CoAtUNet: A symmetric encoder-decoder with hybrid transformers for semantic segmentation of breast ultrasound images

The CoAtNet hybrid deep model yields state-of-the-art performance in image analysis by integrating the strengths of convolutional neural networks (CNNs) along with self-attention mechanisms, combining local feature extraction with global dependency modeling. In this paper, we introduce CoAtUNet, a n...

Full description

Saved in:
Bibliographic Details
Published in:Neurocomputing (Amsterdam) Vol. 629; p. 129660
Main Authors: Zaidkilani, Nadeem, Garcia, Miguel Angel, Puig, Domenec
Format: Journal Article
Language:English
Published: Elsevier B.V 07.05.2025
Subjects:
Breast cancer
CoAtNet
Deep neural networks
Semantic segmentation
Symmetric encoder-decoder
Transformers
Ultrasound images
Deep neural networks
Semantic segmentation
Transformers
Ultrasound images
Breast cancer
CoAtNet
Symmetric encoder-decoder
ISSN:0925-2312
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The CoAtNet hybrid deep model yields state-of-the-art performance in image analysis by integrating the strengths of convolutional neural networks (CNNs) along with self-attention mechanisms, combining local feature extraction with global dependency modeling. In this paper, we introduce CoAtUNet, a novel semantic segmentation network that utilizes a symmetric encoder–decoder architecture in which CoAtNet is applied to both branches. CoAtUNet is specifically designed for tumor segmentation in breast ultrasound (BUS) images, aiming to enhance segmentation accuracy and robustness. By leveraging CNNs for detailed local feature extraction and Transformers for capturing long-range dependencies, CoAtUNet addresses the unique challenges of breast ultrasound images, such as noise, low contrast, and the variability in tumor shapes and sizes. This hybrid approach enables CoAtUNet to outperform traditional asymmetric configurations and other advanced segmentation models. Evaluations conducted on two public breast ultrasound datasets, UDIAT and BUSI, demonstrate the superior performance of CoAtUNet, achieving Dice coefficients of 79.12% and 76.21%, respectively. These results not only surpass existing tumor segmentation methods, but also highlight the potential of CoAtUNet to assist clinicians in precise tumor detection, facilitating more accurate and reliable diagnoses. The CoAtUNet implementation is publicly available at https://github.com/NadeemAlkilani.
ISSN:0925-2312
DOI:10.1016/j.neucom.2025.129660