Automated Bone Age Assessment Across Multi-site U.S. Study: Agreement between AI and Expert Readers

Purpose or Learning Objective: The radiologic determination of bone age (BA) from a left-hand radiograph continues to be the reference standard for skeletal maturity assessment related to short or long stature, premature or delayed puberty, and underlying conditions. Artificial intelligence (AI) algorithms are becoming more prevalent due to the subjectivity and time-consuming nature of BA assessment. In this study we evaluate the agreement between AI and expert readers for BA assessment according to the Greulich and Pyle method.

Methods or Background: Radiographs of 342 patients were analyzed retrospectively (178 males years of age; 165 females 2–16 years of age). Sampling was performed across multiple sites in the United States associated with Washington University School of Medicine in St. Louis, Missouri. Three board-certified pediatric radiologists made blind reads of BA using the Greulich and Pyle (GP) method independently, and AI software was subsequently used to estimate BA from the same images. Agreement of AI with readers was assessed based on comparison of Bland-Altman limits of agreement, orthogonal linear regression, and interchangeability.

Results or Findings: Bland-Altman assessment displayed a mean difference between readers and AI to be − 0.72 (lower confidence interval [LCI], − 1.46; upper confidence interval [UCI], 0.02) months. Using orthogonal linear regression, the slope of agreement between readers and AI software was reported to be 1.02 (LCI, 1.00; UCI, 1.03). No proportional bias was observed. The square root of the absolute value of the equivalence index of the AI software compared with assessments made by readers was observed to be − 5.8 months. This indicates that the AI software is interchangeable with expert readers.

Conclusion: Fully automated AI software showed agreement with expert readers in BA assessment on a multiple-site cohort of U.S. children and adolescents, suggesting AI integration into the radiology workflow is possible and could lead to more efficient BA reading.

Publication History

Article published online:
02 June 2022