Harnessing Lightweight Transformer With Contextual Synergic Enhancement for Efficient 3D Medical Image Segmentation

Xinyu Liu , Zhen Chen , Wuyang Li

IEEE Transactions on Pattern Analysis and Machine Intelligence +

|

December 5, 2025

View abstract on PubMed

Summary

Light-UNETR enhances 3D medical image segmentation by improving model and data efficiency. This lightweight transformer significantly reduces computational costs and data requirements, achieving superior performance with less labeled data.

Area Of Science

Medical Imaging
Artificial Intelligence
Computer Vision

Background

Transformers excel in 3D medical image segmentation but demand high computational resources and extensive labeled data.
Current methods face limitations in efficiency and data requirements, hindering widespread clinical adoption.

Purpose Of The Study

To develop a lightweight transformer (Light-UNETR) for efficient 3D medical image segmentation.
To enhance both model efficiency and data efficiency in transformer-based medical image analysis.

Main Methods

Proposed Light-UNETR with a Lightweight Dimension Reductive Attention (LIDR) module for feature extraction and a Compact Gated Linear Unit (CGLU) for parameter efficiency.
Introduced a Contextual Synergic Enhancement (CSE) strategy using Attention-Guided Replacement and Spatial Masking Consistency to improve learning from unlabeled data.

Main Results

Light-UNETR demonstrated superior performance and efficiency across multiple benchmarks.
Achieved 1.43% higher Jaccard index on Left Atrial Segmentation using only 10% labeled data compared to BCP.
Reduced FLOPs by 90.8% and parameters by 85.8%.

Conclusions

Light-UNETR offers a highly efficient and effective solution for 3D medical image segmentation.
The proposed methods significantly improve data efficiency, enabling robust segmentation with limited labeled data.