Coffee cultivation is of extreme economic importance in many regions of the world. However, diseases and pests pose serious challenges, significantly affecting productivity. To solve this problem, deep neural network techniques are emerging as promising solutions, offering precision and efficiency in identifying plant leaf pathologies under different environmental conditions. This study proposes the analysis of a two-stage methodology, detecting the diseased regions of coffee leaves and classifying the diseases into Miner, Rust, Cercospora and Phoma. The experiments were conducted using two public datasets with a total of 1747 images of Arabica coffee leaves. The complete dataset was used for the detection stage and a subset of data with 4104 cropped images of the diseased region of the leaves was generated for the classification stage. The early stopping technique was used to train the models with a patience of 20 and a total of 300 epochs. The YOLOv8 model was chosen to detect the affected regions on the leaves due to its established real-time detection capability and low computational cost. After detection, the clipped regions of interest were submitted to the InceptionResNetv2, DenseNet169 and Resnet50 models, which are state-of-the-art methodologies used for disease classification. The results show that YOLOv8 obtained an mAP of 85.1% and, for classification, the InceptionResNetv2 model obtained the highest average accuracy with 98.18%, which can be seen in the robustness of this architecture compared to the others. The use of the two-stage methodology makes it possible to optimize each stage separately, making it easier to adjust other architectures for new types of diseases or plants.