Enhancing Radiotherapy Planning with Machine Learning: Correlating Anatomical Features and Planning Difficulty to Guide Optimal Plan Design 📝

Author: Li Chen, Shouliang Ding, Xiaoyan Huang, Lecheng Jia, Hua Li, Hongdong Liu, Yanfei Liu, Zun Piao, Guangyu Wang 👨‍🔬

Affiliation: State Key Laboratory of Oncology in South China, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Shenzhen United Imaging Research Institute of Innovative Medical Equipment 🌍

Abstract:

Purpose: Optimal radiotherapy planning is challenging, influenced by anatomical factors such as surrounding organs and tumor characteristics, which complicate dose distribution and target coverage. While the relationship between anatomical features and planning difficulty remains underexplored, understanding it is crucial for predicting challenges and enhancing plan design. This study aims to predict planning difficulty based on anatomical features and identify key anatomical factors associated with increased difficulty, using breast cancer patients as a case study. These insights will guide the design of optimal plans and improve patient outcomes.
Methods: Data from 98 breast cancer patients were collected, including CT images, delineations of target (PTVcw, PTVsc, PTVim) and organs-at-risk, as well as radiotherapy plans. Using an empirically defined dosimetric criterion, 30 cases were identified as difficult plans. Anatomical features were extracted, and an XGBoost model with five-fold cross-validation was used to classify plan difficulty based on these features, with the importance of anatomical features ranked and analyzed.
Results: In left-sided breast cancer, the model achieved an accuracy, precision, and recall of 0.78, 0.77, and 0.78, respectively. The top five important factors for plan difficulty were: overlap proportion of PTV-heart*, flatness of PTVcw*, least-axial length of ipsilateral lung*, major-axial length of PTVcw, and distance of PTVcw-stomach*. For right-sided breast cancer, the model's accuracy, precision, and recall were all 0.83. The top five important indicators were: least-axial length of ipsilateral lung, major-axial and least-axial* length of PTVcw, flatness of PTVim, and volume of PTVcw* (*p<0.05).
Conclusion: This study identified key anatomical factors influencing radiotherapy plan difficulty and predicted plan difficulty for left- and right-sided breast cancer patients with ~80% accuracy. Adaptable to other cancers, this method estimates difficulty from CT images and delineated structures, highlighting anatomical factors related to plan difficulty to support the creation of optimal treatment plans and improve patient outcomes.

Back to List