Lung-aware dual-branch cross-attention network for explainable lung cancer classification from CT images
TL;DR
Findings show that high diagnostic accuracy and computational efficiency can be achieved alongside anatomically grounded, clinically reliable explainability, supporting the proposed framework’s potential for trustworthy translation into real-world lung cancer screening.
Abstract
Reliable classification of lung cancer from chest CT scans is essential for early diagnosis, yet most deep learning (DL) models lack an anatomical basis and robust, quantitatively validated interpretability, limiting clinical adoption. This study aims to develop and validate a lung-aware, dual-branch DL framework that jointly achieves high diagnostic accuracy and anatomically grounded, quantitatively validated explainability for lung cancer classification from CT images. To conclude, the proposed framework integrates automatic lung region extraction, global-local feature learning via a cross-attention-fused two-branch architecture, and a multi-stage progressive fine-tuning strategy for stable, efficient optimization. On the IQ-OTH/NCCD dataset, the framework achieved 96.36% accuracy, a 95.09% F1-score, and an AUC of 0.99, with malignant cases detected at 0.99 precision and 0.98 recall, training in just 18.4 min. Explainability analysis using gradient- and perturbation-based methods confirmed predictions were predominantly driven by lung regions, with Grad-CAM++ (local branch) achieving the highest mean Lung Focus Score (78.94% ± 12.31%) across a stratified evaluation sample of 108 test images. These findings show that high diagnostic accuracy and computational efficiency can be achieved alongside anatomically grounded, clinically reliable explainability, supporting the framework’s potential for trustworthy translation into real-world lung cancer screening.