Deep learning model for automatic differentiation of EMAP from AMD in macular atrophy

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Distinguishing between age-related macular degeneration (AMD) and extensive macular atrophy and pseudodrusen-like appearance (EMAP) is crucial for patient management.
• Fundus autofluorescence (FAF) imaging provides valuable insights into retinal pigment epithelium and outer retinal health.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) classifier for differentiating EMAP from dry AMD using FAF images.
• To assess the impact of FAF image field of view on classifier performance.

Main Methods:

• Retrospective selection of patients with atrophy secondary to EMAP or dry AMD.
• Training of two ResNet-101 based DL classifiers on 30°×30° and 55°×55° FAF images.
• Independent testing of classifiers on data from a different center.

Main Results:

• The 30°×30° FAF DL classifier achieved 84.6% sensitivity and 85.3% specificity for EMAP diagnosis.
• The 55°×55° FAF DL classifier demonstrated superior performance with 90% sensitivity and 84.6% specificity (p=0.037).
• AI accurately differentiates atrophy causes on FAF, with performance enhanced by wide-field imaging.

Conclusions:

• Deep learning models can effectively distinguish EMAP from AMD using FAF imaging.
• Wide-field FAF acquisitions significantly improve the diagnostic performance of DL classifiers for these conditions.