Essentials of a Robust Deep Learning System for Diabetic Retinopathy Screening: A Systematic Literature Review
Table 1
Study selection criteria. Stage 1 included population, algorithm type, publication category, and image modality and was applied to 747 articles. The full selection criteria were used for the remaining 65 articles, resulting in the final selection of 13 articles. †These were only used as additional search resources. ‡This number was arbitrarily determined.
Inclusion criteria
Exclusion criteria
Population
Individuals with diabetes (type 1 or 2) Patients with any DR stage and/or DMO Populations with DR and other related eye diseases (if data for populations with DR only is separate)
Individuals without diabetes Other retinal diseases
Algorithm type
Deep learning systems Classification tasks (e.g., grading /screening DR) Convolutional neural networks (CNN)
Manual feature construction Expert systems Segmentation tasks (e.g., lesion quantification) Prediction tasks (e.g., future outcomes/prognosis)
Publication category
Peer-reviewed Published
Editorials, letters, opinion pieces, notes or comments Conference abstracts /proceedings Systematic reviews /meta-analyses† Grey literature (e.g., statistics on diabetes /DR, white papers, clinical practice guidelines)
Image modality
Any retinal camera type Field of view: 40 to 45° Retinal colour fundus photographs
Images from: Smartphones /mobiles OCT Fluorescein angiography Stereoscopic imaging Wide/ultrawide field fundus photography
Text availability
Full text available
Full text not available
DR classification
Screening or grading DR DR severity scale
Not screening or grading DR No DR severity scale
Reference standard
Determined by human graders
Not determined by human graders
Outcomes
Sensitivity and specificity measures of DR classification
No sensitivity and specificity measures of DR classification
