FULLY AUTOMATED MEASUREMENT OF PAEDIATRIC CEREBRAL PALSY PELVIC RADIOGRAPHS USING MACHINE LEARNING: EXTERNAL VALIDATION USING A NATIONAL SURVEILLANCE DATABASE

The radiographic analysis required for a national cerebral palsy (CP) hip surveillance programme is resource intensive. BoneFinder® is a machine-learning tool that can automatically calculate Reimer's migration percentage (RMP) from pelvic radiographs. HipScreen is a smartphone application that can partially automate RMP measurement. Three RMP measurement methods were compared across the same set of radiographs: 1) routine manual measurements performed by clinical experts from the CP Integrated Pathway Scotland (CPIPS) database, 2) automated measurements using BoneFinder® and 3) measurements performed by two clinicians using HipScreen. 509 AP pelvic radiographs (1,018 hips; mean age:7.4 years) were selected at random from the CPIPS database. GMFCS levels were I (n=69), II (n=37), III (n=97), IV (n=120) and V (n=186). The absolute mean difference in RMP between BoneFinder® and CPIPS measurements, BoneFinder® and HipScreen and CPIPS and HipScreen measurements was 6.3%, 4.6% and 5.2% respectively. Interobserver reliability (ICC) of RMP measurement across the three methods was excellent (ICC = .92, P<.001, 95% CI .90–.93). Good to excellent ICC and correlation were found between BoneFinder® and CPIPS measurements (ICC = .87, P<.001, 95% CI .75–.93, r=.90) and HipScreen and CPIPS measurements (ICC = .91, P<.001, 95% CI .87–.94, r=.93). The area under the receiver operating characteristic curve for BoneFinder®'s and HipScreen's ability to detect a RMP ≥30/≥40% was .96/.98 and .97/.99, respectively. Fully automated RMP measurements were highly reliable with clinically acceptable measurement error. BoneFinder® appears to perform well in analysis of radiographs in CP children who may have challenging radiographic anatomy.