OCTDL: Optical Coherence Tomography Dataset for Image-Based Deep Learning Methods

• Existing public resources also target narrower tasks: RETOUCH labels three retinal-fluid types, while Duke supplies layer-boundary annotations for severe DME.Source: Open-access datasets · p. 2
• OCTID contains more than 500 images across normal, MH, AMD, CSR, and DR categories, with 25 normal images delineated for segmentation evaluation.Its raster scans have a 2 mm scan length and 512×1024-pixel resolution.Source: Open-access datasets · p. 2
• OCTDL provides 2,064 fovea-centered OCT B-scans labeled across six disease groups and retinal pathologies for research and algorithm development.The groups include AMD, DME, ERM, RAO, RVO, and VID; listed pathologies include MNV, DRIL, drusen, ME, and MH.Source: Background & Summary · p. 1, Open-access datasets · p. 2
• OCTDL is the second-largest open-access OCT image dataset, while Kermany’s dataset remains the largest.Table 1 is identified as a comparative analysis of published OCT datasets.Source: Background & Summary · p. 1, Table 1 · p. 3

Image acquisition

• OCT B-scans were acquired with an Optovue Avanti RTVue XR, centered on the fovea, and interpreted by an experienced retinal specialist.Axial and transverse resolutions were 5 µm and 15 µm, respectively; Figure 1 labels posterior-eye structures from inner to outer retina.Source: Methods · p. 3, Figure 1 · p. 4

Dataset labeling

• Seven trained medical students independently labeled images, disagreements were discussed, ambiguous diagnoses were screened out, and specialists and a clinic head confirmed final diagnoses.Two experienced specialists independently labeled cases and resolved disagreements by consensus; the clinic head confirmed every final diagnosis.Source: Methods · pp. 3–4

Age-related Macular Degeneration

• AMD features include drusen that enlarge or cluster, retinal deformation, and fluid spaces located subretinally, intraretinally, or beneath the RPE.Cuticular drusen can show a saw-tooth appearance; intraretinal fluid may appear as granular hyperreflective cysts.Source: Age-related Macular Degeneration · p. 4

Diabetic Macular Edema

• DME signs include hard exudates, intraretinal fluid, and DRIL, which marks loss of inner-retinal layer boundaries and correlates with visual-acuity prognosis.Hard exudates replace retinal tissue without increasing underlying retinal thickness; DRIL may persist after edema resolves.Source: Diabetic Macular Edema · p. 5

Retinal vascular occlusions

• RVO-associated edema is generally cystic and localized to the inner retina, whereas RAO is associated with ischemic inner-retinal hyperreflectivity and a prominent middle limiting membrane.The presence and severity of DRIL in vein occlusion are indicators of likely visual prognosis.Source: Retinal Vein Occlusion · p. 6, Retinal Artery Occlusion · p. 6

Vitreomacular Interface Disease and Epiretinal Membrane

• VID includes vitreoretinal traction and macular defects, while ERM involves glial tissue on the retinal inner surface and can deform the retina as traction matures.Intraretinal fluid in macular tears is contained within the tear borders; lamellar tears preserve photoreceptor-layer integrity and are often asymptomatic.Source: Vitreomacular Interface Disease · pp. 6–7, Epiretinal Membrane · p. 7

Release contents

• The final release contains 2,064 JPG images from 821 patients, stored in separate folders corresponding to disease labels.Filenames encode disease label, patient ID, and image sequence number.Source: Data records · pp. 7–8

Metadata and cohort

• The labels CSV includes disease, subcategory, condition, patient, eye, sex, year, and image-dimension fields.The reported patients were aged 20–93 years, with a mean age of 63 and a male-to-female ratio of 3:2.Source: Data records · p. 8

Technical Validation

• VGG16 and ResNet50 were selected as established CNN baselines for OCTDL classification.VGG16 has 138 million parameters and 13 convolutional plus three fully connected layers; ResNet50 uses shortcut connections and has 50 layers.Source: Technical Validation · p. 8
• OCTDL was randomly split by patient into training, validation, and test sets at 60:10:30, and experiments used cross-entropy, ADAM, and image augmentation.Augmentations included random crops, flips, rotation, translation, and Gaussian blur; reported metrics included accuracy, F1-score, AUC, precision, and recall.Source: Technical Validation · p. 8
• Dataset-combination experiments used image labels to harmonize categories, including grouping DME with DR and MH with VID.OCTDL and OCTID were mixed and split, while OCTDL served as the test subset for Kermany.Source: Technical Validation · p. 8
• Class-wise balanced accuracy approached 0.979, with AMD highest at 0.963 and RVO lowest at 0.633; recall followed a similar pattern, at 0.975 for AMD and 0.652 for RVO.Combining datasets yielded favorable outcomes and added disease variety and images from different OCT systems.Source: Technical Validation · p. 8
• OCTDL pathology labels may support automated layer segmentation, while disease and pathology labels together and voting ensembles are proposed as routes to increased classification accuracy.Semi- and unsupervised anomaly detection is also described as a promising direction for AI in OCT.Source: Technical Validation · pp. 8–9
• Planned dataset expansion targets rare conditions, including inherited retinal dystrophies and retinopathy of prematurity, and may assist their diagnosis and management.Source: Technical Validation · p. 9

Author contributions statement

• A.N., A.S., and A.K. are credited with data collection, while M.K., A.Z., and A.N. are credited with conceptualization.Source: Author contributions statement · p. 12
• M.K. is credited with software, and M.K., A.N., V.B., and M.R. with preparation of the original draft.Source: Author contributions statement · p. 12
• V.B., M.R., and P.C. are credited with review and editing, while A.M., S.K., and A.B. are credited with supervision.Source: Author contributions statement · p. 12