2507 - GTV-Guided Deep Learning Segmentation of Clinical Target Volumes in Esophageal Cancer Radiotherapy

Purpose/Objective(s): Contouring the Clinical Target Volume (CTV) for esophageal cancer is a time-consuming process that relies heavily on detailed anatomical knowledge of mucosal spread and elective nodal stations. While automated contouring has advanced considerably in other cancer sites, automated CTV segmentation for esophageal cancer remains largely underexplored. This study investigates the performance of a deep learning model in automating esophageal CTV segmentation and evaluates the impact of providing Gross Tumor Volume (GTV) contours as prior guidance.

Materials/Methods: CT scans from 119 esophageal cancer patients were retrospectively analyzed and stratified by primary tumor location: cervical (n=40), middle (n=40), and lower (n=39). The dataset was divided into training (n=105) and validation (n=14) sets. To segment the primary and nodal target volumes (CTVp and CTVn), two nnU-Net-based segmentation models were evaluated: (1) an image-only baseline utilizing standard CT inputs, and (2) a GTV-guided model incorporating multi-channel inputs (CT image, GTVp masks, and GTVn masks). Segmentation performance was quantitatively assessed using the Dice Similarity Coefficient (DSC).

Results: Incorporating GTV guidance improved segmentation accuracy, increasing the overall average CTV DSC from 0.418 to 0.782. Without GTV guidance, the image-only baseline struggled to localize highly variable target locations, yielding an overall CTVp DSC of 0.488 and a CTVn DSC of 0.348. When utilizing GTV guidance, the overall CTVp and CTVn DSC improved to 0.814 and 0.750, respectively. This guidance was particularly beneficial for the lower esophagus, where CTVn accuracy showed the largest gain, increasing from 0.040 to 0.643. Regionally, the middle esophagus achieved the highest accuracy under GTV guidance with a CTV DSC of 0.831, while the lower esophagus remained the most challenging at 0.706 due to the variable location of the CTVp.

Conclusion: GTV-guided deep learning achieves accurate and consistent CTV segmentation across esophageal tumor sublocations, with particularly substantial gains in nodal CTV delineation. Overall, this demonstrates that utilizing GTV context is a highly effective strategy for developing robust, automated contouring tools in esophageal radiotherapy.