Colorectal cancer is the third most common type of cancer diagnosed for men and the second most for women. Today's main methods of examination are expensive, time consuming and intrusive for the patient. Recent technologies, such as CAD and ACD, aims to increase automation in the screening and examination processes. CAD could aid medical professionals during examinations by providing a second opinion, while ACD could be used to screen entire populations, and thus relieving pressure on the health care system. In recent years, neural networks have gained traction among researchers in topics regarding recognition, and we believe it can be utilized in these automated systems. In this thesis, we examine the performance of neural networks for polyp detection. We also explore how data enhancement affect the training and evaluation of the networks, and if it can be used to increase the polyp detection rate. Finally, we experiment with how various training techniques can be used to increase performance. We conclude that neural networks are suitable for polyp detection. We show how data enhancement and training optimization can be used to increase different aspects of the performance. We discuss what aspects are suitable for different scenarios. At the end, we also discuss how our system can be used to detect polyps per frame, per sequence and per polyp, and what the results of our system look like using the different metrics. Detection per frame can be considered a computer science viewpoint, while detection per sequence or per polyp is more of a medical field viewpoint.