Program Information
Learning Knowledge of Radiation Therapy Treatment Plans Using Decision Tree
X Chang*, H Li , D Yang , Washington University School of Medicine, St. Louis, MO
Presentations
TU-C1-GePD-J(A)-3 (Tuesday, August 1, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: To investigate the feasibility of applying a machine-learning method to extract knowledge of radiation therapy treatment planning data.
Methods: 3756 EBRT single-prescription plans from 2007 to 2015 were obtained from MOSAIQ including 14 disease sites: brain, head-and-neck, breast, lung, chestwall, spine, abdomen, pelvic, pelvis, prostate, thorax, TBI, extremity and lymph nodal. Each plan includes seven disease attributes: tumor stage, nodal stage, metastatic stage, diagnosis class, laterality and previous treatment, and five treatment parameters: total dose, fractions, number of fields, technique and modality. A one-class binary decision tree learning method was employed to learn tree-structured knowledge from the historical treatment data. The category data, e.g., site, tumor stage, nodal stage, metastatic stage and technique, were encoded into binary codes with a length the same as the number of categories of each data item. The leaves of the trained decision tree contain the frequency of the historical treatment plans, which are scored by a decision function with respect to the depth of leaves in the tree. The leaves with the smallest score were pruned by applying a 5% rejection rate in the historical data. Decision rules, i.e., knowledge in the radiation therapy treatment domain, were generated by collecting and merging decision conditions from root to each remaining leave.
Results: A decision tree with 128 leaves were generated. 81 leaves were kept after the pruning and 81 corresponding decision rules were extracted. The extract rules were manually reviewed and confirmed.
Conclusion: The one-class decision tree learning method is effective in learning interpretable radiation therapy treatment knowledge. The extracted decision rules are useful for automated detection of errors in treatment plan parameters and for educational purposes.
Funding Support, Disclosures, and Conflict of Interest: Funding: AHRQ R01-HS022888; No conflict of interest; Disclosures: Authors have technology licensing fee from Viewray
Contact Email: