1078 - Detecting and Classifying On-Treatment Patient Shifts, with In-Vivo EPID Images and Deep Learning

L. Moore, and C. Bojechko; University of California San Diego, Department of Radiation Medicine and Applied Sciences, La Jolla, CA

Purpose/Objective(s): EPID images collected in-vivo can detect treatment errors missed by other QA checks. To detect errors a gamma comparison is performed to assess deviations from the treatment intent. However, the clinical impact of gamma failure is often ambiguous and lacks directional information. We propose a model that can detect and classify errors using deep learning. We aim to detect patient shifts on the order of PTV margins (>3mm), which can result in a geometric miss of the tumor. Estimation of the shift direction can provide geometric information about what regions of the target moved out of the beam. In this work, we demonstrate a model which can detect and classify patient shifts that occur on-treatment, after setup imaging, using a measured in-vivo EPID image and predicted image as input. Materials/

Methods: Using a previously developed deep learning model, EPID images with simulated shifts were generated for 938 treatment plans by shifting the input CBCT perpendicular to the beam direction in one of eight possible directions (up, down, left, right and each diagonal) with three possible magnitudes (0.8, 3.2, and 6.4mm). The combinatorial pairing of shifted CBCTs and predicted EPID images for each patient resulted in a training dataset with nearly one million unique model inputs. These synthetic images were then used to train a convolutional neural network, which regressed the two Cartesian coordinates of the patient shift. The trained model was then validated on approximately 260,000 held-out test set inputs with simulated shifts.
Results: The results indicate that the model is accurate at detecting both the direction and magnitude of patient shifts with a mean Euclidian pixel-wise distance error of 0.92, meaning shift simulated for the patient is correctly identified to within 1.5mm on average. We also performed a binary grouping of measured shifts, by labelling shifts larger than 3 mm as ‘shift’ and ‘no shift’ otherwise. This was compared to the traditional gamma pass rate (3%/3mm criteria, 90% pass rate). Used in this manner, we find that our model outperforms the use of a gamma pass rate threshold for classifying the presence or absence of shifts with a specificity/sensitivity of 78/90% compared to 23/99% for the gamma method.
Conclusion: The results from this work indicate that our deep learning model trained on synthetic data can correctly detect and classify patient shifts in external beam radiation treatment. This is superior to the current gamma approach which has reduced specificity and does not give any directional information. Clinical implementation will be performed to provide additional patient-specific quality assurance. Future work will investigate phantom measurements with known shifts to validate the model. Also, this model can easily be trained on other error types to detect and classify other on-treatment errors.