Quantitative Analysis of a General Framework of a CAD Tool for Breast Cancer Detection from Mammograms

In this paper, the quantitative analysis of the various methodologies used in the design steps of a computer aided diagnostic (CAD) tool for breast cancer detection from mammograms is presented and based on results efficient methods for the design of the CAD tool are suggested. The design steps of the CAD tool include preprocessing, segmentation, feature extraction and selection, and classification. The pre-processing steps include manual cropping of original mammograms for removal background details, quantum noise reduction, and contrast enhancements. For quantum noise reduction, a modified TV based filter adapted to Poisson noise is used. For further enhancement of abnormalities, contrast limited adaptive histogram equalization (CLAHE) is used. A modified fuzzy C-means thresholding segmentation is used to segment abnormalities. Eighty eight hybrid features are proposed and extracted for each mammogram in database. Further, two types of feature selection methods viz, unsupervised genetic algorithm (GA) method based on mutual information (MI) and forward sequential feature selection (SFS) based on cross-validated minimum miss-classification errors of various classifiers are proposed and examined. For the selected feature subsets, the performance evaluations of various classifiers have been done for choosing an appropriate classifier. A comparative study of the performance of the proposed CAD tool, encapsulating the better approaches found after investigation, with those of other CAD tools in literature is also presented.