Deep learning models have revolutionized artificial intelligence, most especially in the domain of medical image processing. Nevertheless, the efficiency of these models depends much on hyper parameter modification. The choice of ideal hyperparameters significantly affects computing cost, training efficiency, and model accuracy. This paper addresses the application of four well-known hyper parameter optimization methods Grid Search, Random Search, Bayesian Optimization, and Hyperband to a Custom Convolutional Neural Network (CNN) model for Diabetic Retinopathy (DR) classification. Using the diabetic retinopathy dataset which comprises 3590 retinal fundus images split into five severity levels the study The recommended approach calls for hyperparameter tuning, CNN architectural design, preprocessing of datasets, and model evaluation with a validation accuracy of 99.11%, Hyperband boasts the best efficiency and performance among various optimization techniques. The results show that automated hyper parameter tuning techniques enhance the performance of deep learning models and ought to be applied in next medical picture classification problems.