Diabetic retinopathy (DR) is one of the main global causes of preventable blindness. Its initial sign is red lesions, a word that refers to both hemorrhages (HEs) and microaneurysms (MAs). In typical clinical practice, doctors manually identify these lesions using fundus images. This is a difficult, time-consuming, and effort-intensive task because of the small size and lack of contrast in the lesions.  Hand-crafted feature extraction techniques, including Gabor wavelets, Local Binary Patterns (LBP), and Histogram of Oriented Gradients (HOG), were employed with the support vector machine (SVM) method for classification. Deep learning feature extraction techniques were employed using 16 pre-trained neural network feature extractors through transfer learning. The novelty of this study lies in the utilization and comparison of both hand-crafted and deep learning feature extraction approaches for diabetic retinopathy detection in eye fundus images. This study also explores the effectiveness of hand-crafted feature extraction techniques, which are less computationally expensive and easier to interpret. The study found that both hand-crafted and deep learning feature extraction techniques are effective for diabetic retinopathy detection. ResNet101 was discovered to be the best pre-trained neural network, achieving an accuracy of 95% and an area under the curve (AUC) of 96.0%.  Overall, the study's contributions include the development and evaluation of various CAD systems for diabetic retinopathy detection, insights into the effectiveness of different feature extraction techniques and classification methods, and potential improvements to traditional diabetes diagnosis methods.