Click here to

Session: Motion Compensation in Adaptive Radiotherapy [Return to Session]

Synthetic Bone and Soft-Tissue Images From a Single Exposure Using Deep Learning for Real-Time, Markerless Tracking in Radiation Oncology

Tomi F. Nano1*, Charudatta Manwatkar2*, Hui Lin1*, Yannet Interian2*, and Dante P.I. Capaldi1* (1) University of California, San Francisco, San Francisco, CA (2) University of San Francisco


WE-C930-IePD-F2-5 (Wednesday, 7/13/2022) 9:30 AM - 10:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: Dual-energy (DE) x-ray imaging offers improved visualization of soft-tissue and can provide real-time tracking of hard-to-see tumors during stereotactic-body radiation therapy (SBRT). Conventional bone and soft-tissue DE images require multiple x-ray exposures at different energies at the cost of increased patient dose and treatment times. The purpose of this work is to develop a single-exposure deep-learning framework for generating synthetic DE images and evaluating their real-time tracking capabilities in a clinical radiation oncology workflow.

Methods: Generative Adversarial Network (GAN) pix2pix-models were developed to convert digitally reconstructed chest radiographs (DRRs) from a 120 kV CT scan to bone-only or soft-tissue-only images obtained by DE subtraction (using 80 kV and 140 kV DRRs). An anthropomorphic lung phantom with a soft-tissue target was used to generate a paired-dataset of 920 DRR to DE images with an 80/20 split between training/testing. Generator L1 cost function was used along with discriminator binary cross-entropy cost function for adversarial training. Image quality metrics between real/simulated DE images were measured using signal-difference-to-noise ratio (SDNR) and structural-similarity index (SSIM). Tracking evaluated by measuring a shift error defined as the difference between registration shifts between DE images (real or fake) to the corresponding DRRs over the target regions-of-interests (ROIs).

Results: Synthetic DE bone and soft-tissue GAN model training shows monotonic decrease loss values and results in SDNR and SSIM of 3.5 (x% of ground truth conventional DE images) and 0.84 respectively. Tumor tracking (TT) and spine tracking (ST) error using conventional (TT=0.43±0.24mm/ST=0.03±0.01mm) and synthetic (TT=0.44±0.22mm/ST=0.09±0.01mm) DE images were all within clinical tolerance.

Conclusion: Synthetic DE images generated from a single exposure using a GAN network provide the same tracking accuracy as conventional DE subtracted-images without increasing patient dose from multiple exposures and changing current clinical-workflow. Future work will evaluate tracking accuracy using synthetic DE images retrospectively on patient images.


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email