Dynamic rule refinement strategy of associative classifier for effective mammographic image classification

Computer-aided diagnosis (CADx) has gained significant attention in helping radiologists in the interpretation of mammograms to assist in diagnostic decisionmaking. A more effective CADx increases the probability of cure. An effective mammogram classification technique benefit to the research of computer aided mammography for a better diagnostic assistance. However, the effectiveness of classifiers depends on the training data sets that are often small in data size and static, which does not adapt to changes. The main aim of this thesis is to propose an effective associative classifier using rule refinement technique that adapts changes in databases for building an effective CADx model in the classification of mammogram images. The classifiers using Association Rule (AR) mining gain popularity compared to traditional classifiers due to their nature in reflecting close dependencies among single or multiple features for composing rules with its excellent interpretation. The existing associative classification techniques that are used in Computer Aided Diagnosis (CADx) have proved their efficiency in mammogram classification. The research aims to propose an improved associative classification model with its first step preprocessing that uses segmentation technique with filter that includes certain areas of the image for mammogram peripheral enhancement. The feature extraction is used to extract the most prominent features from mammogram images that represent various classes of the images to be used by classification techniques. A feature selection technique named Correlation Feature Selection (CFS) that involves a heuristic search is adopted for dimensionality reduction of feature space to improve efficiency and at the same time maintain the effectiveness of classification. The thesis discovers useful and interesting relations between features and class in the form of rules to build an efficient associative classifier from a large collection of mammogram images using association rule mining technique. An Associative Classifier that uses rules Highest Average Confidence (ACHAvC) is proposed for an effective classification of mammography. The classifier ACHAvC has achieved high accuracy of 90% and specificity of 90%, however the sensitivity is 78.5% and not commendable in medical domain. The effectiveness of an associative classifier depends largely on the generated rules based on training data. In previous works such as HiCARe, SACMINER, MINSAR, including ACHAvC the training data have been limited, which may produce the classification rules that are static and cannot adapt to a changing charecteristic of test images, as such it may not produce complete and accurate rules for classification. The classification performance can be further improved if the static rules are updated dynamically. The availability of radiologist ground truth for every case could be used to validate the classification result and refine the set of rules generated. A method Rule Refinement based on Incremental Modification (RRIM) is proposed that dynamically refines the rules every time when it is validated with the experts ground truth. As such these refined rules that adapt the changes in the data are then used for classification to further enhance the performance of the classifier ACHAvC with a reduced minimal error and with improved prediction accuracy. The Performance of the proposed methods are evaluated for accuracy, sensitivity and specificity for the mammogram image data set, taken from the digital database for mammography from the University of South Florida, Digital Database for Screening Mammography (DDSM). The proposed method has achieved an overall classification accuracy of 96%, with sensitivity 92.56% and specificity 96.94% in testing stage which is comparatively better than the three benchmark approaches HiCARe, SACMINER, MINSAR that are chosen for the proposed research with accuracy of (91%, 85%, 79%), sensitivity (95%, 84%, 87%) and specificity (84%, 86%, 67%)