The Impact of Filtering for Breast Ultrasound Segmentation using A Visual Attention Model

Abstract

Breast cancer can threaten women's health and become a cause of death. Reducing mortality from breast cancer necessitates early recognition of its signs and symptoms. An essential step in building an early detection system is to segment the breast ultrasound image (BUS). The accuracy of segmentation has a direct bearing on the effectiveness of quantitative analysis and the detection of breast tumor. However, this image segmentation becomes constrained because the BUS image has a shallow quality. Therefore, it is necessary to take preprocessing steps to improve the image. This study aims to compare the efficiency of various filtering techniques for BUS segmentation with the visual attention model. There are 12 filters tested in this study, including Mean, Median, Bilateral, Fast nonlinear, Lee, Lee-enhance, Frost, Kuan, Gamma, Wiener, Speckle Reduction Anisotropic Diffusion Filter (SRAD), and Detail Preserved Anisotropic Diffusion Filter (DPAD). The segmentation process uses a Convolutional Neural Network (CNN) based network architecture, namely Visual Geometry Group architecture with 16 layers (VGG-16). The segmentation results were analyzed using three visual attention models. The results showed that the image before filtering and after filtering showed visually significant results.