Please Don't Move—Evaluating Motion Artifact From Peripheral Quantitative Computed Tomography Scans Using Textural Features

Abstract

Most imaging methods, including peripheral quantitative computed tomography (pQCT), are susceptible to motion artifacts particularly in fidgety pediatric populations. Methods currently used to address motion artifact include manual screening (visual inspection) and objective assessments of the scans. However, previously reported objective methods either cannot be applied on the reconstructed image or have not been tested for distal bone sites. Therefore, the purpose of the present study was to develop and validate motion artifact classifiers to quantify motion artifact in pQCT scans. Whether textural features could provide adequate motion artifact classification performance in 2 adolescent datasets with pQCT scans from tibial and radial diaphyses and epiphyses was tested. The first dataset was split into training (66% of sample) and validation (33% of sample) datasets. Visual classification was used as the ground truth. Moderate to substantial classification performance (J48 classifier, kappa coefficients from 0.57 to 0.80) was observed in the validation dataset with the novel texture-based classifier. In applying the same classifier to the second cross-sectional dataset, a slight-to-fair (κ = 0.01–0.39) classification performance was observed. Overall, this novel textural analysis-based classifier provided a moderate-to-substantial classification of motion artifact when the classifier was specifically trained for the measurement device and population. Classification based on textural features may be used to prescreen obviously acceptable and unacceptable scans, with a subsequent human-operated visual classification of any remaining scans.

Introduction

It is widely acknowledged that computed tomography scans are susceptible to methodological issues such as partial volume effect and beam hardening, operating errors such as positioning errors, and movement of the individual during a scan, the last of which manifests as movement artifact (1). Although some methodological issues are unavoidable, operator errors can be minimized with training, and movement artifacts can be rectified by rescanning. However, rescanning is not always desirable or practical given the additional radiation dose and time required. Moreover, rescanning may occasionally not be required as it is well established that a limited amount of visible motion artifact does not invalidate a scan 1, 2, 3, 4, 5. Anecdotally, children are particularly fidgety (1) and the operator is often left with a scan that has conspicuous signs of motion artifact (streaking and discontinuity of cortical structure 1, 2, 3, 4, 5, 6) and the decision of whether or not to rescan. The acceptable levels of motion artifact have been defined for both high-resolution 2, 3, 4, 5 and regular computed tomography (1). However, the method developed for regular peripheral computed tomography (pQCT) (1) is applicable only to bone shafts and not to distal or proximal bone sites with narrow cortices.

The effects caused by motion artifact on the image reconstruction in computed tomography have been explored by Yang et al (6), but even with this comprehensive understanding of motion-caused artifacts, a consistent standard operating procedure for motion artifact quantification has yet to emerge. The approaches used to detect motion artifact include subjective visual scaling 1, 4, 5, 7, quantification of translation and rotation based on the measured sinogram (measured projections) 2, 3, 4, and exploring analysis results utilizing varying analysis thresholds (1). The objective quantification of translation based on the sinogram can only be done before reconstructing the image with filtered back projection (2). All computed tomography devices measure the sinogram, but the sinogram cannot be extracted from some devices and hence is not an applicable method in all cases. Although the agreement between raters for visual scaling is rather good for normal and high-resolution pQCT 1, 4, 5, an automated method may prove helpful in optimizing consistency and reliability, particularly in very large datasets and multisite studies.

Because visual scaling is based on the appearance of the image after reconstruction, and the motion artifact typically includes streaking and discontinuities of the bone cortex (6), textural analysis could provide a suitable option for the semiquantitative detection of motion artifact from computed tomography scans in the absence of the measured sinogram. Many textural analysis approaches capturing various properties of texture in medical imaging have been presented in the literature (e.g., reviewed in References 8, 9). Of the various approaches, local binary patterns (LBPs) appear particularly well suited for motion artifact detection because LBP capture streaking in images (10) have been successfully applied in an automated radiographic image measurement site annotation in the past (11) and are computationally efficient to implement (10). However, LBP has yet to be tested as a feature to quantify motion artifact.

The purpose of the present study was to develop and validate automated motion artifact classifiers to quantify motion artifact in pQCT scans. Specifically, the aim was to evaluate whether LBP could provide a better classification performance using visual inspection as the ground truth compared to applying current state-of-the art objective motion artifact measures as classification features.