Accurately detecting threats such as plastic firearms presents a complex challenge in modern security systems due to the difficulty in distinguishing these objects from harmless ones when examined using X-ray scanners. This paper explores CNN architecture and image projection methods to compare systems capable of classifying plastic firearms with high accuracy. The results show that integrating data from three sources (a Stream of Commerce dataset, staged images, and synthetically produced images) was crucial for achieving satisfactory classification performance. We also reveal that to improve accuracy and generalization, it is important to expand the training dataset and explore more advanced neural networks, despite the limitations imposed by available computing power. Future work could include exploring the need for multiple views of the baggage examined and the use of more sophisticated imaging technologies, such as CT scanners, to further improve detection capabilities.