Dual-Scale Hybrid Concept Bottleneck Network for Explainable 3D Lung Nodule Malignancy Classification in CT Imaging

Abstract

Accurate differentiation of benign and malignant lung nodules in computed tomography (CT) is important for early lung cancer diagnosis and reliable clinical decision-making. Many existing deep learning methods emphasize either nodule-centred morphology or broader anatomical context and provide limited insight into the radiological information represented by the model. This study proposes a Dual-Scale Hybrid Concept Bottleneck Network (DS-HCBN) for explainable 3D lung nodule malignancy classification. The framework processes a local nodule-centred patch and a larger contextual patch using a shared residual 3D convolutional encoder and a lightweight contextual Transformer. The resulting representations are integrated through gated cross-attention, while eight radiological attributes are learned as supervised intermediate representations within the hybrid classifier. The model was developed on LIDC-IDRI using a leakage-controlled patient-wise split and evaluated on a held-out internal test set. External evaluation was performed on the publicly annotated LNDb cohort using the same frozen model without retraining or fine-tuning. On the LIDC-IDRI internal test set, DS-HCBN achieved 87.34% accuracy, an F1-score of 80.39%, and a ROC-AUC of 94.35%. On LNDb, the model achieved 80.41% accuracy and a ROC-AUC of 78.65%. These results show strong internal discrimination but a clear reduction in cross-dataset performance. The findings support the use of local morphology, anatomical context, and radiological concept supervision for concept-guided lung nodule classification, while highlighting the need for improved domain generalization before broader clinical application.