Lesion-aware multi-view deep learning for benign versus malignant pulmonary nodule classification

Tri-view attention deep learning classified benign versus malignant pulmonary nodules on 869 LIDC-IDRI cases with 93.08% accuracy, AUC 97.06%, and F1 92.68%, outperforming baseline models.

• Evaluation used 869 nodules from LIDC-IDRI (448 benign, 421 malignant); the framework fused axial, coronal, and sagittal CT views with lesion-guided attention and view-weighted fusion.
• Accuracy 93.08%, AUC 97.06%, and F1 92.68% were the highest among the included baseline comparison models in this cohort.
• Ablation analysis showed the combined lesion-guided attention and view-weighted fusion provided the best overall balance; Grad-CAM visualizations concentrated responses in lesion-relevant regions.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.