BEST IN PHYSICS (IMAGING): A GAN-Based Technique for Synthesizing Realistic Respiratory Motion in the 4D-XCAT Phantoms
Y Chang*, Z Jiang, P Segars, Z Zhang, F Yin, L Ren, Duke University Medical Center, Cary, NC
Presentations
TU-C-TRACK 3-2 (Tuesday, 7/27/2021) 1:00 PM - 2:00 PM [Eastern Time (GMT-4)]
Purpose: Synthesize realistic and controllable respiratory motions in the 4D-XCAT phantoms by developing a GAN-based deep learning technique.
Methods: A motion generation model was developed using bicycle-GAN with a novel 4D generator. The model input includes the end-of-inhale(EOI) phase of 4D-CTs and a Gaussian perturbation, and it generates inter-phase deformable-vector-fields(DVFs), which are compounded and applied to the EOI image to generate 4D images. A separate machine-learning respiratory motion amplitude control model was built to predict the perturbation needed to generate a specific motion amplitude. The models were trained and validated using 71 patient 4D-CT data and tested using 6 XCAT to generate 4D-XCAT phantoms with realistic respiration. The simulated 4D images were evaluated by Dice coefficients calculated for the lungs. The generated DVFs were evaluated by deformation energy, the correlations of DVFs and derived ventilation maps with those from the reference images. Results were compared to those from the original 4D-XCAT to demonstrate the improvement of motion realism.
Results: For the simulated 4D-CTs, the lung volume deviation was within 5.8%, and the Dice coefficients of the lungs reached at least 0.95. The cross-correlation of DVFs achieved 0.89±0.10/0.86±0.12/0.95±0.04 along the x/y/z directions in the testing group. The cross-correlation of ventilation maps derived achieved 0.80±0.05/0.67±0.09/0.68±0.13, and the Spearman's correlation was 0.70±0.05/0, 60±0.09/0.53±0.01, respectively, in the training/validation/testing groups. The generated 4D-XCAT phantoms presented similar deformation energy(0.36±0.07) as patient data(training: 0.33±0.17; validation: 0.22±0.14; testing: 0.36±0.14), compared with 0.07±0.01 in the original 4D-XCAT. The Dice coefficient between the original and generated 4D-XCAT phantoms of similar breathing amplitudes was 0.95. The motion amplitude control error was less than 0.5 mm.
Conclusion: The results demonstrate the efficacy of synthesizing realistic controllable respiratory motion in the 4D-XCAT phantom using the proposed method. This crucial development enhances the value of XCAT phantoms for various 4D imaging and therapy studies.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by the National Institutes of Health under Grant No. R01-CA184173 and R01-EB028324.
Handouts
Keywords
Phantoms, Deformation, Modeling
Taxonomy
Contact Email