Predicting Diabetic Retinopathy Severity with Deep Learning: A Survey of Fundus Image Analysis Technique

Diabetic retinopathy (DR) is a significant complication of diabetes mellitus, impacting vision due to retinal abnormalities. Early detection and precise severity assessment are crucial for effective management. Leveraging deep learning techniques and image preprocessing methods, this paper proposes a comprehensive approach to DR classification. Utilizing publicly available datasets like EyePACS, Messidor-2, APTOS, and DDR, preprocessing steps including Gaussian blurring and data augmentation are employed to enhance image quality and address class imbalance. Wavelet decomposition is used for feature extraction to capture multi-resolution information from fundus images. Transfer learning with ResNet variants, coupled with regularization techniques, aids in model generalization. A modified ResNet50 architecture is introduced, featuring custom fully connected layers and additional convolutional layers for improved feature extraction. The model aims to classify retinal diseases into four severity levels: normal, mild, moderate, and severe proliferative. The survey aspect delves into preprocessing methods' effectiveness in improving CNN performance for medical image analysis, specifically in DR detection. The applicability of transfer learning in medical imaging tasks, particularly in DR, is also explored. This study contributes to advancing medical image analysis for DR diagnosis and classification, addressing the critical need for efficient detection and management of this debilitating condition.