Diabetic Retinopathy Recognition: A Comparison of Traditional Image Processing and Deep Learning Approaches

Diabetic retinopathy (DR) is a leading cause of blindness worldwide, affecting a significant portion of the diabetic population. According to a study by the International Diabetes Federation (IDF) [1], individuals with diabetes for over 10 years face a staggering 60% risk of developing retinopathy. Recognizing the severity of this condition, extensive research has been dedicated to developing accurate diabetic retinopathy recognition methods. These methods can be broadly categorized into two primary approaches: traditional image processing and deep learning models.

Traditional image processing techniques for DR recognition involve extracting specific features from retinal fundus images. These features typically include microangiomas, small hemorrhages, hard exudates, cotton wool spots, and other lesion areas. Once extracted, these features are used to train classifiers that can differentiate between healthy and DR-affected images [2]. While traditional methods have demonstrated some success, they encounter difficulties due to the inherent complexity of retinal images. The intricate internal structure of these images makes it challenging to accurately extract pathological features.

Furthermore, traditional methods heavily rely on prior knowledge for lesion segmentation, requiring researchers to possess significant expertise in identifying and isolating relevant features. This dependence on prior knowledge poses a barrier for general researchers, limiting their ability to comprehensively capture all characteristics of the lesion area.