A dual‐prompt framework combining spatial‐mask priors and clinical‐text conditioning for CT‐only GTV delineation in esophageal cancer: A multicenter study
Hongfei Sun, Chengwei Chen, Junqiang Chen, Qifeng Wang, Jie Li, Ziqi An, Yufen Liu, Fan Meng, Lecheng Jia, Zihan Shi, Yan Cui, Yang Liu, Jie Gong, Dong Huang, Lina ZhaoAbstract
Background
Accurate gross tumor volume (GTV) delineation is essential for esophageal cancer radiotherapy, but CT‐only automatic delineation remains difficult because of indistinct tumor boundaries.
Purpose
This study aimed to develop and validate a two‐stage dual‐prompt framework combining a spatial‐mask prior with LLaMA3‐encoded clinical‐text conditioning for primary esophageal GTV delineation.
Methods:
This multicenter study included 857 patients: 500 for development, 125 for internal validation, 50 for ablation analysis, and 182 for external validation. The proposed framework consists of two stages: (1) a CT‐based automatic segmentation network generating an initial GTV as a spatial mask prompt, and (2) a refinement network that jointly incorporates spatial mask prompt and LLaMA3‐encoded clinical text prompt derived from structured clinical information. Performance was evaluated using overlap, surface distance, and volumetric metrics, and compared with multiple state‐of‐the‐art deep learning models. Ablation and radiomics analyses assessed component contributions and downstream consistency.
Results
In internal validation, the framework achieved a DSC of 80.32 ± 6.65%, an HD95 of 11.42 ± 8.90 mm, and an absolute volume difference of 9.99 ± 10.49 mL; all Holm‐adjusted p values were <0.001 versus the comparison models. Mean DSCs were 78.60 ± 13.39% and 78.36 ± 15.54% in the external cohorts. Mean DSCs were 78.60 ± 13.39% and 78.36 ± 15.54% in the two external cohorts, respectively. In external cohort A, ContextUNETR showed a lower mean HD95 than the proposed model (10.93 ± 8.03 vs. 13.44 ± 18.90 mm), although the paired difference was not statistically significant after Holm correction (adjusted p = 0.906). Compared with the direct Stage‐1 output, the complete framework consistently improved overlap, boundary‐distance, and volumetric‐agreement metrics across the internal and both external validation cohorts (all adjusted p < 0.001). LLaMA3 encoding improved DSC versus an MLP using the same variables (adjusted p = 0.009), whereas ASSD did not differ significantly. Automatic and manual GTVs showed high radiomic concordance and comparable prognostic performance.
Conclusion
The framework improved overlap and volumetric agreement for CT‐only primary esophageal GTV delineation across multicenter validation cohorts, although boundary‐distance performance varied across external cohorts.