Estimating the navigational risk of vessels operating in sea and waterway areas is important for waterway risk management and pollution preparedness and response planning. Existing methods relying on a model-informed expert judgment of ship-ship collision risk are of limited practical use because periodic risk monitoring is feasible only when this can be done without extensive use of organizational resources. To alleviate such limitations, this article presents a new approach based Convolutional Neural Networks (CNNs) and image recognition to interpret and classify ship-ship collision risks in encounter scenarios. The specific aim of the article is to investigate whether a CNN-based model can quickly and accurately interpret images constructed based on data from the Automatic Identification System (AIS) in terms of collision risk. To test this, estimates derived from training data are compared to validation data. It is also investigated whether adding additional navigational information based on AIS data improves the model's predictive accuracy. A case study with data from the Baltic Sea area is implemented, where various model design alternatives are tested as a proof-of-concept. The main finding of this work is that a CNN-based approach can indeed meet the specified design requirements, suggesting that this is a fruitful direction for future work. Several issues requiring further research and developed are discussed, with the validity of the risk ratings underlying the image classification seen as the most significant conceptual challenge before a CNN-model can be put to practical use.