2522 - Integrating Endoscopic Clinical Priors into Deep Learning for Precise Gross Tumor Volume Delineation in Esophageal Cancer

Purpose/Objective(s): Accurate Gross Tumor Volume (GTV) delineation is critical for esophageal cancer radiotherapy but remains challenging on CT images due to poor soft-tissue contrast and indistinct boundaries. We hypothesize that integrating the "distance from incisors" extracted from clinical endoscopy reports—as a geometric prior—can significantly refine deep learning-based segmentation and mitigate longitudinal errors. This study aims to validate a multimodal framework combining CT images and textual medical records to improve segmentation accuracy.

Materials/Methods: Data from 41 patients with pathologically confirmed thoracic esophageal squamous cell carcinoma were retrospectively collected at a single institution. GTVs manually delineated by senior radiation oncologists served as the ground truth. A two-stage "segmentation-then-truncation" framework was developed. First, a 3D nnU-Net model was trained on CT images to generate initial segmentation masks. Second, the tumor's upper and lower boundaries (distance from incisors) were extracted from textual endoscopy reports and mapped to the CT coordinate system (Z-axis). These coordinates served as longitudinal hard constraints to truncate the superior and inferior outliers of the model-predicted masks. Performance was evaluated on a held-out test set using the Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (HD95).

Results: The method was validated on 6 independent test cases. The baseline nnU-Net achieved a mean DSC of 0.84 and a mean HD95 of 9.21 mm, exhibiting significant over-segmentation along the longitudinal direction. The post-processing integration of endoscopic textual constraints improved the mean DSC to 0.90. Notably, the mean HD95 decreased significantly to 4.20 mm, reflecting a substantial reduction in boundary errors. Visual inspection confirmed that the proposed strategy effectively corrected Z-axis over-segmentation, aligning the tumor length more closely with the ground truth.

Conclusion: This study demonstrates that integrating endoscopic textual priors into deep learning workflows effectively corrects the longitudinal segmentation uncertainties inherent in CT-only approaches. This method provides a cost-effective and easily deployable solution for clinical practice, enhancing geometric accuracy without requiring additional imaging modalities. A limitation of the current approach is its inability to correct under-segmentation; future work will focus on developing end-to-end multimodal fusion networks to address this challenge.