To study diseases that can affect apple leaves, such as cedar rust and scab, some datasets are available in the literature. These datasets have been annotated and labeled with the assistance of experts in the field. Two main categories of datasets can be distinguished. The first category involves Lab-built Datasets which are constructed under controlled conditions, often featuring images taken against a compact background that does not accurately reflect real-world environments. Models trained on these datasets may perform well in controlled settings but often face challenges in practical applications. The second category includes datasets collected in realistic field conditions, capturing the variability and complexity of natural environments. The performance of deep learning algorithms can be significantly affected by the differences in image quality and context between these two types of datasets. In this study, two specific datasets were utilized: the Plant Village dataset and the Plant Pathology dataset, focusing on apple leaves categorized into three classes: healthy, rust, and scab. The study employed Convolutional Neural Networks (CNNs) for classification, experimenting with both models trained from scratch and transfer learning models, including VGG16, InceptionV3, MobileNetV2, and Xception. The experimental results indicated that the MobileNetV2 model performed exceptionally well when applied to the Plant Village dataset, which is likely due to its efficiency and ability to generalize from the lab-built images. Conversely, the custom CNN model trained from scratch showed superior performance on the Plant Pathology dataset, which consists of more realistic images, suggesting that it was better suited to handle the complexities of real-world data.