Diabetic Retinopathy (DR) is a potential condition of Diabetes Mellitus (DM) that causes lesions on the retina, reducing vision and perhaps leading to disability if not properly treated. A Special Generative Adversarial Network with Ensemble Classification Regression (SGAN-ECR) model was introduced for the effective categorisation of various DR grades with the generation of high-contrast and low-saturated RF images. However, the Multi-Scale Attention Residual Network with Gradient Boosting (MSA-ResNet-GB) model used in SGAN-ECR requires manually assigning values for many hyperparameters. An inappropriate value of the hyper-parameter leads to an increased error rate. This paper presents a method using the Enhanced Mine Blast Algorithm (EMBA) for selecting optimal hyper-parameters of MSA-ResNet-GB, such as the number of convolution layers, number of filters, filter size, number of Fully Connected (FC) layers and the hidden units in the FC layer to improve the architecture. Essential principles of EMBA are derived from the mine bomb explosion in real-time applications. The initial population of shrapnel fragments represents the initial hyper-parameter, and the computation of their subsequent locations represents the search for the best hyper-parameter. SGAN-ECR with parameter-optimised MSA-ResNet is named SGAN-OECR. In SGAN-OECR, optimised MSA-ResNet is trained by training images, and then the trained model is used to recognise the abnormalities in test images. MLP classifiers in the last FC layers of MSA-ResNet classify the severity of DR lesions. Finally, the experimental results of SGAN-OECR on Kaggle-APTOS and IDRiD datasets have 99.03% and 98.63% of accuracy, respectively, which is higher than the existing DR classification models.