Enhanced multi-grade diabetic retinopathy detection and classification via ensembled deep learning model from retinal fundus images

Diabetic retinopathy (DR) represents a primary cause of vision impairment, highlighting the importance of early and precise detection to reduce its advancement. This study presents DiaRetULS-Net, an innovative Ensembled model developed for the automated detection and classification of diabetic retinopathy severity utilizing retinal fundus images. The proposed methodology utilizes advanced preprocessing techniques, such as Contrast Limited Adaptive Histogram Equalization (CLAHE) for image enhancement, alongside robust feature extraction methods including Discrete Wavelet Transform (DWT) and Local Binary Patterns (LBP) to effectively capture essential frequency and texture-based features. The DiaRetULS-Net architecture combines U-Net for accurate segmentation of retinal abnormalities, the Liquid Time Constant Neural Network (LTCN) for the extraction of dynamic spatial and temporal features, and a Multi-Class Support Vector Machine (SVM) for precise classification of diabetic retinopathy severity levels. The model was assessed using the Messidor-2 dataset and a 5-fold cross-validation approach, resulting in notable performance metrics: 98.83% accuracy, 98.87% specificity, and 99.21% sensitivity. Comprehensive analyses, such as the Receiver Operating Characteristic (ROC) curve, confusion matrix, and error histogram, substantiate the model’s reliability and efficiency.