Multiscale hybrid superpixel method for pre-processing and segmentation of breast tumors in ultrasound images

Background The ultrasound test for cancer screening is low-cost and non-invasive. A major drawback is that multiplicative speckle noise can jeopardize the efficiency of the test. This research proposes a new algorithm to reduce noise and segment breast ultrasound. Methods The method is in two stages. Stage one is based on a combination of the multiscale approach, the Wiener filter, and a new combination of wavelet-transform denoising and the anisotropic Perona-Malik-type filter. In the second stage, pre-processed image is transformed into multiscale images, and then, a boundary efficient superpixel decomposition of the multiscale images is created. Finally, the tumor region is generated by the boundary graph cut segmentation method.Results The algorithm has been tested on 50 synthetic images degraded by speckle noise with varying intensity and 250 breast ultrasound (BUS) images from two datasets. The results are compared with selected state-of-the-art filters. The proposed approach shows better performance in terms of standard evaluation measures. The results are compared with ground truth by the DICE coefficient, the Jaccard coefficient, and the Hausdorff distance. The proposed filter also achieves high accuracy in terms of these segmentation measures.Conclusions The proposed two-stage algorithm achieves better accuracy, compared to selected state-of-the-art methods applied to BUS images.