Developing a segmentation cascade deep learning network based on automated prompts.

Background: Accurate auto-segmentation of gross tumor volume in nasopharyngeal carcinoma (GTVnx) remains a significant challenge. This study introduces a novel two-stage deep learning cascade framework designed to address the precision limitations of single-stage models by strategically decoupling coarse localization from boundary refinement, thereby enhancing accuracy while maintaining computational efficiency. Methods: This study developed an end-to-end model comprising a localization prompt generation unit (PGU) and a fine segmentation unit (FSU), integrated via an attention-guided mechanism. Using a dataset of 276 nasopharyngeal carcinoma (NPC) patients, we quantitatively evaluated three prompting strategies within the PGU: prompt-mask, prompt-box, and a dual-prompt combination. Performance was assessed against expert manual contours using the Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), and mean distance to agreement (MDA). Statistical significance was determined via paired t-tests. Results: The proposed framework achieved state-of-the-art performance. The dual-prompt strategy yielded the most significant improvements over a baseline single-stage model: DSC increased by 9.7% (0.8219 Conclusions: This work presents a computationally efficient, two-stage deep learning-based auto-segmentation (DLAS) framework that significantly improves the accuracy and reliability of GTVnx segmentation. It improves the accuracy and reliability of GTVnx segmentation, features a lightweight architecture and high scalability, and highlights a promising pathway for clinical integration.