Similarity Based Characterization and Identification of Breast Ultrasound Lesions in Space-Frequency Subband

Breast Ultrasound (BUS) has gained popularity over mammograms due to its higher sensitivity, accuracy, faster imaging and lower cost. Computerized diagnosis of breast abnormalities using BUS images is crucial for better treatment planning, monitoring and clinical trials. Present work decomposes the original BUS images by using Stationary Wavelet Transform (SWT) to estimate the textural similarity of breast lesions. The proposed detection approach employs a contrast enhancing technique using sigmoid and gradient magnitude filter to reduce artifacts, nipple shadowing and missing structures of BUS images. Spatial Fuzzy c-Means (FCM) clustering algorithm is used to segment the contour of lesions. Following this, textural compositions of lesions have been assessed in the space-frequency plane to study the degree of homogeneity/heterogeneity of lesions, present in different space-frequency subbands. Some similarity measuring quantifiers are used to estimate textural homogeneity of lesions. Finally, support vector machine (SVM) and Hybrid Adaboost classifiers have been designed to categorize the benignancy/malignancy stages of test data. Experimental results are reported on three standard benchmark datasets and compared with other state-of-art detection techniques.