Is two better than one? Super resolution for dual-layer radiography with convolutional neural networks

Abstract

Traditionally the image quality of a particular imaging system has been limited by the particular protocol and hardware. These limits appear as bounds to the resolution, accuracy, acquisition and processing time, and much more. More recently, deep learning and other artificial intelligence methods have emerged to overcome such bounds by outsourcing information from previous scans from the same, or similar imaging systems. Tasks such as segmentation, deconvolution, multi-modality models, and overall image quality have been greatly aided and have found broad clinical application in both therapy and diagnostic imaging. We propose that super resolution (SR) is another potential application, more specifically in the situation where multiple images of the same object are obtained at once. This study investigates the relative performance difference of single image inputs vs multilayer image inputs via a convolutional neural network (CNN), in particular for a dual-layer flat panel detector. We simulate the data acquisition process through the known modulation transfer function and noise power spectrum of the detector and aim to demonstrate that SR may greatly benefit from even a single additional lossy image. To verify the effectiveness and efficiency of multilayer image SR, we benchmark against a variety of classical CNN-based SR algorithms. Our tests demonstrate that SR for radiography is an attractive and readily realisable application to the image processing process.