Detection of Diabetic Eye Disease from Retinal Images Using a Deep Learning Based CenterNet Model.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) and diabetic macular edema (DME) are leading causes of vision loss in diabetic patients.
• Manual screening of retinal images is labor-intensive, costly, and lacks sufficient trained personnel.
• Current automated systems struggle with generalization across diseases and real-world conditions.

Purpose of the Study:

• To develop an automated system for accurate detection and classification of DR and DME lesions.
• To enhance the performance of deep learning models for eye disease screening.
• To improve the localization and classification of disease indicators in retinal images.

Main Methods:

• A novel method combining dataset preparation, feature extraction using DenseNet-100, and a custom deep learning CenterNet model.
• Annotation of suspected retinal samples to precisely locate regions of interest.
• Training the CenterNet model on annotated images for disease localization and classification.

Main Results:

• Achieved high average accuracy: 97.93% on APTOS-2019 and 98.10% on IDRiD datasets.
• Demonstrated superior performance over state-of-the-art methods through cross-dataset validation.
• Effectively recognized small lesions and handled over-fitted training data with enhanced localization power.

Conclusions:

• The proposed framework accurately locates and classifies DR and DME lesions.
• The method excels at extracting key points from low-intensity and noisy images.
• This approach significantly contributes to automated detection and recognition of diabetic eye disease lesions.