Ten percent of the data ( n = 1681) were used as a holdout test set. Images were included only if a pediatric radiologist (R.S.A., with 6 years of experience) determined them to be satisfactory for bone age evaluation, excluding those with congenital anomalies, compromised image quality, and poor patient positioning. Of the radiographs, 41% were multiple radiographs in the same patient, either the same hand at a different timepoint or the contralateral hand at the same timepoint. From the Children’s Hospital of New York (CHONY) in New York City, NY (a level 1 pediatric trauma center), we obtained 16 810 frontal view pediatric hand radiographs (from approximately 10 000 pediatric patients) obtained for trauma between December 14, 2009, and May 31, 2017. Approval from the institutional review boards of both participating institutions was obtained prior to the study. This was a retrospective study comparing several methods for bone age assessment.
BONE AGE AND CHRONOLOGICAL AGE MANUAL
We tested our TDL-BAAM on trauma hand radiographs in a geographically and demographically distinct population and compared the performance to both automated and manual bone age assessment by the GP method.
BONE AGE AND CHRONOLOGICAL AGE HOW TO
This trauma hand radiograph–trained DL-BAAM (TDL-BAAM) learns bone age patterns corresponding to a relatively healthy and diverse population in hopes to be generalizable to other populations and to provide a potential example of how to approach the establishment of a population-specific TDL-BAAM. The purpose of our study was to develop a GP-independent deep learning model for automated bone age assessment (DL-BAAM) by training the algorithm to assess chronological age using bone morphology from a training set of more than 10 000 pediatric trauma hand radiographs in a large academic children’s hospital with a diverse population. Therefore, two questions arise regarding how to best optimize training for automated bone age assessment algorithms: (a) Should ground truth be determined from a pediatric sample clinically requiring a bone age examination for underlying systemic abnormalities, and (b) should ground truth be established from a skeletal age assigned by human interpretation using the GP atlas? Our study proposes an alternative approach to methods used in prior studies by training on data from normal pediatric hand radiographs and using chronological age as ground truth, as opposed to human interpretation using previously described methods such as GP. Inherent to these prior methods is the reliance on GP as both the training and evaluation standards. For example, the Radiological Society of North America (RSNA) Pediatric Bone Age Challenge used a training dataset of more than 12 000 hand and wrist radiographs obtained clinically for bone age assessment where bone age determined by GP was extracted from clinical reports ( 16). Interestingly, several recent studies of automated bone age assessment report methods that draw from ground truth (known as supervised machine learning) established through reference to GP, with many using extracted bone ages from clinical reports ( 11, 15). Recent approaches have used deep learning and convolutional neural networks, which learn imaging features relevant to particular tasks from large datasets, without programming explicit rules or extracting specific features (shape, texture, etc) ( 9– 17). There has been growing interest in the development of automated methods for bone age assessment to improve on the efficiency, accuracy, and reliability of human interpretations. The Tanner-Whitehouse method of bone age assessment is a more reliable alternative to GP however, it is relatively labor intensive and time-consuming ( 4).
Despite its broad acceptance in clinical practice, the GP method is subject to human factor limitations of inter- and intraobserver variability ( 4– 8). In a 2016 survey of pediatric radiologists, more than 97% used the GP atlas ( 3). While several methods of bone age assessment exist, the most widely used is direct visual comparison of an individual patient’s left hand and wrist radiograph with the Greulich and Pyle (GP) Radiographic Atlas of Skeletal Development of the Hand and Wrist ( 3– 5). Radiographic bone age assessment is an important component of the diagnostic workup for a variety of pediatric endocrine, metabolic, and growth disorders ( 1, 2).
0 kommentar(er)
