Related Experiment Video
Updated: Jun 29, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deep local-to-global feature learning for medical image super-resolution
Wenfeng Huang1, Xiangyun Liao2, Hao Chen3
1Guangdong Provincial Key Laboratory of Computer Vision and Virtual Reality Technology, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen 518000, China; Faculty of Engineering and Information Technology, University of Technology Sydney, Broadway, NSW 2007, Australia.
This study introduces LGSR, a novel deep learning framework for enhancing low-resolution medical images. LGSR effectively restores both fine details and global context, improving diagnostic accuracy across various imaging modalities.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Medical image quality is crucial for diagnosis but often compromised by low resolution and artifacts.
- Existing super-resolution (SR) techniques struggle with the unique challenges of medical images, such as diverse anatomical structures and the need for both local and global information.
- Standard convolution neural networks (CNNs) face difficulties in accurately restoring medical images due to variations in organ and tissue appearance.
Purpose of the Study:
- To develop an advanced image super-resolution (SR) method specifically designed for medical imaging.
- To address the limitations of generic SR techniques in capturing both fine local details and essential global context in medical scans.
- To improve the diagnostic utility of low-resolution (LR) medical images by restoring high-resolution (HR) counterparts.
Main Methods:
- A novel CNN-ViT neural network, termed LGSR (Local-to-Global feature learning for medical image Super-Resolution), was developed.
- Incorporated a dynamic-local learning framework with deformable convolutions for capturing diverse anatomical details.
- Integrated pixel-pixel and patch-patch global learning using non-local mechanisms and a vision transformer (ViT) to preserve global context.
Main Results:
- LGSR demonstrated superior performance in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) compared to state-of-the-art methods.
- The method achieved excellent visual quality in restoring both local features and global information across five medical imaging types (Ultrasound, OCT, Endoscope, CT, MRI).
- LGSR showed competitive computational costs, with efficient network parameters, runtime, and Floating Point Operations (FLOPs).
Conclusions:
- The proposed LGSR method effectively enhances medical image super-resolution by integrating local and global feature learning.
- LGSR offers significant improvements for medical image analysis, outperforming existing SR techniques on diverse datasets.
- The enhanced image quality from LGSR positively impacts downstream tasks, such as OCT image segmentation, highlighting its clinical relevance.
