mehmetaytugyuruk/retina-resnet-age-estimation

Retinal Fundus Age Estimation — ResNet

PyTorch checkpoints for the paper:

M. A. Yürük and A. Memiş, "Age Prediction and Categorization from Retinal Fundus Images Using Residual Neural Networks," in 2026 5th International Informatics and Software Engineering Conference (IISEC), Ankara, Türkiye, 2026, pp. 628–633. DOI: 10.1109/IISEC69317.2026.11418414

Code and full reproduction instructions: github.com/mehmetaytugyuruk/retina-resnet-age-estimation

Companion study (Vision Transformers, same dataset): mehmetaytugyuruk/retina-vit-age-estimation

Model description

Five ResNet variants (ResNet-18/34/50/101/152), ImageNet-pretrained and fully fine-tuned, with the final fully-connected layer replaced by a custom regression head predicting a single continuous value: chronological age. The head architecture depends on the backbone's feature vector size (512 for ResNet-18/34, 2048 for ResNet-50/101/152). Age-category classification (Pediatric / Young Adult / Middle Age / Senior / Elderly) is derived from the regression output at evaluation time only — no separate classification head was trained.

Each architecture was trained twice: once on non-filtered (cropped and resized) retinal images, and once on the same images additionally processed with a Ben Graham vessel-enhancement filter.

Files

File	Backbone	Preprocessing
resnet18-nonfiltered.pth / resnet18-filtered.pth	ResNet-18	non-filtered / Graham-filtered
resnet34-nonfiltered.pth / resnet34-filtered.pth	ResNet-34	non-filtered / Graham-filtered
resnet50-nonfiltered.pth / resnet50-filtered.pth	ResNet-50	non-filtered / Graham-filtered
resnet101-nonfiltered.pth / resnet101-filtered.pth	ResNet-101	non-filtered / Graham-filtered
resnet152-nonfiltered.pth / resnet152-filtered.pth	ResNet-152	non-filtered / Graham-filtered

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 estimation. 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).
• Preprocessing: retina disk cropped, padded, and resized to 224×224; optionally followed by a Ben Graham filter.
• Loss: Smooth L1, reweighted per-sample via Label Distribution Smoothing (LDS) to counter age-distribution imbalance.
• Optimizer: AdamW, learning rate 1e-4, weight decay 1e-4.
• Scheduler: ReduceLROnPlateau (factor 0.5, patience 5 epochs).
• Epochs: 80, batch size 32.

Evaluation results

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

Non-filtered images

Model	MAE	Accuracy	F1
ResNet-18	5.17	0.8426	0.7157
ResNet-34	5.11	0.8382	0.7071
ResNet-50	5.23	0.8373	0.7060
ResNet-101	5.09	0.8431	0.7132
ResNet-152	5.27	0.8368	0.7008

Graham-filtered images

Model	MAE	Accuracy	F1
ResNet-18	5.24	0.8369	0.7053
ResNet-34	5.27	0.8451	0.7165
ResNet-50	5.21	0.8313	0.6912
ResNet-101	5.02	0.8422	0.7081
ResNet-152	5.17	0.8418	0.7056

Best overall MAE: ResNet-101, Graham-filtered (5.02 years). Best age-category F1: ResNet-34, Graham-filtered (0.7165).

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.
• 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-resnet-age-estimation", "resnet101-filtered.pth")
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)

model = models.resnet101(weights=None)
model.fc = nn.Sequential(
    nn.Linear(2048, 512), nn.BatchNorm1d(512), nn.ReLU(inplace=True), nn.Dropout(0.4),
    nn.Linear(512, 128), nn.BatchNorm1d(128), nn.ReLU(inplace=True), nn.Dropout(0.3),
    nn.Linear(128, 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{yuruk2026age,
  title     = {Age Prediction and Categorization from Retinal Fundus Images Using Residual Neural Networks},
  author    = {Yürük, Mehmet Aytuğ and Memiş, Abbas},
  booktitle = {2026 5th International Informatics and Software Engineering Conference (IISEC)},
  pages     = {628--633},
  year      = {2026},
  publisher = {IEEE},
  doi       = {10.1109/IISEC69317.2026.11418414}
}