Author: Karyn A Goodman, Yang Lei, Tian Liu, Pretesh Patel, Jing Wang, Kaida Yang, Jiahan Zhang 👨🔬
Affiliation: Icahn School of Medicine at Mount Sinai, Department of Radiation Oncology and Winship Cancer Institute, Emory University 🌍
Purpose: This study aims to improve organ-at-risk (OAR) segmentation in pancreatic cancer stereotactic body radiotherapy (SBRT) by integrating clinical guidelines into deep learning workflows. We use large language models (LLMs) to generate task-specific guideline prompts, enabling the incorporation of textual domain knowledge into the segmentation process.
Methods: Clinical guidelines were extracted via ChatGPT. The process used a custom pipeline designed to identify and encode relevant standards from established sources, including ASTRO, RTOG, and ESTRO. These guidelines were transformed into task-specific prompts and encoded into a shared embedding space using the Contrastive Language–Image Pretraining (CLIP) model. The encoded textual features were combined with imaging data processed through a Swin transformer-based image encoder. The integrated features guided a fully convolutional decoder enhanced with automatic pathway (AP) modules, enabling dynamic, task-specific routing for precise segmentation. The framework was trained and evaluated on planning CT datasets with performance assessed using Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD), and other metrics.
Results: The method achieved robust performance with an average DSC of 0.850 and an MSD of 2.73 mm across all OARs. Strong results were observed for challenging structures such as the duodenum (DSC 0.864, MSD 1.38 mm) and small bowel (DSC 0.805, MSD 4.05 mm). The spinal cord showed exceptional accuracy (DSC 0.877, MSD 0.90 mm), while the stomach and large bowel demonstrated good performance (DSC 0.847 and 0.862, respectively).
Conclusion: This framework demonstrates the value of combining LLM-generated guideline prompts with deep learning for OAR segmentation in pancreatic cancer SBRT. By integrating clinical knowledge into the segmentation process, the method shows promise for enhancing radiotherapy planning precision. Future work will focus on automating prompt generation, expanding guideline repositories, and incorporating multimodal imaging to further improve clinical applicability and potentially enable safer dose escalation in pancreatic cancer treatment.