A Dual-Attention-ResUNet++ for Breast Tumor Segmentation using Ultrasound Images

Document Type : Original Article

Authors

1 1Electronics and Communications Engineering (ECE) Department, Faculty of Engineering, Mansoura University, Mansoura, 35516, Egypt

2 Electronics and Communications Engineering (ECE) Department, Faculty of Engineering, Mansoura University, Mansoura, 35516, Egypt

Abstract

This study introduces DAtt-ResUNet++, an advanced deep learning model specifically designed to enhance breast tumor segmentation in ultrasound images. The model integrates a Dual-Attention mechanism within the ResUNet++ framework, significantly improving its ability to focus on tumor regions while capturing relevant contextual information from surrounding tissue. By combining both spatial and channel-based attention, DAtt-ResUNet++ achieves higher segmentation precision. For evaluation, the model was rigorously tested on a public dataset of 780 breast ultrasound images, utilizing a robust 10-fold cross-validation approach. It achieved impressive results with a Dice similarity coefficient of 90.40 ± 0.88%, Intersection over Union (IoU) of 84.62 ± 1.12%, Sensitivity of 89.82 ± 0.75%, Precision of 93.44 ± 0.56%, and Accuracy of 98.73 ± 0.12%. These results position DAtt-ResUNet++ as a competitive tool against state-of-the-art methods, showcasing its potential to improve breast tumor segmentation in ultrasound imaging. Future research will explore further optimizations and validation on additional datasets.

Keywords

Main Subjects

Files

Share

How to cite

Statistics