A Hybrid Deep Learning Framework for Combining Quantitative Radiomics Features with Serum Numerical Biomarkers in Non-Small Cell Lung Cancer Staging

• Sunday Sunday

  Ladoke Akintola University of Technology

  Author

Non-small cell lung cancer (NSCLC) remains the leading cause of cancer-related mortality worldwide, with accurate staging being paramount for determining optimal treatment strategies and predicting patient outcomes. Traditional staging methods relying on invasive procedures and qualitative assessments are time-consuming, subject to inter-observer variability, and may not capture the full extent of tumor heterogeneity. Recent advances in radiomics and machine learning offer promising non-invasive alternatives, yet existing approaches typically employ unimodal data sources, limiting their discriminative power. This study addresses this gap by proposing a hybrid deep learning framework that synergistically integrates quantitative radiomic features extracted from computed tomography (CT) images with serum numerical biomarkers, including C-reactive protein (CRP), tumor mutation burden (TMB), and lactate dehydrogenase (LDH). The framework employs a multimodal fusion architecture combining a deep neural network for radiomic feature processing with gradient boosting for biomarker integration. Evaluation on a retrospective dataset of 422 NSCLC patients demonstrated superior performance over unimodal approaches, achieving a classification accuracy of 89.4% (95% CI: 86.2-92.1%) and an AUC-ROC of 0.93 (95% CI: 0.90-0.95). The hybrid model significantly outperformed radiomics-only (78.0% accuracy) and biomarker-only (85.2% accuracy) approaches (p < 0.01). Feature importance analysis identified wavelet-transformed texture features and TMB as the most influential predictors. This framework provides a replicable, non-invasive tool for enhancing NSCLC staging accuracy, with implications for personalized treatment planning and improved clinical decision-making.

• PDF

Copyright (c) 2026 Sunday Sunday (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.