Comparing Interactive Machine Learning and Unassisted Segmentation of Geographic Atrophy From Fundus Autofluorescence

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) segmentation from fundus autofluorescence is crucial for monitoring disease progression.
• Manual segmentation is time-consuming and can suffer from inter-annotator variability.

Purpose of the Study:

• To evaluate the OPTIMEyes Interactive Machine Learning (IML) platform for ophthalmology image annotation.
• Assess efficiency, uniformity, and noninferiority of AI-assisted segmentation compared to unassisted segmentation of geographic atrophy (GA).

Main Methods:

• 10 annotators segmented 110 fundus autofluorescence images of GA.
• Ground truth segmentations were provided by an expert retinal specialist.
• Annotation time and DICE scores were used to compare unassisted and AI-assisted segmentation.

Main Results:

• AI assistance reduced average annotation time by 96 seconds.
• OPTIMEyes improved segmentation similarity across annotators (mean DICE difference 0.02).
• Significant improvements in segmentation quality were observed for challenging images (mean DICE improvement: 0.38).

Conclusions:

• OPTIMEyes enhances GA segmentation efficiency and uniformity in fundus autofluorescence images.
• The platform shortens annotation time and improves inter-annotator agreement without sacrificing segmentation quality.
• OPTIMEyes facilitates the creation of high-quality labeled datasets for research and clinical applications.