Geographic Atrophy Segmentation Using Multimodal Deep Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) is a leading cause of vision loss.
• Accurate segmentation of GA lesions is crucial for monitoring disease progression and evaluating treatments.
• Current manual segmentation methods can be time-consuming and subjective.

Purpose of the Study:

• To evaluate deep learning (DL) based methods for precise segmentation of geographic atrophy (GA) lesions.
• To compare the performance of DL models with expert graders using fundus autofluorescence (FAF) and near-infrared (NIR) imaging.

Main Methods:

• Retrospective analysis of imaging data from Proxima A and B natural history studies of GA.
• Utilized two multimodal DL networks, UNet and YNet, for automatic GA lesion segmentation on FAF images.
• Assessed segmentation accuracy using Dice coefficient scores, Bland-Altman plots, and Pearson correlation coefficients, comparing DL networks to expert grader annotations.

Main Results:

• DL networks achieved high segmentation accuracy, with Dice scores ranging from 0.89 to 0.92, comparable to inter-grader scores (0.94).
• High correlations (r) were observed for GA lesion area between DL networks and graders (0.959–0.981).
• Longitudinal correlations for lesion area enlargement were lower than cross-sectional correlations, indicating challenges in tracking changes over time.

Conclusions:

• Multimodal DL networks demonstrate capability for accurate GA lesion segmentation, yielding results comparable to experienced human graders.
• DL-based tools show promise for supporting efficient and personalized GA patient assessments in clinical research and practice.