<p>Diabetic Retinopathy (DR) is a diabetes-related eye condition that affects the blood vessels in the retina. Deep Learning (DL) algorithms play a vital role in the detection and grading of DR. This paper proposes a sequential approach for multi class grading of Diabetic Retinopathy, which consists of a preprocessing block, blood vessel extraction and removal block, followed by an ensemble learning block and a weight optimisation block for DR classification. The Nelder-Mead algorithm was applied for the optimisation of weights learned by the ensemble model. We have also experimented with different loss functions to improve the results of multiclass classification. The study used the publicly available MESSIDOR dataset. The images were pre-processed and morphological extraction of blood vessels was performed from the images to learn the disease lesions effectively. The ensemble block for DR classification was formed with an aggregation of simple DenseNet-121 and complex InceptionV3 networks, as the pre-trained Convolutional Neural Networks (CNNs). Different loss functions were applied to the model, weight optimisation, and hyperparameter tuning of the models was performed. The results indicate that the proposed model works better than recent state-of-the-art models. The model was able to achieve 98.62% of accuracy with the application of the Nelder-Mead optimisation algorithm and adjusted focal loss functions. For the interpretability of the results of the model, a Gradient-weighted Class Activation Mapping (Grad-CAM) representation of different classes was produced. Our results indicate that the final layer Grad-CAM representation of Class 3 and Class 4 was better interpretable than Class 0 and Class 1 of the corresponding ground truth images. The classification ability of the models contributed to the development of computer aided automatic detection and classification of DR</p>