Related Experiment Video
Updated: Aug 9, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
TransLiteUNet: A Lightweight CNN-Transformer Hybrid for Efficient 3D Brain Tumor Segmentation with Sub-0.5 M
Lixin Zhou1,2, Yuanyuan Yang1, Yunfeng Yang1,2
1Laboratory for Medical Imaging Informatics, Shanghai Institute of Technical Physics, Chinese Academy of Sciences, Shanghai 200083, China.
Journal of Imaging
|July 27, 2026
Summary
We developed TransLiteUNet, a novel lightweight 3D deep learning model for accurate brain tumor segmentation. This Transformer-CNN hybrid significantly reduces computational costs while outperforming existing methods.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Accurate 3D brain tumor segmentation requires both local and global feature analysis.
- Existing models often face challenges in balancing high accuracy with computational efficiency.
- Transformer and Convolutional Neural Network (CNN) architectures have complementary strengths for feature extraction.
Purpose of the Study:
- To propose TransLiteUNet, a lightweight 3D deep learning model for accurate brain tumor segmentation.
- To enhance parameter efficiency and reduce computational cost in 3D segmentation models.
- To achieve state-of-the-art performance without requiring model pretraining.
Main Methods:
- Developed TransLiteUNet, a hybrid 3D CNN-Transformer architecture.
- Introduced a 3D axial depthwise separable convolution residual structure (3DRes-ADS block) for parameter efficiency.
- Integrated a lightweight LiteViT module to improve global feature modeling cost-effectively.
Main Results:
- TransLiteUNet (0.43 M parameters, 14.98 GFLOPs) and TransLiteUNet-S (0.31 M parameters, 7.68 GFLOPs) demonstrate significantly reduced model complexity.
- Achieved superior performance compared to leading models on two public datasets under identical conditions.
- Demonstrated orders-of-magnitude reduction in parameter and computational costs, optimizing inference and training.
Conclusions:
- TransLiteUNet offers an effective and efficient solution for 3D brain tumor segmentation.
- The proposed lightweight architecture balances accuracy and computational cost.
- This approach advances the field of medical image analysis for neuro-oncology.