Sub-Masses Detection in Malignant Breast Ultrasound Images

Th is work deals with detection of sub-lesions and major lesion in breast ultrasound (US) images. Most of the recent classificat ion uses normal and abnormal breast images to develop their algorithm. The majority of the current algorithms are interested in the major lesion when detecting the lesion boundary. US images, in first step were roughly preprocessed and classified. A function based on classification parameters is used to select the best segmentation threshold. A Second involved step of US image p rocessing includes: Linear and non-linear filtering, segmentation, morphology operations and lesion classification and detections. Seven gray intensity statistical features and 4(distances) x 22 gray level co-occurrence matrices (GLCM) texture features were calculated from segmented masses and background samples. Support vector machine (SVM) was implemented to classify the segmented suspicious masses features against the background and major lesion features. The GLCM features are reduced to 8, at distance 4, such that 93% of lesions were having classification greater than 80% versus the background. For 20 images one sub-lesion (satellite), on average, was detected beside the major lesion. The ob jective of the paper is to estimate the amount of features similarity between the major lesion and the sub-masses that could be segmented from the same image. The proposed algorithm would effect ively help in the early detection of breast cancer.