Diabetic Retinopathy Detection from Fundus Images of the Eye Using Hybrid Deep Learning Features.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic Retinopathy (DR) is a complication of long-term diabetes, leading to vision impairment if not diagnosed early.
• High blood sugar damages retinal blood vessels, causing changes like Microaneurysms (MAs), Exudates (EXs), and Hemorrhages (HMs).
• Manual DR detection is challenging due to subtle structural changes in the retina.

Purpose of the Study:

• To propose a hybrid technique for the detection and classification of Diabetic Retinopathy (DR) in retinal fundus images.
• To leverage transfer learning (TL) and Convolutional Neural Network (CNN) models for enhanced DR diagnosis.
• To evaluate the performance of the proposed method against existing DR detection approaches.

Main Methods:

• A hybrid feature extraction method using pre-trained CNN models via transfer learning (TL).
• Generation of a hybrid feature vector from combined CNN features.
• Classification of fundus images using various classifiers for binary and multiclass DR detection.

Main Results:

• The proposed hybrid technique demonstrated significant performance improvements in DR detection.
• Achieved a 97.8% accuracy for binary classification of DR.
• Attained 89.29% accuracy for multiclass classification of DR.

Conclusions:

• The developed hybrid method offers a more effective approach to Diabetic Retinopathy detection.
• The integration of TL and CNNs provides a robust framework for analyzing fundus images.
• This technique holds potential for earlier and more accurate diagnosis of DR, aiding in vision preservation.