Glaucoma Detection from Raw SD-OCT Volumes: A Novel Approach Focused on Spatial Dependencies

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Glaucoma is a leading cause of global blindness.
• Artificial intelligence (AI) aids ophthalmologists using fundus and optical coherence tomography (OCT) imaging.
• 3D spectral-domain OCT (SD-OCT) scans offer valuable data for glaucoma detection.

Purpose of the Study:

• To develop a novel deep learning methodology for glaucoma detection using 3D SD-OCT scans.
• To leverage spatial dependencies of features extracted from B-scans for improved diagnostic accuracy.
• To analyze hidden information within 3D SD-OCT scans for early glaucoma identification.

Main Methods:

• A two-stage deep learning approach was employed, involving a slide-level feature extractor and a volume-based predictive model.
• The feature extractor utilized novel residual and attention convolutional modules combined with fine-tuned architectures.
• Long Short-Term Memory (LSTM) networks and a sequential-weighting module (SWM) were used for volume-level prediction, processing features from conditioned SD-OCT volumes.

Main Results:

• The slide-level feature extractor achieved Area Under the Curve (AUC) values exceeding 0.93 on primary and external test sets.
• The end-to-end system, combining CNN and LSTM, reached an AUC of 0.8847 in the prediction stage, surpassing state-of-the-art methods.
• Class Activation Maps (CAMs) were generated to visualize critical regions in B-scans for distinguishing between healthy and glaucomatous eyes.

Conclusions:

• The proposed model effectively extracts B-scan features and integrates latent space information for volume-level glaucoma prediction.
• The model, incorporating residual and attention blocks with an SWM to refine LSTM outputs, demonstrates superior performance compared to existing 3D deep learning architectures for glaucoma detection.