A lightweight multi-deep learning framework for accurate diabetic retinopathy detection and multi-level severity...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) detection requires timely analysis of complex fundus images.
• Developing accurate automated algorithms for DR diagnosis remains a significant challenge.
• Early detection and classification of DR are critical for effective patient management and preventing vision loss.

Purpose of the Study:

• To present a novel, lightweight deep learning network for the detection and severity subclassification of diabetic retinopathy.
• To develop a two-stage framework that first identifies the presence of DR and then classifies its severity.
• To demonstrate the efficacy of transfer learning using a pre-trained model for enhanced DR subclassification.

Main Methods:

• A lightweight deep learning network was designed for DR detection.
• A two-stage classification approach was implemented: presence of DR and then DR severity subclassification (mild, moderate, severe, proliferative).
• Transfer learning was utilized in the second stage, reusing the designed model for severity classification on correlated fundus images.

Main Results:

• The proposed model is lightweight with fewer learnable parameters compared to existing methods.
• The two-stage framework achieved a 99.06% classification rate for DR detection.
• The model attained 90.75% accuracy for DR severity identification on the APTOS 2019 dataset.

Conclusions:

• The developed lightweight deep learning network provides an efficient solution for DR detection.
• The two-stage approach significantly enhances classification performance for both DR presence and severity.
• This framework shows promise for improving automated analysis of fundus images in clinical settings.