Early Detection Of Diabetic Retinopathy Using Optimized Deep Learning Framework.

Diabetic Retinopathy (DR) is a serious eye complication of diabetes that can lead to
blindness if not detected early. Timely diagnosis of DR is crucial to prevent or delay vision loss.
This paper presents a deep learning-based approach for the early detection of DR. The
methodology includes data collection, augmentation, preprocessing, segmentation, feature
extraction, feature selection, and classification. Data augmentation techniques such as rotation,
flipping, and rescaling are applied, followed by preprocessing steps like Gaussian filtering,
contrast normalization, and resizing. The proposed framework uses an Enhanced U-Net model
for precise retinal segmentation. For feature extraction, methods like Scale-Invariant Feature
Transform, Local Ternary Patterns, and Haralick features are employed to capture detailed
descriptors of retinal structures. Feature selection is optimized using a hybrid Particle Swarm
Optimization and Whale Optimization Algorithm, reducing dimensionality and retaining
significant features. Classification is performed using a fine-tuned Residual Neural Network. The
proposed method achieves high performance, with a precision of 0.974, accuracy of 0.968,sensitivity of 0.971, specificity of 0.955, and F1-score of 0.970. The Matthews Correlation
Coefficient is 0.962, with false-negative and false-positive rates of 0.029 and 0.021, respectively.
Comparative analysis with state-of-the-art models demonstrates the effectiveness of the proposed
approach in providing reliable DR detection. The framework, implemented in Python, offers a
scalable and efficient solution for early DR diagnosis in clinical settings.
Keywords: Diabetic Retinopathy, Deep Learning, Enhanced U-Net, Feature Selection, Hybrid
Optimization, Residual Neural Network, Early Diagnosis, Medical Image Analysis.