Cross-modality Training Approach for CT Super-resolution Network

In this work, we propose a U-Net-based super-resolution neural network, SRU-Net, to create emulated high spatial resolution (eHR) CT images from low spatial resolution (LR) CT images. As resolution could be defined by the modulation transfer function in CT reconstruction, we propose the novel approach based on CT reconstruction kernels to create realistic multi-detector CT (MDCT) synthetic LR images from high-resolution cone-beam CT (CBCT) scans. Keeping a constant sampling grid size of 0.20 × 0.20mm2, we reconstruct two types of MDCT-like LR images and one corresponding HR image from the same CBCT raw data and train two models respectively. We validated the performance of the trained models on unseen LR CBCT images. We then applied the trained network to MDCT images. Mean squared error, structural similarity index measures and peak signal-to-noise ratio of two models show significant improvements (p < 0.001) in the eHR images.