Swiss study compares two bone age algorithms

The accuracy of bone age determination by BoneXpert and Panda in 188 images was reported at ECR.

Federica Zanca from Leuven, together with four co-authors from Switzerland, presented a study comparing two bone age algorithms, BoneXpert from Visiana and Panda (based on deep learning) from ImageBiopsyLab. The study included 188 images taken under real-world conditions across 11 centres in Switzerland. The ground truth was provided by an exceptionally reliable manual rater. BoneXpert’s intended use includes autonomous use, while Panda’s is intended to only assist the radiologist. However, in this study, both algorithms were used without human interference.

The mean absolute deviation between the algorithm and the ground truth was 0.36 y for BoneXpert and 0.42 y for Panda, and this difference was significant with p=0.01.

In the Bland Altman plots below one can clearly see that the agreement is better with BoneXpert.

Figure 1: BoneXpert versus radiologist

 

Figure 2: Panda versus radiologist

The plots reveal that there are markedly fewer large deviations with BoneXpert.

So what is the clinical signficance of the different performance? The authors adressed this by defining the clinically acceptable limit of agreement to be ±1 year, and they found twice as many such significant deviations for Panda.

The table below summarises all the findings.

The poster is available though myESR (requiring log in)

Table: The deviation between the bone age algorithm and the radiologist

BoneXpert Panda
Mean Absolute Deviation 0.36 y 0.42 y
Root Mean Square Deviation 0.47 y 0.55 y
Number of deviations > 1 y 7 14
Largest deviation 1.2 y 1.9 y

 

BoneXpert – the most widely used radiology AI in the Netherlands in 2022

At ECR, Kicky van Leeuwen of Radboud University Medical Center gave the talk “The rise of artificial intelligence solutions in radiology departments in the Netherlands”.

Van Leeuwen and her co-workers have asked the 69 radiology departments in the Netherlands about their adoption of AI.

They found that the most widely used methods are the ones shown here:

The most popular method is BoneXpert from Danish Visiana, used in 9 departments.

The second use case is detection of lung nodules in CT scans, used by 8 departments – the product Veye Lung Nodules was developed by the Dutch company Aidence (recently acquired by RadNet).

The third-most used method is stroke – left-ventricle occlusion or intracranial haemorrhage detection on CT and CT angiography – probably StrokeViewer from the Dutch vendor Nicolab.

 

The investigators also inquired about the obstacles for adopting AI and found that the most important were “cost” and “IT”. So, one can infer that the most widely used methods have a favourable balance between cost and benefit, and carry a low IT burden.

Only 3 departments reported use of an “AI platform”.

This talk was also highlighted by AuntMinnie

 

Note that the survey did not obtain 100% response rate, and indeed Visiana’s records show that there are 14 customers of BoneXpert in the Netherlands, each performing between 100 and 1300 analyses per year.

United States patent granted to Visiana

The US patent office has granted Visiana a patent for the invention of an image processing method.

A key element of BoneXpert is to locate each bone accurately as a first step in the bone age analysis. Visiana has invented a new method to locate a landmark, e.g. the tip of a bone. About 100 examples are needed for the machine learning method, which trains a decision tree to predict the location from a nearby image patch. The predictions from more than 300 patches are placed in a voting map, and the peak of votes yields the landmark location with subpixel accuracy. The figure shows the voting map for finding the tip of a distal phalanx.

The invention could also be useful in land vehicles and smartphones.