Background
Rheumatoid arthritis (RA) is a chronic disease, in which inflammation at the joint may lead to erosions (i.e. cortical interruptions) [
1,
2]. Interruptions in the cortical bone surface are often accompanied with underlying trabecular bone loss [
3‐
5]. The presence, size and number of cortical interruptions within a joint, and the number of joints affected, are each associated with poor functional outcome and predictors of further progression of structural damage [
3,
6,
7]. The quantification of interruptions on conventional radiography (CR) is considered the gold standard in clinical practice, however, it has a lower sensitivity compared to ultrasound, computed tomography (CT) and magnetic resonance imaging (MRI) [
8‐
10].
High-resolution peripheral quantitative CT (HR-pQCT) is a low-dose imaging technique that is able to assess the three-dimensional (3D) bone structure at the micro-scale (82 μm nominal isotropic voxel size) of the peripheral skeleton in vivo [
11]. Multiple studies have reported results on the visual inspection of the presence, number and size of interruptions with underlying trabecular bone voids in finger joints of patients with RA using HR-pQCT [
12‐
22]. Excellent intra- and inter-rater reliability have been reported, but in all these studies only relatively large cortical interruptions were scored (mean diameter > 1.5 mm) [
5,
13‐
15,
21,
23,
24]. In an earlier study, we showed that the inter-operator reliability is fair when visually scoring smaller cortical interruptions [
25].
In addition, the quantification of interruption volume was primarily based on simple distance measures on a two-dimensional (2D) slice [
5,
14,
15,
17,
18,
23]. A more extensive method is the 3D automated volume determination developed by Töpfer et al. [
21]. However, in this method the location of the interruption still has to be visually identified by an operator. In addition, this volume determination was performed in large interruptions (average volume: 9.3mm
3).
We have therefore developed a semi-automated algorithm that reliably detects small cortical interruptions (with a diameter ≥ 0.246 mm) [
26]. In addition, we showed that interruptions with a diameter of ≥0.41 mm detected on HR-pQCT were also detected on μCT, the 3D gold standard [
27]. However, this algorithm only analyzed the presence of an interruption in the cortex and did not consider the underlying trabecular bone loss as part of the total interruption volume. This is important because not only the presence but also the size of cortical interruptions (which includes the trabecular bone void) are associated with poor functional outcome and predictors of further progression of structural damage [
3,
6,
7].
Furthermore, the reproducibility of our algorithm on scan/re-scan with repositioning in-between the scans has not yet been tested in the standard workflow of the HR-pQCT scanner. Only the effect of the operator was investigated and not the influence of re-positioning of the hand nor the effect of image quality (noise and motion artifacts) in addition to the effect of the operator. Two previous studies tested the reproducibility on scan/re-scan data in the standard workflow of the HR-pQCT scanner for structural and density parameters in metacarpal heads [
13,
28]. The density parameters showed precision errors of ≤2%, but for trabecular and cortical structural parameters, precision errors up to 33% were found [
13,
28]. However, precision errors of the cortical parameters were only tested in healthy controls and not in early arthritis patients. Moreover, the phalangeal base was not included as part of the metacarpophalangeal (MCP) joint.
Therefore, the aims of this study were: 1) to extend our algorithm for detection of cortical interruptions with underlying trabecular bone void volume detection, 2) to evaluate the precision errors of our algorithm to detect cortical interruptions and its volume using scan/re-scan data in the standard workflow of the HR-pQCT scanner, and 3) to evaluate the precision errors of the trabecular and cortical density and micro-structure parameters in the standard workflow of the HR-pQCT scanner.
Discussion
In this study, we calculated the precision errors of our extended algorithm for detection of cortical interruptions and underlying trabecular bone void volume in MCP joints on scan/re-scan HR-pQCT data with repositioning in-between the scans in early arthritis patients. In addition, we calculated the precision errors for the bone density and micro-structure parameters. Reproducibility of our algorithm was excellent (ICCs ≥0.82), especially for the interruption volume (ICC 0.99). Reproducibility for the bone density and micro-structure parameters was also excellent (ICCs ≥0.84). Bland-Altman plots showed no systematic error in the reproducibility of our algorithm, bone density and bone micro-structure parameters. The reproducibility LSCSD value per joint was 4.2 for number of interruptions, 5.8 mm2 for interruption surface, and 3.2 mm3 for interruption volume.
