ReconFormer: Accelerated MRI Reconstruction Using Recurrent Transformer
Problem
Reconstructing high-fidelity MRI images from highly under-sampled k-space data is a challenging ill-posed inverse problem.
Why it matters
Under-sampling enables accelerated MRI, but excessive under-sampling makes image reconstruction ill-posed.
What the AI did
ReconFormer, a recurrent Transformer using Recurrent Scale-wise Attention and Recurrent Pyramid Transformer Layers, iteratively reconstructs high-fidelity MR images from highly under-sampled k-space data.
Results
Up to 8× acceleration; 1.1M trainable parameters; validated on multiple datasets (abstract truncated before full metrics)
Impact
Abstract supports a lightweight model validated across multiple datasets; clinical adoption not stated.
Record
| Venue | IEEE Transactions on Medical Imaging |
|---|---|
| Published | 2023 |
| Work type | article |
| Authors | Pengfei Guo, Yiqun Mei, Jinyuan Zhou, Shanshan Jiang, Vishal M. Patel |
| Citations | 127 (OpenAlex, at import time) |
| References | 81 |
| Model or method | ReconFormer (recurrent Transformer, RPTLs, RSA) |
| Organization | not stated in source |
| Primary source | https://doi.org/10.1109/tmi.2023.3314747 |
| Open access | green |
| Keywords | MRI reconstruction, transformer, recurrent attention, k-space undersampling, deep learning |
Importance rating
significant — thousands to millions affected
Accelerated MRI reconstruction could benefit many patients via faster scans, but the abstract gives no scale of clinical impact; it shows up to 8x acceleration with a lightweight model.