Many methods for estimating remaining useful life (RUL) for predictive maintenance rely on sensor data. In environments where sensors might be economically unviable the condition of a machine or its parts is often assessed visually. Wear-and-tear is difficult to quantify and therefore makes it challenging to build statistical models for RUL estimation. This study investigates whether accuracy in RUL estimation models of wear-and-tear parts can be improved using inspection pictures and artificial neural networks. Using a combined approach of design science research and a case study, RUL estimation models for bearings of ABB turbochargers were evaluated. First, a benchmark model was built that relies only on historical data of turbocharger inspections in tabular form. Then this model was concatenated with various convolutional neural networks such as ResNet50 trained on images of the same inspections. The analysis shows that this concatenating leads to a reduction of the mean absolute error in RUL prediction by up to 18.4%. I conclude that using convolutional neural networks and inspection pictures improves the overall RUL estimation for wear-and-tear parts.