---
license: mit
library_name: glimpse-ct
tags:
- computed-tomography
- inverse-problems
- image-reconstruction
- implicit-neural-representation
- sparse-view-ct
pipeline_tag: image-to-image
datasets:
- AmirEhsan1995/lodopab-ct-glimpse
---

# GLIMPSE — Generalized Locality for Scalable and Robust CT

Coordinate-based CT reconstruction from sparse-view sinograms
([paper](https://ieeexplore.ieee.org/abstract/document/11018464),
[arXiv](https://arxiv.org/abs/2401.00816),
[code](https://github.com/swing-research/Glimpse)), published in *IEEE Transactions
on Medical Imaging*. Instead of reconstructing a whole image at once, GLIMPSE
predicts **one pixel at a time** from only the sinogram data *local* to that
pixel's coordinate. This locality makes it resolution-agnostic and gives strong
**out-of-distribution generalization** — e.g. train on natural images / faces and
reconstruct medical brain scans without retraining.

## This checkpoint

| | |
|---|---|
| Image size | 128 |
| Projection angles (views) | 50 |
| Noise (SNR, dB) | 45 |
| Forward operator | `odl` (circle=False) |
| Parameters | ~1.3 M |

## Results (LoDoPaB-CT, 50 views, calibrated)

GLIMPSE substantially outperforms classical filtered back-projection (FBP), both
in-distribution and out-of-distribution:

| Set | FBP PSNR / SSIM | GLIMPSE PSNR / SSIM |
|---|---|---|
| In-distribution (LoDoPaB-CT test) | 30.8 dB / 0.79 | **38.0 dB / 0.93** |
| Out-of-distribution (brain CT) | 26.1 dB / 0.51 | **31.6 dB / 0.88** |

## Usage

```python
import numpy as np, torch
from glimpse import GlimpseModel, Config
from glimpse.operators import build_operator
from glimpse.reconstruct import make_coordinate_grid, reconstruct_image

device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = GlimpseModel.from_pretrained("AmirEhsan1995/Glimpse").eval().to(device)

# Build the matching ODL parallel-beam operator (50 views over [0, 180) deg).
cfg = Config.from_yaml('configs/lodopab.yaml')        # from the GitHub repo
_, init_angles = cfg.resolve_angles()
operator = build_operator(cfg, np.deg2rad(init_angles))

volume = torch.as_tensor(my_image[None], dtype=torch.float32, device=device)  # (1, H, W)
sino = operator.project(volume)                       # sparse-view sinogram
coords = make_coordinate_grid(cfg.image_size).unsqueeze(0).to(device)
recon = reconstruct_image(sino, coords, 1, model, chunk_size=1024)
recon = recon.reshape(cfg.image_size, cfg.image_size)
```

See the [demo notebook](https://github.com/swing-research/Glimpse/blob/main/notebooks/inference_demo.ipynb)
for an end-to-end example (data download, FBP baseline, PSNR/SSIM, figures).

## Citation

```bibtex
@article{khorashadizadeh2025glimpse,
  title   = {GLIMPSE: Generalized Locality for Scalable and Robust CT},
  author  = {Khorashadizadeh, AmirEhsan and Debarnot, Valentin and Liu, Tianlin and Dokmani{\'c}, Ivan},
  journal = {IEEE Transactions on Medical Imaging},
  year    = {2025},
  doi     = {10.1109/TMI.2025.3568017}
}
```