DETECTION OF DIAGNOSTIC FEATURES FOR CLASSIFICATION OF MAMMOGRAMS

Abstract

cancer. Mammography is a procedure of obtainning mammogram i.e. X-Ray image of breast strctures. Among all the mammogram abnormalities, microcalcifications are the most difficult type of tumour to detect. Area of thesis is mammogram image analysis for extraction of diagnostic information from it. Thesis work proposed a scheme which focuses on enhancement of mammogram mak-ing use of morphological operators, segmentation of microcalcification using minimum cross entropy thresholding technique and extraction of number of diagnostic features for classification of mammogram into malignant or benign case. Enhancement algorithms are developed using top-hat and h-dome filters. Both makes use of morphological operators. Top-hat filtering is dependent on selection of structuring element so is prone to be image feature dependent while h-dome extracts microcalcification irrespective of shape variation in microcalcifications present on mam-mogram. Traditional CLAHE enhancement algorithm is used for comparison. For the evaluation of . the performance of enhancement algorithms; detail variance to back-ground variance ratio and contrast improvement index these two performance indices are used. After enhancement image background almost -become uniform. Minimum Cross en-tropy based algorithm is developed for thresholding and extracting microcalcifications from the mammogram. This algorithm segments image for different thresholds and after each thresholding checks cross entropy between original image and segmented im-age. Threshold giving least cross entropy value is taken as final threshold. Qualitataive analysis of enhancement as well as segmentation alorithms are done by presenting pro-cessed images to expert radiologist. It is found that h-dome enhancement is the best method among all stated methods for enhancement of microcalcifiactions without dis-torting the background parencymal tissue. Shape and texture of an image objects gives most of information about the nature of objects. Therefore we extracted 44 features providing diagnostic information based on shape and texture of image for classification. SVM is used for classification as it is outperforming other conventional techniques as stated in literature. SVM with Radial Basis Function kernel with tuned parameters gave 82 % accuracy of classification. McGill University database images are used for quantitative and qualitative analysis with the assistance of radiologist.