Segmentation And Classification Of Breast Lesions In Ultrasound Images

This paper proposes a new approach for computer-aided diagnosis (CAD) system with automatic contouring and texture analysis to aid in the classification of breast lesions using ultrasound. First, the goal is to remove the speckle noise while preserving important information from the lesion boundaries, anisotropic diffusion filtering is applied to the ultrasonic image. A morphological watershed transform is used for BUS image segmentation, automatically extracts the precise contour of breast lesions. 32 GLCM features are extracted from the segmented lesion. Support vector machine (SVM) classifier utilizes the selected feature vectors to identify the breast lesion as benign or malignant. Database consists of 50 images (38 Benign and 12 Malignant) and the computer-delineated margins were compared against manual outlines drawn by radiologist. The area under receive operating Characteristic (ROC) curve for proposed CAD systems using all textural features is 0.89. The classifier performance is evaluated by 4 parameters, Accuracy = 92.00, Sensitivity = 94.73, Specificity = 83.34, Matthew's correlation coefficient (MCC) = 0.78.