GradeEye seg-unet-bcedice

This repository contains a retinal lesion segmentation UNetVessel with an EfficientNet-B4 encoder, trained at 384x384 using BCE+Dice loss on ddr+idrid. The model outputs one logit channel; apply sigmoid to obtain a continuous probability map.

Checkpoint

File	Encoder	Image size	Loss	Best validation Dice	Best epoch
best.safetensors	efficientnet_b4	384	bce_dice	0.2519	22

Preprocessing and output

Use RGB input, scale to [0,1], apply ImageNet mean=(0.485, 0.456, 0.406) and std=(0.229, 0.224, 0.225), and resize to 384x384. The output is (batch, 1, 384, 384) logits; apply torch.sigmoid for probabilities.

The masks produced by this model feed: gradeeye/four-ch-soft, gradeeye/four-ch-morph.

Loading

import json
import torch
from modeling import load_model

config = json.load(open('config.json'))
model = load_model('best.safetensors', config)
with torch.no_grad():
    probability = torch.sigmoid(model(rgb_tensor))

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.

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