Prediction of infiltration degree of ground-glass nodules using a fusion of CT radiomics and deep learning

This study explores a novel pre- and post-fusion strategy for assessing the infiltration degree of pulmonary ground-glass nodules (GGNs). The infiltration degree was defined according to the pathological progression along the adenocarcinoma spectrum, ranging from atypical adenomatous hyperplasia and adenocarcinoma in situ to minimally invasive adenocarcinoma and invasive adenocarcinoma. A retrospective analysis was conducted on 517 GGNs patients, collecting clinical data, CT imaging parameters, and CT scans. The dataset was randomly divided into training and testing cohorts at a ratio of 7:3. Clinical and imaging features with P < 0.05 were selected using the chi-square test and Mann-Whitney U test to construct a logistic regression-based clinical imaging model. Radiomic features were extracted via PyRadiomics, and deep learning features were obtained using the ResNet-101 framework. Feature selection involved observer consistency tests, Mann-Whitney U tests, and LASSO regression. Radiomics and deep learning features were respectively input into 11 machine learning classifiers to build individual models, which were subsequently integrated through a post-fusion strategy to construct the radiomics and deep learning model. Clinical, radiomic, and deep learning features were integrated using pre-fusion, followed by post-fusion of classification results to construct the fusion model. The clinical-imaging model achieved an AUC of 0.745(95% CI: 0.662–0.823), while radiomics and deep learning models reached 0.803(95%CI: 0.728–0.870) and 0.830(95% CI: 0.764–0.892), respectively. The fusion model showed competitive performance, with an AUC of 0.865(95%CI: 0.805–0.916, p < 0.05), demonstrating the effectiveness of the proposed pre- and post-fusion strategies.