An automated algorithm for the detection of cortical interruptions on high resolution peripheral quantitative computed tomography images of finger joints

Objectives: To introduce a fully-automated algorithm for the detection of small cortical interruptions (≥0.246mm in diameter) on high resolution peripheral quantitative computed tomography (HR-pQCT) images, and to investigate the additional value of manual correction of the automatically obtained contours (semi-automated procedure). Methods: Ten metacarpophalangeal joints from seven patients with rheumatoid arthritis (RA) and three healthy controls were imaged with HR-pQCT. The images were evaluated by an algorithm according to the fully- and semi-automated procedure for the number and surface of interruptions per joint. Reliability between the fully- and semi-automated procedure and between two independent operators was tested using intra-class correlation coefficient (ICC) and the proportion of matching interruptions. Validity of single interruptions detected was tested by comparing it to visual scoring, as gold standard. The positive predictive value (PPV) and sensitivity were calculated. Results: The median number of interruptions per joint was 14 (range 2 to 59) and did not significantly differ between the fully- and semi-automated procedure (p = 0.37). The median interruption surface per joint was significantly higher with the fully- vs. semi-automated procedure (respectively, 8.6mm2 vs. 5.8mm2 and 6.1mm2, p = 0.01). Reliability was almost perfect between the fully- and semi-automated procedure for both the number and surface of interruptions (ICC≥0.95) and the proportion of matching interruptions was high (≥76%). Also the inter-operator reliability was almost perfect (ICC≥0.97, proportion of matching interruptions 92%). The PPV ranged from 27.6% to 29.9%, and sensitivity from 69.7% to 76.3%. Most interruptions detected with the algorithm, did show an interruption on a 2D grayscale image. However, this interruption did not meet the criteria of an interruption with visual scoring. Conclusion: The algorithm for HR-pQCT images detects cortical interruptions, and its interruption surface. Reliability and validity was comparable for the fully- and semi-automated procedures. However, we advise the use of the semi-automated procedure to assure quality. The algorithm is a promising tool for a sensitive and objective assessment of cortical interruptions in finger joints assessed by HR-pQCT.