An Attention-Based Swin U-Net-Based Segmentation and Hybrid Deep Learning Based Diabetic Retinopathy Classification Framework Using Fundus Images

Diabetes mellitus frequently occurs with diabetic retinopathy (DR), which results in lesions mostly in the retinas that damages the vision. Earlier DR identification and treatment can deeply lower the threat of visual loss. Compared to computer-aided diagnosis methods, the manual diagnosing of DR by ophthalmologists using fundus images of the retina leads to high-cost consumption, time consuming, and more effort. This study employs an attention-based deep learning framework for classifying DR using fundus images, and the severity level is estimated to provide better treatment. The required fundus images are obtained through real-world benchmark datasets. Then, the attained images are pre-processed to remove unnecessary distortions in the images. Subsequently, the resultant pre-processed image is sent through the attention-based swin U-net for segmentation, where the affected diseased part is clearly visualized using the segmentation process. Here, the hybrid optimization algorithm named spiral bacterial colony optimization (SBCO) is utilized to optimize the features within the U-net. Furthermore, fundus images are classified using deep structures, where the segmented image is directly undergone to the 3-dimensional convolutional neural network (3D-CNN), and the optimally selected features from the segmentation stage are applied to the dual attention recurrent neural network (DA-RNN). For instance, the parameters inside the deep structures are optimized via the same developed SBCO. Finally, the severity level of the DR is calculated. Several experimental observations are done to evaluate the effectiveness of the developed deep structure-based DR detection approach, considering numerous performance metrics.