---
license: cc-by-nc-4.0
tags:
  - medical-imaging
  - diabetic-retinopathy
  - domain-generalization
  - pytorch
pipeline_tag: image-classification
library_name: pytorch
---

# GradeEye four-ch-morph

This repository contains GradeEye CORN ordinal diabetic-retinopathy classifiers using the `convnext_tiny` backbone at 384x384 resolution, with 4-channel input and a 4-threshold ordinal head. The `_ema.safetensors` file is the **primary** artifact: the paper's reported evaluation metrics were generated using the EMA state dict. The unsuffixed `.safetensors` file is the corresponding raw `model_state_dict` secondary artifact.

## Checkpoints

| Primary EMA weights | Raw secondary weights | Architecture | Held-out fold | Best QWK | Epoch |
|---|---|---|---|---:|---:|
| `lodo_aptos_convnext_tiny_best_ema.safetensors` | `lodo_aptos_convnext_tiny_best.safetensors` | `convnext_tiny` | varies | 0.7789 | 12 |
| `lodo_ddr_convnext_tiny_best_ema.safetensors` | `lodo_ddr_convnext_tiny_best.safetensors` | `convnext_tiny` | varies | 0.7702 | 7 |
| `lodo_eyepacs_convnext_tiny_best_ema.safetensors` | `lodo_eyepacs_convnext_tiny_best.safetensors` | `convnext_tiny` | varies | 0.7833 | 3 |
| `lodo_messidor2_convnext_tiny_best_ema.safetensors` | `lodo_messidor2_convnext_tiny_best.safetensors` | `convnext_tiny` | varies | 0.7395 | 2 |

## Preprocessing

1. Resize the RGB fundus image to 384x384 using the same offline preprocessing pipeline.
2. Convert RGB to float in [0,1] and apply ImageNet normalization: mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225).
3. Append one auxiliary channel in [0,1] after RGB normalization. It is not ImageNet-normalized.

### Fourth-channel construction

- Source pool: `segmentation_pooled_morph`.
- Producer: `gradeeye/seg-unet-bcedice`.
- Exact method: The BCE+Dice U-Net raw sigmoid probability map after 3x3 elliptical morphological opening, 3x3 elliptical closing, and Gaussian blur (sigma=0.5). It remains a continuous [0,1] soft-probability map, not a morphological gradient or edge channel.
- The resulting tensor is `(4, 384, 384)` in channel-first layout.
- This is the only one of the three published four-channel variants with morphological post-processing; it is not a gradient or edge channel.

## Loading

```python
import json
from modeling import load_model

config = json.load(open('config.json'))
model = load_model('lodo_eyepacs_convnext_tiny_best_ema.safetensors', config)
# model(x) returns CORN logits with shape (batch, 4)
```

Install `torch`, `timm`, and `safetensors`, and make the GradeEye source repository available on `PYTHONPATH`.

## Intended use and limitations

These weights are released for research and reproducibility only. They are not validated for clinical diagnosis or treatment decisions. Performance varies substantially by held-out dataset and should not be interpreted as clinical-grade generalization.

Source code and paper materials: [https://github.com/ahmed-farhanur-rashid/gradeeye](https://github.com/ahmed-farhanur-rashid/gradeeye).