A deep learning framework for the early detection of multi-retinal diseases

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal fundus images are crucial for diagnosing ocular conditions.
• Early detection and treatment of eye diseases are facilitated by rapid image analysis.
• Deep learning algorithms offer potential for automated diagnosis from retinal images.

Purpose of the Study:

• To develop a non-invasive method for early detection and treatment of multiple eye diseases using Convolutional Neural Networks (CNNs).
• To evaluate the performance of different CNN models in classifying retinal fundus images representing various ocular conditions and healthy eyes.

Main Methods:

• Utilized the Retinal Fundus Multi-disease Image Dataset (RFMiD) containing images of Media Haze (MH), Optic Disc Cupping (ODC), Diabetic Retinopathy (DR), and healthy (WNL) eyes.
• Applied pre-processing techniques including data augmentation, cropping, resizing, dataset splitting, image-to-array conversion, and one-hot encoding.
• Experimented with three CNN models, extracting features for predictive diagnostic decisions and assessing performance using accuracy, F1 score, recall, and precision.

Main Results:

• A 12-layer CNN achieved up to 89.81% validation and 88.72% testing accuracy with data augmentation.
• A 20-layer CNN reached 90.34% validation and 89.59% testing accuracy with augmented data, though it exhibited overfitting.
• The achieved accuracy rates indicate the CNN models effectively distinguish between different eye diseases and healthy images.

Conclusions:

• The developed deep learning framework demonstrates reliable and efficient capabilities for the simultaneous detection of multiple eye diseases.
• CNNs can accurately analyze color fundus images for improved diagnostic decision-making in ophthalmology.
• This approach supports timely treatment planning through early and accurate identification of various ocular conditions.