Evaluation of Full-Reference Image Quality Assessment Metrics for Artifact Sensitivity in Lung CT Images for Radiotherapy

Abstract

In adaptive radiotherapy, the quality of images acquired during treatment has to be determined. For this purpose, the sensitivity of full-reference image quality assessment metrics to CT artifacts in lung images is investigated. From an examplary patient image, distortion-free and distorted images were created by simulating artifacts in the image and sinogram space. Two experiments were conducted to analyze the metric sensitivity to realistically distorted images and the behavior for different distortion levels. The results show that pixel-level intensity difference (PID) metrics are sensitive to global transformations, such as rotation, shift, and motion. Gradient-based metrics like gradient magnitude similarity deviation (GMSD) react to blurring, truncation and aliasing artifacts, whereas a histogram-based metric like Jensen–Shannon divergence (JSD) can be used for noise, breathing, metal and other artifacts. A combination of a PID metric, GMSD and JSD provides a comprehensive quality assessment in the context of longitudinal image monitoring along radiotherapy treatment.