A Lightweight 3D Distillation Volumetric Transformer for 3D MRI Super-Resolution

Jianwei Zhao , Tao Hong , Hao Qi

IEEE Journal of Biomedical and Health Informatics +

|

April 1, 2025

View abstract on PubMed

Summary

This study introduces a lightweight 3D Transformer for magnetic resonance imaging (MRI) super-resolution, improving image quality while reducing computational costs. The Transformer-based dual-attention feature distillation (TDAFD) network balances performance and efficiency for 3D MRI reconstruction.

Area Of Science

Medical Imaging
Artificial Intelligence
Computer Vision

Background

Traditional 2D and existing 3D super-resolution (SR) methods for 3D MRI face challenges with increased network parameters and computational costs for higher accuracy.
There is a need for efficient 3D MRI SR methods that can effectively utilize volumetric information without prohibitive resource demands.

Purpose Of The Study

To propose a lightweight 3D multi-scale distillation volumetric Transformer, named Transformer-based dual-attention feature distillation (TDAFD) network, for 3D MRI super-resolution.
To address the trade-off between reconstruction accuracy and computational cost in 3D MRI super-resolution.

Main Methods

The TDAFD network incorporates dual-attention feature distillation (DAFD) modules and recursive volumetric Transformers (RVT).
DAFD modules include multi-scale feature distillation (MSFD) for global feature extraction and feature enhancement dual attention blocks (FEDAB) for focusing on key features.
RVT adapts 2D Transformers to 3D and uses recursion to capture long-range dependencies efficiently, reducing network parameters.

Main Results

The TDAFD network effectively extracts deeper features through multi-scale distillation and Transformer mechanisms.
Experiments demonstrate superior reconstruction performance compared to popular 3D MRI SR methods.
The proposed method achieves a balance between performance and network parameters, reducing weights and FLOPs.

Conclusions

The proposed TDAFD network offers an efficient and effective solution for 3D MRI super-resolution.
It successfully addresses the limitations of existing methods by optimizing feature extraction and network architecture.
The TDAFD network provides a promising approach for enhancing 3D MRI quality with reduced computational overhead.