SynFundus-1M Synthetic Fundus Dataset

• SynFundus-1M is a large-scale synthetic fundus dataset comprising 1,000,018 images with 15 annotation types covering 11 disease categories and 4 readability labels.
• The dataset is generated using a DDPM-based SynFundus-Generator trained on 1.3 million real images, enabling controlled synthesis of realistic retinal imagery.
• Experiments indicate that adding SynFundus-1M improves diagnostic model performance and enhances pretraining efficiency compared to conventional natural image datasets.

Searching arXiv for the SynFundus-1M paper and its generation framework. SynFundus-1M is a publicly available synthetic fundus image dataset introduced as a million-scale resource for intelligent medical imaging research, particularly in settings where large-scale public datasets with high-quality annotations are limited by privacy constraints and annotation cost. It contains 1,000,018 images and provides fifteen types of annotation comprising 11 disease types and 4 readability labels. The dataset was generated by a Denoising Diffusion Probabilistic Model called SynFundus-Generator, trained on over 1.3 million private authentic fundus images collected from health examinations, outpatient, and inpatient hospital scenes. The central claim of the work is that synthetic fundus images can approximate authentic retinal imagery closely enough to support both downstream fine-tuning and pretraining for retinal disease diagnosis, while reducing the release burden associated with patient data (Shang et al., 2023).

1. Dataset definition and scope

SynFundus-1M is described as a high-quality million-scale synthetic fundus images dataset with fifteen types of annotation (Shang et al., 2023). Its scale is explicitly given as 1,000,018 images, and it is presented as the largest publicly available fundus dataset in the source description, with the broadest annotation coverage among the listed comparators.

The dataset’s annotation space is divided into two categories. The first is disease annotation, covering 11 disease types. The second is image interpretability annotation, covering 4 readability labels. This dual structure is important because it supports not only disease classification but also filtering by clinical usability. A plausible implication is that the dataset is intended for both diagnostic modeling and image-quality-aware pipeline design, since readability labels are assigned to the whole image and to key retinal regions.

The disease labels are:

• DR: Diabetic Retinopathy, with an enum scale 0–4
• AMD: Age-related Macular Degeneration
• AON: Anomalies of the Optic Nerve
• CRP: Choroidal Retinal Pathology
• DM: Degenerative Myopia
• DME: Diabetic Macular Edema
• EM: Epimacular Membrane
• GC: Glaucoma
• HtR: Hypertensive Retinopathy
• PM: Pathological Myopia
• RVO: Retinal Vein Occlusion

The readability labels are:

• Readable: whether the whole fundus image is clinically interpretable
• RO: readability of the Optical Disc region
• RR: readability of the Retinal region excluding optic disc
• RM: readability of the Macular region

With the exception of DR, which uses a severity scale, the disease labels are Boolean; the readability labels are also Boolean (Shang et al., 2023).

2. Annotation schema and clinical semantics

The annotation design combines pathology presence with regional visibility. In practical terms, this means SynFundus-1M does not treat fundus images solely as class-conditioned retinal photographs; it encodes whether clinically significant regions are readable. That distinction matters because fundus image interpretation is sensitive to local quality degradation, and a global “usable/unusable” flag is often insufficient for downstream systems that focus on the optic disc, retinal background, or macula separately.

The disease schema is heterogeneous. DR is represented as Enum: 0–4; 0=none, 1=mild, 2=moderate, 3=severe, 4=proliferative, whereas the remaining disease categories use Boolean presence/absence labels (Shang et al., 2023). This asymmetry reflects the fact that diabetic retinopathy grading is commonly modeled as an ordinal task, while the other listed conditions are framed in the source as binary labels.

The readability schema is similarly stratified. Readable applies to the full image, while RO, RR, and RM refer to clinically meaningful subregions. This suggests the dataset can support tasks such as region-specific quality assessment, disease diagnosis under quality constraints, and curriculum-style data selection based on interpretability.

