mehmetaytugyuruk/retina-vit-age-prediction

Retinal Fundus Age Prediction — Vision Transformer

PyTorch checkpoints for the paper:

M. A. Yürük and A. Memiş, "Decoding Chronological Age from the Retinal Fundus Images: A Deep Learning-based Analysis with Vision Transformers," in 2026 34th Signal Processing and Communications Applications Conference (SIU), İstanbul, Türkiye, Jul. 2026, pp. 1–4. doi: 10.1109/SIU71813.2026.11636734

Code and full reproduction instructions: github.com/mehmetaytugyuruk/retina-vit-age-prediction

Companion study (ResNet backbones, same dataset): mehmetaytugyuruk/retina-resnet-age-prediction

Model description

Four Vision Transformer variants (ViT-B/16, ViT-B/32, ViT-L/16, ViT-L/32), ImageNet-1K-pretrained and fully fine-tuned, with the classification head replaced by a single-neuron regression head predicting chronological age. Age-category classification (Pediatric / Young Adult / Middle Age / Senior / Elderly) is derived from the regression output at evaluation time only. All four variants were trained on Graham-filtered retinal images only (the only preprocessing variant used in this paper).

Files

File	Backbone	Patch size	Layers	Hidden size	Heads
vit-b16-filtered.pth	ViT-B/16	16×16	12	768	12
vit-b32-filtered.pth	ViT-B/32	32×32	12	768	12
vit-l16-filtered.pth	ViT-L/16	16×16	24	1024	16
vit-l32-filtered.pth	ViT-L/32	32×32	24	1024	16

Each checkpoint is a dict with keys: model_state_dict, mean_age, std_age, epoch, val_loss, val_mae. Predicted age is recovered as pred = model(image) * std_age + mean_age.

Intended use

Research and reproducibility for the associated paper — benchmarking, ablation studies, or extension work on retinal-fundus age prediction. Not intended for clinical or diagnostic use; the underlying dataset and models have not been clinically validated.

Training details

• Dataset: Retina Age Analysis Dataset (Kamran, 2025), 9,857 fundus images, patient-level split (6,902 train / 1,493 validation / 1,462 test) — same split as the companion ResNet study.
• Preprocessing: retina disk cropped, padded, and resized to 224×224, followed by a Ben Graham filter.
• Loss: Smooth L1, reweighted per-sample via Label Distribution Smoothing (LDS).
• Optimizer: AdamW, weight decay 0.05. Learning rate 3e-5 for ViT-B/16, ViT-B/32, and ViT-L/32; 2e-5 for ViT-L/16 (reduced for training stability given its larger size).
• Warmup: 5 epochs training only the regression head before unfreezing the full backbone.
• Scheduler: ReduceLROnPlateau (factor 0.5, patience 5 epochs).
• Epochs: 80.

Evaluation results

Values below are exactly as reported in the paper's Table V (test set, 1,462 images, Graham-filtered). MAE in years; Accuracy/F1 are for the derived 5-class age categorization.

Model	MAE	Accuracy	F1
ViT-B/16	4.99	0.8368	0.6974
ViT-B/32	5.52	0.8241	0.6764
ViT-L/16	4.86	0.8497	0.7219
ViT-L/32	5.59	0.8211	0.6684

ViT-L/16 is the best-performing model overall. The paper additionally reports a point-estimate comparison against the companion ResNet study's results (ResNet-101, MAE 5.01 as cited in this paper's own Table VI).

Limitations

• Trained and evaluated on a single public dataset; generalization to other populations, imaging devices, or acquisition protocols is untested.
• Not clinically validated — do not use for diagnosis, screening, or any medical decision-making.
• Age-category boundaries are derived post-hoc from a regression output, not directly optimized as a classification objective.
• ViT-L/16's lower learning rate (2e-5) versus the other three variants (3e-5) is a real methodological detail not explicitly called out in the paper's own methods section; documented here and in the GitHub repo for transparency.
• The dataset's demographic composition and any consent/de-identification details are governed by the original dataset authors, not verified independently here.

Example usage

import torch
import torch.nn as nn
from torchvision import models
from huggingface_hub import hf_hub_download

ckpt_path = hf_hub_download("mehmetaytugyuruk/retina-vit-age-prediction", "vit-l16-filtered.pth")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)

model = models.vit_l_16(weights=None)
model.heads.head = nn.Linear(model.heads.head.in_features, 1)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()

# pred_age = model(image_tensor).item() * ckpt["std_age"] + ckpt["mean_age"]

Full preprocessing and inference code: see the GitHub repository.

License

Checkpoints released under the MIT License. The training dataset is separately MIT-licensed by its original authors.

Citation

@inproceedings{yuruk2026decoding,
  title     = {Decoding Chronological Age from the Retinal Fundus Images: A Deep Learning-based Analysis with Vision Transformers},
  author    = {Yürük, Mehmet Aytuğ and Memiş, Abbas},
  booktitle = {2026 34th Signal Processing and Communications Applications Conference (SIU)},
  year      = {2026},
  pages     = {1--4},
  address   = {İstanbul, Türkiye},
  publisher = {IEEE},
  doi       = {10.1109/SIU71813.2026.11636734}
}