M. A. Yürük and A. Memiş, "Exploring the Impact of Alternative Color Spaces in Deep Retinal Age Prediction from Fundus Images," ATEEC 2026. Under review.

Code, split manifests, per-model predictions, and full reproduction instructions: github.com/mehmetaytugyuruk/retina-color-spaces-age-prediction

Companion studies on the same dataset: retina-resnet-age-prediction · retina-vit-age-prediction

Model description

22 ResNet-50 age regressors, ImageNet-pretrained and fully fine-tuned, that differ only in the color representation of the input or in the training seed. Everything else — the patient-disjoint splits, the architecture, the optimizer, the schedule, the loss, and the checkpoint-selection rule — is identical across every checkpoint. This is what makes the color representation the isolated variable in the study.

The final fully-connected layer is replaced by a regression head predicting a single continuous value: chronological age.

Single-channel models (e.g. hsv_h) receive their one channel replicated three times, so the input shape expected by the pretrained backbone is preserved without any architectural change. The Fusion Model is the one exception: its conv1 accepts 12 channels.

Files

Checkpoints are grouped by color-space family. Every file is <family>/<model>.pt.

File	Input representation	Seed
`rgb/rgb_seed42.pt`	RGB (all channels)	42
`rgb/rgb_seed43.pt`	RGB (all channels)	43
`rgb/rgb_seed44.pt`	RGB (all channels)	44
`rgb/rgb_seed45.pt`	RGB (all channels)	45
`rgb/rgb_seed46.pt`	RGB (all channels)	46
`rgb/rgb_r_seed42.pt`	R channel	42
`rgb/rgb_g_seed42.pt`	G channel	42
`rgb/rgb_b_seed42.pt`	B channel	42
`lab/lab_seed42.pt`	CIELAB (all channels)	42
`lab/lab_l_seed42.pt`	L (lightness)	42
`lab/lab_a_seed42.pt`	a (green–red)	42
`lab/lab_b_seed42.pt`	b (blue–yellow)	42
`hsv/hsv_seed42.pt`	HSV (all channels)	42
`hsv/hsv_h_seed42.pt`	H (hue)	42
`hsv/hsv_s_seed42.pt`	S (saturation)	42
`hsv/hsv_v_seed42.pt`	V (value)	42
`ycrcb/ycrcb_seed42.pt`	YCrCb (all channels)	42
`ycrcb/ycrcb_y_seed42.pt`	Y (luma)	42
`ycrcb/ycrcb_cr_seed42.pt`	Cr (red-difference)	42
`ycrcb/ycrcb_cb_seed42.pt`	Cb (blue-difference)	42
`grayscale/grayscale_seed42.pt`	Grayscale structural control	42
`fusion/rgb_lab_hsv_ycrcb_seed42.pt`	12-channel RGB+Lab+HSV+YCrCb (Fusion Model)	42

Checkpoint contents

Each .pt file is a dict. It is self-describing — the preprocessing constants needed for inference travel with the weights, so no config file is required:

Key	Contents
`model_state_dict`	ResNet-50 weights including the regression head
`normalization.statistics`	`channel_mean` / `channel_std` for this representation, computed from the train split only
`target_statistics`	`mean`, `sample_std`, `min_age`, `max_age`, `n` for age denormalization
`model_selection`	Selection metric and its value (`validation_mae`)
`epoch`	Epoch the selected checkpoint came from
`representation`	Representation name and contract fingerprint
`resolved_config`	The complete resolved training configuration
`runtime`	Device, seed, and exact library versions used for the run
`config_fingerprint`	Hash pinning the run to its configuration

The model is trained on normalized age targets. Recover years with:

predicted_age = raw_output * target_statistics["sample_std"] + target_statistics["mean"]

Intended use

Research and reproducibility for the associated paper — benchmarking, ablation studies, or extension work on retinal-fundus age prediction and color-representation ensembling. Not intended for clinical or diagnostic use; neither the dataset nor the models have been clinically validated.

Training details

Dataset: Retina Age Analysis Dataset (Kamran, 2025), 9,857 fundus images from 5,393 patients, patient-level split (6,902 train / 1,493 validation / 1,462 test). No patient appears in more than one split.
Preprocessing: threshold-based bounding-box crop of the retinal field, aspect-preserving resize, zero-pad to 224×224. Color conversion is applied after this, so all representations share identical image geometry.
Normalization: per-representation channel statistics computed from the train split only.
Loss: Smooth L1 (β = 1.0), reweighted per sample via Label Distribution Smoothing (Gaussian, σ = 2.0) to counter age-distribution imbalance.
Optimizer: AdamW, learning rate 1e-4, weight decay 1e-4, gradient-norm clipping at 1.0.
Scheduler: ReduceLROnPlateau (factor 0.5, patience 5, min LR 1e-7).
Augmentation: horizontal flip p = 0.5, rotation within ±10° (training only).
Epochs: 80, batch size 32. Best checkpoint selected by validation MAE.
Hardware: NVIDIA L4 GPU, PyTorch 2.11, CUDA 12.8, fp16 mixed precision.

