Lungnet-RiskX: A Hybrid Deep Learning Framework for Lung Cancer Detection, Nodule Classification, and Personalized Risk Assessment

Abstract

Background: Lung cancer remains one of the most common causes of cancer-related death globally, yet catching it early is still a significant challenge in clinical practice. Standard imaging tools such as CT scans and chest X-rays have real limitations: low-contrast lesions are easy to miss, and even experts can disagree on what a single scan reveals. Small pulmonary nodules slip through even in experienced hands, pushing back diagnosis and worsening outcomes. Better, automated tools are needed to help clinicians catch these nodules earlier and assess individual patient risk more reliably. Methods: This study proposes LungNet-RiskX, a hybrid deep learning framework that combines Convolutional Neural Networks (CNN) for local feature extraction, Vision Transformers (ViT) for global context, and Extreme Gradient Boosting (XGBoost) for classification and personalized risk scoring. The model was Trained on 53,853 lung CT images (normal, benign, malignant categories) with preprocessing (resizing to 64x64, CLAHE contrast enhancement, median filtering, augmentation), We evaluated performance using accuracy, ROC-AUC, precision, recall, and F1-score. Results: LungNet-RiskX achieved 94.21% overall accuracy, outperforming both standalone and partially hybrid models. The model demonstrated strong classification performance across all classes. ROC-AUC scores were strong across all classes, particularly for malignant detection (0.998). Probability-based risk scoring provided interpretable probability outputs, improving interpretability and supporting clinical decision-making. Conclusion: LungNet-RiskX offers a robust, hybrid, and scalable framework for early lung cancer detection, nodule categorization, and personalized risk assessment, with potential to enhance screening efficiency and inform preventive strategies. Future multi-center validation is recommended. Bangladesh J Medicine 2026; 37(3): 271-275.