A two-stage progressive deep segmentation network for tumor detection in breast ultrasound images

Identifier:  imarina:9471737

Handle: https://hdl.handle.net/20.500.11797/imarina9471737

Authors:  Zaidkilani; N; Abdel-Nasser; M; García; MA; Puig; D

Abstract:
Segmenting tumorous regions in breast ultrasound images is a challenging problem due to several factors, including the relatively low contrast of the available images, the presence of speckle noise, and the considerable variations in breast mass sizes and shapes. Current methods are not precise enough and prone to misdetections. An efficient deep neural model is proposed for automatically segmenting tumorous regions in breast ultrasound images. The model is constituted by two consecutive encoder-decoder (autoencoder) networks. The first autoencoder extracts a preliminary binary mask from the given image. The second autoencoder refines that mask after concatenating it with the original image. The encoders within each autoencoder can be defined by applying any state-of-the-art network. In addition, cost-sensitive learning has been used in order to focalize training on the segmentation errors of the minority class (tumor). Semantic segmentation based on advanced deep learning methods is thus applied in order to enhance tumor segmentation in breast ultrasound images. The proposed model offers advanced capabilities for automated segmentation with the aim of helping physicians identify and diagnose tumors using state-of-the-art techniques. This model outperforms recent tumor segmentation methods in the experiments conducted on two public datasets of breast ultrasound images (UDIAT and BUSI). The largest improvement for both datasets was achieved by using CoAtNet as baseline model (Dice index equal to 84.49% and 78.94%, respectively). © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.