APTOS 2019: Diabetic Retinopathy Dataset

• APTOS 2019 is a public fundus image dataset for five-level diabetic retinopathy grading, using the standard ICDR scale for clinical severity.
• It underpins tasks like binary DR detection, referral triage, and self-supervised representation learning while highlighting challenges such as class imbalance and image heterogeneity.
• Researchers employ diverse preprocessing pipelines and stratified splits, with evaluation metrics like quadratic weighted kappa to capture adjacent-grade confusions.

Searching arXiv for papers mentioning APTOS 2019 to ground the article in published work. APTOS 2019, usually referred to in the literature as the APTOS 2019 Blindness Detection dataset, is a public color fundus image dataset used for five-level diabetic retinopathy (DR) grading from retinal photographs. Across recent arXiv studies, it functions both as a supervised benchmark for ordinal disease staging and as a substrate for derived tasks such as binary DR detection, referable-DR triage, cross-dataset transfer, and self-supervised representation learning (Doshi et al., 13 May 2026, Hashmi, 19 Apr 2026, Florindo, 26 May 2026). The label space is consistently the standard five-grade DR severity scale—Class 0: No DR, Class 1: Mild, Class 2: Moderate, Class 3: Severe, Class 4: Proliferative DR—although secondary descriptions are not fully uniform about dataset size: several papers report 3,662 fundus images, whereas one paper reports 3,622 retinal fundus photographs, which suggests a reporting inconsistency across papers rather than a change in the grading task (Doshi et al., 13 May 2026, Shakibania et al., 2023, Kumar et al., 18 Nov 2025).

1. Identity and clinical annotation scheme

The dataset is described as a public fundus image dataset from the Kaggle APTOS 2019 Blindness Detection competition and is used for diabetic retinopathy grading from color fundus photographs (Doshi et al., 13 May 2026, Zhang et al., 9 May 2025). In task formulations that retain the original label space, each image is assigned one of five ordered DR grades. Multiple papers explicitly align these labels with the International Clinical Diabetic Retinopathy Disease Severity Scale or the standard five-level ICDR/ICDRSS staging scheme (Kumar et al., 18 Nov 2025, Hashmi, 19 Apr 2026).

The ordered label semantics are central to how APTOS 2019 is used. In ordinal-regression papers, the dataset is treated not as an arbitrary 5-way classification corpus but as a clinically ordered severity continuum, where confusing adjacent grades is less severe than collapsing healthy eyes into proliferative disease or vice versa (Hashmi, 19 Apr 2026, Kumar et al., 18 Nov 2025). This motivates cumulative-threshold formulations such as CORAL, scalar-regression formulations with clamping and rounding, and evaluation with quadratic weighted kappa rather than accuracy alone (Hashmi, 19 Apr 2026, Kumar et al., 18 Nov 2025).

APTOS 2019 is also routinely re-labeled for deployment-oriented tasks. One paper collapses the five classes into binary DR detection, with No DR versus DR, where the positive DR class is formed by merging labels 1 through 4 (Shakibania et al., 2023). Another paper defines a binary referral task for edge triage, where Classes 0 and 1 are mapped to Non-referable DR and Classes 2, 3, and 4 to Referable DR, and then reports deployment results in a derived 4-class output space consisting of Class 0–1, Class 2, Class 3, and Class 4 (Doshi et al., 13 May 2026). These regroupings preserve the underlying APTOS grading ontology while adapting it to screening workflows.

2. Size, image heterogeneity, and class imbalance

Most papers using the Kaggle release describe APTOS 2019 as containing 3,662 retinal fundus images (Doshi et al., 13 May 2026, Florindo, 26 May 2026, Shakibania et al., 2023). One 2025 paper instead describes 3,622 retinal fundus photographs (Kumar et al., 18 Nov 2025). This suggests a discrepancy in secondary reporting. The papers are otherwise consistent that APTOS 2019 is a five-class DR grading dataset of retinal fundus imagery.

The dataset is repeatedly characterized as heterogeneous. One detailed description states that images are large and heterogeneous in size, ranging from $474 \times 358$ up to $4288 \times 2848$, and that they often contain dark borders around the circular retinal field of view (Doshi et al., 13 May 2026). Other papers motivate cropping, masking, or normalization by citing variable field of view, off-center framing, lighting variation, and background artefacts in public fundus datasets of the APTOS type (Hashmi, 19 Apr 2026, Kumar et al., 18 Nov 2025).

