Related Experiment Video
Updated: May 29, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
PMFSNet: Polarized multi-scale feature self-attention network for lightweight medical image segmentation
Jiahui Zhong1, Wenhong Tian1, Yuanlun Xie2
1School of Information and Software Engineering, University of Electronic Science and Technology of China, Chengdu, 610054, PR China.
Computer Methods and Programs in Biomedicine
|February 1, 2025
Summary
PMFSNet offers efficient medical image segmentation by balancing global and local features with fewer parameters. This lightweight model achieves superior performance on diverse datasets, ideal for resource-constrained environments.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Artificial Intelligence
Background:
- Current medical image segmentation models are precise but computationally intensive and large.
- Large models risk overfitting on limited medical data and neglect local feature representation.
- Transformer-based models offer wide receptive fields but are computationally demanding.
Purpose of the Study:
- To develop a novel, lightweight medical image segmentation model (PMFSNet).
- To balance global and local feature processing while minimizing computational redundancy.
- To enable efficient deployment on edge devices and in resource-constrained settings.
Main Methods:
- Proposed PMFSNet, a streamlined UNet-based hierarchical model.
- Incorporated a plug-and-play multi-scale feature enhancement (PMFS) block.
- Simplified self-attention mechanisms for reduced computational complexity.
Main Results:
- Achieved superior segmentation performance with fewer than one million parameters.
- Obtained high IoU scores on 3D CBCT Tooth, ovarian tumors, skin lesions, and gastrointestinal polyps.
- Demonstrated strong DSC scores on retinal vessel segmentation datasets (DRIVE, STARE, CHASE-DB1).
Conclusions:
- PMFSNet shows competitive performance with significantly reduced parameters and inference time.
- Offers an optimal balance between efficiency and performance for medical image analysis.
- Suitable for integration and deployment in resource-constrained clinical environments.
