Automated deep learning-based AMD detection and staging in real-world OCT datasets (PINNACLE study report 5)

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal optical coherence tomography (OCT) scans are abundant in eye care centers.
• Electronic health records alone are insufficient for high-quality retrospective data analysis.
• Classifying stages of age-related macular degeneration (AMD) is crucial for patient management.

Purpose of the Study:

• To develop a deep learning classifier for age-related macular degeneration (AMD) stages.
• To efficiently identify early/intermediate (iAMD), atrophic (GA), and neovascular (nAMD) stages in retrospective OCT data.
• To generate classification uncertainty estimates for improved reliability.

Main Methods:

• A two-stage convolutional neural network was trained on Topcon OCT images.
• Stage 1: A 2D ResNet50 classified individual OCT B-scans.
• Stage 2: Four ResNet models classified entire OCT volumes using concatenated B-scan outputs.
• Monte-Carlo dropout was used for uncertainty estimation.

Main Results:

• The model was trained on 3765 scans from 1849 eyes.
• The classifier achieved an average ROC-AUC of 0.94 on a real-world test set.
• The model accurately distinguished between Normal, iAMD, GA, and nAMD classifications.

Conclusions:

• Deep learning effectively classifies AMD stages from retrospective OCT data.
• The developed classifier provides a robust tool for analyzing large-scale OCT datasets.
• This technology can significantly aid in the retrospective analysis of AMD progression and treatment outcomes.