Synthetic Lumbar fracture CT images from X-ray base on PR-GAN

Background: Lumbar fractures (LF) is a common disease that can cause spinal nerve damage. X-ray and CT are common methods to detect LF, but each has its own disadvantages. Objective: To develop a neural network for CT image generationform X-ray and to verify whether the synthetic CT images of LF affect clinical diagnosis. Materials and Methods: A total of 2528 of X-ray and CT images from 650 patients were used to develop the Positional Attention Resnet GAN ( PR-GAN) (two X-ray images and two CT images for each patient). Result: For the sagittal lumbar fractures (SLF) group, the SSIM and PSNR values were 0.773 and 28.463. For the coronal lumbar fractures (GLF) group, the SSIM value was 0.753 and the PSNR value was 27.0911. The SSIM value and PSNR value of sagittal region of interest (SROI) was 0.8979and 29.6234respectively. There was no statistically significant difference between the CT images synthesized by PR-GAN and real CT images in diagnosing LF (sagittal P=0.5847, coronal P=0.1686). Additionally, there was no significant difference in diagnosing fractures protruding into the spinal canal (P=0.3616). Conclusions: The present study developed a disease-specific PR-GAN which yielded markedly enhanced image quality compared to the traditional CycleGAN. Spine surgeons can achieve accurate diagnoses of LF based on the synthesized CT images.