Dosimetric features-driven machine learning model for DVH prediction in VMAT treatment planning

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Current dose-volume histogram (DVH) prediction models lack patient-specific dosimetric features, hindering accurate input-output parameter correlation.
• Estimating achievable treatment plan quality requires incorporating patient dosimetric properties into DVH prediction frameworks.
• Machine learning offers a promising approach to develop advanced DVH prediction models.

Purpose of the Study:

• To develop a machine learning-based DVH prediction framework utilizing dosimetric metrics from planning target volume (PTV)-only plans.
• To establish a correlative relationship between PTV-only plan DVHs and clinical treatment plan (CTP) DVHs.
• To estimate the potentially achievable quality of VMAT treatment plans for prostate cancer patients.

Main Methods:

• A support vector regression (SVR) model was employed for DVH prediction.
• A database of volumetric modulated arc therapy (VMAT) plans from 63 prostate cancer patients was utilized.
• PTV-only plans were generated, and their DVHs were used as input to predict CTP DVHs using 53 training cases and validated on 10 cases.

Main Results:

• The model achieved high accuracy in predicting dosimetric endpoints (DEs) for the bladder (98% within 10% error in training) and rectum (85% within 10% error in training).
• Validation tests showed 92% and 96% of DEs within 10% error bounds for bladder and rectum, respectively.
• Eight of ten validation plans (80%) met the 10% error bound for both organs, with low sum of absolute residuals (SAR).

Conclusions:

• A novel machine learning model driven by dosimetric features from PTV-only plans enables accurate DVH prediction.
• This framework efficiently generates the best achievable DVHs for VMAT planning.
• The approach enhances the prediction of treatment plan quality by incorporating patient-specific dosimetric information.