Development and internal validation of a high-resolution computed tomography radiomics and three-dimensional deep learning diagnostic prediction model for preoperative differentiation of minimally invasive and invasive adenocarcinoma in subsolid nodules.
Authors
Affiliations (1)
Affiliations (1)
- Department of Radiology, Jiaxing Hospital of Traditional Chinese Medicine, Jiaxing, China.
Abstract
Preoperative differentiation between minimally invasive adenocarcinoma (MIA) and invasive adenocarcinoma (IAC) in pulmonary subsolid nodules (SSNs) is important for surgical planning, but conventional computed tomography (CT) assessment and single-modality models have limited ability to characterize tumor heterogeneity. This study aimed to develop and internally validate high-resolution CT (HRCT)-based radiomics-deep learning (DL) fusion diagnostic prediction model for differentiating MIA from IAC. This retrospective single-center diagnostic prediction model development and internal validation study included 374 patients with surgically confirmed SSN-associated lung adenocarcinoma between September 2015 and June 2025, including 124 MIA and 250 IAC cases. Patients with single SSNs, preoperative HRCT, adequate image quality, and pathological confirmation were included, whereas those with solid nodules, multiple SSNs, poor-quality images, unclear pathology, or incomplete data were excluded. Three-dimensional (3D) lesions were manually segmented on HRCT. Radiomics features were extracted using PyRadiomics, and DL features were obtained from a pretrained 3D MedicalNet model. Radiomics-only, DL-only, and fusion models were developed using ElasticNet-regularized logistic regression and internally validated by stratified five-fold cross-validation and 1,000 bootstrap resampling. The MIA and IAC groups had comparable age, sex, and smoking history, whereas IAC showed more frequent lobulation, spiculation, pleural indentation, larger nodule size, and higher solid component proportion. In internal validation, the radiomics-only, DL-only, and fusion models achieved area under the curves (AUCs) of 0.833, 0.887, and 0.937, respectively. The fusion model showed the best performance, with a sensitivity of 0.864, specificity of 0.807, and accuracy of 0.845, and demonstrated favorable calibration and decision-curve performance. This study provides preliminary proof-of-concept evidence that integrating HRCT-based radiomics and DL features may improve internal validation performance for differentiating MIA from IAC in SSNs.