MAR-GAN: Multi attention residual generative adversarial network for tumor segmentation in breast ultrasounds

Abstract

Introduction

Ultrasonography is among the most regularly used methods for earlier detection of breast cancer. Automatic and precise segmentation of breast masses in breast ultrasound (US) images is essential but still a challenge due to several causes of uncertainties, like the high variety of tumor shapes and sizes, obscure tumor borders, very low SNR, and speckle noise.

Method

To deal with these uncertainties, this work presents an effective and automated GAN based approach for tumor segmentation in breast US named MAR-GAN, to extract rich, informative features from US images. In MAR-GAN the capabilities of the traditional encoder-decoder generator were enhanced by multiple modifications. Multi-scale residual blocks were used to retrieve additional aspects of the tumor area for a more precise description. A novel boundary and foreground attention (BFA) module is proposed to increase attention for the tumor region and boundary curve. The squeeze and excitation (SE) and the adaptive context selection (ACS) modules were added to increase representational capability on encoder side and facilitates better selection and aggregation of contextual information on the decoder side respectively. The L1-norm and structural similarity index metric (SSIM) were added into the MAR-GAN’s loss function to capture rich local context information from the tumors’ surroundings.

Results

Two breast US datasets were utilized to evaluate the effectiveness of the suggested approach. Using the BUSI dataset, our network outperformed several state-of-the-art segmentations models in IoU and Dice metrics, scoring 89.27 %, 94.21 %, respectively. The suggested approach achieved encouraging results on UDIAT dataset, with IoU and Dice scores of 82.75 %, 88.54 %, respectively.