Performance Comparison of Knowledge-Based Dose Prediction Techniques Based on Limited Patient Data.

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Accurate dose prediction is critical for advancing knowledge-based planning and automated treatment planning techniques in radiation therapy.
• Evaluating different dose prediction methods is essential for optimizing treatment planning accuracy.

Purpose of the Study:

• To compare the dose prediction accuracy of three methods: statistical voxel dose learning, spectral regression, and support vector regression.
• To assess these methods using limited patient training data for various radiation therapy plans.

Main Methods:

• Three methods (statistical voxel dose learning, spectral regression, support vector regression) were applied to predict doses for noncoplanar and volumetric-modulated arc therapy plans across head and neck, lung, and prostate sites.
• k-fold cross-validation (k=4) with 20 cases per site was used. Statistical voxel dose learning utilized voxel distance to planning target volume, while spectral regression and support vector regression employed 28 features including volume data.
• Principal component analysis (PCA) was explored for dimension reduction, and separate models for organs at risk were investigated.

Main Results:

• Statistical voxel dose learning, using separate models per organ at risk, achieved the lowest root mean squared error across all sites and modalities.
• PCA demonstrated significant error reduction for spectral regression and support vector regression, particularly for noncoplanar plans.
• Training separate models for each organ at risk consistently yielded higher accuracy than collective voxel models.

Conclusions:

• Statistical voxel dose learning provides the most accurate dose prediction method, proving more robust to patient variability than advanced machine learning techniques with dimension reduction.
• The findings support the utility of statistical voxel dose learning for improving automated and knowledge-based radiation therapy planning.