Refining a Deep Learning Model’s Ability to Identify GA Area

The GEODE study evaluated the use of 3D volumetric OCT data to accurately segment the area.

By: Simrat K. Sodhi, MSc, MB BChir (Cantab), Hasenin Al-Khersan, MD, David S. Boyer, MD, Charles C. Wykoff, MD, PhD, Netan Choudhry, MD, FRCSC

Retinal Physician March 1, 2025 Vol 22, Issue March 2025 Page(s): 19-21

Clinical Report: Refining a Deep Learning Model’s Ability to Identify GA Area

Overview

The GEODE study demonstrates the efficacy of a deep learning model in accurately segmenting geographic atrophy (GA) areas using 3D volumetric OCT data. The model achieved a high average Dice coefficient (DSC) of 0.826, indicating strong agreement with manual grading.

Background

Geographic atrophy (GA) is a significant cause of irreversible vision loss in patients with age-related macular degeneration (AMD). Accurate segmentation of GA areas is crucial for monitoring disease progression and evaluating treatment efficacy, especially with the advent of new therapies. The development of automated tools for GA assessment can enhance clinical decision-making and patient outcomes.

Data Highlights

Metric	Value
Average R²	0.906
Average DSC	0.826
DSC with nIR	0.829

Key Findings

The GEODE study utilized a U-Net–based architecture for direct segmentation of 3D volumetric OCT data.
Average R² for Spectralis OCT scans was 0.906, indicating high accuracy in area measurement.
The average Dice coefficient (DSC) was 0.826, demonstrating substantial overlap with manual grading.
The addition of near infrared (nIR) images did not significantly improve segmentation performance (DSC=0.829).
Systematic case reviews identified factors affecting DSC, such as data quality and segmentation errors.

Clinical Implications

The findings support the use of deep learning models for automated segmentation of GA in clinical settings, potentially improving monitoring and treatment decisions. Clinicians may consider integrating such tools into routine practice to enhance diagnostic accuracy and patient management.

Conclusion

The GEODE study highlights the potential of advanced deep learning techniques in improving the segmentation of geographic atrophy, paving the way for more effective monitoring and treatment strategies in AMD.

PentaVision

Retinal Physician