A Computer-Aided Diagnosis System of Breast Lesion Classification Based on Multi Angle Fusion Strategy in Ultrasound Images
Yijun Zhao, Tobore Igbe, Wenbin Yan, Dashun Que · 2020
Ultrasound is one of the most widely applied imaging modalities for breast lesion assessment. The accurate breast lesion diagnosis can improve patients’ survival rate. The purpose of this study is to develop a computer-aided diagnosis (CAD) system that can acquire information from grayscale and elastic ultrasound images to classify benign and malignant breast tumors. The U-Net was used for automatic segmentation of grayscale ultrasound images. After reconstructing elastic ultrasound images, 218 breast features (morphological, gray, calcification, texture and elastic features) were extracted. We proposed a feature selection method based on multi-angle fusion strategy (MAFS) to select important breast features. In this method, least absolute shrinkage and selection operator (LASSO), mutual information (MI) and random forest (RF) feature selectors were fused to establish a robust and accurate feature selector. A support vector machine (SVM) classifier was used for breast lesion classification. The ultrasound images in our study contained 77 malignant cases out of 199 breast lesion cases, in which the results of core biopsy or fine-needle aspiration were regarded as the golden standard. The area under the receiver operating characteristic curve (AUC) of the combined breast set based on MAFS, LASSO, MI and RF were 0.897, 0.871, 0.880 and 0.869, respectively. The accuracy, sensitivity and specificity of the breast lesion classification performance based on MAFS were 91.0%, 84.0% and 95.4%, respectively. Our results demonstrate the feasibility that the CAD system based on MAFS can be applied in breast lesion diagnosis. Also, the CAD system based on MAFS performed well in grayscale and elastic ultrasound image sets, showing it had robust and generalization ability in ultrasound images.