Democratizing Deep Learning Research Through Large Publicly Available Datasets and Tools

Abstract

Purpose : Deep learning technologies hold great potential to transform and optimize decision support systems in healthcare. As we advance deep learning research and its applications in ophthalmology, massive amounts of high-quality pre-labelled data will be required for model training. Yet, publicly available large databanks for this purpose are limited and the process of generating such datasets can be prohibitively costly both in time and effort. To ease this challenge and to support the democratization of deep learning research, we built and make publicly available a large data set of high-quality labelled optical coherence tomography images along with a repository of supporting analytical toolkits.

Methods : One-hundred and fifty volumes split evenly between intermediate age-related macular degeneration (AMD), diabetes[MVC1] retinopathy, and normal controls were obtained using the Heidelberg Spectralis. From this dataset, 20 volumes (19-69 B-scans/volume; 1668 total scans) from each disease class were labeled for the location of the ILM, INL/OPL boundary, top of the IS/OS, the inner and the outer boundary of the RPE using a custom-built image labelling platform. All images were graded independently three times by a team of experienced graders. A subset of 200 images were graded twice for presence/location of fluid and classes of drusen. All images were adjudicated and corrected if needed to ensure all grades match layer definitions. Agreement between graders was determined using DICE, Intersection over Union (IoU) and Interclass Correlation Coefficients (ICC).

Results : A total of 5004 grades were produced for the presence and location of 5 retinal interfaces and corresponding 6 layers. The agreement between graders for the retinal layers is shown in Table 1, separated by disease. Labels for fluid presence at the B-scan and bounding box tags are also generated and available.

Conclusions : This dataset and subsequent web listing (AI4eyes.org) serve as the beginning of a providing a common database by which to benchmark novel tool development. The platform also provides tools for uniform labelling, thereby lowering the burden on new entrants. The hope is that this platform will continue to expand through our work and others to provide an ever growing depository of images and labels for groups looking to develop artificial intelligence tools, as well as a source for tools to help new groups begin.

This is a 2021 ARVO Annual Meeting abstract.