Room: Exhibit Hall
Purpose: To demonstrate breathing phase identification using the cine-mode frames of the electronic portal imaging device (EPID) acquired during radiation beam delivery and deep networks with transfer learning.
Methods: We set up a 4D lung phantom with programmed cyclic target movement in 3cm range and 4 sec period to simulate breathing motion and collected 10 phases in 4DCT of the phantom. We delivered lung SBRT treatment beams (6X, 5cmÃ—5cm) using Varian TrueBeam to the phantom at various gantry angles with cine-mode EPID frame acquisition at 6Hz.We labeled each EPID frame by the phase number for network training and validation. The phase numbers were tracked using the sinusoidal patterns in the time domain-concatenated EPID frames. We applied the well-trained natural image classification network, AlexNet, but replaced and re-learned the last 3 layers to fit in our study. The input to the network was a set of 3 consecutively-acquired EPID frames filling the RGB channels, and the outputs were the activation values of 10 breathing phases based on the last EPID frame.
Results: In 10 epochs of training on 70% of the EPID frames, the network achieved >90% training accuracy and >70% prediction accuracy on the test images for all projection angles. Both training and prediction accuracies become 99.5%, if error is defined as deviation by more than one phase. In fact, since the motion is continuous, EPID frames acquired consecutively are highly similar but may be labeled as two consecutive phases due to discretization.
Conclusion: Transfer learning is a powerful method to utilize well-trained ImageNet models for medical applications. The learned deep network encodes phase features and circumvents EPID modeling challenges in resolving phase information. With accurate and fast prediction, the deep learning network is a promising tool for real-time volumetric imaging.
Funding Support, Disclosures, and Conflict of Interest: with funding support through Varian MRA