Category	Label set	Type
Disease	DR	Enum $[0,1,2,3,4]$
Disease	AMD, AON, CRP, DM, DME, EM, GC, HtR, PM, RVO	Bool
Readability	Readable, RO, RR, RM	Bool

3. Generation pipeline: SynFundus-Generator

SynFundus-1M was generated using SynFundus-Generator, a DDPM-based model trained from authentic fundus imagery (Shang et al., 2023). The paper states that the generator is based on the MedFusion DDPM architecture and is trained in two stages: a VAE stage and a diffusion stage.

In the VAE stage, the model minimizes the reconstruction objective

$$\mathcal{L}_{VAE} = \mathbb{E}_{q(f_0|x)} \left[ \| x - \tilde{x} \|^2 \right]$$

where $x$ is the input RGB image and $\tilde{x}$ is the reconstructed image. The VAE compresses each 512 × 512 RGB image into a 64 × 64 latent space (Shang et al., 2023).

In the diffusion stage, the latent encoding $f_0$ is corrupted with Gaussian noise over $T=1000$ steps,

$f_T = f_0 + \textit{Gaussian noise}$

and a U-Net-based diffusion model learns to reverse this process while conditioning on disease and readability labels. The stated training objective is

$\mathcal{L}_{\text{diffusion} = \mathbb{E}_{t, f_0, \epsilon} \left[ \| \epsilon - \epsilon_\theta(f_t, t, \text{condition}) \|^2 \right]$

where $\epsilon$ is random noise and $\epsilon_\theta$ is the predicted noise (Shang et al., 2023).

The released images are generated under predefined conditions, and the source description states that this allows control over disease/readability conditions while maintaining diversity. All generated images are then automatically labeled using the same AI-diagnose platform used in the broader data curation pipeline. The reported computational cost is approximately 120 hours on A100 GPUs for image generation and approximately 550 hours on V100 GPUs for automated annotation.

4. Training data provenance and privacy framing

The model was trained on over 1.3 million authentic fundus images drawn from health examinations, outpatient, and inpatient hospital scenes (Shang et al., 2023). The source description emphasizes broad disease and acquisition diversity in the training corpus. It also states that labels were bootstrapped and completed using a proprietary AI-diagnose platform, with the annotation models reporting >90% sensitivity and specificity.

Because the public release consists of synthetic rather than authentic images, the dataset is positioned as a privacy-conscious alternative to direct clinical image release. The source notes several privacy-related steps, including removal of personal information, OCR checks for text, use of synthesized images, and strict data handling for model training, and states that no personal or identifiable info is present in the released data (Shang et al., 2023).

This suggests that SynFundus-1M occupies an intermediate position between fully private institutional datasets and conventional public benchmarks: its statistical content is derived from large-scale clinical data, but the distributable artifact is synthetic. A plausible implication is that the dataset was designed not only to improve model performance but also to change the data-sharing constraints under which ophthalmic ML research is conducted.

5. Validation of realism and fidelity

The paper validates the synthetic images through both human assessment and distributional comparison (Shang et al., 2023). In the human study, four experienced annotators (5+ years) classified 250 synthetic and 250 real EyePACS images, balanced by disease. Their F1-scores were ~0.6—barely above random (0.5), which the source interprets as evidence that the synthetic images are difficult to distinguish from authentic fundus images.

The work also reports Fréchet Inception Distance (FID) comparisons:

• SynFundus-1M vs. EyePACS: 29.75
• MedFusion synthetic images vs. EyePACS: 63.05
• EyePACS vs. APTOS: 33.69

These values are used to argue that SynFundus-1M is quantitatively closer to EyePACS than a prior synthetic baseline and that its distance to a real dataset is comparable to the distance between two real datasets (Shang et al., 2023). The accompanying qualitative claim is that disease-related vision features (e.g. lesions) are well simulated in the synthetic images.

The realism claim should be read carefully. The source does not assert identity with authentic distributions in every respect; rather, it reports that experienced annotators can hardly distinguish the images and that FID is favorable relative to the stated comparators. This suggests high perceptual and distributional plausibility, not a formal proof of clinical equivalence across all uses.

