Impact of Despeckling on Segmentation of Breast Ultrasonographic Images

Abstract

Breast ultrasonographic imaging is a crucial imaging technique for diagnosing breast tissue abnormalities, especially in cases of dense breast tissue. However, speckle noise in breast ultrasonographic (BUS) images leads to inaccurate segmentation of anatomical structures and lesions, which further leads to inaccurate classification of breast tissue abnormalities. This study presents the impact of despeckling methods on BUS image segmentation based on image quality metric evaluation and clinical validation. In this work, first, speckle noise is reduced from BUS images of the hybrid dataset (BUSI+PGI Rohtak, HR, India), and then an efficient DL based segmentation technique is applied. The results obtained are compared with the segmentation of original images, and it is noticeable that segmentation performance in terms of mean IOU and accuracy increased from 0.904 to 0.920 and from 0.937 to 0.945, respectively, when segmentation is performed on despeckled BUS images, in spite of the original BUS images. The BUS images are carefully marked under the guidance of a Senior Professor of radiology, and the segmented BUS images are compared with the original BUS images by evaluating the overlap region in terms of mean IOU. It is observed that segmentation with despeckled images preserves diagnostic information more precisely.