Diabetic retinopathy prediction based on vision transformer and modified capsule network

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy is a leading cause of blindness in working-age adults.
• Early detection and accurate severity grading are crucial for sight preservation.
• Manual diagnosis by ophthalmologists is time-consuming and requires significant expertise.

Purpose of the Study:

• To develop an automated hybrid deep learning method for diabetic retinopathy severity prediction.
• To improve the efficiency and accuracy of diabetic retinopathy diagnosis.

Main Methods:

• Proposed a hybrid deep learning model combining a fine-tuned Vision Transformer and a modified Capsule Network.
• Implemented advanced computer vision techniques: power law transformation and contrast-limiting adaptive histogram equalization for preprocessing.
• Evaluated the model on four diverse datasets: APTOS, Messidor-2, DDR, and EyePACS.

Main Results:

• Achieved high test accuracy scores across all datasets: 88.18% (APTOS), 87.78% (Messidor-2), 80.36% (DDR), and 78.64% (EyePACS).
• Demonstrated superior performance compared to existing state-of-the-art methods.
• The hybrid model effectively predicts diabetic retinopathy severity levels.

Conclusions:

• The proposed hybrid deep learning approach offers a promising automated solution for diabetic retinopathy severity prediction.
• This method can significantly aid ophthalmologists in early detection and management, potentially reducing blindness.
• The approach shows excellent potential for clinical application in diabetic eye care.