Evaluation results

Test-set MAE in years (1,462 images). Full-precision values; the paper rounds to two decimals.

Combination strategies

Configuration	Members	Test MAE
Ensemble 1	RGB + Lab + HSV + YCrCb (seed 42)	4.5983
Ensemble 2	RGB seeds 43–46	4.7436
Fusion Model	12-channel early fusion	5.2039

Ensemble 1 beats the equal-cost Ensemble 2 control by 0.1453 years (3.06% relative). A paired patient-level cluster bootstrap (100,000 repetitions, 809 test patients resampled as clusters) gives a 95% CI of [0.0133, 0.2773] years, with 98.417% of repetitions positive.

Combining different color spaces helps; combining more RGB seeds helps less; merging the color spaces at the input instead of at the prediction (the Fusion Model) helps least. The four color spaces are invertible transformations of one another, which limits what a single jointly-trained model can extract from their concatenation.

Full representations (seed 42)

Representation	Validation MAE	Test MAE
RGB	5.4090	4.9067
YCrCb	5.4046	5.1084
Lab	5.5419	5.2448
HSV	5.4661	5.2852
Grayscale	6.1552	5.9976

No non-RGB full representation beats RGB on the test split.

Single-channel ablations (seed 42)

Representation	Validation MAE	Test MAE
YCrCb-Y	6.0825	5.6944
RGB-G	6.1317	5.8333
HSV-S	6.3637	5.9151
YCrCb-Cb	6.5453	5.9632
Lab-b	6.5460	6.0295
Lab-a	6.7317	6.0522
RGB-B	6.5042	6.0592
Lab-L	6.4377	6.0677
HSV-H	6.5357	6.1345
YCrCb-Cr	6.4902	6.1461
RGB-R	7.0720	6.7047
HSV-V	7.1904	6.7435

Every single-channel model is worse than full RGB. YCrCb-Y is the strongest single channel.

Individually seeded RGB models

Model	Test MAE
RGB seed 42	4.9067
RGB seed 43	5.0659
RGB seed 44	4.9930
RGB seed 45	4.9849
RGB seed 46	5.1115

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.
The primary claim concerns two fixed ensembles, not a population of arbitrary seeds. Only RGB was trained with four additional control seeds; seed-population robustness is not claimed for Lab, HSV, or YCrCb.
Ensembles were equally weighted; learned or weighted ensembling was not evaluated.
No explainability analysis was performed, so which retinal structures each color space attends to remains an open question.
The dataset's demographic composition and consent/de-identification details are governed by the original dataset authors and were not independently verified here.

Example usage

python

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

REPO = "mehmetaytugyuruk/retina-color-spaces-age-prediction"
ckpt = torch.load(
    hf_hub_download(REPO, "rgb/rgb_seed42.pt"),
    map_location="cpu",
    weights_only=False,
)

model = models.resnet50(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()

# Preprocessing constants travel with the checkpoint.
stats = ckpt["normalization"]["statistics"]
mean = torch.tensor(stats["channel_mean"]).view(3, 1, 1)
std = torch.tensor(stats["channel_std"]).view(3, 1, 1)
target = ckpt["target_statistics"]

# image: float32 tensor, shape (3, 224, 224), values in [0, 1],
# already cropped to the retinal field and zero-padded (see the GitHub repo).
# image = (image - mean) / std
# with torch.no_grad():
#     raw = model(image.unsqueeze(0)).item()
# predicted_age = raw * target["sample_std"] + target["mean"]

Loading a single-channel model (e.g. hsv/hsv_h_seed42.pt) is identical — replicate the selected channel three times before normalizing. The Fusion Model needs a 12-channel conv1; see build_resnet50_12ch_regressor() in the GitHub repository.

Full preprocessing, inference, and evaluation code: github.com/mehmetaytugyuruk/retina-color-spaces-age-prediction.

License

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

Citation

bibtex

@inproceedings{yuruk2026colorspaces,
  title     = {Exploring the Impact of Alternative Color Spaces in Deep Retinal
               Age Prediction from Fundus Images},
  author    = {Yürük, Mehmet Aytuğ and Memiş, Abbas},
  booktitle = {ATEEC 2026},
  year      = {2026},
  note      = {Under review}
}

mehmetaytugyuruk/retinal-color-transfer

Retinal Fundus Age Prediction — Color-Space Study