nodules and blindly assessed the morphological features of the high⁃risk nodules.The sensitivity,positive predictive value,and false⁃positive rate of nodules detected by AI in the three groups of images were calculated and compared.The Bland⁃Altman plot was used to compare the mean difference in length diameter of the common nodules identified by AI in the VNC⁃A,VNC⁃V,and TNC,respectively.The concordance of morphological features of high⁃risk nodules based on AI detection and subjective assessment by radiologists was also evaluated.In addition,CT values,background noise,and signal⁃to⁃noise ratio of the thoracic aorta,pulmonary artery trunk,scapular muscles,and chest wall fat were measured and compared in the three sets of images at the AW workstation to objectively assess image quality.The radiation dose for each scan period was recorded.Results There was no significant difference in the sensitivity of AI in detecting nodules between the three sets of images(P=0.345),whereas the difference in positive predictive value and false⁃positive rate was statistically significant(P=0.007、0.002),with VNC⁃A showing a significantly lower positive predictive value and a higher false⁃positive rate.The morphological features of high⁃risk nodules in VNC⁃A and VNC⁃V were evaluated based on AI or radiologists alone,which were comparable to TNC.However,when the radiologists'recognition results were taken as the gold standard,the number of Vascular convergence sign identified by AI was significantly less than that of radiologists(P<0.001),and there was no statistical difference in the recognition performance of other morphological features between AI and radiologists(P>0.05).The Bland⁃Altman plot analysis indicated that the mean difference of common nodules length diameter between TNC and VNC⁃A,VNC⁃V were 0.151 mm and 0.057 mm,respectively.In the objective evaluation of image quality,apart from fat,VNC images showed similar CT values to TNC images in evaluating different tissues of the chest(P>0.05),with lower background noise and higher signal⁃to⁃noise ratio.Using VNC images in place of TNC resulted in a 31.65%reduction in the total effective radiation dose to the patient.Conclusion Based on AI⁃assisted diagnosis system combined with spectral CT virtual non⁃contrast technology,VNC in the venous phase is recommended.While significantly reducing radiation dose,it guarantees high⁃quality images,offers superior nodule detection performance,and essentially restores morphological features.