Breast cancer detection by using associative classifier with rule refinement method based on relevance feedback

Abstract

Computer-aided diagnosis system that uses classification process for an automated detection of breast cancer could provide a second opinion that improves diagnosis. Several researchers have proposed the use of associative classifier that generates strong associations between features and reveals hidden relationship that can be missed by other classification algorithms. However, the effectiveness of an associative classifier depends largely on the generalized rules based on training data. Often, the number of training data is limited, which may further produce the classification rules that are stagnant and cannot adapt to a changing distribution of test images, as such it may not produce complete and accurate rules for classification for future cases. This paper aims to address this issue by refining rules that are static using dynamic rule refinement technique. This technique helps to adapt to the changes in the new evidences that can be used for classification to further enhance the performance of the associative classifier. A method named the rule refinement based on incremental modification is proposed that dynamically refines the rules after the suggested term is validated by the experts. Once the initial classification is performed using the generalized rules for each test example, the results are validated using the experts feedback. Based on the validated classification result either correct or incorrect, the rules that are responsible for classification are refined in three phases. These refined rules are used for classification of future test examples that leads to improved prediction accuracy in comparison with the classifier that uses generalized static rules. The performance of the proposed method evaluated on the digital database for screening mammography dataset is promising and has achieved an overall classification accuracy of 96% in the biased setting and an accuracy of 95.24% in the unbiased setting as compared to the accuracy 90.48% of baseline classifier with static rules.