Deep Learning-Based Segmentation of Geographic Atrophy: A Multi-Center, Multi-Device Validation in a Real-World...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

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

Purpose of the Study:

• To develop and validate a deep learning algorithm for automated GA segmentation.
• To assess the algorithm's performance using optical coherence tomography (OCT) images.
• To evaluate the algorithm's applicability in routine clinical practice.

Main Methods:

• A 3D U-Net deep learning architecture was employed for model construction.
• The algorithm was trained and validated on OCT scans from patients with GA, with and without neovascular AMD (nAMD).
• Model accuracy was quantified using Dice Similarity Coefficient (DSC) and correlation (r²), comparing automated segmentation to manual labels.

Main Results:

• The algorithm achieved a mean DSC of 0.83 (r²=0.91) for Spectralis OCT data and 0.82 (r²=0.88) for Cirrus OCT data.
• The model demonstrated strong agreement with manual GA grading across two different OCT devices.
• Performance remained robust even in cases with concurrent nAMD.

Conclusions:

• The deep learning algorithm provides accurate and automated segmentation of GA in AMD patients.
• The model's consistent performance across devices and patient subgroups suggests significant clinical utility.
• This automated approach has the potential to streamline GA assessment in routine clinical settings.