Results — inside field-of-view
From-scratch U-Net (default) on the 2-channel CLAHE+Frangi input; it beat the pretrained smp ResNet34 alternative at every patch size tried. Both clear the > 0.81 DRIVE benchmark.
0.834
F1 (Dice) · target > 0.81
0.845
Sensitivity
0.972
Specificity
0.980
AUC
Example
Method
RGB fundus image └─ CLAHE green channel # channel 0 — contrast enhancement └─ Frangi vesselness # channel 1 — Hessian tubularity prior └─ U-Net (pretrained ResNet34 encoder, or from-scratch) └─ 1×1 conv → sigmoid → vessel probability map
- Input: 2-channel
[CLAHE green, Frangi]stack, padded to 608×608. - Loss:
0.5·BCE + 0.5·Dice— the Dice term handles the ~10:1 background/vessel imbalance. - Training: 64×64 random patches (8000/epoch), Adam + cosine schedule on an RTX 5060; best-validation-F1 checkpoint saved.
- Evaluation: F1 / sensitivity / specificity / AUC computed strictly inside the FOV mask, as in the DRIVE literature.
Goodness scores (no labels needed)
- Vessel coverage % — share of the retina marked as vessel (healthy ≈ 8–13%).
- Confidence % — how decisive the model is across the retina; high means it rarely sits on the fence.
Run it
pip install -r requirements.txt
python train.py --epochs 150 --patch-size 48 --patches-per-epoch 8000
python evaluate.py --checkpoint checkpoints/unet_drive.pth
uvicorn app:app --port 8000 # then open http://localhost:8000/
Dataset
DRIVE — Digital Retinal Images for Vessel Extraction, via Kaggle (original: grand-challenge).