A global review of publicly available datasets for ophthalmological imaging: barriers to access, usability, and generalisability

Abstract

Health data that are publicly available are valuable resources for digital health research. Several public datasets containing ophthalmological imaging have been frequently used in machine learning research; however, the total number of datasets containing ophthalmological health information and their respective content is unclear. This Review aimed to identify all publicly available ophthalmological imaging datasets, detail their accessibility, describe which diseases and populations are represented, and report on the completeness of the associated metadata. With the use of MEDLINE, Google's search engine, and Google Dataset Search, we identified 94 open access datasets containing 507 724 images and 125 videos from 122 364 patients. Most datasets originated from Asia, North America, and Europe. Disease populations were unevenly represented, with glaucoma, diabetic retinopathy, and age-related macular degeneration disproportionately overrepresented in comparison with other eye diseases. The reporting of basic demographic characteristics such as age, sex, and ethnicity was poor, even at the aggregate level. This Review provides greater visibility for ophthalmological datasets that are publicly available as powerful resources for research. Our paper also exposes an increasing divide in the representation of different population and disease groups in health data repositories. The improved reporting of metadata would enable researchers to access the most appropriate datasets for their needs and maximise the potential of such resources.

The Lancet Global Health Commission on Global Eye Health: vision beyond 2020

“…This Commission did a global review to identify publicly available datasets of ophthalmic images. 285 We identified 94 (including six with data from two world regions) open access datasets with more than 500 000 images from 23 countries; there were 13 datasets for which the country was not specified. 34 datasets originated from populations in Europe, 21 from southeast or east Asia, 16 from North America, nine from north Africa and the Middle East, four from south Asia, two from Latin America, and one from sub-Saharan Africa.…”

Health data poverty: an assailable barrier to equitable digital health care

“…15 And, focusing just on the field of ophthalmology, a global review of ophthalmic imaging datasets revealed disparities in the representation of different populations and disease groups in publicly available health data repositories. 16 These imbalances lead to datasets that underrepresent key segments of the overall population. As these same datasets are used to develop and validate digital health technologies, a possible extreme scenario is that datadriven interventions are safe and effective for some people, but dangerous and ineffective for others.…”

Annotated retinal optical coherence tomography images (AROI) database for joint retinal layer and fluid segmentation

“…Besides the RETOUCH challenge database, the only publicly available dataset, which some authors use for evaluation of their methods, is the DUKE dataset [28] containing 110 B-scans of 10 DME (diabetic macular oedema) patients acquired with the Spectralis OCT. In a recent review paper, Khan et al [29] gave an extensive overview of publicly available databases in ophthalmology (up to May 2020) in which authors raise serious concerns about "data poverty" and argue about challenges in data collection.…”