A Hybrid Wavelet Filter for Medical Image Compression
Image compression addresses the problem of reducing the amount of data required to represent a digital image. The basis of the reduction process is the removal of redundant data without affecting the quality of the image. This can be achieved by reducing or, whenever possible, eliminating various types of redundancy that exist in the imaging data. Mammogram, or generally, most medical images are particularly attractive for image compression techniques due to their enormous storage requirements (250-300MB/patient/visit). In contrast to other type of images where a compression algorithm can be applied on the image as a whole, mammogram images must preserve all of the image's diagnostically critical information that help radiologists give an accurate diagnosis. In other words, the mammogram image can not be compressed as a whole, but rather as a segmented image, broken down into smaller images (containing the critical details), with each region/sub-image encoded differently. A mammogram image is usually partitioned into two regions: Region of Interest (ROI) and Background (BG). Lossless and lossy compression, or a combination of the two methods, might just be the answer to optimal mammogram compression, or any content-based segmented images. The region containing the micro-calcification (in the breast) is compressed using a lossless algorithm to preserve all the details, whereas the rest of the image could be encoded using a lossy algorithm. Lossless compression methods allow perfect reconstruction, however, their overall performance is modest, leading to only 1-3:1 compression rates. Lossy compression allows much higher compression rates (30-80:1); however, at the cost of degraded image quality, which in most medical images is not acceptable. The research work presents a hybrid, content-based lossless-lossy combination, compression method for mammogram images. The image is first morphologically segmented into two regions, Region of Interest and Background, with each region being differently encoded. Huffman coding and Differential Huffman coding are used to compress the ROI area, and Wavelet Packets using sub-band filters are used for compressing the BG area. A total of sixty wavelets types were tested, and the best matched, optimal for mammogram image compression, Daubechies quadrature filter (db1) was modeled and described in terms of its order, number of taps, orthogonality, compactness, and smoothness. The corresponding best case decomposition level was also investigated. Finally, an entropy encoder (lossless) compresses the combined ROI and BG regions even further. The overall performance of the method is compared to other established methods such as Discrete Cosine Transform, Discrete Fourier Transform, and JPEG2000. The visual quality of the resulting compressed images is evaluated using strict, widely accepted objective and subjective criteria. Under the objective criteria means, the CR, RMSE, PSNR measures are employed. Subjectively, a group of thirty engineering students (non specialists in the field) were asked to rate a set of five images at different compression rates. The images were displayed both in random as well as in increasing compression rate order. By using a combination of objective and subjective evaluation criteria, the overall performance of the dissertation could be more accurately described and analyzed. Finally, the effectiveness and performance of the proposed method was finally tested on fifty three grey scale mammogram images in terms of CR, MSE, and PSNR. A total of fifty four 8-bit grey scale mammograms were used in this study. The images are freely available for downloading from the University of South Florida, Digital Mammography Database server.  Two different methods were proposed for the lossless encoder and entropy encoder: Huffman Coding and Differential Huffman Coding. The overall performance, in terms of compression rate, of the proposed algorithm was in one case similar to most current methods, and significantly better in the second case. The compressed image reproduced the original image by 100% in the ROI and above 95% in the BG region. Since all of the relevant, diagnostically important details are contained in the ROI, they were fully preserved during the compression process.
Wavelet, Image Compression, Hybrid
February 8, 2006.
A Dissertation Submitted to the Department of Electrical and Computer Engineering in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy.
Includes bibliographical references.
Florida State University