A frequently reported class histogram is the following (Doshi et al., 13 May 2026, Shakibania et al., 2023):

Class	Stage	Count
0	No DR	1805
1	Mild DR	370
2	Moderate DR	999
3	Severe DR	193
4	Proliferative DR	295

These counts make the imbalance structure explicit: the dataset is skewed toward No DR and Moderate DR, while Severe DR is markedly underrepresented (Doshi et al., 13 May 2026). One paper further notes that 59.4% of APTOS images are non-referable, which directly constrains how much cloud usage can be reduced in a referral cascade (Doshi et al., 13 May 2026). Other studies do not always print the histogram, but they still identify imbalance as an experimental concern and address it through weighted random sampling, complement cross entropy, augmentation, or selective enrichment from external datasets (Hashmi, 19 Apr 2026, Shakibania et al., 2023).

This imbalance is not merely a descriptive property. It shapes error patterns reported on APTOS 2019: performance is typically strongest on No DR, intermediate on Mild and Moderate, and weakest on Severe and Proliferative DR, with adjacent-stage confusions dominating classwise failure modes (Florindo et al., 6 May 2026, Shakibania et al., 2023).

3. Split construction and benchmark protocol variability

APTOS 2019 does not appear in these papers as a single standardized evaluation protocol. Instead, authors commonly construct internal stratified splits from the publicly labeled portion of the dataset. One study states that the full public release contains 3662 labeled training images and 1928 test images with private labels; because the official Kaggle test labels were unavailable, the study excluded the competition test set and split only the 3662 labeled training images into 70%/10%/20% train/validation/test subsets (Shakibania et al., 2023). In that protocol, the per-class split counts are explicitly reported, and the test subset contains 20% of each category from the original APTOS distribution (Shakibania et al., 2023).

Another paper is equally explicit that its evaluation is based on an internal split created by the authors: 20% test, then from the remaining 80%, 80% train / 20% validation, yielding final proportions of 64% train, 16% validation, and 20% test, with a fixed random seed of 42 and stratification by ICDR class (Doshi et al., 13 May 2026). The resulting held-out test split contains 733 images and is described as a held-out APTOS test split, not as an official blind benchmark (Doshi et al., 13 May 2026).

A third paper uses APTOS 2019 for training and internal validation via stratified sampling while preserving label balance, but does not report the number of images used, the train/validation ratio, or per-class counts (Hashmi, 19 Apr 2026). Two recent self-supervised learning papers are even sparser: they define APTOS as a 5-class downstream benchmark but do not state the exact split, whether an official split was used, or how many images were assigned to each subset (Florindo et al., 6 May 2026, Florindo, 26 May 2026).

The practical consequence is that APTOS 2019 functions less as a fixed leaderboard-style benchmark and more as a common data source from which multiple internal protocols are derived. This makes direct metric comparison nontrivial even when the label space nominally remains five-class DR grading.

4. Preprocessing conventions and retinal content standardization

The literature does not present a single canonical preprocessing recipe for APTOS 2019. Instead, APTOS-specific pipelines vary according to model design and deployment assumptions, although several recurring objectives are visible: removal of dark non-retinal background, standardization of retinal field geometry, and enhancement of lesion visibility (Hashmi, 19 Apr 2026, Kumar et al., 18 Nov 2025, Doshi et al., 13 May 2026).

In a cross-dataset robustness study, APTOS 2019 is the reference domain for a dual-stream preprocessing pipeline. Images first undergo circular cropping to isolate the retinal region, remove the dark non-retinal background and edge artefacts, and impose a more consistent spatial layout. Two complementary enhanced views are then produced: one branch uses Ben Graham normalization, described as emphasizing vessels and improving overall contrast, and the other uses CLAHE, which enhances local contrast and is intended to make subtle lesions easier to detect. These views are resized to $224 \times 224$ and $300 \times 300$ for dual EfficientNet backbones (Hashmi, 19 Apr 2026). The same paper is precise that histogram matching is applied to Messidor-2 images, not to APTOS images; an APTOS training image provides the reference distribution for matching the external domain (Hashmi, 19 Apr 2026).

