September 2019: BoneXpert version 3.0 released

The long-awaited version 3.0 of BoneXpert is now rolling out to all customers.
It represents the most significant update, since we launched BoneXpert 10 years ago.
The main features are:

  1. The accuracy of the bone age measurement is dramatically improved.
  2. Carpal bone age determination
  3. Bone age and BHI determination down to age zero
  4. Fewer rejections of images and the bones are located more accurately

Here is a more detailed account of the four new features in version 3:

1. Accuracy

The root mean square error (RMSE) of BoneXpert relative to manual bone age rating has decreased from 0.72 y in version 2 to 0.63 y in version 3. This is reported at ESPE in Vienna in the talk by David Martin who performed this validation on 8250 images from Tübingen. The improvement in accuracy is more dramatic than these numbers indicate, because the error 0.63 y is mainly due to the manual rater variability. If one instead compares BoneXpert to the average of six manual ratings, the RMSE is only 0.45 y. In fact, BoneXpert has the same accuracy as the average of four manual ratings. This is described in detail in the post about how RSNA boosts the accuracy of automated bone age rating.

2. Carpal bone age

Version 2 ignored carpals, because there is a consensus in some countries that carpals are less important. Bone age should furthermore be determined from bones with epiphyses, i.e. radius, ulna and the 19 short bones. When version 3 now analyses the carpals, it provides carpal bone age as a separate number. This is because different hormones control the maturation of carpals and short bones. E.g. short bones respond more to sex steroids than carpals. The main bone age result still only takes the bones with epiphyses into account. Carpal bone age deviating from this is potentially a sign of hormonal imbalance. In some disorders, e.g. precocious puberty, this could be the effect of a treatment.

carpal bone age is displayed in the results as Carpal BA

3. Infants

Version 2 did not analyse hands with bone age less than approx. 2 years. Version 3 now extends down to new-borns. The model is based on the Paris study of 425 normal children followed longitudinally with hand X-rays at ages 1, 3, 6, 9, 12, 18, 24, 36, 48 months etc. This study helped establish reference curves for bone age and BHI, so that the measurements of BA and BHI on infants associate with BA SDS and BHI SDS values. The BHI reference curves were reported at the ICCBH conference in May.

4. Finding bones

Version 2 rejected some images due to postprocessing (edge enhancement), which is common in modern digital detectors. BoneXpert version 3 is much better adapted to these variations and rejects far less images. However, it is still advisable to not use strong postprocessing, because it can lead to skewed measurements or even rejection of images.
Version 3 also locates the bones more accurately than version 2; the accuracy is SD = 0.1 mm. This is important for two reasons. Firstly, locating the bone contour accurately leads to more accurate bone age assessment. Secondly, it gives the user a visual feedback on bone assessment, thereby giving confidence in the measurement. For instance, the accurate delineation of the epiphyses shows that BoneXpert understands this crucial aspect of bone age rating.

BoneXpert now finds all bones in the hand