---
license: apache-2.0
pipeline_tag: image-to-image
tags:
- image-restoration
- defocus-deblurring
- microscopy
- pathology
- neural-operator
- pytorch
---

# DGNO — Discontinuous Galerkin Neural Operator

Pretrained checkpoints **and** the preprocessed BBBC006 dataset for
**"Discontinuous Galerkin Neural Operator for Pathology Defocus Deblurring"** (ICML 2026).

- **Code & full usage:** https://github.com/Duane245/DGNO
- **Task:** microscopy / pathology defocus deblurring

DGNO comes in two operator variants — **DGNO-Face** (face-wise discontinuous Galerkin
operator) and **DGNO-Cell** (cell-wise) — both produced by a single training architecture
and loaded by a unified evaluation architecture for inference.

## Checkpoints

Six checkpoints — three datasets × two variants. Each `.pth` stores both the raw weights
(`params`) and the EMA weights (`params_ema`).

| Dataset | DGNO-Face | DGNO-Cell |
|---|---|---|
| BBBC006 w1 | `pretrained/DGNO_BBBC006_w1_Face.pth` | `pretrained/DGNO_BBBC006_w1_Cell.pth` |
| BBBC006 w2 | `pretrained/DGNO_BBBC006_w2_Face.pth` | `pretrained/DGNO_BBBC006_w2_Cell.pth` |
| 3DHistech  | `pretrained/DGNO_3DHistech_Face.pth`  | `pretrained/DGNO_3DHistech_Cell.pth`  |

Download all six into `pretrained/`:

```bash
pip install -U huggingface_hub
huggingface-cli download Duane245/DGNO --include "pretrained/*" --local-dir .
```

Or a single checkpoint:

```bash
huggingface-cli download Duane245/DGNO pretrained/DGNO_3DHistech_Face.pth --local-dir .
```

## BBBC006 dataset

`BBBC006.tar.gz` (~3.4 GB) is the preprocessed BBBC006 set used in the paper — fluorescence
microscopy, single-channel grayscale, with both the `w1` (Hoechst-stained nuclei) and `w2`
(Phalloidin-stained actin) channels in one archive.

```bash
huggingface-cli download Duane245/DGNO BBBC006.tar.gz --local-dir datasets
tar -xzf datasets/BBBC006.tar.gz -C datasets && rm datasets/BBBC006.tar.gz
```

Extraction yields `datasets/BBBC006/{train,test}/{GT,blur}/*.tif`, ready for training and
testing with no further preprocessing; the `w1` / `w2` split is selected at run time.

## Results

| Dataset | Variant | PSNR ↑ | SSIM ↑ | LPIPS ↓ |
|---|---|---|---|---|
| BBBC006 w1 | DGNO-Cell | 37.22 | 0.959 | 0.103 |
| BBBC006 w1 | DGNO-Face | 37.09 | 0.958 | 0.104 |
| BBBC006 w2 | DGNO-Face | 32.66 | 0.847 | 0.323 |
| BBBC006 w2 | DGNO-Cell | 32.54 | 0.848 | 0.322 |
| 3DHistech  | DGNO-Face | 34.02 | 0.890 | 0.095 |
| 3DHistech  | DGNO-Cell | 34.00 | 0.890 | 0.093 |

## License

Apache-2.0. BBBC006 imagery originates from the
[Broad Bioimage Benchmark Collection](https://bbbc.broadinstitute.org/BBBC006).
See the [GitHub repository](https://github.com/Duane245/DGNO) for the citation and details.