A CT-based multi-scale fusion model with SHAP interpretation for preoperative differentiation between lung adenocarcinoma in situ/minimally invasive adenocarcinoma and invasive adenocarcinoma: a multicenter study

Objective To develop a multi-scale model integrating radiomics, habitat, deep learning, and clinical information for preoperative precise differentiation between lung adenocarcinoma in situ/minimally invasive adenocarcinoma and invasive adenocarcinoma in ground-glass nodules (GGNs). Methods A total of 621 patients with GGNs from two centers were retrospectively enrolled. Radiomics, habitat, deep learning (2D, 2.5D, and 3D), and clinic models were constructed separately. Early fusion (feature-level fusion) and late fusion (decision-level fusion) strategies were systematically compared. The SHapley Additive exPlanations method was applied for interpretability analysis of the optimal model. Results The early fusion model demonstrated superior performance, achieving areas under the receiver operating characteristic curves of 0.988, 0.903, and 0.872 in the training, internal validation, and external testing cohorts, respectively. Among the deep learning architectures, the 2.5D ResNet50 achieved the best performance. Conclusions This multi-scale early fusion model provides robust support for preoperative precise diagnosis and individualized treatment decision-making for GGNs.