D-S-Net: an efficient dual-stage strategy for high-precision segmentation of gross tumor volumes in lung cancer CT images.

Background: Accurate delineation of Gross Tumor Volume (GTV) in lung cancer is critical for effective radiotherapy and surgical planning. However, segmentation of GTV in high-resolution CT images remains challenging, particularly when tumors are small or have indistinct boundaries.

Methods: We propose D-S-Net, a novel dual-stage strategy to enhance both the accuracy and efficiency of lung cancer GTV segmentation. In the first stage, a simplified detection network is used to locate candidate regions in high-resolution (512×512) CT slices, reducing input size and computational demand. In the second stage, a modified U-Net variant is applied to perform fine segmentation within the detected regions. The architecture incorporates a spatial attention mechanism and employs a combined loss function (binary cross-entropy and Dice loss) to address class imbalance.

Results: On the lung cancer GTV dataset, D-S-Net achieved a Dice coefficient of 78.52%, representing an improvement of 5.49% over SwinU-Net and outperforming several mainstream models. On the second dataset, D-S-Net reached a Dice coefficient of 86.56%, surpassing the second-best model by 13.19%. Ablation studies demonstrated that the detection stage, spatial attention, and combined loss function effectively improved performance, while computational complexity analysis confirmed the model's efficiency.

Conclusion: The proposed D-S-Net offers a robust and efficient solution for the segmentation of lung cancer GTV in CT images. Its dual-stage design and attention-based enhancements contribute to both accuracy and computational gains, highlighting its potential for clinical applications in radiotherapy planning.