z-retina

Multi-label fundus (retinal) image classification: MetaFormer CAFormer-B36 backbone + ML-Decoder head, trained with asymmetric loss and patient-level splits (see the code repository for full training and Table S1 evaluation).

Source code: github.com/ATMZR/z-retina

Files in this repo

File	Description
best.pth	PyTorch checkpoint (best validation metric during training).
config.json	num_labels, class_names, id2label, label2id — logit index i matches class_names[i].

Label semantics

There are 15 pathology logits. Normal fundus is represented in the training codebase as an all-zero target vector (no positive class). The strings in config.json must stay in this order for metrics and thresholding to match the reference implementation.

How to load weights (with the z-retina codebase)

Clone and install from GitHub, then load the checkpoint using the project’s utilities (see src/z_retina/model.py — load_caformer_checkpoint expects either a raw state_dict or a dict containing a "model" key).

Example evaluation entrypoint from the README:

poetry run z-retina-eval --config configs/default.yaml --checkpoint path/to/best.pth --table_s1

Point --checkpoint at a local copy of best.pth downloaded from this Hub repo.

Model I/O

• Input: RGB, 224×224, ImageNet normalization (see configs/default.yaml and EyeDataset in the code repo).
• Output: 15 logits (multi-label); apply sigmoid per logit for probabilities.

Citation

If you use this model, please cite or link the z-retina repository and respect third-party licences (e.g. vendored ML-Decoder under third_party/ml_decoder in that repo).