Empirical analysis on retinal segmentation using PSO-based thresholding in diabetic retinopathy grading.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Diabetic retinopathy (DR) is a leading cause of blindness, necessitating early diagnosis.
• Deep learning models aid DR diagnosis but can be hindered by background pixels in retinal images.
• Automated feature extraction and grading are crucial for efficient DR detection.

Purpose of the Study:

• To propose an auto-thresholding algorithm for retinal segmentation to mitigate background pixel impact.
• To enhance the accuracy and efficiency of diabetic retinopathy grading using deep learning.
• To analyze the importance of retinal segmentation using Explainable AI (XAI).

Main Methods:

• A Particle Swarm Optimization (PSO)-based thresholding algorithm was developed for retinal segmentation.
• The proposed algorithm was compared against Otsu, histogram-based sigma, and entropy thresholding methods.
• ResNet50 was utilized for grading accuracy evaluation, with XAI employed for feature importance analysis.

Main Results:

• The PSO-based retinal segmentation approach significantly improved accuracy compared to non-segmented methods.
• The model achieved a substantial accuracy of 83.70% on unseen data from the IDRiD fundus dataset.
• Explainable AI confirmed the impact of segmentation on model performance.

Conclusions:

• The proposed PSO-based auto-thresholding effectively determines optimal threshold values for retinal segmentation.
• This approach demonstrably improves the accuracy of diabetic retinopathy grading models.
• The study highlights the importance of accurate retinal segmentation for reliable DR diagnosis.