Comparison of "Human" and Artificial Intelligence Hand-and-Wrist Skeletal Age Estimation in an Epiphysiodesis Cohort.

BACKGROUND

We previously demonstrated that the White-Menelaus arithmetic formula combined with skeletal age as estimated with the Greulich and Pyle (GP) atlas was the most accurate method for predicting leg lengths and residual leg-length discrepancy (LLD) at maturity in a cohort of patients treated with epiphysiodesis. We sought to determine if an online artificial intelligence (AI)-based hand-and-wrist skeletal age system provided consistent readings and to evaluate how these readings influenced the prediction of the outcome of epiphysiodesis in this cohort.

METHODS

JPEG images of perioperative hand radiographs for 76 subjects were independently submitted by 2 authors to an AI skeletal age web site (http://physis.16bit.ai/). We compared the accuracy of the predicted long-leg length (after epiphysiodesis), short-leg length, and residual LLD with use of the White-Menelaus formula and either human-estimated GP or AI-estimated skeletal age.

RESULTS

The AI skeletal age readings had an intraclass correlation coefficient (ICC) of 0.99. AI-estimated skeletal age was generally greater than human-estimated GP skeletal age (average, 0.5 year greater in boys and 0.1 year greater in girls). Overall, the prediction accuracy was improved with AI readings; these differences reached significance for the short-leg and residual LLD prediction errors. Residual LLD was underestimated by ≥1.0 cm in 26 of 76 subjects when human-estimated GP skeletal age was used (range of underestimation, 1.0 to 3.2 cm), compared with only 10 of 76 subjects when AI skeletal age was used (range of underestimation, 1.1 cm to 2.2 cm) (p < 0.01). Residual LLD was overestimated by ≥1.0 cm in 3 of 76 subjects by both methods (range of overestimation, 1.0 to 1.3 cm for the human-estimated GP method and 1.0 to 1.6 cm for the AI method).

CONCLUSIONS

The AI method of determining hand-and-wrist skeletal age was highly reproducible in this cohort and improved the accuracy of prediction of leg length and residual discrepancy when compared with traditional human interpretation of the GP atlas. This improvement could be explained by more accurate estimation of skeletal age via a machine-learning AI system calibrated with a large database.

LEVEL OF EVIDENCE

Prognostic Level III. See Instructions for Authors for a complete description of levels of evidence.