Despite the fact that cryo-electron microscopy has become a dominant technique in structural biology, many software procedures still rely on manual input of the user. A current bottleneck of routine data acquisitions is the initial screening setup, where a great amount of time and human effort is spent on the identification of suitable target areas of the sample mesh grid due to ice thickness and contamination. In this thesis a two-pass method based on state-of-the-art Deep Neural Network architectures is presented for detection, classification and ranking of the target areas. A training set comprising 5280 grid squares has been assembled to train a model with five data augmentation techniques and a custom classification partition. Additionally, extensive searches on the hyperparameter spaces have been conducted to optimize the final model performance, which is evaluated by visually analyzing the model attention. These experiments have resulted in a classification model with an accuracy score of 93.14%. The method can also be applied to the identification of grid holes and represents a key step towards fully automated cryo-electron microscopy.