The predictive value of 2.5D deep learning model based on contrast-enhanced CT for pneumonia-mimicking peripheral lung cancer

Differentiating pneumonia-mimicking peripheral lung cancer (PLC) remains clinically challenging because conventional radiomics methods show limited diagnostic performance. This study aimed to develop an interpretable 2.5D deep learning–based multi-feature fusion model to improve PLC prediction on contrast-enhanced CT. This multicenter retrospective study included 302 patients from two medical centers. A 2.5D dataset was generated from venous-phase contrast-enhanced CT by extracting seven slices centered on the largest axial cross-section of each lesion. Multiple deep learning architectures were trained using the generated 2.5D data, followed by multi-instance learning and ensemble-based fusion strategies. Deep learning features, PLH and BoW features were further integrated with radiomics and clinical variables to construct the combined model. Model performance and interpretability were evaluated using ROC analysis, calibration curves, DCA, SHAP, and Grad-CAM. Sex, venous-phase CT attenuation values, and carcinoembryonic antigen levels were identified as independent clinical predictors. Among the evaluated deep learning architectures, ResNet101 achieved an AUC of 0.809 (95% CI: 0.717–0.900) in the independent testing cohort. The combined model yielded AUCs of 0.874 (95% CI: 0.830–0.918) in the training cohort and 0.855 (95% CI: 0.774–0.936) in the independent testing cohort. Calibration analysis showed good agreement between predicted and observed outcomes, while DCA indicated a higher net clinical benefit. SHAP and Grad-CAM analyses provided additional insight into model predictions. An interpretable 2.5D deep learning–based multi-feature fusion model was developed for reliable prediction of pneumonia-mimicking PLC, providing effective decision support for early diagnosis and improved clinical management.