ENSEMBLE DEEP LEARNING WITH RADIOMICS FEATURES FOR LUNG CANCER CLASSIFICATION

Abstract

Lung cancer remains one of the leading causes of cancer-related mortality, with accurate early classification being critical for guiding clinical interventions. Traditional deep learning frameworks often struggle to generalize across heterogeneous CT datasets, while purely radiomics-based models may not capture complex spatial representations. To address these limitations, we propose an ensemble strategy that integrates deep features from a pre-trained 3D ResNet-34 with handcrafted radiomics features extracted via PyRadiomics, fused through an XGBoost classifier. Nodule candidates are first localized using YOLOv7, followed by feature extraction and dimensionality reduction with LASSO re-gression, supported by SMOTE-Tomek resampling to alleviate class imbalance. The proposed ensemble was evaluated using NLST (NCDB subset) and the LUNGx Challenge datasets, achieving an AUC of 0.937, precision of 91.2%, and recall of 89.4%. Comparative assessment against existing approaches, including HRDEL (2023) and EMLC (2025), demonstrates superior discriminative ability, particularly in cases with small or heterogeneous nodules. Unlike earlier ensemble radiomics-based frameworks that show AUC values in the range of 0.86–0.90, our pipeline consistently surpasses these benchmarks under cross-validation. These findings establish the robustness of hybrid deep radiomics fusion for clinical decision support and highlight its potential to outperform current state-of-the-art ensemble pipelines.