Texture Modelling
Diagnostically lossless video compression for angiogram data using a wavelet-based texture modeling approach
Traditionally, most approaches to the compression of medical images have been lossless in nature, resulting in perfect reconstruction of the original image data. In contrast, lossy approaches make no assurances of perfect reconstruction, although in the context of medical imaging, such approaches usually aim for the target of being diagnostically lossless. This implies that any imperfections introduced by the compression process have no tangible detrimental effect on the diagnostic process. Such a system can achieve compression ratios in excess of 8:1, in contrast with lossless methods which often give barely 2:1 compression.
Why are traditional lossy compression techniques are bad for angiograms?
A sub-band decomposition of angiogram images using a wavelet transform reveals that the majority of the image information is contained in the lower frequency sub-bands. The high frequency sub-bands of the images are generally of a low magnitude, and demonstrate poor spatial and temporal correlation. Much of this high frequency data is additive noise which is produced as part of the capture process. Consequently, the lack of spatial and temporal correlation makes efficient compression of these sub-bands difficult. One option would be to remove this high frequency data by filtering. Unfortunately, superimposed on this high frequency noise-like texture are the diagnostically important high frequency elements mainly associated with the edges of the arteries. Hence the removal of the high frequency sub-bands results in a loss of diagnostic information, as the precise position of an arterial boundary is crucial to the detection of medical conditions such as stenosis (narrowing of the arteries). Even in the regions of the image where the high frequency sub-bands are diagnostically not important, the effect of filtering results in an synthetic appearance to the image.
Approaches
The above problems can be takled by first segmenting the image into regions which are textured, but contain no diagnostically important information in the high frequency sub-bands, and regions which contain diagnostically important data. The regions labeled as containing diagnostically important high frequency data are encoded using a standard wavelet coding algorithm, where as the remaining regions are modeled as texture patterns, using an auto-regressive texture modeling approach. The modeling is performed in the wavelet domain, and so the edges between different regions are not visible due to the smoothing effect of the inverse wavelet transform.
Advantages of Texture-Based Modelling techniques
The effect of the texture modeling is to significantly reduce the bit rate in the regions where it is applied, as only the texture modeling parameters are required to be transmitted.
The effectiveness if the algorithm at different bit rates is assessed by a consultant cardiologist with the key aim of identifying any degradation in the diagnostic content of the images. In addition, comparisons of image quality are made with more standard DCT and wavelet based coding approaches, as well as with the H.263 video coding standard.
Last update August 7, 2000