2638 - A Simple Model to Assist Radiation Oncologists with Identification of Patients for Focal Boosted EBRT for Medium to High Risk Prostate Cancer
Presenter(s)
D. Provenzano1, J. Y. Wang2, and Y. J. Rao3; 1Biomedical Engineering, George Washington University School of Engineering and Applied Science, Washington, DC, 2Department of Radiation Oncology, George Washington University School of Medicine and Health Sciences, Washington, DC, 3Department of Radiation Oncology, Stanford University School of Medicine, Stanford, CA
Purpose/Objective(s): The FLAME trial demonstrated very promising results for EBRT for medium and high risk prostate cancer by treating intra-prostatic tumors at a higher dose, or focal boosting, which improved biochemical progression free survival (bDFS) while maintaining toxicity. The patient selection process however still requires manual processing by radiation oncologist. This study sought to determine if a machine learning model could instead identify ideal patients candidates. Additional care was taken to create a simple machine learning model that could be run locally and with less computational power such as a ResNet.
Materials/Methods: This study collected T2-Weighted Magnetic Resonance Imaging Data (MRI) for a subset of 63 subjects with clinically significant prostate cancer from the Prostate-X cohort on the TCIA database. Patients were selected in accordance with FLAME guidelines and internal local (GW) clinical trial selection criteria including: distance from bladder, distance from rectum, size of tumor, presence of metastatic disease, and node positive status. Residual Neural Network (ResNet) was created using a hybrid-transfer learning process with initial ImageNet trained model, and additional model layers added to optimize model build on MRI data, dataset was trained on an 80% subset of the data, tested on a 20%, and processed using 5-fold cross validation and a shuffle test to ascertain statistical significance.
Results: ResNet model achieved 96% accuracy at a p < 0.01 after training, 5-fold cross validation, and shuffle test on the validation subset of the data to select patients for focal boosting.
Conclusion: A simple machine learning model was able to mirror Radiation Oncologist selection process for candidate patients for Focal Boosted EBRT. This represents a simple and easy way to ease Radiation Oncologist workload while implementing FLAME trial for prostate cancer. Future testing is being conducted on more data.