Deep Learning-Based Fast Volumetric Imaging for Treatment Setup During Lung Radiotherapy
Y Lei*, M Axente, Z Tian, T Wang, J Roper, A Kesarwala, K Higgins, J Bradley, T Liu, X Yang, Winship Cancer Institute of Emory University, Atlanta, GA
Presentations
TH-E-BRC-2 (Thursday, 7/14/2022) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Ballroom C
Purpose: To demonstrate the feasibility of 3D image volume formation from a pair of simultaneously acquired perpendicular 2D projection images (kV-MV pair). This would provide a potential solution of fast imaging guidance for daily treatment setup of breath-hold (BH) radiotherapy (RT).
Methods: We introduce VolNet, a novel deep learning-based method capable of inferring a 3D image from a perpendicular kV-MV pair. VolNet consists of four modules: 1) a feature extractor for kV-MV feature extraction, 2) an alignment module to map projection images through a Cartesian coordinate system, 3) a downsizing module to include more global semantic information, and 4) an estimation module for volumetric image projection. Fifteen lung patients received both a 4D CT and a deep inspiration BH CT at simulation. This patient-specific model was trained by ray tracing through each of the phases in the 4D CT datasets. The information used to simulate orthogonal 2D simultaneous kV-MV paired projections, which was then used to generate 3D images across the motion range captured in the 4DCT. The 3D dataset was then compared to the simulation BH CT.
Results: Within the body contour, the mean absolute error (MAE), peak-signal-to-noise-ratio (PSNR) and normalized cross correlation (NCC) were 64.4 ± 17.5 HU, 25.3 ± 3.2 dB and 0.95 ± 0.02 between the VolNet-generated CT 3D datasets and the ground truth BH CT images.
Conclusion: This proof-of-concept study provides an alternative to CBCT for fast volumetric imaging to facilitate the daily treatment setup of BH RT. Future work includes using the VolNet model to generate the treatment BH CBCT based on the acquired kV-MV pair, and testing against the actual setup scan used in the patient alignment for BH RT.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- RT X-Ray Imaging: combined KV/MV imaging
Contact Email