IVCM DINO

DINO for Tortuosity Grade Classification of Corneal Nerve Fibers.

1. Installation

a) Request access to the data from zenodo.
b) Copy the data to the data/ directory.
c) In data/ we provide the data_split.json file that contains indices for training/validation/testing splits without the duplicated files.
To use the subsets without the duplications, rename your data folders to CORN1500_noD and CORN-3_noD.
d) Install dependencies:
Experiments were run on Python 3.12.12

pip install -r requirements.txt

2. Training

For Linear Probing run:

python run_ssl_linear_probing.py

For Fine-tuning run:

python run_sl_dino.py

Citation

If you used this work, please cite:

@misc{ouan_self-supervised_2026,
	title = {Self-{Supervised} {ImageNet} {Representations} for {In} {Vivo} {Confocal} {Microscopy}: {Tortuosity} {Grading} without {Segmentation} {Maps}},
	url = {http://arxiv.org/abs/2603.15269},
	doi = {10.48550/arXiv.2603.15269},
	urldate = {2026-05-18},
	publisher = {arXiv},
	author = {Ouan, Kim and Moreau, Noémie and Bozek, Katarzyna},
	month = may,
	year = {2026},
	note = {arXiv:2603.15269 [cs]},
	keywords = {Computer Science - Computer Vision and Pattern Recognition},
}