(Tuesday, 7/16/2019) 10:00 AM - 10:30 AM
Room: Exhibit Hall | Forum 5
Purpose: On-treatment simultaneous kV/MV imaging have been proven to be a promising method to track intrafraction tumor motion. However, one of the challenges, especially for markerless tracking, has been the low quality of MV images. In this work, we developed a machine learning based algorithm to improve the MV image quality.
Methods: For data generation, 1376 kV/MV image pairs were acquired with a lung phantom using TrueBeam. Three 3D-ARC plans generated with different isocenters were used for treatment delivery. During beam delivery, kV/MV image pairs were continuously taken and automatically grabbed with VARIAN iTools Capture software. For the neural network model, a convolutional neural network (CNN) based architecture was adopted, specifically designed as ResNet structure with Inception as the basic modules. Each Inception block contained multi-scale convolution using kernels from size 1 to size 13. The acquired MV images were used as input, and the output was compared with the corresponding kV images at the same gantry angle, with mean squared error as the loss function. Adam algorithm was used for the optimization, and 200 epochs were run for the network training. 80% of the total data were used for training the network, and 20% for the validation. Image quality was analyzed with contrast-to-noise ratio (CNR), where all images were normalized to maximum value of 1 before analysis.
Results: After the CNN model was applied to the test MV images, the quality of the output images was significantly improved, with much less noise and strong contrast by visual judgment. The time cost for processing one image was 20ms, introducing little latency on real-time tracking workflow. Quantitative analyses showed 48±13% increase in CNR around the embedded tumor.
Conclusion: The noise and contrast of MV images have been significantly improved after the CNN processing, which could lead to more robust motion tracking.
Funding Support, Disclosures, and Conflict of Interest: Research is supported by Varian, and from MSK Cancer Center Support Grant/Core Grant (P30 CA008748).
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
