Impact of Synthetic and Real CT Training Datasets for Deep Learning Super Resolution Models

Super resolution can enhance the sharpness and resolution of low-dose micro-computed tomography (micro-CT) scans. This allows larger physical samples to be imaged, acquisition times to be reduced and radiation exposure to be minimized. However, acquiring training data for supervised learning can be tedious and time-consuming. In this paper, we present two methods for generating low-resolution (LR) samples from high-resolution (HR) reference data efficiently. Experimental results, evaluated using the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), Feature Similarity (FSIM) and Learned Perceptual Image Patch Similarity (LPIPS) metrics, suggest that synthetic images, generated using a lightweight Convolutional Neural Network (CNN), can provide bone structure image features comparable to those derived from real data. This supports their use in training super resolution models.