A Low Complexity Efficient Deep Learning Model for Automated Retinal Disease Diagnosis.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early identification and treatment of retinal diseases are vital for preserving vision.
• Optical coherence tomography (OCT) is a key imaging technology for diagnosing eye conditions.
• Deep learning offers potential for automated analysis of OCT images.

Purpose of the Study:

• To develop and evaluate a deep learning model for classifying human retinal OCT images into four categories.
• To enhance image quality and address data imbalance using preprocessing and generative adversarial networks (GANs).
• To compare the performance of the proposed model against established transformer-based architectures.

Main Methods:

• Human retinal OCT images were preprocessed and augmented using GANs, resulting in 130,649 images.
• A novel lightweight optimized compact convolutional transformers (OCCT) model was developed.
• The OCCT model was trained and evaluated on 32x32 images, alongside Vision Transformer (ViT), Swin Transformer, and eight transfer learning models.

Main Results:

• The OCCT model achieved a high accuracy of 97.09% in classifying retinal conditions.
• OCCT demonstrated superior performance compared to ViT and Swin Transformer models.
• The model's stability was confirmed, maintaining performance even with reduced training data.

Conclusions:

• The proposed OCCT model is effective for accurate and reliable classification of retinal diseases from OCT images.
• This deep learning approach holds promise for improving early diagnosis and treatment of vision-threatening conditions.
• The OCCT model offers a robust and efficient solution for automated retinal image analysis.