VSSR: View-Specialized Sequential Refinement for Accelerated MRI

Post-Reconstruction Volumetric Refinement for Accelerated MRI via Cross-Plane Consistency

IEEE Engineering in Medicine and Biology Conference (EMBC) 2026 — Paper ID: 4868

Alexander Nazarov, Nahum Kiryati, Dani Roizen, Gahl Greenberg, Arnaldo Mayer Tel-Aviv University & Sheba Medical Center

---

Overview

VSSR is a post-reconstruction refinement framework for accelerated MRI. Three view-specialized expert networks refine the volume sequentially — one per anatomical plane (sagittal, axial, coronal) — each correcting the residual artifacts left by the previous stage.

Code & instructions: github.com/MR-Nazarov/VSSR

---

Model Weights

Lexer1/VSSR/
├── vssr_stage1_sagittal.pth      ← RDUNet SAC cascade, Stage 1
├── vssr_stage2_axial.pth         ← RDUNet SAC cascade, Stage 2
├── vssr_stage3_coronal.pth       ← RDUNet SAC cascade, Stage 3 (final output)
├── baselines/
│   ├── sunet_sagittal_8x.pth     ← Single-view SUNet baseline
│   ├── sunet_axial_8x.pth
│   └── sunet_coronal_8x.pth
├── cascades_8x/
│   └── SAC_final.pth             ← Best cascade 8× (PSNR 31.28 dB on IXI)
└── cascades_4x/
    └── SAC_final.pth             ← Best cascade 4×

Download

from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="Lexer1/VSSR", filename="cascades_8x/SAC_final.pth")

Or download all weights at once:

hf download Lexer1/VSSR --local-dir ./models

---

Training Data

• IXI dataset — publicly available at brain-development.org/ixi-dataset, 8× retrospective undersampling
• Sheba Medical Center — prospectively accelerated in-house dataset (not publicly available)

---

Citation

@inproceedings{nazarov2026vssr,
  title     = {Post-Reconstruction Volumetric Refinement for Accelerated {MRI} via Cross-Plane Consistency},
  author    = {Nazarov, Alexander and Kiryati, Nahum and Roizen, Dani and Greenberg, Gahl and Mayer, Arnaldo},
  booktitle = {2026 IEEE Engineering in Medicine and Biology Conference (EMBC)},
  year      = {2026},
  note      = {Paper ID: 4868}
}

---

License

Model weights are licensed under CC BY-SA 4.0. Code is licensed under MIT — see the GitHub repository.