Background <p>Diabetic retinopathy (DR) is a progressive diabetic complication that may result in irreversible vision loss and blindness in both the young and adult population. DR should be diagnosed early in order to avoid visual impairment. Nevertheless, the process of hand-grading of fundus images by ophthalmologists is time-consuming and they might not detect tiny retinal lesions, making it imperative to have the right automated diagnostic systems.</p> Objective <p>This study aims at creating an automated system to identify small lesions and categorise the stages of diabetic retinopathy severity given fundus images.</p> Methods <p>An improved histogram equalization method increases the fundus images at the regions of interest. A deep Residual SqueezeNet model is used to segment the regions of lesions. The areas where these lesions are segmented are then used to extract features using the Slime Mould Algorithm. Lastly, the Support Vector machine is used to classify the stages of diabetic retinopathy severity and compare them with other available classifiers. The IDRiD and EyePACS datasets are used to test the proposed framework.</p> Results <p>The proposed framework had 99% accuracy, 100% precision, and 100% recall using strict fivefold cross-validation with no overlap between training and test data on the IDRiD dataset. It also attained 99% accuracy, 100% precision, and 100% recall using strict fivefold cross-validation on the EyePACS dataset, proving better than the currently used classification methods.</p> Conclusion <p>The experiment findings suggest that the proposed automated system is very helpful in recognizing the presence of small lesions in the retina and correctly identifying the stages of diabetic retinopathy severity.</p>