Deep learning-based solid component measuring enabled interpretable prediction of tumor invasiveness for lung adenocarcinoma

Background: The nature of the solid component of subsolid nodules (SSNs) can indicate tumor pathological invasiveness. However, preoperative solid component assessment still lacks a reference standard.Methods: In this retrospective study, an AI algorithm was proposed for measuring the solid components ratio in SSNs, which was used to assess the diameter ratio (1D), area ratio (2D), and volume ratio (3D). The radiologist measured each SSN's consolidation to tumor ratio (CTR) twice, four weeks apart. The area under the receiver -operating characteristic (ROC) curve (AUC) was calculated for each method used to discriminate an Invasive Adenocarcinoma (IA) from a non-IA. The AUC and the time cost of each measurement were compared. Furthermore, we examined the consistency of measurements made by the radiologist on two separate occasions.Results: A total of 379 patients (the primary dataset n = 278, the validation dataset n = 101) were included. In the primary dataset, compared to the manual approach (AUC: 0.697), the AI algorithm (AUC: 0.811) had better predictive performance (P =.0027) in measuring solid components ratio in 3D. Algorithm measurement in 3D had an AUC no inferior to 1D (AUC: 0.806) and 2D (AUC: 0.796). In the validation dataset, the AI 3D method also achieved superior diagnostic performance compared to the radiologist (AUC: 0.803 vs 0.682, P =.046). The two measurements of the CTR in the primary dataset, taken 4 weeks apart, have 7.9 % cases in poor consistency. The measurement time cost by the radiologist is about 60 times that of the AI algorithm (P