Related Experiment Video
Updated: Jul 8, 2025
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
406
Transformers in medical image segmentation: a narrative review
Rabeea Fatma Khan1, Byoung-Dai Lee1, Mu Sook Lee2
1Department of Computer Science, Graduate School, Kyonggi University, Suwon, Republic of Korea.
Quantitative Imaging in Medicine and Surgery
|December 18, 2023
Summary
This survey reviews transformer networks for medical image segmentation, finding they match convolution neural networks and benefit from hybrid approaches. Future work should address data limitations and explore transfer learning.
Area of Science:
- Computer Vision
- Medical Imaging
- Artificial Intelligence
Background:
- Transformers, initially for NLP, are now key in computer vision.
- They are increasingly vital for medical image segmentation, rivaling convolution neural networks.
- Hybrid methods combining transformers and CNNs excel at capturing local and global contexts.
Purpose of the Study:
- To survey innovative transformer networks for efficient medical image segmentation.
- To analyze architectures and attention mechanisms enhancing medical data processing.
- To identify research gaps and future directions in transformer-based medical imaging.
Main Methods:
- Literature search across major academic databases (Google Scholar, arXiv, etc.).
- Focus on English-language research published between 2021 and 2023.
- Qualitative and quantitative analysis of transformer architectures and attention mechanisms.
Main Results:
- Transformers demonstrate competitive performance with CNNs in medical image segmentation.
- Hybrid models show significant improvements by integrating local and global feature extraction.
- Various transformer architectures and attention mechanisms are effective for complex medical data.
Conclusions:
- Transformer networks are a powerful tool for medical image segmentation.
- A key challenge is the scarcity of annotated medical data for training deep learning models.
- Future directions include transfer learning and foundation models for specialized segmentation tasks.
