GLANCE: Enhanced Pulmonary Nodule Segmentation via Global-Local Interaction

Overview

GLANCE introduces a dual-stream architecture combining a global context transformer and a multi-receptive grouped atrous mixer to improve pulmonary nodule segmentation from CT scans. Validated on four public datasets, it achieves state-of-the-art performance by continuously fusing complementary global and local features at multiple scales.

Background

Early detection of lung cancer relies heavily on accurate segmentation of pulmonary nodules in CT images, which is challenging due to nodules' diverse sizes, shapes, and low contrast against surrounding tissues. Conventional CNNs struggle to capture long-range dependencies, while transformers may overlook fine local details. Existing hybrid models often fuse features late in the network, causing representational clashes and suboptimal learning. A robust segmentation model must integrate both global context and local detail continuously to handle scale disparity and subtle nodule characteristics.

Data Highlights

Key Findings

• GLANCE employs a dual-stream encoder with a global context transformer and a multi-receptive grouped atrous mixer to capture complementary features.
• Continuous cross-scale consensus fusion integrates global and local features at every hierarchical level, preventing representational clashes.
• A dual-head pyramid refinement decoder enables simultaneous nodule segmentation and center heatmap detection.
• Extensive ablation studies confirm the critical role of continuous fusion in achieving superior segmentation and detection performance.
• Validated on four public benchmarks, GLANCE consistently outperforms existing state-of-the-art models.

Clinical Implications

GLANCE's improved segmentation accuracy facilitates earlier and more reliable detection of pulmonary nodules, potentially enhancing lung cancer screening outcomes. Its ability to integrate global context with fine local details addresses challenges posed by small or ambiguous nodules, supporting more precise clinical decision-making. Adoption of such advanced models may reduce false negatives and improve patient prognosis through timely intervention.

Conclusion

GLANCE represents a significant advancement in pulmonary nodule segmentation by harmoniously combining global and local feature representations through continuous fusion. This approach sets a new benchmark in accuracy and robustness across multiple public datasets, offering promising clinical utility for early lung cancer diagnosis.

References

1. Selvadass et al. -- SAtUNet: Atrous Convolution Modules for Multi-scale Context
2. Wu et al. -- Channel Residual UNet for Iterative Feature Refinement
3. Xu et al. -- Dual Encoding Fusion Model for Atypical Nodules