Room: AAPM ePoster Library
Purpose: The purpose of this study is to develop a deep-learning technique to predict dual-energy (DE) X-ray subtraction images from single-energy kV X-ray fluoroscopic images for improving tumor localization accuracy during radiation therapy.
Methods: We first generated 36,000 training datasets of artificial dual-energy X-ray fluoroscopic images of lung area at two energy levels with different imaging angles and regions to simulate possible variations in imaging conditions for radiation therapy. Recurrent residual convolutional neural networks based on U-Net were then trained to predict DE subtraction images of bone and soft tissue respectively from single-energy X-ray images. We calculated the peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) values for evaluating the predicted DE images using 10,000 test datasets. For further application, we also evaluated tumor tracking performance using the predicted soft tissue images.
Results: The experimental results showed that the predicted DE subtraction images were highly similar to the original DE subtraction images: PSNR = 32.08 ± 2.36 dB and SSIM = 0.9664 ± 0.0075 for bone images, while PSNR = 41.67 ± 1.71 dB and SSIM = 0.9931 ± 0.0005 for soft-tissue images respectively. The tumor tracking errors were 0.64 ± 0.57 pixel for the predicted soft-tissue images and 12.38 ± 18.86 pixel for the single-energy raw images, respectively.
Conclusion: The proposed deep-learning-based method can predict high-quality DE subtraction images with a standard single energy kV X-ray imaging device. The predicted DE subtraction images can thus improve target tumor localization accuracy without specialized hardware and extra imaging dose.