No-Reference Image Quality Metrics for Structural MRI

Abstract

Neuroimagery must be visually checked for unacceptable levels of distortion prior to processing. However, inspection is time-consuming, unreliable for detecting subtle distortions and often subjective. With the increasing volume of neuroimagery, objective measures of quality are needed in order to automate screening. To address this need, we have assessed the effectiveness of no-reference image quality measures, which quantify quality inherent to a single image. A data set of 1001 magnetic resonance images (MRIs) recorded from 143 subjects was used for this evaluation. The MRI images were artificially distorted with two levels of either additive Gaussian noise or intensity nonuniformity created from a linear model. A total of 239 different quality measures were defined from seven overall families and used to discriminate images for the type and level of distortion. Analysis of Variance identified two families of quality measure that were most effective: one based on Natural Scene Statistics and one originally developed to measure distortion caused by image compression. Measures from both families reliably discriminated among undistorted images, noisy images, and images distorted by intensity nonuniformity. The best quality measures were sensitive only to the distortion category and were not significantly affected by other factors. The results are encouraging enough that several quality measures are being incorporated in a real world MRI test bed.

The spatial statistics of structural magnetic resonance images: application to post-acquisition quality assessment of brain MRI images

“…They can be categorized into reference methods [12], [13], [14], [15], [16] and no-reference methods [17], [18], [19], [20], [21]. Detailed review of image quality evaluation for a general class of images and MRI images can be found in [22], [23], [24], [25], [26], [27], [28], [29].…”

“…Many current quality assessment methods such as [17], [19] adopt specific type of distortion, considered as common, to evaluate the image under consideration. This approach can be said to be biased towards specific distortion because all possible distortions combine with ideal features to manifest as image attribute [18].…”

Bayesian framework inspired no-reference region-of-interest quality measure for brain MRI images

“…The use of phantom is not an effective approach for comparative performance evaluation because it lacks the natural anatomical variability and image acquisition artifacts that are usually encountered in real images [82]. Furthermore the algorithms for these proposed methods are not readily available from the authors and it is difficult to faithfully implement the techniques without direct interaction with the authors [41]. Having regard to these shortcomings we describe three characteristics that distinguish our proposed methods from existing post-quality assessment methods for brain MRI images.…”

“…The report in [41] apply analysis of variance (ANOVA) algorithm to assess the variation of several quality measures with different levels of distortions. The authors in [42] combine the detection of artifacts and estimation of noise level to measure image quality.…”

“…There is risk of ambiguity in quality measures that are based on the use of ANOVA models [48]. ANOVA limits the performance of the technique proposed in [41] to the detection of distortion. It cannot transform the different levels of distortion into a quality index.…”