Deep learning-based detection and classification of geographic atrophy using a deep convolutional neural network...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Geographic atrophy (GA) is a leading cause of vision loss.
• Accurate detection and classification of GA in fundus autofluorescence (FAF) images are crucial for patient management.
• Current methods may be time-consuming and subjective.

Purpose of the Study:

• To develop and validate a deep learning algorithm for automated detection and classification of GA in FAF images.
• To differentiate between diffuse-trickling GA (dt-GA) and other GA patterns (ndt-GA).

Main Methods:

• A multi-layer deep convolutional neural network (DCNN) was trained using FAF images from patients with GA, healthy individuals, and those with other retinal diseases (ORDs).
• Two classifiers were developed: one for GA detection (GA vs. healthy/GA vs. ORD) and another for differentiating GA patterns (dt-GA vs. ndt-GA).
• The DCNN model performance was evaluated using accuracy and cross-entropy metrics.

Main Results:

• The GA classifiers achieved high training (99/98%) and validation (96/91%) accuracies.
• The classifier for differentiating GA patterns showed 99% training accuracy and 77% validation accuracy.
• The algorithm demonstrated statistically significant probability scores for GA and dt-GA classifications (p < 0.001).

Conclusions:

• This study presents the first deep learning-based algorithm for automated GA detection and classification in FAF images.
• The developed classifiers exhibit excellent performance.
• This AI model holds potential for predicting GA progression risk and informing therapeutic strategies.