Version 2 of BoneXpert from January 2013 introduced an increased robustness towards image post-processing – more specifically: edge enhancement and contrast amplification – so that bone age and BHI (Bone Health Index) are largely insensitive to these operations. This “Topic of Interest” tells the story behind this improvement.

Image post-processing in radiology

The radiographs produced in modern digital radiology departments have often been subjected to some amount of “image post-processing” by software associated with the modality itself in order to render the details more visible for the human eye. Twenty years ago, the method of choice was unsharp masking , which enhances edges at a certain scale (i.e. distance) e.g. 0.4 mm. However, this is useful only when used moderately; otherwise it leaves clearly visible dark “halos” around white objects. This is illustrated in the figure below, showing a raw image and a version with unsharp masking.

In 1994 Agfa invented multiscale image contrast amplification (Musica), which is far better than unsharp masking. It enhances edges at several scales at once, so that the halo effect is spread out and not bothering the human eye. In addition, it can reduce edges at very large scales, which is the same as reducing contrast variations over large distances. This in turn allows the image grey levels to be scaled up to improve contrast. The figure above also shows an example of multiscale contrast amplification, which compared to unsharp masking gives a much more pleasant image – it avoids focussing on one particular spatial frequency and instead gives room for details at many scales. The last example in the figure shows extreme use of contrast amplification, where the “long distance” contrast differences have been removed almost completely: the grey level inside the hand (at A) is now much darker, while it is much lighter in the background (at B). Such a strong amplification seems unnecessary for hand radiographs – but it could be useful in e.g. thorax images.

Other vendors of CR and DR systems have introduced similar multiscale techniques as an integral part of their modalities, because this is perceived to add a considerable value to these systems. On Siemens the post-processing is called Diamond View, on Fujifilm MFP (Multi Frequency processing), on Philips Unique, and on Canon MLT.

Variations in image quality

When developing methods like BoneXpert for quantitative image analysis in clinical practice, one needs to have a strategy for coping with the variations in image quality.

It is easy to make the image analysis invariant to uniform changes in brightness and contrast described by linear rescaling of grey levels.

The variation of signal-to-noise ratio stemming from variations in dose (mAs) and voltage (kVp) presents a larger challenge. It turns out that the BoneXpert bone age measurement is very insensitive to variations in noise level; this is because the used image features are relatively coarse-grained – BoneXpert is not looking for tiny lesions as in cancer screening.

The variation in sharpness was the largest challenge for BoneXpert before version 2. BoneXpert’s measurements depended slightly on the image sharpness: blurring the image or sharpening it by edge enhancement led to a slight change in bone age. As the child matures, the growth plates gradually fuse, and finally they are completely resolved. Edge enhancement therefore makes the growth zones appear less mature, and so BoneXpert arrived at a slightly smaller bone age, and edge enhancement consequently had to be kept at a low strength before version 2.

As the post-processing algorithms become more sophisticated, they can alter the images more dramatically without creating unpleasant artefacts, and as a result, image post-processing is the strongest source of variation in image quality in CR and DR images seen today.

Coexistence with image post-processing

A quantitative image analysis method like BoneXpert can take two approaches to coexist with the image postprocessing. One is to configure the modality to generate two images: a raw image for the image analysis, and a post-processed image for display. This approach is used with methods for Computer Assisted Detection (CAD) for mammography, where the two versions of are often denoted “image for processing” and “image for presentation”.

Setting up such a dual image data flow was not considered practical for bone age hand X-rays, which are just a small fraction of the images in general radiology. Instead a friendly coexistence with image post-processing was devised, composed of three elements:

1) The radiology department is advised to set up a protocol for hand X-rays for bone age, where the contrast amplification is turned off, or at least kept at a moderate level.

2) BoneXpert estimates the sharpness of each image, and if it is above some threshold, the image is rejected for analysis. Thus extreme contrast amplification as in the last example in the figure is not accepted

3) The BoneXpert algorithm for bone age and BHI is enhanced so that it compensates for variations in sharpness. With this correction, it was possible to allow a larger amount of postprocessing than before.

Radiologists accept that strong post-processing is not needed for these hand X-rays. Indeed it seems best to stick to a “classical” look of these radiographs to facilitate manual rating with reference to the Greulich Pyle atlas, which consists of plain X-rays with no contrast amplification.

Author: Hans Henrik Thodberg (c). Published 7 Feb 2014,