Research map / Image generation / Super-resolution
MFTN: Multi-Level Feature Transfer Network Based on MRI-Transformer for MR Image Super-resolution
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MFTN, a pyramid multi-scale reconstruction network plus a multi-level MRI-Transformer feature-transfer branch that uses a reference MR modality to super-resolve low-resolution MR images, trained with a loss that adds a perceptual contrastive term.
Problem
Super-resolution of low-resolution MR images suffers from insufficient detail reconstruction, and subpixel misalignment between LR images and multi-contrast reference images causes artifacts and edge blurring.
Approach
A pyramid multi-scale feature reconstruction network fed by multi-level transfer features, where each level is produced by an MRI-Transformer module that builds cross-attention correlations between LR features and reference-image features (including a degraded reference) so that long-range matching reduces sensitivity to subpixel displacement; plus a perceptual contrastive loss.
Key results
- Achieves higher PSNR and SSIM than EDSR, MCSR, NEU, MINet and MSDT on FastMRI and AXA for 2x and 4x SR (per-table averages, numbers not extracted in text)
- Ablation shows MRI-Transformer outperforms the T2Net Task Transformer variant (RT), and combining it with perceptual contrastive loss gives the best results among four cases
- Joint loss (Rec+Per+Com+SSIM) and multi-scale structure each outperform their ablated variants on FastMRI 4x SR
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MFTN instead of EDSR, MCSR, NEU, MINet, or MSDT | MR image super-resolution · FastMRI, 2× and 4× | PSNR | Helped |
| Use MFTN instead of EDSR, MCSR, NEU, MINet, or MSDT | MR image super-resolution · AXA, 2× and 4× | PSNR | Helped |
| Replace the Task Transformer module with the MRI-Transformer module (ablation) | MFTN · FastMRI, 2× | PSNR | Helped |
| Add perceptual contrastive loss (ablation) | MFTN with MRI-Transformer · FastMRI, 2× | PSNR | Helped |
| Use the joint loss function instead of the tested alternative combinations of loss terms (ablation) | MFTN · FastMRI, 4× | PSNR | Helped |
| Use a multi-scale network structure instead of low-scale or high-scale feature reconstruction (ablation) | MFTN · FastMRI, 4× | PSNR | Helped |
Limitations
- Evaluated only on two datasets (FastMRI PD/PDFS and a self-built AXA T2/FLAIR set)
- Self-built AXA dataset is private and small (50 patients)
- Relies on a well-aligned reference image; subpixel misalignment is only addressed implicitly
Experimental setup
Datasets: FastMRI AXA
Benchmarks: FastMRI AXA
Metrics: PSNR SSIM
Baselines: EDSR MCSR NEU MINet MSDT MSPN WavTrans McMRSR
Builds on: T2Net AECR-Net perceptual loss VGG-19 perceptual loss Transformer multi-head attention
Closest papers by meaning
- Transformer-Empowered Multi-Scale Contextual Matching and Aggregation for Multi-Contrast MRI Super-Resolution (CVPR 2022 · similarity 0.88)
- MFTN: A Multi-scale Feature Transfer Network Based on IMatchFormer for Hyperspectral Image Super-Resolution (ICML 2024 · similarity 0.84)
- Boosting ViT-based MRI Reconstruction from the Perspectives of Frequency Modulation, Spatial Purification, and Scale Diversification (AAAI 2025 · similarity 0.83)
- Rethinking Diffusion Model for Multi-Contrast MRI Super-Resolution (CVPR 2024 · similarity 0.83)
- Null Space Matters: Range-Null Decomposition for Consistent Multi-Contrast MRI Reconstruction (AAAI 2024 · similarity 0.82)
- A Unified Model for Compressed Sensing MRI Across Undersampling Patterns (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.