Clinically relevant deep learning for detection and quantification of geographic atrophy from optical coherence...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Geographic atrophy (GA) is a severe complication of age-related macular degeneration (AMD) and a leading cause of global blindness.
• Current methods for GA detection and quantification lack speed, reliability, and objectivity, hindering disease monitoring and therapeutic development.
• Objective and automated detection of GA from optical coherence tomography (OCT) scans is crucial for clinical trials and patient care.

Purpose of the Study:

• To develop and validate a fully automated deep-learning method for detecting and quantifying geographic atrophy from OCT retinal scans.
• To assess the performance of the developed AI model against expert human graders.

Main Methods:

• A modified U-Net deep-learning architecture was employed to create models for segmenting GA and its features from Heidelberg Spectralis OCT scans.
• Model development utilized a manually segmented dataset of 5049 B-scans from 984 OCT volumes.
• External validation was performed on an independent dataset of 884 B-scans from 192 OCT volumes from patients receiving routine care.

Main Results:

• The deep-learning model achieved high agreement with expert consensus grading on the external validation dataset, with a median Dice Similarity Coefficient (DSC) of 0.96 and an intraclass correlation coefficient (ICC) of 0.93.
• The AI model outperformed the agreement between human graders (DSC 0.80, ICC 0.79).
• Independent models accurately segmented key GA features: retinal pigment epithelium loss (DSC 0.95), photoreceptor degeneration (DSC 0.96), and hypertransmission (DSC 0.97).

Conclusions:

• A validated deep-learning model demonstrates performance comparable to manual specialist assessment for geographic atrophy segmentation.
• Automated analysis of OCT scans using AI offers a promising approach for the diagnosis and prognosis of GA in clinical practice and research.
• Further clinical validation is recommended for widespread adoption in patient care.