Presenter(s)
H. Peng, Y. Liang, S. Wei, Q. Liu, X. Chen, J. Dai, Y. Tang, and K. Men; Department of Radiation Oncology, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China
Purpose/Objective(s): Clinical target volume (CTV) delineation for rectal cancer is essential yet labor-intensive and subject to inter-observer variability. Image-only auto-segmentation models remain limited by boundary ambiguity and the absence of explicit clinical semantic guidance. We developed a text-guided segmentation network (TG-SegNet) that integrates CT imaging with clinical text prompts to improve accuracy, clinical acceptability, and contouring efficiency.
Materials/Methods: We retrospectively collected 567 preoperative rectal cancer cases (2017–2025), split into training (n=452) and independent testing (n=115). TG-SegNet integrates CT-derived visual features with text prompts encoded by PubMedBERT and fused via cross-attention and a fine-grained image–text alignment module. Text prompts contained (i) patient case information (age, sex, tumor stage, tumor location, CT scan position) and (ii) guideline-derived descriptions indicating which CTV subsites should be included. Performance was compared with nnU-Net and two ablations: TG-SegNet without text prompts and TG-SegNet with a simplified alignment module. Additional analyses included text-prompt and module ablations. Quantitative metrics were Dice similarity coefficient (DSC), 95% Hausdorff distance (HD95), mean surface distance (MSD), surface DSC (S-DSC), and average path length (APL). Two senior radiation oncologists performed a blinded review of 20 random test cases, scoring target coverage, boundary appropriateness, guideline adherence, and overall acceptability (5-point Likert). Efficiency was assessed by comparing auto-contour correction time and manual contouring time. All quantitative metrics, scores, and efficiency outcomes were statistically analyzed (p < 0.05).
Results: TG-SegNet achieved the best performance across all quantitative metrics: DSC 0.927±0.022, HD95 7.01±6.05 mm, MSD 1.94±1.08 mm, S-DSC 0.799±0.074, and APL 7372±4452 (all p<0.01). In clinical evaluation, TG-SegNet significantly improved target coverage, guideline adherence, and overall clinical acceptability compared with nnU-Net and ablations (p<0.05), with boundary appropriateness comparable to nnU-Net. TG-SegNet had the shortest correction time (3.39±1.10 minutes), corresponding to 82.1% time savings versus manual delineation. Text-prompt ablations suggested that the CTV subsite prompt component contributed more to performance improvement than patient case information, while stage, tumor location, and scan position were relatively more important among the case-information components. Module ablations showed that both cross-attention and fine-grained alignment were beneficial.
Conclusion: By integrating clinical semantics with CT imaging, TG-SegNet improved rectal cancer CTV auto-segmentation accuracy, clinical acceptability, and efficiency, with the potential to reduce clinician workload and support more standardized, consistent contouring.