Conv-ViT: A Convolution and Vision Transformer-Based Hybrid Feature Extraction Method for Retinal Disease Detection.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Current deep learning models for retinal disease detection often rely on either texture or shape features alone.
• This limitation hinders model robustness and accurate classification of diverse retinal conditions.

Purpose of the Study:

• To develop a hybrid deep learning model, Conv-ViT, integrating both texture and shape feature extraction for improved retinal disease detection.
• To classify four retinal disease classes: choroidal neovascularization (CNV), diabetic macular edema (DME), DRUSEN, and NORMAL.

Main Methods:

• Developed a fusion model (Conv-ViT) combining transfer learning-based Convolutional Neural Networks (CNNs) like Inception-V3 and ResNet-50 for texture analysis.
• Integrated a Vision Transformer (ViT) model for shape-based feature extraction, analyzing long-distance pixel correlations.
• Utilized foveal cut optical coherence tomography (OCT) images for training and validation.

Main Results:

• The Conv-ViT model achieved a weighted average classification accuracy, precision, recall, and F1 score of approximately 94%.
• The fusion of texture and shape features significantly enhanced classification performance compared to models using single feature types.

Conclusions:

• The proposed Conv-ViT model demonstrates superior performance in retinal disease classification by effectively leveraging both texture and shape information.
• This hybrid approach offers a more robust and accurate method for detecting various retinal diseases from OCT images.