Author: Thomas L. Hayes, Nicholas C. Koch, Han Liu, Qingyang (Grace) Shang, Benjamin J. Sintay, Caroline Vanderstraeten, David B. Wiant 👨🔬
Affiliation: Fuse Oncology, Cone Health, Cone Health Cancer Center 🌍
Purpose:
This study evaluates the accuracy of a deep learning-based automatic breast planning script in predicting beam energy for breast cancer treatments. The script was validated and implemented for clinical use in May 2023 and its performance to date is compared to approved treatment plans.
Methods:
The automatic planning script utilizes a deep learning algorithm trained on breast separation and beam energy data from 2018–2022. Upon successful completion, the script-generated plan details were stored as JSON messages in a SQL database. Dosimetrists were free to modify the resulting plan in any way to satisfy clinical goals. Data from 414 patients (782 individual beams) approved between May 2023 and January 2025 were extracted and cross-referenced with the Varian database of clinically approved plans. The analysis focused on whether the energy predicted by the deep learning algorithm matched the energy ultimately used in treatment.
Results:
The automatic planning script accurately predicted the treatment energy for 660 beams (84.4%) out of 782 analyzed. In 52 cases (6.6%), the treatment plan required a higher energy than predicted, while in 72 cases (9.2%), the final plan required a lower energy. These discrepancies highlight the importance of clinical review for cases where patient-specific factors may necessitate deviations from algorithm-generated plans.
Conclusion:
The deep learning-based automatic breast planning script demonstrates a high degree of accuracy in predicting beam energy, with an 84.4% match rate to approved treatment plans. While most plans align with the algorithm's predictions, instances of energy adjustment underscore the necessity of integrating clinical expertise into the planning process. This study supports the utility of deep learning algorithms in enhancing the efficiency and consistency of breast cancer treatment planning, with opportunities for further refinement of predictive accuracy.