A combined convolutional and recurrent neural network for enhanced glaucoma detection.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a primary cause of irreversible blindness globally.
• Current glaucoma detection methods primarily rely on spatial features from static fundus images.
• Limitations exist in capturing dynamic pathological features present in fundus videos.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for enhanced glaucoma detection.
• To integrate spatial and temporal feature extraction for improved diagnostic accuracy.
• To compare the performance of a combined CNN-RNN model against a traditional CNN model.

Main Methods:

• Utilized a dataset of 1810 fundus images and 295 fundus videos.
• Developed a hybrid model combining Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory Recurrent Neural Networks (RNNs) for temporal feature analysis.
• Trained and tested the combined CNN-RNN model against a standalone CNN model.

Main Results:

• The combined CNN-RNN model achieved a high average F-measure of 96.2% in distinguishing glaucoma from healthy eyes.
• The base CNN model, using only spatial features, attained an average F-measure of 79.2%.
• The hybrid model demonstrated a significant enhancement in glaucoma detection accuracy.

Conclusions:

• Integrating spatial and temporal features from fundus videos markedly improves glaucoma detection accuracy.
• The combined CNN-RNN approach offers a promising advancement for early and accurate glaucoma diagnosis.
• This proof-of-concept study highlights the potential of video-based AI analysis in ophthalmology.