Related Experiment Video
Updated: Aug 6, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Dynamic Aware Biopsy Needle Identification in Ultrasound Images Using Temporal Prior Guided U-Net Cross Transformer
Myeongjin Lee1, Dong Gyu Beom1, Eun Hui Bae2
1Department of Intelligent Electronics and Computer Engineering, Chonnam National University, Gwangju, Republic of Korea.
Ultrasound in Medicine & Biology
|July 22, 2026
Summary
This study introduces UXFormer, a novel deep learning model that improves ultrasound-guided needle detection by combining classical background subtraction with a U-Net and Vision Transformer fusion. The enhanced model achieves superior accuracy, especially with limited training data.
Area of Science:
- Medical Imaging and Image Analysis
- Deep Learning in Healthcare
- Minimally Invasive Procedures
Background:
- Ultrasound-guided needle placement is crucial for biopsies, anesthesia, and drug delivery.
- Current deep learning models like U-Net have limitations in receptive field, while Vision Transformers require extensive data.
- Needle detection accuracy is vital for patient safety and procedural success.
Purpose of the Study:
- To enhance deep learning frameworks for more reliable ultrasound-guided needle detection.
- To overcome limitations of existing models by integrating classical signal processing with advanced deep learning.
- To enable accurate needle localization even with small training datasets.
Main Methods:
- Proposed U-Net Cross Transformer (UXFormer), a hybrid architecture combining U-Net and Vision Transformer.
- Integrated classical background subtraction to enrich inductive bias and capture temporal prior information.
- Employed temporal-to-spatial cross-attention and global-to-local cross-convolutional blocks for enhanced feature fusion.
Main Results:
- UXFormer demonstrated significant improvements over competing methods.
- Achieved a 14.2% increase in Jaccard index, 9.0% in Dice score, and 8.5% in recall.
- Reduced tip position error by 44.0% and trajectory angle error by 17.7%, even with variable needle visibility.
Conclusions:
- The unified framework effectively bridges classical signal processing and deep learning.
- UXFormer strengthens inductive bias and significantly reduces data requirements for transformer-based models.
- Superior needle detection performance was achieved across diverse conditions and limited data volumes.
