DA-LUNGNET: a multi-stage deep framework with adaptive attention for early detection of subcentimeter pulmonary nodules.

Early and reliable detection of subcentimeter pulmonary nodules remains a major bottleneck in low-dose CT-based lung cancer screening due to high miss rates, vascular-adhesion-induced false positives, and insufficient multi-scale feature fusion. To address these limitations, we propose DA-LungNet, a multi-stage deep framework with adaptive attention that integrates dynamic candidate generation, attention-guided fine segmentation, and 3D contextual verification. The first stage introduces dynamic focal RetinaNet with cross-scale feature interaction to maximize recall under extreme class imbalance. The second stage employs an attention-guided U-Net++ augmented with a dense attention bridging module (DABM) for enhanced edge representation and gradient propagation. The final stage integrates a 3D contextual pyramid module (3D-CPM) to model inter-slice spatial continuity and suppress vascular false positives. Extensive experiments on LIDC-IDRI and DSB2017 datasets demonstrate that DA-LungNet achieves state-of-the-art performance with 92.7% Dice, 93.4% sensitivity for < 6 mm nodules, and an FP/scan rate of 1.4, outperforming existing models (e.g., nnU-Net, DeepLabV3 + , TransUNet) by up to 21.1% in Dice improvement and 87.3% in FP reduction. The model generalizes robustly across multi-center datasets with < 3% performance variance, while maintaining real-time inference (2.3 s/case). These findings indicate that DA-LungNet effectively redefines the sensitivity-specificity trade-off in early lung cancer screening by unifying dynamic loss optimization, dense attention refinement, and contextual 3D reasoning-offering a clinically viable paradigm for precision pulmonary nodule detection.