RESUMO
AIMS: Diabetic retinopathy (DR) becomes a complicated type of diabetic that causes damage to the blood vessels of the retina's light-sensitive tissue. DR may initially cause mild symptoms or no symptoms. But prolonged DR results in permanent vision loss, and hence, it is necessary to detect the DR at an early stage. METHODS: Manual diagnosing of DR retina fundus image is a time-consuming process and sometimes leads to misdiagnosis. The existing DR detection model faces few shortcomings in case of improper detection accuracy, higher loss or error values, high feature dimensionality, not suitable for large datasets, high computational complexity, poor performances, unbalanced and limited number of data points, and so on. As a result, the DR is diagnosed in this paper through four critical phases to tackle the shortcomings. The retinal images are cropped during preprocessing to reduce unwanted noises and redundant data. The images are then segmented using a modified level set algorithm based on pixel characteristics. RESULTS: Here, an Aquila optimizer is employed in extracting the segmented image. Finally, for optimal classification of DR images, the study proposes a convolutional neural network-oriented sea lion optimization (CNN-SLO) algorithm. Here, the CNN-SLO algorithm classifies the retinal images into five classes (healthy, moderate, mild, proliferative and severe). CONCLUSION: The experimental investigation is performed for Kaggle datasets with respect to diverse evaluation measures to deliberate the performances of the proposed system.