Novel Prediction Based Technique for Efficient Compression of Medical Imaging Data

The annual volume of imaging data in modern paperless hospitals can approach up to 10 terabytes, heavily pressing the storage and transmission requirements (Choong et al., 2007). Utilizing efficient compression techniques for those data in order to reduce associated costs is very attractive from both viewpoints: financial and organizational (Sanchez, Abugharbieh & Nasiopoulos, 2009; Sanchez et al., 2008). Although lossy techniques can yield better compression results, due to possible compression artifacts in the compressed image, they are less favored compared to lossless compression techniques in certain medical applications such as image-based diagnosis, archival etc. Compression itself helps in alleviating storage requirements for medical imaging system. Additionally, it also helps in accommodating the on-line transmission and availability of patient diagnostic imaging data which is essential for future electronic health frameworks. Moreover, new approaches in medical imaging such as 3D and 4D imaging and bio–modeling produce even greater amounts of image data. For efficient storage and transmission of those data and utilization of systems that exploit 3D and 4D imaging technologies, compression is inevitable. In this field, at least certain parts of images are required to be stored and transmitted without any loss of information. The lossless compression algorithm that we propose can also be efficiently employed for at least those vital parts of interest in this kind of applications (Zagar et al., 2007). Important property of image data is high degree of correlation among neighboring pixels which is crucial for any compression technique since it makes it possible to decorrelate the samples using some sort of prediction–based modeling. If employed modeling technique effectively models the spatial correlation among neighboring pixels, remaining data will be mostly decorrelated and easily coded with an entropy coder. On the other hand, it is well known that image data are nonstationary, i.e. properties of image regions vary all over the image (Memon & Wu, 1999). Accordingly, it is necessary to adapt the model to the changing image characteristics. Another assumption of local stationariness is very well applicable to the image data. This means that for arbitrarily small image regions, the model adapted to the dominant local property will be effective inside the region. Predictive image coding in which the prediction error of the current pixel is coded has shown to be the most effective technique in lossless image compression. Using prediction, image data are decorrelated prior 7