Discrimination of retinal images containing bright lesions using sparse coded features and SVM

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetic Retinopathy (DR) is a leading cause of vision loss due to microvasculature damage.
• Accurate DR diagnosis relies on identifying retinal lesions like microaneurysms, exudates, and drusen from fundus images.
• Distinguishing between similar-looking bright lesions (exudates, drusen) is crucial for improved screening.

Purpose of the Study:

• To apply sparse coding techniques for classifying retinal images.
• To discriminate between normal retinal images and those with exudates or drusen.
• To evaluate the efficacy of dictionary learning and sparse coding for DR lesion classification.

Main Methods:

• Utilized sparse coding techniques for retinal image classification.
• Employed dictionary learning to extract discriminant features from retinal images.
• Compared the proposed sparse coding method with the Bag-of-Visual-Word approach using a linear SVM classifier.

Main Results:

• Dictionary learning effectively captured retinal image structures, yielding discriminant sparse coded features.
• The sparse coding method outperformed the Bag-of-Visual-Word approach in image classification.
• Achieved high sensitivity and specificity across normal (96.50%/97.70%), drusen (99.10%/100%), and exudates (97.40%/98.20%) classes with an 828-image dataset.

Conclusions:

• Sparse coding techniques provide a robust method for classifying retinal images in diabetic retinopathy screening.
• The proposed approach demonstrates superior performance in discriminating between normal, drusen, and exudates, aiding in early detection.
• This method offers a promising advancement for automated analysis of retinal images, contributing to better vision preservation.