Algorithm Appears to Aid Diabetic Retinopathy Accuracy
Ophthalmologists can increase the confidence in the accuracy of their diabetic retinopathy diagnoses by using deep learning algorithms, according to a recent evaluation involving nearly 1,800 retinal fundus images.
Five general ophthalmologists, four retina specialists and one retina fellow read each image under one of the following circumstances: unassisted, grades only (a histogram of diabetic retinopathy predictions from a deep-learning model), and grades plus explanatory heatmaps. Among the results:
- Clinicians made more accurate diagnoses in the grades-only scenario, vs unassisted.
- Adding the heatmap improved diabetic retinopathy diagnostic accuracy, but lessened it for non-diabetic retinopathy diagnoses.
- Reader sensitivity increased with both types of assistance. For moderate non-proliferative diabetic retinopathy or worse, unassisted showed 79% accuracy, vs 88% for grades only and 89% when the heatmap was added; specificity remained high across all three scenarios.
- Retina specialists experienced greater accuracy with model assistance than unassisted readers or just the model.
- Clinician confidence in their diagnosis and duration of task increased with model assistance.
- Grades + heatmap and grades only were equally effective in most cases.
- Duration of task lessened over time for all conditions, but made the biggest drop for grades + heatmap.
Sayres R, Taly A, Rahimy E, et al. Using a deep learning algorithm and integrated gradients explanation to assist grading for diabetic retinopathy. [Published online ahead of print December 13, 2018]. Ophthalmology. 2018;44(5):557-565. doi: https://doi.org/10.1016/j.ophtha.2018.11.016.
