MMCT-Net: a Multi-Modal Hybrid CNN-Transformer Fusion Network for Preoperative Prediction of Malignant Invasion in...

The accurate preoperative prediction of invasive adenocarcinomas in pulmonary ground-glass nodules (GGNs) is critical for determining the prevention and subsequent treatment of lung cancer. Our goal is to enhance the performance of preoperative prediction through a novel artificial intelligence model to reduce the surgical mismatch rates. We propose a multi-modal hybrid CNN-Transformer fusion network (MMCT-Net) capable of extracting multi-level deep learning features that encompass both local-to-global contextual information and 2D to 3D spatial representations for precise differentiation between preinvasive and invasive lesions. The model also incorporates an adaptive feature integration mechanism to combine these deep learning features synergistically with complementary clinical parameters and radiomics signatures. In this multicenter retrospective study, we analyzed 1-mm thin-section computerized tomography scans and clinicopathological data from 421 patients undergoing GGN surgeries across three centers, all confirmed by histopathology. The experimental results demonstrated that the proposed method achieved an AUC of 92.65% ± 4.85% in internal validation and maintained robust performance (AUC = 89.30% ± 6.33%) in cross-site validation. Our method exhibits superior robustness and outperforms existing approaches in diagnostic performance. Compared with current standard surgical procedures, the non-invasive group showed a 10.85% reduction in mismatch rate, while the invasive group demonstrated a 32.21% decrease. This method improves the success rate of identifying invasive adenocarcinomas in GGNs and has the potential to reduce surgical mismatch rates while facilitating more accurate personalized treatment for patients.