Breast cancer classification from ultrasonic images based on sparse representation

Abstract

This thesis presents a novel Sparse Representation-based Classifier (SRC) that provides superior performance in terms of high Area Under the roc Curve (AUC) in classifying benign and malignant lesions of breasts captured in ultrasound images. Although such a classifier was initially proposed for face recognition, the use of this has been proposed in medical diagnosis from ultrasonic images in this dissertation for the first time. The classifier is based on `1-norm based sparse representation of a patient’s test data in terms of linear combination of the features of the benign and malignant test lesions available in the training set. The proposed classifier uses an index called Sparsity Rank (SR) for the classification obtained from the normalized energy of the weights as a linear combination of the global sparse representation of the ultrasound images of the training set. The performance of the classifier is further enhanced to a great extent by two ways; first, by the use of a method that intelligently combines the features extracted from the multiple ultrasound scan of the same patient, and the second, by using the reduced feature set. The combining principle of the multiple data scans is based on averaging of the SRs of all the scans. The near-to-optimal feature set is obtained by a suboptimal strategy to evade the time exhaustive brute force approach that has a combinatorial search space.With all the enhancements an AUC of 0:9754 has been achieved, when training and testing sets are chosen by leave-one-out approach from the data set.