Related Experiment Video
Updated: May 27, 2026
Radiation Planning Assistant - A Streamlined, Fully Automated Radiotherapy Treatment Planning System
Published on: April 11, 2018
A hybrid cluster-then-predict machine learning radiotherapy knowledge-based planning framework for similarity
Trent Benedick1, J Daniel Chu2, Stephanie Zhou2
1Image Processing and Informatics Laboratory, Department of Biomedical Engineering, University of Southern California, 1042 Downey Way, Los Angeles, CA, 90089, USA. benedick@usc.edu.
Purpose:
Radiotherapy treatment planning is currently premised on individual clinical experience and use of many dose based optimization and knowledge-based planning (KBP) models. This study introduces and validates a novel KBP algorithm that matches cases based on a holistic model of the target to Organ at Risk (OAR) constellation geometry to assist during treatment plan creation by identifying retrospective cases with similar target-OAR constellation geometry for use as reference cases to serve as planning templates.
Methods:
We developed a multi step algorithm to calculate geometric similarity between radiotherapy cases. The method extracts quantitative geometric features (Overlap Volume Histograms and Spatial Target Signatures) to characterize the spatial relationships between all target volumes and OARs. Component, partial, dissimilarities between corresponding features are calculated with an Earth Mover's Distance (EMD). A final, clinically relevant similarity score is computed using a hybrid cluster-then-predict machine learning approach (a hard-gated Mixture-of-Experts architecture), where cases are first partitioned by spectral clustering, then separate supervised models are trained within each cluster to predict case similarity.
Results:
The algorithm was validated with a knowledge base of 192 retrospective Head and Neck Cancer cases. A leave-one-out cross-validation demonstrated a strong correlation (r = 0.8009) between our KBP model's geometry based similarity score and actual dose similarity. Additionally, a clinical evaluation with an ROC analysis yielded a high area under the curve (AUC) of 0.8797, confirming the model's ability to identify cases the clinicians determined were clinically relevant as reference cases to serve as planning templates.
Conclusions:
This study validates our novel KBP algorithm that successfully identifies clinically relevant retrospective cases using a comprehensive geometric model of target-OAR constellations. By providing personalized patient specific, knowledge base best practice reference case data before beam configuration and treatment plan creation, this method provides a useful new tool to increase standardization, improve plan quality, and assist clinicians during initial plan creation.
