Atlas-based Machine Learning for Pediatric Bone Age Assessment

Bone age assessment (BAA) for evaluating skeletal maturation is performed exclusively during childhood. It is typically based on x-ray images of the left hand and wrist. It is widely used in general pediatrics and pediatric endocrinology as a safe and cost-effective tool for diagnosis, treatment planning, and prognosis, with minimal radiation risk. In the United States (U.S.), about 4.7% of the children need BAA. In the U.S., the Greulich-Pyle (GP) method is widely used for BAA, and it consists of a set of manual interpretation guidelines and two x-ray image Atlases (i.e., male and female GP Atlases) of White children in Cleveland, Ohio. Following the GP method, a radiologist (or an endocrinologist) memorizes the GP atlases and the corresponding interpretation guidelines as much as possible. The radiologist/endocrinologist then estimates a child’s bone age by visually comparing the child’s hand and wrist x-ray with the reference x-ray images in the GP Atlases, following the interpretation guidelines. Due to the subjective nature of human judgment and the qualitative nature of the guidelines, interreader variability in bone age estimation can be as high as 5.8 months. Another known issue is that the GP method causes significant misinterpretation of bone ages in Non-White children (e.g., Asian and Hispanic children) because the GP Atlases only have data of White children and the guidelines were created based on such data. Consequently, it is imperative to develop a new BAA method with more comprehensive atlases and quantitative guidelines. Machine learning (ML) for BAA has gained considerable attention due to its potential to automate the BAA process. However, the current ML methods for BAA rely on data annotated by humans with the GP method. As a result, these ML methods cannot perform better than the GP method, and the deficiencies of the GP method are inherited by the ML methods developed on the GP method-derived training data. Also, the current ML methods for BAA use end-to-end trained “black-box” models, making it difficult for human users to understand and verify decisions from ML, which limits their acceptance in clinical practice. In this project, we will develop a novel Atlas-based ML approach for BAA, and it will overcome the limitations of the above methods and therefore achieve better performance by major advantages: (1) intuitive to and verifiable by humans, and (2) reconfigurable for use with appropriate atlases (better than the GP Atlases) for different racial/ethnical groups. This approach will be developed and validated on internal and external datasets. Project Number: 1R21HD119531-01A1 | Fiscal Year: 2026 | NIH Institute/Center: Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) | Principal Investigator: Liang Liang | Institution: UNIVERSITY OF MIAMI CORAL GABLES, CORAL GABLES, FL | Award Amount: $395,006 | Activity Code: R21 | Study Section: Special Emphasis Panel[ZRG1 CDMA-E (90)] View on NIH RePORTER: https://reporter.nih.gov/project-details/11372619