TranSpeed: Transformer-based Generative Adversarial Network for Speed-of-sound Reconstruction in Pulse-echo Mode

Longitudinal speed of sound (SoS) serves as a valuable biomarker for tissue characterization and facilitates phase aberration correction to enhance ultrasound image quality. While deep learning (DL) methods have shown promise for SoS reconstruction, traditional convolutional neural networks (CNNs) h...

Full description

Saved in:
Bibliographic Details
Main Authors: Chen, Haotian, Zhu, Yuanbin, Zuo, Jingyi, Kabir, MD Rizwanul, Han, Aiguo
Format: Conference Proceeding
Language:English
Subjects:
Breast
Deep learning
Encoding
Frequency control
Generative adversarial networks
Image quality
Image reconstruction
Neural networks
Numerical models
Numerical simulation
Pulse echo
Speed of Sound Reconstruction
Transformer
Transformers
Ultrasonic imaging
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Longitudinal speed of sound (SoS) serves as a valuable biomarker for tissue characterization and facilitates phase aberration correction to enhance ultrasound image quality. While deep learning (DL) methods have shown promise for SoS reconstruction, traditional convolutional neural networks (CNNs) have limitations in capturing long-range dependencies. This study introduces TranSpeed, a transformer-based generative adversarial network designed to reconstruct complex SoS distributions with a large receptive field. TranSpeed is designed to capture long-range dependencies by patchifying input phase-shift maps and incorporating positional encoding. A model was trained on numerical simulation samples and evaluated on an open-access breast ultrasound computed tomography (USCT) dataset. The results demonstrated the feasibility of TranSpeed in reconstructing complex SoS patterns.
ISSN:2375-0448
DOI:10.1109/UFFC-JS60046.2024.10794018