OCTDL: An Open-Access OCT Dataset for Deep Learning in Ophthalmology

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Apr 20 2024

In an article published in the journal Nature, researchers introduced an open-access optical coherence tomography (OCT) dataset for image-based deep learning (DL) methods (OCTDL) consisting of over 2000 labeled OCT images of retinal diseases, including age-related macular degeneration (AMD), diabetic macular edema (DME), and other conditions.

The images were acquired using an Optovue Avanti RTVue XR and interpreted by an experienced retinal specialist. The paper applied DL classification techniques to the dataset for advancing the diagnosis and monitoring of ocular conditions.

Background

OCT is a non-invasive imaging modality critical in clinical ophthalmology for its ability to provide detailed cross-sectional images of the retina. Initially focused on imaging the macular area, OCT technology has since advanced to encompass the visualization of vascular structures and neural tissue. OCT works based on low coherent light interferometry, detecting backscattered near-infrared light to construct depth profiles of biological tissues.

Improved resolution and image quality have enabled OCT to monitor and diagnose sight-threatening ocular conditions, such as glaucoma, diabetic retinopathy, and AMD. Previous work in OCT research has led to the development of various open-access datasets and DL models for the classification and segmentation of retinal conditions.

Studies such as the Kermany dataset and other public datasets have provided extensive data on retinal pathologies. However, these datasets often focus on a limited range of diseases or lack comprehensive annotations for different pathological conditions. To address these gaps, this paper presented a new open-access dataset, OCTDL, comprising over 2000 OCT images labeled according to various retinal diseases and conditions.

The dataset included macular scans of AMD, DME, and other conditions, offering a rich resource for developing advanced DL models. By providing a diverse and well-annotated dataset, this work supported the progress of automatic processing and early disease detection, aiding in monitoring and managing retinal diseases effectively.

Methodological Approach for OCT Imaging and Retinal Disease Classification

This study investigated various retinal diseases using high-resolution OCT images, which provided detailed views of the retina's internal structures. The images were obtained using the Optovue Avanti RTVue XR system and subsequently analyzed by a team of trained medical students and experienced clinical specialists. The main diseases examined included AMD, DME, retinal vein occlusion (RVO), retinal artery occlusion (RAO), vitreomacular interface disease (VID), and epiretinal membrane (ERM).

In AMD, OCT images revealed the presence of drusen, which were rounded deposits between Bruch’s membrane and the retinal pigment epithelium (RPE). As AMD progressed, these drusen could lead to disruptions in the foveal architecture and potential fluid accumulation beneath the retina. DME was characterized by the presence of hard exudates, intraretinal fluid, and disorganization of retinal inner layers. These abnormalities could compromise visual acuity and signal potential barriers in the blood-retinal interface.

RVO was linked with secondary macular edema, manifesting as cystic changes, and increased hyperreflectivity in the inner retina due to ischemia. RAO, on the other hand, led to acute tissue ischemia and pronounced hyperreflectivity of the inner retina. VID encompassed conditions such as vitreomacular traction syndrome and macular retinal holes, often resulting from pathologic adhesion between the retina and the vitreous body. ERM, characterized by the proliferation of glial tissue, might cause wrinkling of the retina and a reduction in visual acuity.

Data Records and Technical Validation Overview

The OCTDL comprised 2,064 images from 821 patients, featuring a range of retinal diseases. Stored in joint photographic experts group (JPG) format, the images were organized into folders by disease label, with associated metadata such as patient ID and condition. An accompanying CSV file provided detailed information about each image, including disease, subcategory, and patient demographics.

In validating the dataset, the performance of visual geometry group (VGG16) and residual network (ResNet50) DL architectures was tested on OCTDL, as these architectures were recognized for their proficiency in OCT image classification. VGG16, known for its simplicity, and ResNet50, featuring shortcut connections to address the vanishing gradient problem, provided strong baselines for evaluating the dataset's performance.

The dataset was split into training, validation, and test subsets with a 60:10:20 ratio. Various data augmentation techniques were used, including cropping, flipping, and blurring, to enhance model training. Metrics such as accuracy, F1-score, and area under the curve demonstrated high performance across most categories, particularly in detecting AMD. Combining OCTDL with other datasets like OCT image database (OCTID) and Kermany expanded disease variety and improved training reliability. Future research might explore applications such as automated segmentation and anomaly detection to advance artificial intelligence's role in ophthalmology. 

Conclusion

In conclusion, the OCTDL presented a comprehensive resource for DL research in ophthalmology by offering over 2,000 labeled OCT images of various retinal diseases. The dataset supported advancements in diagnosing and monitoring ocular conditions, particularly AMD and DME.

By validating the dataset using VGG16 and ResNet50 architectures, researchers demonstrated its high performance and potential for further development. Combining OCTDL with other datasets expanded disease coverage and improved training reliability. Future research could leverage OCTDL for automated segmentation and anomaly detection, enhancing artificial intelligence applications in ophthalmology.