Improving Non-Small Cell Lung Cancer Classification Through Radiogenomics and Transformer-Based Deep Learning Fusion Strategies

Abstract

Getting the right subtype of non–small cell lung cancer (NSCLC) is essential for choosing the best treatment, but it is tough when looking at tissue and genetic data alone. In this study, we set out to improve NSCLC subtype identification—specifically through transformer-based deep learning. We introduce two fusion techniques to fuse CT/PET scans with clinical and genetic data. The first, Intermediate Fusion, combines these data streams partway through the model. The second, Late Fusion, lets each data type run through its own processing pipeline before merging their predictions at the end. Then we compare our approaches with an earlier Intermediate Fusion strategy. Results demonstrated that the proposed Late Fusion achieved superior performance, with 96.12% accuracy, and the proposed Intermediate Fusion achieved 95.64% accuracy, outperforming an earlier Intermediate Fusion which achieved 94.04% accuracy. Both proposed (late fusion, intermediate fusion) methods pick up true cases equally well with 95.5% sensitivity, but Late Fusion’s boosting its precision to 96.22% and F1-score to 95.86%. Model interpretability is evaluated using SHAP for tabular data and attention-based analysis for imaging to reveal modality-specific contributions. These findings indicate that late fusion not only improves classification performance but also supports more transparent and clinically meaningful decision-making.