Image Quality Assessment for Magnetic Resonance Imaging

Abstract

Image quality assessment (IQA) algorithms aim to reproduce the human's perception of the image quality. The growing popularity of image enhancement, generation, and recovery models instigated the development of many methods to assess their performance. However, most IQA solutions are designed to predict image quality in the general domain, with the applicability to specific areas, such as medical imaging, remaining questionable. Moreover, the selection of these IQA metrics for a specific task typically involves intentionally induced distortions, such as manually added noise or artificial blurring; yet, the chosen metrics are then used to judge the output of real-life computer vision models. In this work, we aspire to fill these gaps by carrying out the most extensive IQA evaluation study for Magnetic Resonance Imaging (MRI) to date (14,700 subjective scores). We use outputs of neural network models trained to solve problems relevant to MRI, including image reconstruction in the scan acceleration, motion correction, and denoising. Seven trained radiologists assess these distorted images, with their verdicts then correlated with 35 different image quality metrics (full-reference, no-reference, and distribution-based metrics considered). Our emphasis is on reflecting the radiologist's perception of the reconstructed images, gauging the most diagnostically influential criteria for the quality of MRI scans: signal-to-noise ratio, contrast-to-noise ratio, and the presence of artifacts.

Pytorch Image Quality: Metrics for Image Quality Assessment

“…Previous studies [36,9] showed non-linear relation between IQM scores and human accessors' scores. A common solution is to apply a non-linear regression by adopting the five-parameter modified logistic function [40].…”

“…Previous studies of IQMs showed their efficiency in reflecting the human perception of visual quality [35,36,37,9]. That brings a temptation to use the best-performing differentiable IQMs as the loss functions for a direct optimization of the image processing models [38].…”

“…Most methods are evaluated using a single feature extractor (typically, Inception Net [33]). A recent study on IQMs for Magnetic Resonance Imaging [9] showed that the choice of feature extractor plays a critical role, heavily influencing the performance of DB metrics. These result may inspire our users to experiment with various feature extractor options.…”

“…Instead, they are intended to be used to compare distributions of two image sets. However, we investigate DB metrics for pair-wise comparison of images following the computation strategy proposed in [9]. To encounter for the initial setting, we represent each image in a pair as a set of overlapping patches of size 96 × 96 with stride = 32, which allows us to reformulate the pair-wise comparison of images as a comparison of patch distributions.…”

“…At the time of writing this manuscript, our GitHub project receives over 4000 monthly views, having accumulated 640 stars and having been forked over 57 open-source projects. This tool has been actively used in the research community, including the development of new neural network architectures [8] and the large-scale medical image quality study [9].…”