WAVELET BASED MEDICAL IMAGE ANALYSIS FOR DISEASE DIAGNOSIS

Abstract

The area of the dissertation is medical image analysis for disease diagnosis with different algorithms which are mainly based on wavelet transformation and morphological image processing. This dissertation report specifically gives a detailed description of the different image pre-processing and image enhancement techniques as applied to digital mammograms. The performance of different enhancement algorithms is evaluated by finding different quantitative measures such as D-Index and DV/BV ratio. Finally, the algorithms developed for detecting extremely subtle calcifications in mammograms are presented, which are very crucial to achieve the end goal i.e. classification of disease. MIAS Mini-mammographic database . is used to validate different enhancement algorithms. Because of the minute sizes of the microcalcification, the image resolution needs to be very high. The image size in mini-mammographic database has been clipped and padded to become 1,024*1,024 pixels. The dynamic range of the pixel is eight bit- that is, a gray scale of o to 255. Background tissues included in this database include fatty, fatty glandular, and dense glandular. The database yields abnormalities such as calcification, well defined circumscribed masses, masses with specules, ill defined masses, architectural distortion, asymmetry and normal. The ground truth of each mammogram is also included in the database, which provides the location where the abnormality is situated in terms of (x, y, r) within a given mammogram, where, (x, y) are the x and y axis coordinates, and r is the approximate radius of the circle enclosing the abnormality. A cropping operation is applied to the image to prune the images with the help of crop operation in image processing. Cropping cuts of the unwanted portions of the image, thus all the unnecessary background information and most of noise are eliminated by clipping large number of background pixels from the images, storage requirements, I/O time, and image processing time is significantly reduced. After, the ROI has been extracted different enhancement algorithms like, histogram equalization, contrast limited adaptive histogram equalization, contrast stretching, pseudo-coloring, difference-image technique, unsharp masking, statistical based sub-band filtering and morphological enhancement are applied . The results obtained after the use of combination of techniques like unsharp masking and contrast stretch, and morphological enhancement and contrast stretch is also demonstrated. To validate the performance of different enhancement algorithms a number of quantitative measures are proposed in the literature. The D-Index measure as proposed by Singh et. al. [29] is the only quantitative measure that uses both standard deviation and entropy in the image to evaluate the enhancement. D-Index and DV/BV ratio are used to quantitatively rank different enhancement techniques as applied to digital mammograms. It is found that morphological enhancement is the best method for enhancement of microcalcifiactions without distorting the background parencymal tissue. For microcalcification detection, we are interested in microcalcifications of 0.05-i.o mm in diameter, i.e. calcifications of size 1 to 5 pixels in diameter are targets to be detected. The morphologically enhanced ROI image is then presented to the algorithm that is used for detection of calcifications. Both morphological and wavelet based processing are used for detection of extremely subtle calcifications in mammograms. Developing modules for feature extraction and classification stage are future directions in this area to develop a fully automated CAD based system for disease diagnosis. The desire to use computers in place of second radiologist, or as a pre-screener to separate out clearly normal mammograms, is the motivation for computer aided detection research.