Visual detection of cortical breaks in hand joints: reliability and validity of high-resolution peripheral quantitative CT compared to microCT

Peri-articular cortical breaks are one of the characteristic features of bone involvement in rheumatoid arthritis (RA) and predictors of further radiographic progression [1, 2]. Early detection of cortical breaks is an important indicator for intensifying treatment in order to modify the disease course [3]. In daily clinic, conventional radiographs (CR) are considered the gold standard for detection of cortical breaks in the hand joints in rheumatic diseases. CR is widely available, fast to perform, relatively cheap, and extensively validated, however its sensitivity to detect structural bone changes is low compared to computed tomography (CT), MRI and ultrasound [4‐7]. A novel, sensitive imaging technique is High-Resolution peripheral Quantitative Computed Tomography (HR-pQCT) [8]. HR-pQCT allows analysis of the cortical and trabecular microarchitecture of peripheral bones with an isotropic resolution of 82 micrometer (μm). This technique is now also applied for 3D assessment of the bone microarchitecture in the hand joints [8‐10]. A study by Stach et al. has demonstrated that HR-pQCT is more sensitive than CR in detecting cortical breaks in the hand joints in RA and also in healthy controls [8]. However, the resolution of the HR-pQCT images can be of the same order as the thickness of the cortical bone in finger joints. Due to partial volume effects, it is possible that thin cortices are falsely identified as breaks. Therefore, in particular with thin cortices, the reliability, sensitivity and specificity of the measurements might be impaired and depend on the reader’s perception.

The aims of this study were 1). to investigate the intra- and interreader reliability, and 2). to determine the sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of HR-pQCT in detecting cortical breaks in hand joints, using μCT images with a much higher resolution (18 μm) as gold standard. We were particularly interested in the methodology of identifying cortical breaks by HR-pQCT, not to study the clinical value of these cortical breaks. We hypothesized that HR-pQCT is a reliable and sensitive imaging method for identifying cortical breaks in hand joints compared to μCT.

The mean age of the donors was 85.1 ± 9.6 years, the medical history was unknown. Average vBMD was 203 mgHA/cm³ for MCP joints, 293 mgHA/cm³ for PIP joints and 245 mgHA/cm³ for the total joints. The scans of ten MCP and nine PIP joints with in total 152 quadrants were available for analysis. One PIP joint could not be evaluated due to a missing μCT scan. Furthermore, Reader 1 considered the quality of the images of the metacarpal head in one MCP joint as too low on μCT due to a protocol error during scanning. Therefore four quadrants were excluded in the analyses of the μCT images.

Table 1 shows the total number of cortical breaks each reader found on HR-pQCT and μCT images. The differences in scores between the first and second reading of Reader 1 were not statistically significant on HR-pQCT (139 vs 118 breaks, p = .064) and μCT (142 vs 156 breaks, p = .163). However, the difference in the mean score between HR-pQCT versus μCT was statistically significant (respectively, 129 and 149 breaks, p < 0.05). The total number of cortical breaks on HR-pQCT scored by Reader 1 (first reading) versus Reader 2 was not statistically significant (respectively, 139 versus 151 breaks, p = .288). On μCT, Reader 2 found significantly more breaks than Reader 1 (first reading) (241 vs 142 breaks, p < .001). In total 4 quadrants with large discrepancies between the readers were re-examined. Several reasons for discrepancy were identified:

Table 2 shows the intra- and interreader reliability based on the 152 quadrants on HR-pQCT and 148 quadrants on μCT that were evaluated. Intrareader reliability was moderate to substantial for the presence of breaks (HR-pQCT: k = 0.52 and μCT: k = 0.71) and for the number of breaks (HR-pQCT: ICC = 0.61 and μCT: ICC = 0.75).

Interreader reliability was fair to moderate for the presence of breaks (HR-pQCT: k = 0.37 and μCT: k = 0.45,) and for the number of breaks (HR-pQCT: ICC = 0.55 and μCT: ICC = 0.54). The values of PABAK were comparable (Table 2).

Sensitivity, specificity, PPV and NPV of HR-pQCT in the detection of cortical breaks with μCT as gold standard were calculated with the mean scores of Reader 1 (reading 1 and 2) and the score of Reader 2. The sensitivity was 81.6 %, specificity 64 %, PPV 81.6 %, and NPV 64 % respectively for Reader 1, and sensitivity 68.9 %, specificity 69.4 %, PPV 82.6 % and NPV 51.5 % for Reader 2 (Table 3 and Additional file 5).

In Fig. 1, several examples of cortical breaks on corresponding HR-pQCT and μCT images are presented. Panel A and B show a cortical break on both HR-pQCT and μCT. In panel C, a discontinuity of the cortex is found on HR-pQCT. However, it did not meet the definition of a cortical break applied in this study, because it was visible on one slice only, leading to discrepancy with the results from μCT. In panel D, a cortical break was detected on HR-pQCT, but not on μCT, where a thin cortical lining was seen.

