Labeled Retinal Optical Coherence Tomography Dataset for Classification of Normal, Drusen, and CNV Cases

Published: 6 October 2021| Version 1 | DOI: 10.17632/8kt969dhx6.1


This dataset consists of more than 16,000 retinal OCT B-scans from 441 cases (Normal: 120, Drusen: 160, CNV: 161) and is acquired at Noor Eye Hospital, Tehran, Iran. Images are labeled by a retinal specialist. The structure of the folders are as below: - CNV, DRUSEN, NORMAL folders - Within each class, folders are separated patient-wise with numbers from 1 to <number_of_patients>. - Within each patient folder, images (B-scans) are labeled with <0XX_LABEL> format where <XX> is the B-scan number, and <LABEL> is the specialist's selected label for that specific B-scan. The excel spreadsheet (data_information.csv) includes information such as "Patient ID", "Class", "Eye", "B-scan", "Label", and "Directory" for all images (16823 rows, 6 columns). The python code ( includes code for loading images and labels as NumPy arrays. The written function outputs the input data as an array with shape (number_of_images, imageSize, imageSize, 3) and output data as a list of labels (Normal: 0, Drusen: 1, CNV: 2). There are two different options for reading the files: - Option 1: Reading all images. This would result in 16822 images. - Option 2: Reading the worst-case condition images for each volume (i.e., if a patient was detected as a CNV case, only CNV-appearing B-scans were included for training procedure and normal and drusen B-scans of that patient are excluded from the dataset). This would result in 12649 images.