A different APTOS-oriented ordinal-regression paper emphasizes background masking, green channel isolation, median filtering for denoising, CLAHE, three-channel replication, and resizing to $224 \times 224$ (Kumar et al., 18 Nov 2025). The green channel is described as offering the strongest contrast for retinal vasculature and lesion structures, thereby making microaneurysms, hemorrhages, and related DR signs easier to discern (Kumar et al., 18 Nov 2025).

A deployment-oriented cascade paper uses deterministic cropping and resizing rather than lesion-specific enhancement. For its edge tier, the red channel is thresholded to define foreground pixels, the retinal region is autocropped, the image is resized with aspect ratio preserved, zero-padded to $512 \times 512$, and normalized with ImageNet mean and standard deviation. The cloud tier uses a separate bounding-box crop over non-background retinal pixels, resizing to $392 \times 392$ and ImageNet normalization (Doshi et al., 13 May 2026). Another dual-branch study adopts a simpler APTOS pipeline: all images are resized to $224 \times 224 \times 3$ and augmented with blur, vertical flip, horizontal flip, random rotation, sharpening, CLAHE, embossing, fancy PCA, and random brightness and contrast (Shakibania et al., 2023).

Several papers omit these details altogether for APTOS 2019, including image resolution, normalization constants, and dataset-specific augmentation policy (Florindo et al., 6 May 2026, Florindo, 26 May 2026). This omission is itself a salient feature of the benchmark’s literature, because preprocessing choices appear to be a substantial determinant of reported performance.

5. Roles of APTOS 2019 in contemporary model design

APTOS 2019 is used in recent work not merely as a generic image dataset but as an organizing domain around which distinct learning paradigms are built. In the dual-resolution ordinal-regression framework, it is the primary in-domain dataset, the source of supervised labels for model development, the basis for internal validation and checkpoint selection, and the reference domain for image normalization when aligning Messidor-2 (Hashmi, 19 Apr 2026). The model combines EfficientNet-B0 at $224 \times 224$ and EfficientNet-B3 at $300 \times 300$, fuses their features with a learnable channel-wise gate, and predicts the five ordered grades with CORAL ordinal regression (Hashmi, 19 Apr 2026).

APTOS 2019 also serves as a downstream target for self-supervised learning. In one paper, a ConvNeXt-Tiny encoder is pretrained through a Chaotic Denoising Autoencoder, where the input image normalized to $4288 \times 2848$0 is corrupted by the pixelwise logistic-map transformation $4288 \times 2848$1 with $4288 \times 2848$2, reconstructed with MSE, then fine-tuned for 5-class APTOS classification and fused with an ImageNet-pretrained ConvNeXt-Large through attentive feature fusion (Florindo et al., 6 May 2026). A related paper uses Chaos-SSL, a SimCLR-style contrastive pretraining scheme with Logistic, Tent, and Sine chaotic maps applied to augmented views of APTOS images before downstream fine-tuning and fusion with an ImageNet branch (Florindo, 26 May 2026).

Other papers exploit APTOS 2019’s ordered grading semantics without cumulative ordinal heads. One study replaces the final layer of a pretrained ResNet50 with Dropout and a single Dense output neuron, trains with mean squared error to predict a continuous severity score, and obtains the discrete grade by clipping to $4288 \times 2848$3 and rounding (Kumar et al., 18 Nov 2025). Another uses APTOS 2019 as the principal evaluation benchmark for a dual-branch ResNet50–EfficientNetB0 transfer-learning model trained with complement cross entropy, sometimes after selective class enrichment from Messidor-2 and IDRiD (Shakibania et al., 2023).

APTOS 2019 also supports system-level work. In an edge-cloud screening architecture, the dataset is the sole experimental dataset used to define a two-stage cascade: MobileNetV3-small performs local binary triage between referable and non-referable DR, while RETFound-DINOv2 ViT-L/14 performs cloud-side ordinal grading only on the forwarded subset (Doshi et al., 13 May 2026). In this setting, the dataset’s class imbalance and referable/non-referable composition directly determine the possible reduction in cloud calls (Doshi et al., 13 May 2026).

6. Evaluation norms and reported benchmark behavior

The dominant evaluation metric for APTOS 2019 severity grading is quadratic weighted kappa (QWK), chosen because it penalizes distant-stage disagreements more heavily than adjacent-stage errors and therefore better matches ordinal DR grading than flat accuracy (Hashmi, 19 Apr 2026, Kumar et al., 18 Nov 2025, Shakibania et al., 2023). Accuracy, macro-F1, AUC, sensitivity, and specificity are also reported in some studies, especially when APTOS is regrouped into binary or 4-class outputs (Florindo et al., 6 May 2026, Doshi et al., 13 May 2026).

