能谱CT虚拟平扫在肺结节AI辅助诊断系统预测亚实性结节恶性概率中的应用

The application of virtual non-contrast images of energy spectrum CT combined with AI-assisted diagnosis system in predicting the malignant probability of sub-solid pulmonary nodules

目的:通过与能谱平扫(TNC)图像对比,探究能谱CT虚拟平扫(VNC)图像在AI肺结节辅助诊断系统预测亚实性结节恶性概率中的效能表现。方法:本研究共纳入86例因肺内亚实性结节而行手术切除的患者,其中男26例、女60例,年龄(61.33±11.66)岁。按病理组织学结果将结节分为3组:A组为浸润前病变;B组为微浸润腺癌;C组为浸润性腺癌。将患者术前TNC和VNC图像上传至AI肺结节辅助诊断系统进行结节检测并记录结节的恶性概率预测值、体积及CT值,进一步行三组间结节数值的非参数检验(Kruskal-Wallis H检验)及每组TNC与VNC图像配对样本的非参数检验(Wilcoxon检验)。结果:入组病例中共切除88个亚实性结节,其中A组、B组和C组分别有27个、28个及33个结节。在TNC和VNC图像中3组结节均可被AI系统检出。利用TNC图像时,AI系统对A组、B组和C组中结节恶性概率的预测值分别为74.60%±19.76%、89.97%±8.55%和94.25%±7.04%;在利用VNC图像时,对三组中恶性概率预测值分别为70.01%±23.43%、88.20%±10.35%和94.51%±5.17%;2种图像上三组间预测值的差异均有统计学意义(P<0.001)。在TNC和VNC图像上,三组间结节的CT值及体积的差异亦有统计学意义(P<0.05)。每组结节在TNC和VNC图像上恶性概率预测值之间的差异无统计学意义(P>0.05)。结论:利用肺结节AI辅助诊断系统预测亚实性结节的恶性概率预测值时,VNC图像与TNC图像的预测效能相似。

Objective:To evaluate the performance of visual non-contrast(VNC)images combined with the deep learning(DL)-based artificial intelligence(AI)-assisted pulmonary nodule diagnosis system in predicting the malignant probability of pulmonary sub-solid nodules by comparing with true non-contrast(TNC)images.Methods:A total of 86 patients with mean age of(61.33±11.66)years who underwent surgical resection of pulmonary sub-solid nodules were enrolled in this study,including 26 men and 60 women.The sub-solid nodules were divided into 3 groups according to the histopathological results.Pre-invasive lesions were classified as group A,and minimally invasive adenocarcinoma and invasive adenocarcinoma as group B and group C,respectively.DL-based AI-assisted diagnosis system was then utilized to detect sub-solid nodules on the preoperative VNC and TNC images.The malignant probability predictive value,CT value,and volume of the nodules were recorded.Kruskal-Wallis H-test was used to examine the differences in TNC or VNC-derived CT values,nodule volume,and the predicted malignant probability among these three groups.Then paired-samples Wil-coxon test was performed to analyze the differences of these measurements and predicting values between VNC and TNC for each group.Results:A total of 88 sub-solid nodules were resected in the enrolled patients,in which 27,28,and 33 cases were included in group A,B and C,respectively.All the sub-solid nodules were detected by the AI-assisted pulmonary nodule diagnosis system on either VNC or TNC images.The predicting malignant probability of sub-solid nodules were 74.60%±19.76%,89.97%±8.55%and 94.25%±7.04%respectively on TNC images,while 70.01%±23.43%,88.20%±10.35%and 94.51%±5.17%on VNC images for group A,B and C,respectively.Statistically significant differences were observed in both TNC and VNC images-derived CT value,nodule volume and the predicted malignant probability among these three groups.Of special note was no significant difference of predicted malignant probability in each group between TNC and VNC images.Conclusion:VNC images can replace TNC images when using AI-assisted pulmonary nodule diagnosis system to predict the malignant probability of pulmonary sub-solid nodules.