The use of pre-trained, widely known deep learning models for object detection and classification in medical images is beneficial across a wide range of applications. For coronary artery disease, deep learning-based methods have been implemented to segment the arteries and detect stenosis with promising results. However, how to automatically filter and select valid frames from entire video sequences remains underexplored, as this is usually a manual task. This step, also known as image screening, consists of selecting frames in which the radiocontrast has perfused to achieve sufficient contrast quality for viewing the coronary arteries correctly. This task is crucial for preparing datasets for detector and segmentation models. In this work, an open-access dataset of invasive coronary angiographies (CADICA) has been used to evaluate the impact and feasibility of automatic screening using three architectures: CNN and transformer-based. Additionally, stratified cross-validation and Bayesian hyperparameter optimization have been used to obtain robust and reliable results. The results indicate that a transform-based architecture outperforms CNN-based models for screening key frames, demonstrating great potential as a component of automated ICA image analysis.