2008 - Automatic Recognition of Post-Radiotherapy Tumor Regions in Esophageal Cancer Based on Medical Physics Prior Information: A Multi-Center Study

02:30pm - 04:00pm PT

Hall F

Screen: 25

POSTER

Presenter(s)

Ziqi An, - Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi'an, Shaanxi

Z. An¹, H. Sun², and L. N. Zhao¹; ¹Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi'an, Shaanxi, China, ²Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, xi'an, Shaanxi, China

Purpose/Objective(s):

Cancer patients mainly rely on CT images for diagnosis during follow-up exams. In the absence of PET images, AI-assisted deep learning models face challenges in accurately identifying esophageal cancer tumor regions with high heterogeneity. This study leverages medical physics prior information to enhance model precision in CT-based post-radiotherapy tumor regions recognition.

Materials/Methods:

We retrospectively collected 249 esophageal cancer cases from three centers for model training and internal/external testing. CT images obtained before and after radiotherapy were registered elastically for correction, with physician-delineated post-radiotherapy tumor regions as the ground truth. In this study, we integrated pre-radiotherapy dose (95% prescription volume) and target volume into the nnUNet framework to develop a model for predicting post-radiotherapy tumor regions in CT images. The pre-radiotherapy dose information defined a feature learning scope, while the target volume information highlighted key regions for the model to focus on. After element-wise operations with post-radiotherapy CT, they served as model inputs. Model accuracy was assessed using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and F1-score.

Results:

In the internal test set, the proposed model achieved DSC and IoU accuracies of 0.77±0.12 and 0.64±0.15. In the external set, these were 0.75±0.11 and 0.61±0.13. The numerical results demonstrate that the new model outperforms current automatic delineation models for esophageal cancer tumor regions in terms of both accuracy and stability. Furthermore, the average F1-scores of the proposed model in the internal and external test sets were 0.78 and 0.76, respectively, representing improvements of 22% and 27% in prediction accuracy compared to the initial model trained solely based on CT images.

Conclusion:

Multi-center experiments show that the proposed model accurately and stably predicts post-radiotherapy esophageal tumor regions on CT images, providing a new diagnosis option during follow-up exams for cancer patients.