Validation of Automatic Greulich-Pyle Bone Age on Children with short stature of various diagnoses

Bone age (BA) determination is routinely determined for the diagnostic evaluation of children affected by short stature. Most investigators use the Greulich-Pyle (GP) Atlas because of its intuitive simplicity and rapidity. The intra- and inter-individual variations of BA assessment in clinical practice impose a major handicap on its usefulness in clinical trials and research. As a solution, a fully automatic bone age determination method, called BoneXpert, was developed. The aim of this study is to validate BoneXpert in a large set of children with short stature of diverse diagnoses. We surveyed the coverage of BoneXpert by analyzing characteristics of images accepted and rejected respectively by BoneXpert, the accuracy beyond BoneXpert's intended age ranges as well as its interoperability in use of digital or printed X-rays. Subsequently we analyzed the reliability of BoneXpert. We investigated the validation of BoneXpert compared with other determination methods. We used 5865 X-rays of the left hand of 1152 patients showing a manual GP-BA between 0.5 and 19 years. The X-rays were taken between 1973 until 2005 at the University’s Children Hospital Tuebingen. BoneXpert succeeded to analyze 5668 images of 1009 patients. The difference between printed X-rays and their original digital format were insignificant (correlation coefficient 0.999; P < 0.0001). The core of the study was comparing the automatic bone age values with the original manual bone age determinations. The BoneXpert BA deviated by more than 2.05 years (2 standard deviations, SD) from the manual operator BA in 1.6% (n=92) of X-rays. Those outliers were re-rated by three investigators not knowing chronological age, original manual BA or the result of the automatic BA rating. All the new ratings except 3 were within threshold T of 2.05 years. The SD between manual and automatic BA was 0.73 years [0,66;0,84] 95% CI, the mean deviation was 0.09 years. The variance was 0.967 in boys and 0.957 in grils. There was no significant difference between various diagnoses. Also we could compare BoneXpert’ results with the semiautomatic method CASAS. Compared with the original manual BA using GP (manBA) the deviation of BoneXpert’s rating (autBA) were less than the BA scores of CASAS (CBA), while the standard deviation of the differences (accuracy) and the bias were insignificant. The mean deviation between manBA and autBA was 0.28 years vs. 1.17 years between manBA and CBA. The accuracy (SD of difference) was 0.80 years [95% CI 0.69; 0.88], bias 0.04 y/y and 0.95 years [95% CI 0.85; 1.08], bias 0.1 y/y respectively. Direct comparison between CASAS and BoneXpert yields an average deviation of 0.87 years (n.s.), the accuracy was 0.84 years. [95% CI 0.75; 0.94]. BoneXpert is able to process all X-rays meeting the external requirements of the software. It can avoid errors in inter- and intraindividual inconsistency. BoneXpert’s ability to process virtually all images automatically, to avoid errors, and to obtain good agreement with manual operators suggests that the method could be efficient and reliable in the spectrum of children with short stature that are presented to endocrine clinics.

Author

• Nyembo, Dorothee

Identifier

hdl:10900/45891

OAI record oai:identifier

https://publikationen.uni-tuebingen.de/oai/request?verb=GetRecord&metadataPrefix=mets&identifier=oai%3Apublikationen.uni-tuebingen.de%3A10900%2F45891