2-Net: a two-network ensemble for multi-label fundus classification

Trained weights for the paper "A Two-Network Ensemble for Multi-Label Fundus Classification" by Théodore Cousin¹, Coralie Hemptinne², Sébastien Jodogne¹ (¹ ICTEAM, UCLouvain, Louvain-la-Neuve, Belgium; ² Cliniques Universitaires Saint-Luc, Brussels, Belgium).

Code: https://forge.uclouvain.be/ThodoreCousin/2-net-implem

The 2-Net detects retinal conditions from a colour fundus photograph as a multi-label problem. It combines two networks that fail differently: a general-purpose DINOv2 ViT-B/14 adapted with LoRA, and an EfficientNetV2-L that sees the image at 896 px to capture fine-grained detail. Each network is calibrated per class on validation data, and the two are fused by a geometric mean of their probabilities. The 2-Net performs on par with the 30-member winner of the RIADD challenge, with 4.5 times fewer parameters and a CPU inference time of 4.4 s instead of 77.5 s per image.

Files

One checkpoint per network and dataset, all trained with seed 42 (one of the five seeds of the paper).

File	Network	Input	Parameters
DINOv2-LoRA/{MuReD,RFMiD}/DINOv2-LoRA-<Dataset>.pt	DINOv2 ViT-B/14, LoRA rank 16, token-pooling head	518 × 518	87.8 M (1.2 M trained)
EfficientNetV2-L/{MuReD,RFMiD}/EfficientNetV2-L-<Dataset>.pt	EfficientNetV2-L, linear head	896 × 896	117.3 M
RETFound/{MuReD,RFMiD}/RETFound-<Dataset>.pt	RETFound ViT-L/16, fully fine-tuned, token-pooling head	384 × 384	≈ 304 M
calibration/2-Net-<Dataset>.json	2-Net calibration: per-class Platt parameters of both members and decision thresholds, fitted on the validation split

• 2-Net = DINOv2-LoRA + EfficientNetV2-L (the main system of the paper).
• 2-Net (RETFound) = RETFound + EfficientNetV2-L, the variant used to compare against retina-specific pretraining.
• The ViT checkpoints are PyTorch state_dicts. The EfficientNetV2-L checkpoints are dicts with the state_dict and its metadata (architecture, class names, seed, …). All of them are loaded with the code of the repository above.

Outputs

One probability per class, in this order:

• MuReD (20): DR, NORMAL, MH, ODC, TSLN, ARMD, DN, MYA, BRVO, ODP, CRVO, CNV, RS, ODE, LS, CSR, HTR, ASR, CRS, OTHER.
• RFMiD (29): Disease_Risk, DR, ARMD, MH, DN, MYA, BRVO, TSLN, ERM, LS, MS, CSR, ODC, CRVO, TV, AH, ODP, ODE, ST, AION, PT, RT, RS, CRS, EDN, RPEC, MHL, RP, OTHER (the 29 labels of the RIADD challenge; OTHER groups the remaining 18 RFMiD labels).

Training data

• MuReD (2,208 images, 20 labels). No official split: we use a class-stratified split of 1,504 / 260 / 444 images (train / validation / test), provided in the code repository.
• RFMiD (3,200 images). Official split of 1,920 / 640 / 640 images, 29 labels.

Images are cropped to the fundus disc and enhanced with CLAHE. Training uses oversampling of rare-class images, data augmentation (flips, small rotations, photometric changes, cutout, random crops) and MixUp / CutMix. DINOv2-LoRA and RETFound use a distribution-balanced loss, EfficientNetV2-L an asymmetric loss.

Results

Test set, macro F1 / macro mAP / Recall@n. Our models use per-class Platt calibration and decision thresholds tuned on the validation split.

These checkpoints (seed 42)

Model	F1 MuReD	F1 RFMiD	mAP MuReD	mAP RFMiD	R@n MuReD	R@n RFMiD
EfficientNetV2-L	.637	.535	.720	.551	.791	.854
RETFound, full fine-tuning	.621	.481	.683	.534	.764	.864
DINOv2, LoRA	.675	.516	.726	.581	.791	.872
2-Net (RETFound)	.671	.510	.745	.601	.800	.877
2-Net (DINOv2)	.706	.544	.747	.607	.804	.873

Paper (mean ± 1 SD over five seeds)

Model	F1 MuReD	F1 RFMiD	mAP MuReD	mAP RFMiD	R@n MuReD	R@n RFMiD
EfficientNetV2-L	.629 ± .009	.511 ± .031	.704 ± .012	.556 ± .010	.779 ± .008	.858 ± .006
RETFound, full fine-tuning	.618 ± .015	.489 ± .012	.677 ± .012	.546 ± .010	.751 ± .020	.866 ± .003
DINOv2, LoRA	.655 ± .018	.506 ± .030	.711 ± .009	.565 ± .012	.792 ± .011	.869 ± .003
2-Net (RETFound)	.658 ± .013	.511 ± .003	.730 ± .011	.602 ± .010	.799 ± .011	.876 ± .001
2-Net (DINOv2)	.673 ± .020	.527 ± .012	.735 ± .008	.589 ± .011	.806 ± .004	.875 ± .003

Comparisons with C-Tran, RetExpert, RET-CLIP and the RIADD winner are in the paper.

How to use

git clone https://forge.uclouvain.be/ThodoreCousin/2-net-implem && cd 2-net-implem
pip install -r requirements.txt
hf download TheodoreCousin/2-Net --local-dir weights

python infer.py --dataset MURED --images path/to/images/ --out predictions.csv

--dataset selects the label set and the networks trained on it (MURED or RFMiD). Each image goes through DINOv2-LoRA and EfficientNetV2-L; their probabilities are calibrated with calibration/2-Net-<Dataset>.json, fused by a geometric mean and thresholded per class. The CSV has one row per image with the probability of every class and the predicted labels. Preprocessing (fundus crop, CLAHE, resizing) is done by the script. The code repository also explains how to reproduce every result of the paper.

Intended use and limitations

• Research use only. These models are not medical devices and have not been validated in clinical practice. They must not be used for diagnosis.
• They were trained and evaluated on MuReD and RFMiD only; performance on other cameras, populations or acquisition protocols is unknown, and no cross-dataset evaluation was done.
• Rare diseases remain hard: macro F1 on the rarest classes stays around 0.3.
• Calibration and thresholds are fitted on each dataset's validation split and do not transfer as is to new data.
• The label sets mix pathological signs and diseases, and the models give probabilities without an explanation of the evidence behind them.

License

The weights are released under CC BY-SA 4.0: you may share and adapt them, including commercially, provided you give appropriate credit and distribute your contributions under the same license.

The RETFound checkpoints are fine-tuned from RETFound, itself released under CC BY-NC 4.0, so its non-commercial restriction still applies to them. DINOv2-LoRA starts from Meta's DINOv2 and EfficientNetV2-L from ImageNet weights distributed with timm.

Citation

TO BE COMPLETED