Room: Karl Dean Ballroom B1
Purpose: To develop a machine learning model to detect errors in HDR brachytherapy treatment plans.
Methods: With IRB approval, our clinical HDR patient database was anonymized. We extracted the plan, structure, dose, and image DICOM data for 237 cylinder applicator cases. These cases were divided into 15 separate classes based on cylinder diameter, treatment length, prescribed dose, prescription depth, and presence of known errors. We selected 14 features from the treatment plan DICOM data. K-nearest neighbor (KNN) classification models were trained on this dataset using MATLAB, and a script was written to classify new treatment plans.
Results: The trained model was able to correctly classify 235 out of the 237 original treatment plans (99.2% accuracy). Additionally, it was able to accurately classify new sample treatment plans and could also correctly identify treatment plans containing errors. Detected errors include incorrect prescription, catheter path reconstruction, and catheter orientation.
Conclusion: Machine learning is a powerful tool that can be applied to detecting errors in HDR brachytherapy. While detecting errors for cylinder applicator plans is relatively straightforward, the same process is applicable to more complex treatments as well. Future work will include applying our methods to identifying errors in tandem and ring, and interstitial HDR treatment plans. This method of error detection, when coupled with existing safety checks for plan errors, is an improved technique for ensuring safe and accurate treatment delivery for HDR patients.