人工智能辅助诊断系统对肺腺癌亚型的预测价值

 目的 探讨人工智能(artificial intelligence, AI)辅助诊断系统分析磨玻璃结节(ground glass nodule, GGN)的CT定量参数对肺腺癌亚型的预测价值。方法 回顾性选取新疆医科大学第一附属医院昌吉分院2017-01至2021-12经手术病理证实的肺内磨玻璃结节患者97例,根据病灶的浸润程度分为非浸润组(44例)和浸润组(53例)。提取GGN的AI定量参数特征,采用独立样本t检验比较两组间统计学差异;预测GGN病灶侵袭程度用受试者特征曲线(receiver operator characteristic curve, ROC)和二元Logistics回归模型评估AI定量参数的诊断效能。结果 非浸润组和浸润组比较,GGN最大径、体积、平均CT值和实性成分所占比均有差异(P<0.05),CT定量参数的预测价值从高到低依次为实性成分占比、平均CT值、最大径、体积。Logistics回归分析显示实性成分占比(OR=1.262,P<0.05)及平均CT值(OR=1.010,P<0.05)在预测GGN侵袭中的诊断价值较高,可作为独立预测因子,诊断阈值为1.085%和-557.00 HU。结论 AI可通过分析GGN的实性成分占比和平均CT值对肺腺癌亚型做出有效预判。

Value of artificial intelligence analysis of CT quantitative parameters of ground-glass nodules in predicting subtypes of lung adenocarcinoma

Objective To explore the predictive value of the CT quantitative parameters of ground-glass nodules (GGN) by artificial intelligence (AI) assisted diagnosis system for lung adenocarcinoma subtypes.Methods A total of 97 cases of GGN confirmed by surgery and pathology were retrospectively analyzed, and they were divided into non-invasive group and invasive group according to the degree of infiltration of the lesions. The characteristics of AI quantitative parameters were extracted, and the independent sample T test was used to compare the statistical differences between the two groups. Receiver operator characteristic curve (ROC) and binary logistic regression model were used to predict the degree of invasion of GGN lesions to evaluate the diagnostic performance of AI quantitative parameters.Results The maximum diameter GGN, volume, average CT value and proportion of solid components in the non-invasive group and the invasive group were different between the two groups (P<0.05). The prediction value of CT quantitative parameters from high to low was: solid component proportion, average CT value, maximum diameter, and volume. Logistic regression analysis showed that the proportion of real components (OR=1.262, P<0.05) and mean CT value (OR=1.010, P<0.05) had high diagnostic value in predicting GGN invasion and could be used as independent predictors, diagnostic thresholds 1.085% and -557.00 HU.Conclusions AI can effectively predict lung adenocarcinoma subtypes by analyzing the proportion of solid components and average CT value of GGN.