The intra-operator reproducibility LSC
SD value of the algorithm for the interruption volume was higher in our study than the intra- and inter-operator LSC
SD reported by Töpfer et al. for single interruptions (LSC 3.2 mm
3 versus 1.4 mm
3 and 2.1 mm
3, respectively) [
21]. The study from Töpfer et al. differed in several aspects from ours. They analyzed a selection of interruptions in one dataset on its volume by two operators. In contrast, we used scan/re-scan data and included all interruptions, irrespective whether they were detected on the first scan but not on the second scan and vice versa. These aspects will lead to higher reproducibility errors. By excluding the effect of the rescanning (i.e. intra-operator reliability), the LSC
SD value was comparable to the study of Töpfer et al. (LSC 1.9mm
3 versus 1.4mm
3) [
21].
In our study, we also included the phalangeal base, thus creating a larger scan region for analyzing bone density and micro-structure parameters. This did not affect the precision errors, except for Ct.Po, which was substantially lower compared to a previous study (8.7% versus 27.7%) [
28]. The precision errors of the other parameters obtained in our study were similar as in previous studies [
13,
28]. In our study, the precision errors (CV
RMS) were generally below 2% for the bone density parameters (except for Tb.BMD), below 5% for the trabecular bone parameters (except for Tb.SpSD), below 10% for the cortical bone parameters. The precision errors are also comparable as observed in radius and tibia scans [
11].
The mean number of cortical interruptions and interruption surface per joint detected in this study (3.1 and 4.2 mm
2, respectively) were substantially lower than in our previously reported study using the same algorithm (9.5 and 13.5 mm
2, respectively) [
27]. In our previous study, anatomic specimens from high-aged subjects (mean 85.1 years) were used with a low BMD (vBMD of the joints: 245 mgHA/cm
3 versus 338 mgHA/cm
3 in this study). In the present study, the bone is better mineralized and therefore these voxels are less likely to represent non-bone voxels after segmentation and, hence, a lower number of interruptions was found.
The mean volume of the interruptions detected with the algorithm was substantially lower compared to previous studies that investigated volumes of interruptions in 3D (1.1 mm
3 versus > 4 mm
3), confirming that our algorithm enables the detection of much smaller interruptions [
13,
21,
24].
Our study has several limitations. First, with our algorithm, the trabecular void volume underlying the cortical interruption that can be detected is limited to a depth of 4 mm. This means that the algorithm underestimates the volume of interruptions with a depth greater than 4 mm. However, 4 mm is approximately half the width of the metacarpal head. Hence, such interruptions are not the primary focus of research with HR-pQCT, because these large interruptions can also be detected by other imaging techniques with lower resolution. For example, “small” interruptions, i.e. < 10 mm
3 are occasionally missed with MRI [
12]. Thus, our algorithm can best be used for studies with HR-pQCT aiming at early detection of structural damage, i.e. small interruptions, in patients with RA. Second, our algorithm requires manual correction of the outer margin of the contours in case of large cortical interruptions and motion artifacts which can make the analysis time consuming [
26]. However, this correction is also advised by the manufacturer for the standard evaluation protocol for assessment of the bone density and micro-structure parameters. Further automation of the outer contour would improve the applicability. The strength of our algorithm is that it is developed within the scanner software (IPL). Therefore, the algorithm can be easily implemented to other scanners.
The current investigation of the reproducibility of the algorithm and the extension of underlying trabecular bone void detection was a next step in the validation of our algorithm in the detection of small cortical interruptions in finger joints by HR-pQCT. We found that the algorithm was highly reproducible, but still had substantial precision errors compared to the mean value detected. Therefore, the next step is to test this algorithm in clinical studies in order to determine its potential value in monitoring patients with RA, and discriminating patients with RA, preferably early in the disease course, from healthy controls.