Bone age validation study from Leipzig

Radiologists at the University Hospital in Leipzig have published a comparison of three automated bone age methods: BoneXpert, BoneView form Gleamer and PANDA from Image BiopsyLab

The study collected images of 306 children covering the age range 1–18 years. This wide range allowed the study to reveal how the methods behaved at low and high bone ages.

A reference bone age (denoted  “Ground Truth”) was formed as the average of three human ratings.

The overall agreement between each AI and the reference was almost the same for the three methods, although BoneXpert still showed the best correlation

BoneXpert is the only method that covers the full bone age range 0-19 yr, and it is interesting to see how the other methods handled the low- and high-bone age ends.

BoneView rejected images if the chronological age of the DICOM file was below 3 y and also rejected images with a bone age 17 and above.

PANDA accepted all images, but gave large errors below 5 years as is clearly visible. Also, for females with reference bone age above 15 y, PANDA tended to give predicted bone age not much larger than 15 y – a kind of saturation effect.

The corresponding author is Dr Daniel Gräfe and the full text is available online