Using a Deep Learning Algorithm and Integrated Gradients Explanation to Assist Grading for Diabetic Retinopathy.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Accurate and timely diagnosis of DR is crucial for effective management.
• Computer-assisted diagnosis holds promise for improving DR screening efficiency.

Purpose of the Study:

• To evaluate the impact of deep learning (DL) algorithms on ophthalmologists' diagnostic performance for DR.
• To assess how different levels of algorithmic assistance (grades only vs. grades plus heatmap) affect reader accuracy, confidence, and grading time.

Main Methods:

• Ten ophthalmologists evaluated 1796 retinal fundus images from 1612 diabetic patients.
• Images were assessed under three conditions: unassisted, with DL-generated DR grades, and with DL grades plus explanatory heatmaps.
• Performance metrics included sensitivity, specificity, accuracy, confidence ratings, and grading time.

Main Results:

• Model assistance significantly improved reader accuracy compared to unassisted grading (P < 0.001).
• Assistance increased sensitivity for moderate-to-worse DR detection without compromising specificity.
• Heatmap assistance improved accuracy for DR-positive cases but reduced it for DR-negative cases; overall effectiveness was similar to grades-only assistance.

Conclusions:

• Deep learning algorithms enhance DR diagnosis accuracy and reader confidence in assisted settings.
• Algorithmic assistance can increase initial grading time, but this effect diminishes with experience.
• DL tools show potential for improving DR screening and patient outcomes.