This study is the first that reports on aspects of reliability and validity of detecting cortical breaks in hand joints using HR-pQCT with μCT as gold standard. Cortical breaks were found in all joints with both imaging modalities. Intrareader reliability of HR-pQCT and μCT was moderate to substantial, while interreader reliability was fair to moderate. The sensitivity of HR-pQCT in detecting cortical breaks was high (81.6 %).

In our study, only cortical discontinuities meeting the definition of a break, ie clearly visible cortical interruptions on at least 2 HR-pQCT (or 9 μCT) consecutive slices in two planes, were scored as a cortical break. They may have different pathological or physiological backgrounds, such as erosions or vascular channels [8, 9]. A formal classification system for defining breaks visualized on HR-pQCT and μCT is lacking. Histological examination is needed to provide more insight in the nature of these cortical breaks and for developing definitions.

In a previous study by Stach et al. an almost perfect and substantial intra- and interreader reliability was reported using HR-pQCT for grading bone lesions and discrimination between healthy individuals and RA patients (k = 0.82 and k = 0.75 respectively), but reliability on the presence and number of cortical breaks was not reported [8]. The precision in scoring abnormalities visualized with several imaging techniques varies widely, even by experienced readers, as has been demonstrated for example for scoring radiographs in RA (ICC ranged from 0.65 to 0.99) [14]. In general, lower values for interreader reliability in comparison with intrareader reliability are reported [15], corresponding to our findings. In our study, the breaks were scored visually, which is reader dependent. An automatic scoring algorithm, with detection of pre-defined definitions of breaks and executed automatically by the computer, could potentially improve reliability by minimizing reader interventions.

We investigated the sensitivity of HR-pQCT in detecting cortical breaks with μCT as the gold standard and found a high sensitivity (81.6 %). Unfortunately, no comparative studies are available. In contrast, two studies used HR-pQCT as the reference method for investigating the sensitivity to detect cortical breaks of other imaging modalities [9, 16]. These studies reported a sensitivity of 85.7 % for MRI, 60.9 % for CR, and 83–100 % for ultrasound with HR-pQCT as the reference method [9, 16]. We found a lower specificity of HR-pQCT in detecting cortical breaks (64 %) in comparison to sensitivity. A possible explanation for this could be a phenomenon attributed to a partial volume effect leading to a reduced cortical signal on HR-pQCT, giving the impression that a cortical break is present, whereas on μCT the cortex is intact. An example of this is shown in Fig. 1, panel D.

There are several limitations of this study. First, we evaluated whether the total number of breaks counted per quadrant corresponded between the two imaging modalities, but did not consider correspondence in exactly the same location. This might have led to an overestimation of the reliability. Second, we used fingers from cadaver specimens with unknown medical history and a relatively high mean age (85.1 years). Due to the old age of the donors, and the preservation in formalin, the cortices might become less mineralized [17]. The average vBMD of the specimens was 245 mgHA/cm³, which is some 20 % lower than the average in the normal population (>300 mgHA/cm³) [10, 18]. This may hamper the scoring of a cortical break on HR-pQCT. It is also possible that thin regions were falsely identified as a cortical break. However, the use of cadaveric specimens was essential as in vivo human subjects cannot be measured by μCT because of a long scanning time. Third, the cadaveric specimens had slightly different orientations in the HR-pQCT versus the μCT scanner. Despite the careful visual matching of the regions of interest on HR-pQCT and μCT, the angle at which the transversal images were viewed was slightly different in some joints and a cortical break might therefore be missed. Fourth, a discrepancy between the readers regarding the number of cortical breaks identified on μCT was noticed. μCT images provide much detail, and in particular very small cortical interruptions were not always picked up by Reader 1. This indicates that, when visually analyzing μCT images, more stringent definitions are necessary than when using HR-pQCT because of the higher resolution.

Cortical breaks were commonly visualized in hand joints with HR-pQCT and μCT. Reliability of both HR-pQCT and μCT was fair to moderate. HR-pQCT was sensitive to detect cortical breaks with μCT as gold standard. In spite of the limitations of our study, including the discrepancy of μCT results between the readers, we have shown that HR-pQCT is highly sensitive to detect cortical breaks with a fair to moderate reliability compared to μCT. Our findings need further evaluation, preferably with focus on histological analyses to clarify the nature of the breaks and to establish more reliable definitions and a classification system for analyzing cortical breaks on high-resolution CT images.

CR, conventional radiographs; CT, computed tomography; DICOM, digital imaging and communications in medicine; HR-pQCT, high-resolution peripheral quantitative computed tomography; ICC, intraclass correlation coefficient; MCP, metacarpophalangeal; PIP, proximal interphalangeal; PPV, positive predictive value; RA, rheumatoid arthritis; k, Cohen’s Kappa; μCT, microCT; μm, micrometer