Introducing a Novel Layer in Convolutional Neural Network for Automatic Identification of Diabetic Retinopathy

Area of Science:

• Ophthalmology
• Computer Science
• Medical Imaging

Background:

• Diabetic retinopathy is a leading cause of blindness, necessitating accurate early detection.
• Color fundus images are crucial for diagnosing diabetic retinopathy.
• Convolutional neural networks (CNNs) show promise in automated detection but can be further optimized.

Purpose of the Study:

• To propose a novel CNN architecture for enhanced diabetic retinopathy identification.
• To evaluate the impact of integrated image preprocessing techniques on CNN performance.
• To compare the effectiveness of different image enhancement methods within the CNN framework.

Main Methods:

• Developed a novel CNN architecture incorporating a preprocessing layer before the first convolutional layer.
• Integrated two image enhancement techniques: Contrast Enhancement (CE) and Contrast-limited Adaptive Histogram Equalization (CLAHE).
• Compared the performance of the proposed framework with a standard CNN for identifying exudates, hemorrhages, and microaneurysms.

Main Results:

• The proposed CNN framework achieved a total accuracy of 87.6% with CE and 83.9% with CLAHE.
• The standard CNN without the preprocessing layer achieved a total accuracy of 81.4%.
• The novel architecture demonstrated a significant improvement in classification performance.

Conclusions:

• Embedding a preprocessing layer with image enhancement techniques notably boosts CNN performance for diabetic retinopathy detection.
• Contrast Enhancement proved more effective than CLAHE in this specific application.
• The proposed CNN architecture offers a promising advancement for automated analysis of retinal images.