Purpose: The real-time tumor-tracking radiotherapy system consists of two x-ray tubes and color image intensifiers. The imaging dose for real-time tumor monitoring was a problem. The purposes of this study were to develop synthetic color fluoroscopic image generation model using the low dose images and evaluate the tracking accuracy for synthetic images.
Methods: Two patients who underwent SBRT for a lung tumor were enrolled in this study. Seven fields were created for treatment plan. For each field, high and low dose color fluoroscopic images were acquired at 30 fps for 10 s from two directions at treatment simulation. For each direction, color fluoroscopic images were randomly selected 200 images for train and 100 images for test data. Using these data, the model that generates the synthetic color fluoroscopic images was created based on the deep learning. Second, using created model, the synthetic color fluoroscopic images were generated using low dose fluoroscopic images acquired at 1fx treatment. Structural similarity (SSIM) were calculated to evaluate the differences between high, low dose and synthetic fluoroscopic images and the difference between the position for synthetic and low dose images was calculated to evaluate the tracking accuracy.
Results: For all fields and directions, the mean Â± standard deviation (S.D.) of SSIM between high and synthetic high dose images was 0.91 Â± 0.03. Those between low and synthetic high dose images was 0.62 Â± 0.06. The synthetic high dose fluoroscopic images were significant similar to a high dose images. The mean Â± S.D. of tracking accuracy 0.10 Â± 0.05 mm, 0.18 Â± 0.14 mm and 0.14 Â± 0.13 mm in LR, AP and SI directions, respectively.
Conclusion: Our results showed the tracking of fiducial markers using synthetic images might be possible. Further, reduction of the imaging dose for real-time tumor monitoring would be expected.