Assessing the efficacy of 2D and 3D CNN algorithms in OCT-based glaucoma detection

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Neuroscience and Vision Science

Background:

• Glaucoma is a leading cause of irreversible blindness, necessitating early and accurate diagnosis for effective management.
• Optical coherence tomography (OCT) provides high-resolution imaging of the optic nerve head (ONH) and macula, crucial for glaucoma assessment.
• Convolutional neural networks (CNNs) are increasingly used for analyzing OCT images, but the comparative efficacy of 2D versus 3D CNNs for glaucoma detection is not well-established.

Purpose of the Study:

• To compare the diagnostic performance of 2D-CNN models, 3D-CNN adaptations, and a custom 3D-CNN-Encoder for glaucoma detection using volumetric OCT images.
• To evaluate the efficacy of these algorithms on datasets focused on both the macula and the optic nerve head (ONH).
• To assess the utility of 2D and 3D CNNs in advancing automated glaucoma diagnostic systems, especially considering the limited availability of pre-trained 3D models.

Main Methods:

• Utilized state-of-the-art (SOTA) 2D-CNN models and their 3D adaptations with specific weight transfer techniques.
• Developed and tested a custom 5-layer 3D-CNN-Encoder algorithm.
• Evaluated model performance on two distinct datasets comprising macular and ONH OCT scans.

Main Results:

• The 2D-CNN algorithm consistently outperformed its 3D counterparts in glaucoma detection across both datasets.
• Achieved high area under the curve (AUC) values: 0.960 for macular OCT images and 0.943 for ONH OCT images.
• Demonstrated the robustness and effectiveness of 2D CNNs for glaucoma detection using volumetric OCT data.

Conclusions:

• 2D-CNN algorithms offer a robust and effective approach for glaucoma detection from volumetric OCT images, comparable or superior to tested 3D CNNs.
• The findings highlight the practical utility of 2D CNNs in clinical settings, particularly given the current scarcity of extensively pre-trained 3D models.
• This comparative analysis supports the integration of 2D CNNs into advanced ophthalmic diagnostic systems for improved glaucoma screening and management.