Accuracy and self-validation of automated bone age determination

Area of Science:

• Radiology
• Medical Imaging
• Pediatric Endocrinology

Background:

• The BoneXpert method, established in 2009, automates bone age determination from hand X-rays.
• It is currently utilized in over 200 hospitals worldwide.

Purpose of the Study:

• To introduce and validate version 3 of the BoneXpert method.
• To assess the accuracy and self-validation mechanism of the automated system.

Main Methods:

• Training utilized 14,036 images from the RSNA Bone Age Challenge, 1642 normal children's images, and 8250 images from Tübingen patients.
• Validation involved cross-validation on Tübingen images and testing on the RSNA dataset.
• The self-validation mechanism was evaluated for its ability to reject incorrectly analyzed images.

Main Results:

• Version 3 achieved a cross-validated root mean square (RMS) error of 0.62 years on Tübingen images, an improvement over the previous 0.72 years.
• The RMS error on the RSNA test set was 0.45 years compared to six manual ratings.
• The self-validation mechanism rejected 0.4% of RSNA images, with significantly fewer severe bone age errors (12x less) in accepted images compared to manual ratings.

Conclusions:

• BoneXpert version 3 demonstrates improved accuracy in bone age determination.
• The self-validation mechanism effectively identifies images at risk of incorrect analysis, enhancing safety.
• The automated method shows potential to be safer and more accurate than single manual bone age ratings.