Assessment of texture feature extraction to classify the benign and malignant lesions from breast ultrasound images

Abstract

Detection of Breast cancer which is growing fast is crucial. Mammography procedures are uncomfortable involving ionizing radiation. Ultrasound, a broadly popular medical imaging modality which is noninvasive, real time, convenient, and of low cost is preferable. But speckle noise corrupts, diminishing the excellence of ultrasound image. curvelet, shearlet, and tetrolet methods are a prerogative to reduce the speckle noise and enhance signal-to-noise ratio (SNR) and contrast resolution of the image before segmentation. After segmentation, to classify the breast lesion, a set of 15 texture features are extracted from the original image and also from curvelet, shearlet, and tetrolet filtered images. Optimal features are selected to increase the classification performance by using a Releiff algorithm and four best features are taking into account for feature ranking. These features used to classify the lesions from breast ultrasound images by using SVM with polynomial kernel and FKNN algorithms. The proposed algorithm tested on 178 samples, of which 85 samples are benign masses, and 93 samples are malignant masses. The results show that SVM classifier outperforms Fuzzy KNN with 87.39% accuracy, 88.89% sensitivity, and specificity of 88.46% for the texture features from the tetrolet transform.