An image-based technique to assess the perceptual quality of clinical chest radiographs

Current clinical image quality assessment techniques mainly analyze image quality for the imaging system in terms of factors such as the capture system modulation transfer function, noise power spectrum, detective quantum efficiency, and the exposure technique. While these elements form the basic underlying components of image quality, when assessing a clinical image, radiologists seldom refer to these factors, but rather examine several specific regions of the displayed patient images, further impacted by a particular image processing method applied, to see whether the image is suitable for diagnosis. In this paper, the authors developed a novel strategy to simulate radiologists' perceptual evaluation process on actual clinical chest images.

Ten regional based perceptual attributes of chest radiographs were determined through an observer study. Those included lung grey level, lung detail, lung noise, rib-lung contrast, rib sharpness, mediastinum detail, mediastinum noise, mediastinum alignment, subdiaphragm-lung contrast, and subdiaphragm area. Each attribute was characterized in terms of a physical quantity measured from the image algorithmically using an automated process. A pilot observer study was performed on 333 digital chest radiographs, which included 179 PA images with 10:1 ratio grids (set 1) and 154 AP images without grids (set 2), to ascertain the correlation between image perceptual attributes and physical quantitative measurements. To determine the acceptable range of each perceptual attribute, a preliminary quality consistency range was defined based on the preferred 80% of images in set 1. Mean value difference (μ(1) - μ(2)) and variance ratio (σ(1) (2)/σ(2) (2)) were investigated to further quantify the differences between the selected two image sets.

The pilot observer study demonstrated that our regional based physical quantity metrics of chest radiographs correlated very well with their corresponding perceptual attributes. The distribution comparisons, mean value difference estimations, and variance ratio estimations of each physical quantity between sets of images from two different techniques matched our expectation that the image quality of set 1 should be better than that of set 2.

The measured physical quantities provide a robust reflection of perceptual image quality in clinical images. The methodology can be readily applied for automated evaluation of perceptual image quality in clinical chest radiographs.