A feature agnostic approach for glaucoma detection in OCT volumes

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for glaucoma diagnosis, measuring retinal nerve fibre layer (RNFL) and ganglion cell-inner plexiform layer (GCIPL) thickness.
• Current machine learning methods often rely on segmented imaging features like peripapillary RNFL thickness and cup-to-disc ratio.

Purpose of the Study:

• To develop and evaluate a deep learning technique for direct glaucoma classification from raw OCT volumes of the optic nerve head (ONH).
• To compare the performance of this deep learning approach against traditional feature-based machine learning algorithms.

Main Methods:

• A 3D Convolutional Neural Network (CNN) was employed to classify eyes as healthy or glaucomatous directly from unsegmented OCT volumes.
• The deep learning model's accuracy was compared with classical machine learning algorithms, including logistic regression.
• Class Activation Maps (CAM) were utilized to identify critical regions within the OCT volumes for glaucoma detection.

Main Results:

• The deep learning approach achieved a superior Area Under the Curve (AUC) of 0.94, compared to 0.89 for the best classical method (logistic regression).
• The CNN identified key anatomical regions, including the neuroretinal rim, optic disc cupping, and lamina cribrosa (LC), as significant for glaucoma classification.
• These identified regions align with established clinical markers for glaucoma, such as cup volume, cup diameter, and rim thinning.

Conclusions:

• Direct analysis of OCT volumes using a 3D CNN offers a more accurate and insightful method for glaucoma detection compared to traditional machine learning.
• The deep learning model's ability to highlight important anatomical regions provides valuable insights for glaucoma diagnosis and monitoring.