External validation of a deep-learning based automatic contouring algorithm for the radius and tibia on first- and...

Current contouring software for high-resolution peripheral quantitative CT (HR-pQCT) requires manual correction, which can be subjective and time-consuming. In this study, we externally validated a deep learning (DL) model for automatic periosteal and endosteal contouring in second-generation HR-pQCT scans of the distal radius and tibia. Additionally, we compared the DL model's contours with uncorrected standard automatic periosteal and endosteal HR-pQCT contours (AUTO) in second-generation HR-pQCT scans and evaluated its performance in first-generation HR-pQCT scans. In 939 second-generation HR-pQCT scans, median Dice (DSC) and Jaccard (JSC) similarity coefficients for periosteal and endosteal DL contours were >0.92 and median average (ASSD) and maximum (HD) symmetric surface distances <1.15 mm when compared with our reference contours (AUTO with manual periosteal correction). Most HR-pQCT parameters differed significantly when comparing DL and reference contours, but differences were small (median between -0.6% and +6.6%) and correlations high (>0.78). Comparison with AUTO gave considerably larger maximum HDs and differences in HR-pQCT parameters - predominantly due to erroneous AUTO contours. In 300 first-generation HR-pQCT scans, median DSC and JSC were >0.98 and median ASSD and HD <0.49 mm when comparing periosteal DL contours with manually-guided snake-based periosteal contours (SNAKE). Most HR-pQCT parameters from standard morphological and extended cortical analysis differed significantly when comparing periosteal DL and SNAKE contours combined with standard automatic endosteal contours, but differences were small (median between -1.2% and +4.1%) and correlations high (>0.95). To conclude, the DL model performed well for automatic periosteal and endosteal contouring on second-generation HR-pQCT and for automatic periosteal contouring on first-generation HR-pQCT.