Improving Sparse-view CT Image Quality using fractal images with deep learning

Computed tomography (CT) has brought tremendous benefits to medicine, but it has also increased medical radiation exposure. Sparse-view CT, which reduces the number of projections, is one way to reduce radiation exposure. However, reducing the number of projections causes streak-like artifacts in the reconstructed image. Studies of sparse-view CT using deep learning have reported higher image quality than conventional reconstruction methods. In general, deep learning requires many real images, which can be difficult to collect for medical images due to legal and ethical issues. The purpose of this study is to investigate whether fractal images can be used to improve image quality even when the number of real images is reduced. Fractal images were generated using IFS in Python, and the CT images were abdominal CT images obtained from the CHAOS dataset. They were collected 180° with parallel beams to obtain 36-projection sinograms. Input images (sparse-view images) were then generated by FBP reconstruction. The number of CT images was varied from 10 to 1000 and compared using three different networks: FBPConvNet, W-Net, and W-Net with pre-training. The peak signal-to-noise ratio (PSNR) and the structural similarity index (SSIM) were used to evaluate the accuracy of the reconstructed images. The use of fractal images was found to provide better image quality than training from CT images alone. W-Net with pre-training had the best results, suggesting that the number of real images can be reduced.