SURF features based classifiers for mammogram classification

Breast cancer became second major reason for cancerous deaths in women. Computer-aided diagnosis of mammogram images is essential for primeval identification of cancer. Authors use SURF (Speeded-Up Robust Features) local descriptors to obtain feature vector and different classifier for mammogram classification. SURF features extracted from mammogram images are high in dimension, and very large in number. So, PreARM [1] algorithm is used to optimize SURF features. Optimized SURF feature vectors and the class of training mammograms form the transaction database, and then it is given to Apriori algorithm to mine association rules. Authors use ESAR [2] algorithm to get optimized and strong rules. Mammogram classification is carried out using the filtered and strong association rules. Proposed scheme is tested on standard MIAS and DDSM data set. Algorithms performance is measured with respect to area under ROC (Receiver Operating Characteristic) curve and classification accuracy. Results of associative classifier are compared with classification using SURF descriptor and distance measure and random forest method. Experimental results reveal that SURF outperforms other methods with regard to distinctiveness, repeatability, and robustness. SURF is computed and compared much faster by maintaining its performance. For SURF based associative classifier, accuracy values for MIAS and DDSM database are 92.3076% and 96.875% respectively, and area under ROC curve values for MIAS and DDSM database are 0.9535 and 0.9221.