Quality assessment of compressed and resized medical images based on pattern recognition using a convolutional neural network

Abstract

Given the explosive growth of the amount of medical image data being produced and transferred over networks every day, employing lossy compression and other irreversible image operations is inevitable. As expected, irreversible image coding may decrease image fidelity by introducing undesired artifacts, which may lead to an invalid diagnosis.The purpose of this study is to propose a no-reference model of assessing the quality of a degraded medical image resulting from irreversible coding, based on pattern recognition with the use of a convolutional neural network (CNN). This deep neural network consists of six convolutional layers followed by two fully connected ones for the final image classification. Such network geometry is a common choice for image classification problems nowadays. We aim to construct a model that is specialized for medical images and could serve as a predictor of image quality for algorithm performance analysis. This technique uses a CNN to classify shapes of randomly chosen grayscale intensities.The shapes and grayscale shadings were chosen with the intention to mimic structures and edges appearing in a medical image. Using the accuracy of a classifier, we attempt to quantitatively measure how the information content in an image deteriorates after applying irreversible operations and how this loss of information affects the ability/inability of the neural network to recognize the shapes. The technique may be used to study the performance of irreversible image coding techniques. Two irreversible operations are employed for image degradation: compression and interpolation. We show the difference of image quality resulting from JPEG and JPEG2000 compression algorithms followed by scaling using several interpolation techniques. The main result of this work is the development of a model to quantitatively measure image quality based on pattern recognition using a deep neural network. The presented model of quantitative assessment of medical image quality may be helpful in determining the thresholds for irreversible image post-processing algorithms parameters (i.e. quality factor in JPEG) in order to avoid misdiagnosis. Further investigation of this problem will involve a connection of the introduced method with specific pathologies and various medical image modalities.