Semantic segmentation of retinal exudates using a residual encoder-decoder architecture in diabetic retinopathy

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss, characterized by retinal exudates.
• Early detection of exudates is crucial for timely treatment and prevention of visual impairment.
• Manual detection of exudates from fundus images is labor-intensive and prone to errors due to low contrast and small lesion size.

Purpose of the Study:

• To compare deep convolutional neural network (CNN) architectures for semantic segmentation of exudates.
• To propose a novel residual CNN with skip connections for improved exudate segmentation accuracy.
• To develop a computer-assisted diagnosis tool for efficient and accurate diabetic retinopathy screening.

Main Methods:

• Implemented and compared various deep CNN architectures for exudate segmentation.
• Developed a residual CNN incorporating skip connections to reduce model parameters and enhance performance.
• Utilized image augmentation techniques to optimize network architecture and improve robustness.
• Evaluated the proposed method on three benchmark datasets: E-ophtha, DIARETDB1, and Hamilton Ophthalmology Institute's Macular Edema.

Main Results:

• The proposed residual CNN achieved high performance across all benchmark datasets.
• Achieved precision scores of 0.95, 0.92, and 0.97.
• Achieved accuracy scores of 0.98, 0.98, and 0.98.
• Achieved sensitivity scores of 0.97, 0.95, and 0.95.
• Achieved specificity scores of 0.99, 0.99, and 0.99.
• Area under the curve (AUC) values were 0.97, 0.94, and 0.96.

Conclusions:

• The developed residual CNN demonstrates robust and accurate semantic segmentation of exudates.
• The proposed method is highly suitable for automated diabetic retinopathy screening.
• This computer-assisted diagnosis approach can significantly aid clinicians in early detection and management of DR.