Artificial intelligence-based density proportion analysis in predicting the invasiveness of neoplastic ground-glass...

A positive correlation has been observed between computed tomography (CT) value and the invasiveness of neoplastic ground-glass nodules (GGNs). However, the traditional mean CT value cannot reflect the density heterogeneity of nodules. This study aimed to explore the value of artificial intelligence (AI)-based density proportion analysis in predicting the invasiveness of neoplastic GGNs.

Between January 2019 and May 2023, a total of 996 (687 in the training cohort and 309 in the validation cohort) neoplastic GGNs [352 adenocarcinomas in situ (AISs), 334 minimally invasive adenocarcinomas (MIAs), and 310 invasive adenocarcinomas (IACs)] in 963 patients were retrospectively analyzed. AI software was used to obtain the density histograms of the nodules, and the proportions of the components in lesions with higher density at different density thresholds were subsequently recorded. The optimal density threshold and the corresponding proportion cutoff value for determining invasive lesions (ILs) (MIAs and IACs) and IACs were respectively explored and validated. Comparison of diagnostic efficacy between AI density parameters and radiological features of GGNs for predicting ILs and IACs was also conducted respectively.

For determining the ILs and IACs, the optimal density thresholds and the cutoff values for the proportion of components with density higher than the threshold were ≥-350 Hounsfield units (HU) and 17.22% [area under the curve (AUC): 0.801; 95% confidence interval (CI): 0.769-0.830; sensitivity: 51.59%; specificity: 93.52%; P<0.001] and ≥-250 HU and 5.64% (AUC: 0.882; 95% CI: 0.855-0.905; sensitivity: 85.65%; specificity: 76.15%; P<0.001) in the training cohort, respectively. Compared with other radiological features, the predictive performance of density proportion (AUC: 0.801 for ILs, 0.882 for IACs) was comparable to that of nodule size (AUC: 0.820 for ILs, 0.869 for IACs) but significantly higher than that of the remaining features. In the validation cohort, the AUCs of these parameters for determining ILs and IACs were 0.814 and 0.885 (each P<0.001), respectively. In combination with these indicators, the AUCs of the morphological features in predicting ILs and IACs increased from 0.794 to 0.849 and from 0.843 to 0.902 (each P<0.001) in the training cohort, respectively.

AI-based density analysis has a potential role in determining the invasiveness of neoplastic GGNs.