MULTI-RESOLUTION TEXTURE ANALYSIS OF MAMMOGRAMS USING NEAREST NEIGHBOR CLASSIFICATION TECHNIQUES
B Prathibha, V.R. Sadasivam · International Journal of Information Acquisition · 2010
Mammograms are the most reliable and cost effective method for showing tissues abnormalities of breast. The proposed method classifies the breast tissues by extracting multi texture properties with multi scale wavelet transformations on regions of interest (ROI). The ability of each texture property in discriminating normal and abnormal ROI is analyzed individually and collectively. The method is tested on 217 mammogram images from the mini-MIAS database. Results indicate that multi-resolution image parametrization becomes inevitable when improvement of classification accuracy in textural domains is required. Further, this paper finds the classification accuracy of the nearest neighbor (NN) classification techniques which exploits the underlying density structure of dataset. The study reveals that the variants of the nearest neighbor classifier perform well individually and they significantly enhance the classification accuracy when combined. Finally, the performance of the proposed statistical classifier is compared with radial basis classifier Support Vector Machines (SVM). The receiver operating characteristic (ROC) curve analysis is used as the performance measure to justify the result. It yields an area under the ROC curve (AZ) of 0.946 for proposed scheme, against 0.924 of SVM.