Sample size requirements for knowledge-based treatment planning.

Area of Science:

• Medical Physics
• Radiation Oncology
• Machine Learning in Healthcare

Background:

• Knowledge-based treatment planning (KBP) aims to automate radiation therapy planning.
• Accurate KBP models are crucial for consistent and efficient treatment delivery.
• Understanding the impact of training set size is essential for developing robust KBP models.

Purpose of the Study:

• To investigate the effect of training set size on the accuracy of various knowledge-based treatment planning (KBP) models.
• To determine the minimum sample size required for different KBP prediction tasks.

Main Methods:

• Four KBP models predicting dose-volume histogram (DVH) points, DVH curves, and objective function weights were evaluated.
• Models were trained using datasets ranging from 10 to 200 patients.
• A validation set of 100 patients was used to assess prediction accuracy.
• Statistical testing determined the minimum sample size for achieving stable model performance.

Main Results:

• The minimum required sample size varied significantly across KBP models and prediction endpoints.
• DVH point prediction required over 200 samples for consistent accuracy for all metrics.
• DVH curve prediction needed at least 75 samples for the bladder and 20 for the rectum.
• Objective function weight prediction required at least 10 samples for logistic regression and 150 for K-nearest neighbors.

Conclusions:

• The optimal training set size for KBP models in prostate cancer is endpoint-dependent.
• These findings provide a baseline for sample size requirements in more complex tumor sites.
• Model complexity and prediction target directly influence the necessary data volume for accurate KBP.