6. Utility for downstream training and pretraining

The dataset is evaluated for two distinct training roles: as additional data for downstream tasks and as a pretraining corpus (Shang et al., 2023). The downstream disease diagnosis experiments use ResNet-50 and ViT-B/16 and focus on Diabetic Retinopathy Grading (IDRiD) and Glaucoma Diagnosis (REFUGE2).

For diabetic retinopathy grading with ResNet-50, the reported results are:

Training set	QWK	Accuracy / F1
IDRiD only	0.559	0.553 / 0.445
IDRiD+EyePACS	0.750	0.641 / 0.512
IDRiD+SynFundus-1M	0.744	0.660 / 0.517
IDRiD+EyePACS+SynFundus	0.850	0.806 / 0.776

For glaucoma diagnosis with ResNet-50, the reported results are:

Training set	Acc.	AUC / F1
REFUGE2	0.808	0.833 / 0.733
REFUGE2+AIROGS	0.895	0.835 / 0.805
REFUGE2+SynFundus-1M	0.880	0.860 / 0.801
REFUGE2+AIROGS+SynFundus-1M	0.928	0.876 / 0.879

These results support two claims stated in the source. First, adding SynFundus-1M as extra training data yields similar improvements to adding authentic external datasets such as EyePACS and AIROGS. Second, combining SynFundus-1M with authentic data yields the best performance in the reported experiments (Shang et al., 2023).

The pretraining experiments compare models trained from scratch, pretrained on ImageNet, and pretrained on SynFundus-1M before fine-tuning. The summarized results reported in the source are:

• Scratch: QWK (IDRiD) 0.006, Acc. (REFUGE2) 0.800, F1-Score (IDRiD) 0.110
• ImageNet: -0.006 (ViT), 0.825-0.800, 0.103
• SynFundus-1M: 0.741-0.661, 0.880-0.845, 0.510-0.524

The associated conclusion is that models pretrained on SynFundus-1M converge faster and achieve better performance than those pretrained on ImageNet or trained from scratch (Shang et al., 2023). This is consistent with a domain-specific pretraining hypothesis: synthetic retinal imagery may encode features more aligned with ophthalmic tasks than natural-image pretraining corpora.

7. Position within public ophthalmic datasets and research significance

The source compares SynFundus-1M with several public ophthalmic datasets: EyePACS (88,702), AIROGS (101,442), APTOS (5,590), PALM (1,200), and REFUGE2 (2,000) (Shang et al., 2023). In that comparison, SynFundus-1M is distinguished by the combination of size, annotation breadth, and readability labeling.

Dataset	Images	Coverage
EyePACS	88,702	DR only
AIROGS	101,442	Glaucoma only
APTOS	5,590	DR only
PALM	1,200	Pathological myopia
REFUGE2	2,000	Glaucoma only
SynFundus-1M	1,000,018	11 diseases, 4 readability labels

In this framing, SynFundus-1M is not merely a synthetic augmentation corpus; it is intended as a foundational resource for open ophthalmic ML research. The dataset is stated to be publicly available for the research community, with a GitHub release at https://github.com/parap1uie-s/SynFundus-1M (Shang et al., 2023).

Several broader implications follow from the reported evidence, though they should be treated as interpretations rather than direct claims. First, the dataset supports the view that high-quality synthetic medical images can function as effective training data when generated from sufficiently large and diverse authentic corpora. Second, the inclusion of both disease and readability annotations suggests a move toward datasets that reflect diagnostic and acquisition realities simultaneously. Third, the reported gains from SynFundus-1M pretraining suggest that synthetic domain-specific corpora may serve as practical substitutes for natural-image pretraining in specialized medical imaging pipelines.

At the same time, a common misconception would be to treat synthetic scale alone as sufficient. The source’s contribution is specifically tied to high-quality annotations, a DDPM-based conditional generator, and validation against both human raters and downstream tasks (Shang et al., 2023). The significance of SynFundus-1M therefore lies in the conjunction of scale, conditional controllability, annotation richness, and empirical utility, rather than in synthetic image volume alone.