Program Information
Harnessing Machine Learning to Identify Beneficial Anatomy for Proton Therapy
D Hall1*, Z Liao2 , R Mohan2 , H Paganetti1 , (1) Massachusetts General Hospital & Harvard Medical School, Boston, MA, (2) UT MD Anderson Cancer Center, Houston, TX
Presentations
MO-RAM-GePD-J(A)-5 (Monday, July 31, 2017) 9:30 AM - 10:00 AM Room: Joint Imaging-Therapy ePoster Lounge - A
Purpose: The dosimetric advantages of proton beam therapy (PBT) over photon therapy (e.g. IMRT) are currently evaluated via treatment planning, which is resource-intensive and requires software and clinical expertise typically confined to PBT centers. This study utilizes machine learning to predict the patient-specific treatment plan achievable with each treatment modality and subsequently evaluates the expected benefits of PBT.
Methods: Using a training cohort of existing treatment plans, a random forest of decision trees was trained to infer point dose from 5 geometric features (e.g. distance-to-target). It could then predict the 3-dimensional dose distribution for a new patient, based upon the geometric features of each voxel. The number of trees was optimized by observing the out-of-bag error curve. Generalization errors were evaluated with 5-fold cross-validation, by validating the predicted equivalent uniform dose (EUD) and normal tissue complication probability (NTCP) to OARs and computing gamma index pass rates.
Results: Random forests were independently trained for PBT and IMRT, using a cohort of NSCLC patients enrolled in a trial that randomized treatment between PBT (36 pts) and IMRT (68 pts). The EUD and NTCP predicted for lung, heart and esophagus were found to be in good agreement with the actual plan. The coefficient of determination (R-squared) was between 66% and 88% for EUD predictions, demonstrating that the models successfully captured interpatient variation. The median pass rate of a 3%/3mm gamma test for the predicted 3-dimensional dose was between 62% and 89%, depending upon OAR and modality.
Conclusion: We developed a model that can identify whether a patient’s anatomy permits increased dosimetric advantages from PBT. This could enable better-informed PBT referral decisions, improving the cost-effectiveness of this expensive therapy. It could also estimate the expected benefits at an earlier stage in the clinical workflow, or provide a high-throughput patient pre-selection for model-based PBT trials.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by National Institutes of Health grant U19 CA21239.
Contact Email: