UC-stack: a deep learning computer automatic detection system for diabetic retinopathy classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) diagnosis relies on subjective interpretation of optical images, leading to variability.
• Automated systems are needed to improve the accuracy and consistency of DR screening.

Purpose of the Study:

• To develop and evaluate a novel, automated deep learning system for early detection and grading of diabetic retinopathy.
• To provide an objective alternative to manual DR assessment, supporting clinical decision-making.

Main Methods:

• Utilized advanced image preprocessing: contrast-limited adaptive histogram equalization and Gaussian filtering.
• Developed a deep learning system comprising feature segmentation, deep learning feature extraction, and ensemble classification modules.
• Validated the system on four public retinal image databases (APTOS 2019, Messidor, DDR, EyePACS).

Main Results:

• The proposed automated system demonstrated promising performance in binary and multi-class DR classification tasks.
• Achieved superior performance compared to existing segmentation methods, CNN architectures, Swin Transformer models, and recent literature.
• Performance was evaluated using nine metrics, including AUC and quadratic weighted Kappa score.

Conclusions:

• The automated system offers superior performance and accuracy for diabetic retinopathy screening.
• Provides reliable support to clinicians, reducing reliance on subjective interpretations.
• Contributes to more consistent and dependable diabetic retinopathy evaluations.