Multi-view attention-based deep learning for benign–malignant classification of pulmonary nodules

Abstract

Accurate classification of pulmonary nodules is essential for early lung cancer diagnosis, yet the limited spatial information provided by a single computed tomography slice may hinder the characterization of complex nodule morphology. Although multi-view strategies can capture complementary anatomical information, conventional fusion methods often fail to account for lesion-specific features and differences in the contribution of individual views. To address these limitations, we propose a tri-view pulmonary nodule classification framework that combines Lesion-Guided Attention (LGA) with View-Weighted Fusion (VWF). The framework extracts complementary representations from axial, coronal, and sagittal slices. LGA introduces lesion masks as weak spatial priors to refine lesion-related feature responses while preserving contextual information, whereas VWF learns sample-dependent weights to adaptively aggregate the three view representations. The framework was evaluated on 869 nodules (448 benign and 421 malignant) derived from the public Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset. It achieved an accuracy of 93.08%, an AUC of 97.06%, and an F1-score of 92.68%, representing the highest values for these three metrics among the models included in the baseline comparison. Ablation analysis further showed that the combined use of LGA and VWF provided the best overall balance among the evaluated configurations. In the displayed test cases, Gradient-weighted Class Activation Mapping visualizations showed that the learned responses were concentrated mainly within or adjacent to lesion-relevant regions. These findings support the potential value of combining lesion-aware feature refinement with adaptive multi-view fusion for pulmonary nodule classification within the evaluated LIDC-IDRI cohort.