OctNET: A Lightweight CNN for Retinal Disease Classification from Optical Coherence Tomography Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal diseases pose a growing health challenge, necessitating early detection to prevent vision impairment.
• Diabetes is a major risk factor, with expected exponential growth in affected individuals.
• Optical Coherence Tomography (OCT) provides high-resolution retinal imaging for quantitative analysis.

Purpose of the Study:

• To develop a deep neural network-based classifier for automated detection of key retinal pathologies.
• To classify Diabetic Macular Edema (DME), drusen, and Choroidal NeoVascularization (CNV) using OCT images.
• To improve computer-aided diagnosis of retinal diseases.

Main Methods:

• A deep convolutional neural network with six convolutional blocks was employed for image classification.
• Gradient-based class activation mapping was utilized for result interpretability.
• The model was trained and validated on a large public dataset of 83,484 retinal OCT images.

Main Results:

• The classifier achieved high performance metrics: 99.69% precision, 99.69% recall, and 99.69% accuracy.
• Out of 968 test cases, only three were misclassified.
• The model demonstrated superior classification performance compared to the ResNet-50 model.

Conclusions:

• The proposed network significantly reduces trainable parameters (6.9% of ResNet-50) through downsampling and weight sharing, enhancing training efficiency.
• Class activation mapping results visually align with actual retinal OCT images.
• The model's reduced complexity and high accuracy make it suitable for real-time clinical applications.