Towards Clinically Deployable Lung Nodule Classification Using a Hybrid Multi-Scale 3D CNN–Radiomics Framework with Explainable Artificial Intelligence

Lung cancer is one of the leading causes of cancer-related deaths worldwide, making early diagnosis essential for improving patient outcomes. Although Low-Dose Computed Tomography (LDCT) supports early detection, manual interpretation of volumetric CT scans is time-consuming and subject to considerable inter-observer variability. Existing deep learning approaches also face challenges related to model interpretability, limited use of radiomic information, varying nodule sizes, and computational complexity. This paper presents a hybrid framework that combines multi-scale 3D Convolutional Neural Networks (3D CNNs), handcrafted radiomic features, a Convolutional Block Attention Module (CBAM), and 3D Grad-CAM for explainable lung nodule classification. The model was evaluated on 2,661 annotated nodules from 1,018 CT scans in the publicly available LIDC-IDRI dataset using a strict patient-level data split. It achieved 93.8% accuracy, 91.4% sensitivity, 95.2% specificity, and an AUC of 0.9489, while requiring only ~450K parameters and an average inference time of 58 ms. Quantitative evaluation of the generated 3D Grad-CAM heatmaps produced an Intersection-over-Union (IoU) of 0.77 with expert annotations, indicating good agreement between model explanations and clinically relevant regions. These results demonstrate that combining multi-scale feature learning with radiomic descriptors can provide accurate and computationally efficient lung nodule classification.