Retinal OCT Image Classification — C8
24,000 retinal OCT scans across eight conditions, split into balanced folds — a benchmark you can trust to be fair before you trust your model to be good.
Images
24,000 · JPEG
Classes
8 retinal conditions
Splits
18,400 / 2,800 / 2,800
Licence
CC BY-NC-SA 4.0
01 · What it contains
IndexChapter 01
What the dataset is
Optical coherence tomography scans of the retina, compiled from multiple reputable sources and sorted into eight clinical conditions. Every class holds exactly 3,000 images, split 2,300 train / 350 validation / 350 test — so a model that reports 90% accuracy has genuinely earned it rather than learned the class distribution.
Chapter 02
Why it exists
The companion research needed a multi-class OCT benchmark, and the widely used public OCT datasets at the time were smaller in their class taxonomy. Consolidating eight conditions with equal representation made the comparison between architectures meaningful.
Chapter 03
The eight classes
Perfectly balanced by construction — 3,000 images per class. Select a class to read its condition; the tiles are illustrative, never real scans.
Class distribution · 24,000 images · 8 classes
Selected class
AMD
Age-related Macular Degeneration
3,000 images · 12.5% of corpus
Illustrative tiles — not real samples
Image dimensions are deliberately left non-uniform — resizing is left to the consumer, since the right input size depends on the backbone.
Chapter 04
How people use it
Beyond the companion paper, the dataset is indexed in the openmedlab Awesome-Medical-Dataset collection, and Kaggle hosts 45 public notebooks using it for transfer-learning baselines such as VGG16.
Chapter 05
Responsible use
Cite this
Obuli Sai Naren. (2021). Retinal OCT Image Classification - C8 [Data set]. Kaggle.
doi.org/10.34740/KAGGLE/DSV/2736749 ↗Lessons
What it taught me
- Publishing this as an undergraduate was the first time my work left my hands and kept going without me. That is still the goal.
- A benchmark’s value is its fairness, not its size. Equal class counts cost weeks and saved every downstream comparison.
- Documentation determines adoption. The README is why people cite it correctly.