Reported APTOS-centered performance spans several protocol-dependent regimes. A dual-resolution CORAL model reports validation QWK $4288 \times 2848$4 on an APTOS validation split and QWK $4288 \times 2848$5 on an unseen Messidor-2 test set, using APTOS as the main training and validation domain (Hashmi, 19 Apr 2026). The attentive CDAE + Attention model reports accuracy $4288 \times 2848$6 and macro-F1 $4288 \times 2848$7 on APTOS 2019 (Florindo et al., 6 May 2026). The best Chaos-SSL APTOS configuration—Tent map with 30 SSL pretraining epochs—reports accuracy $4288 \times 2848$8 and macro F1-score $4288 \times 2848$9 (Florindo, 26 May 2026). A scalar ordinal-regression model with ResNet50 reports peak Validation QWK $224 \times 224$0 and validation MSE loss $224 \times 224$1 (Kumar et al., 18 Nov 2025). A dual-branch transfer-learning model evaluated on a fixed APTOS test split after selective merged training reports 5-class QWK $224 \times 224$2 and accuracy $224 \times 224$3, while also reporting binary DR detection accuracy $224 \times 224$4 (Shakibania et al., 2023). In a deployment-style 4-class cascade, APTOS yields accuracy $224 \times 224$5 and QWK $224 \times 224$6 with 49.52\% cloud-call rate, compared with a cloud-only baseline at 80.76\% accuracy and 0.8184 QWK (Doshi et al., 13 May 2026).

These numbers should be read in light of differing protocols. Some are validation results, some come from internal test splits, some use merged training sources, and some operate in derived label spaces rather than the original five-class task (Hashmi, 19 Apr 2026, Shakibania et al., 2023, Doshi et al., 13 May 2026). Even so, a recurrent empirical pattern is stable across papers: APTOS errors are dominated by adjacent-grade confusions, especially Mild vs Moderate and Moderate vs Severe/Proliferative, while No DR is usually the most separable class (Florindo et al., 6 May 2026, Shakibania et al., 2023).

7. Limitations, reproducibility issues, and dataset disambiguation

APTOS 2019’s research value is accompanied by recurring limitations in how it is used. Several papers explicitly note or implicitly reveal that the dataset is class-imbalanced, heterogeneous in acquisition conditions, and prone to adjacent-grade ambiguity because DR stage boundaries can be subtle even for human graders (Hashmi, 19 Apr 2026, Kumar et al., 18 Nov 2025, Shakibania et al., 2023). One cross-dataset study treats APTOS as a useful but insufficiently diverse source domain, showing that strong in-domain validation performance does not eliminate performance loss on Messidor-2 despite preprocessing and limited multi-domain training (Hashmi, 19 Apr 2026). The edge-cloud paper frames its results as retrospective feasibility evidence, not clinical validation, because the reported 733-image test set is an internal stratified split and no external-domain evaluation is provided (Doshi et al., 13 May 2026).

Reproducibility is uneven across the APTOS literature. Some papers provide explicit class counts, split ratios, training schedules, hardware, and optimization details (Doshi et al., 13 May 2026, Shakibania et al., 2023). Others omit core protocol components such as split construction, image resolution, preprocessing, augmentation policy, batch size, and random seed (Florindo et al., 6 May 2026, Florindo, 26 May 2026). In one study, even the number of APTOS images used and the train/validation ratio are not reported, despite APTOS being the anchor domain for model selection (Hashmi, 19 Apr 2026). A plausible implication is that APTOS 2019 remains easy to access as a dataset but comparatively difficult to standardize as a benchmark.

A separate source of confusion is nomenclature. A 2025 paper on OCT4DME / APTOS2021 Dataset explicitly distinguishes its own OCT-based anti-VEGF treatment-response dataset from the earlier APTOS 2019 Blindness Detection dataset and notes that APTOS 2019 is the 2019 Kaggle challenge for DR grading (5-level classification) from CFP with 3,662 images (Zhang et al., 9 May 2025). This distinction is important: APTOS 2019 is a fundus-photo diabetic retinopathy grading dataset, whereas the later APTOS competition dataset is an OCT benchmark for a different clinical task.