AEyeDB: Curated Retinal Fundus Dataset

• AEyeDB is a curated, high-resolution fundus image dataset that supports both supervised ViT-based classification and GANomaly-based anomaly detection.
• It employs rigorous preprocessing and systematic augmentation methods to achieve perfect classification accuracy and reliable anomaly detection metrics.
• The dataset facilitates clinical decision support with explainable error maps and calibrated risk assessments for retinal disease screening.

AEyeDB is a curated, high-resolution fundus image dataset developed for advancing automated retinal disease detection and explainable anomaly detection using deep learning. Designed to facilitate robust benchmarking, AEyeDB offers a consistently labeled collection focused on healthy and diabetic retinopathy (DR) cases, with rigorous curation, preprocessing, and augmentation pipelines. It is structured to support both supervised classification and unsupervised anomaly detection research, with documented use in Vision Transformer (ViT) classification and GANomaly-based anomaly detection (Ruhland et al., 23 Nov 2025).

1. Cohort Composition and Data Acquisition

AEyeDB comprises data from 103 participants (age ≥ 18 years, mean 26.93 ± 7.18, 21 female, 82 male), all of whom provided informed consent. A total of 256 fundus photographs were obtained via a Rodenstock FundusScope at high native resolution (several thousand pixels per side). Following exclusion of duplicate and suboptimal images, the final dataset consists of 204 fundus photographs. The images underwent a standardized preprocessing workflow yielding 224×224 pixel RGB inputs suitable for integration with ViT-based architectures (Ruhland et al., 23 Nov 2025).

Disease labeling was performed using participant questionnaires, which collected information on retinal risk factors such as diabetes and hypertension. The resulting annotated categories for AEyeDB are:

• Normal retina
• Diabetic retinopathy (determined by self-reported diabetes history)

No images of age-related macular degeneration (AMD), glaucoma, or other rare pathologies are included. There are no pixel-wise lesion masks or expert grading annotations; labels are based on clinical history and manual quality control.

2. Preprocessing and Augmentation Methodologies

Quality assurance involved exclusion of images marked as suboptimal. Preprocessing steps included intensity normalization (scaling pixel values to [0,1]) and center-cropping to produce a circular field of view with subsequent resizing to 224×224 pixels.

A comprehensive set of data augmentation methods was systematically assessed:

• Geometric augmentations: Random horizontal flips ($p=0.5$), uniform random rotations ($\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$), and translations ($\Delta_{x,y}\sim[-0.1W, +0.1W]$ where $W$ is image width or height).
• Color and blur augmentations: Brightness jitter ($I' = I\cdot\alpha$, $\alpha\sim\mathrm{Uniform}[0.8,1.2]$), contrast jitter ($I' = I\cdot\beta + (1-\beta)\mu$, $\beta\sim\mathrm{Uniform}[0.8,1.2]$, $\mu$ mean intensity), and Gaussian blur ($\sigma\sim\mathrm{Uniform}[0.1,2.0]$).
• Contrast and contour enhancement: Histogram equalization (applied per channel/patch) and Laplacian contour sharpening ($\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$0 with $\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$1).

Geometric and color augmentations were observed to yield the most stable improvements across broader datasets, while histogram equalization specifically benefited tasks involving structural subtleties. Laplacian enhancement consistently reduced downstream performance (Ruhland et al., 23 Nov 2025).

3. Supervised Classification Performance

A ViT-B/16 classifier, fine-tuned from ImageNet weights, was evaluated using a stratified 70/15/15% split (Train/Val/Test) on AEyeDB. The training regime employed the Adam optimizer ($\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$2, cosine learning rate decay, 5-epoch warmup, 30 epochs total).

On the AEyeDB test set, the ViT-B/16 achieved:

• Accuracy: 1.000 (across no augmentation, geometric, and color augmentation setups)
• Weighted F1-score: 0.94–0.97
• Matthews correlation coefficient (MCC): 0.95–0.97

No statistically significant performance change was observed with augmentations, as baseline classification was already perfect (paired $\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$3-test, $\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$4), highlighting high intra-dataset consistency (Ruhland et al., 23 Nov 2025).

4. GANomaly-Based Anomaly Detection

AEyeDB serves as a benchmark for explainable anomaly detection via a GANomaly architecture. The pipeline is composed of:

• Encoder ($\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$5): Maps input $\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$6 to latent $\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$7
• Decoder ($\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$8): Reconstructs $\theta\sim\mathrm{Uniform}[0^\circ,360^\circ]$9 from $\Delta_{x,y}\sim[-0.1W, +0.1W]$0
• Re-encoder ($\Delta_{x,y}\sim[-0.1W, +0.1W]$1): Maps reconstructed $\Delta_{x,y}\sim[-0.1W, +0.1W]$2 back to $\Delta_{x,y}\sim[-0.1W, +0.1W]$3
• Adversarial discriminator: Operates on features from intermediate generator layers

The loss formulation includes:

1. Reconstruction loss: $\Delta_{x,y}\sim[-0.1W, +0.1W]$4
2. Adversarial feature loss: $\Delta_{x,y}\sim[-0.1W, +0.1W]$5
3. Latent consistency: $\Delta_{x,y}\sim[-0.1W, +0.1W]$6
4. (Optional) KL regularization: $\Delta_{x,y}\sim[-0.1W, +0.1W]$7

The weighted total loss employs $\Delta_{x,y}\sim[-0.1W, +0.1W]$8. Training was conducted from scratch for 100 epochs using only healthy AEyeDB images (≈170), with $\Delta_{x,y}\sim[-0.1W, +0.1W]$9 and no observed benefit from attention-mask weighting (Ruhland et al., 23 Nov 2025).

On AEyeDB, the raw anomaly score ($W$0) yielded an uncalibrated AUC of 0.76. Explainability was achieved via 90th-percentile absolute reconstruction error maps, which reliably emphasized retinal vessels and cotton-wool-spot-like lesions in abnormal images.

5. Calibration and Clinical Decision Support

For probabilistic calibration, the GUESS framework was used to estimate pathology probability via Bayes’ rule:

$W$1

where $W$2 is the anomaly score, $W$3 are kernel density estimates in abnormal/healthy cohorts, and $W$4 are prior prevalences. This calibration renders the output $W$5, supporting threshold-independent decision-making; for example, clinicians may select referral cutoffs based on risk. The ordering quality, as measured by AUC (0.76), facilitates robust ranking of suspicious fundus cases. This approach avoids issues associated with single-threshold tuning for clinical risk stratification (Ruhland et al., 23 Nov 2025).

6. Significance and Benchmark Contributions

AEyeDB’s high-resolution, quality-controlled fundus images with verified healthy/DR annotations establish a reference for both conventional classification and interpretable anomaly detection. The dataset yielded perfect ViT classifier results (Acc = 1.000, F1 ≈ 0.95) in supervised setups and demonstrated the utility of systematic augmentation ablations. The GANomaly anomaly detection pipeline—paired with GUESS calibration and error mapping for interpretability—underscores AEyeDB’s value for evaluating both algorithmic accuracy and potential clinical deployment strategies. A plausible implication is that AEyeDB’s strict curation and explicit methodological documentation make it suitable for future research on cross-domain generalization, data efficiency, and explainable AI in ophthalmic imaging (Ruhland et al., 23 Nov 2025).