Mammogram Compression Algorithm Based On Encoding Technique For Region Of Interest

The size of the digital mammography is generally about 40Mb. The big size usually delays the implementation of computer-aided diagnosis and teleradiology in some degree. Hence, an efficient data-compression scheme to reduce digital data storage cost and transmission time without significant degradation of medical image is needed. Based on the basic principles of set partitioning in hierarchical trees (SPIHT) algorithm and the encoding techniques for region of interest (ROI) in JPEG2000 standard, an interactive encoding scheme for mammographic ROI is established.Due to the features obtained from lifting 5/3 wavelet transform on mammograms, some effective improvements on original SPIHT algorithm are presented. In general, high-frequency subbands have strong similarity. When one high-frequency subbands is zero trees, the other two corresponding high-frequency subbands are usually zero trees. Based on this characteristic, the objects in the list of insignificant sets (LIS) could be encoded at one time. Therefore, the cost of the insignificant information is reduced. Human's vision is often more sensitive to low-frequency information than high-frequency. In addition, the energy of image exists in low-frequency part. Hence, the wavelet coefficients denoted low-frequency parts are encoded with high priority, while the other coefficients denoted high-frequency parts are encoded with low priority. In implements, a right-shift value about the wavelet coefficients corresponding to high-frequency would be selected. This stage of the proposed algorithm balances the peak signal to noise ratios (PSNR) of ROI and background region. In order to ensure the rebuilding quality of ROI, ROI should be encoded with high priority. In implements, a left-shift value about the wavelet coefficients corresponding to ROI should be selected.The preliminary experiments have chosen optimal right-shift and left-shift values. The results show that the PSNRs of ROI and background region obtained by the proposed method are better than the vqSPIHT algorithm.