Author: Ming Dong, Carri K. Glide-Hurst, Behzad Hejrati, Joshua Pan, Yuhao Yan 👨🔬
Affiliation: Department of Computer Science, Wayne State University, Departments of Human Oncology and Medical Physics, University of Wisconsin-Madison, Department of Human Oncology, University of Wisconsin-Madison 🌍
Purpose: Upright patient positioners and vertical CT reduce tumor motion and stabilize internal anatomy during treatment delivery. Yet, to fully exploit the advantages of upright, translation of standard of care supine datasets will be required. We introduce UprightVision, a deep learning model generating 3D upright datasets and associated contours from inputted supine datasets.
Methods: UprightVision is a paired cycle-consistency conditional generative adversarial network (cGAN) that inputs supine datasets to predict displacement vector fields (DVFs), transforming supine datasets and contours to the upright orientation. To address shape changes, a novel loss function was implemented incorporating guiding contours (e.g., external and bladder for pelvis), paired supervision, DVF smoothness regularization, and cycle-consistency. UprightVision was compared to conventional cGAN that directly predicts upright datasets from supine. Matched pair supine and upright pelvis spoiled-GRE MRI of 11 healthy volunteers (6 Male, 5 Female) were acquired within 40 minutes on a 0.6-Tesla MRI. Upright MRIs were rigidly registered to supine based on the lumbar spine. Six-fold cross validation was performed. Model performance was evaluated via peak-signal-to-noise ratio (PSNR), structural similarity (SSIM), and learned perceptual image patch similarity (LPIPS) between predicted and real upright MRI followed by Mann-Whitney U-tests. Dice similarity coefficients (DSCs) of predicted upright contours were calculated.
Results: Subjects oriented upright demonstrated anterior-posterior elongation of bladder which was reflected in the synthetic upright MRI for both models. PSNR (d.B.) was 15.3±1.1, 14.9±1.2, SSIM was 0.44±0.05, 0.44±0.06, and LPIPS was 0.15±0.04, 0.11±0.03 for cGAN and UprightVision, respectively. UprightVision had comparable PSNR (Δ~3%, p=0.01) but better LPIPS (Δ~27%, p=0.001). However, UprightVision retained image features and DSCs were 0.96±0.01 and 0.71±0.18 for external and bladder, respectively. UprightVision applied to CT data was also demonstrated.
Conclusion: Feasibility of UprightVision has been demonstrated. Next steps include validation in larger cohorts and implementing modality-agnostic models for widespread implementation in upright radiotherapy.