This work is framed in the context of the project ’automatic labeling of ultrasound images’ developed by the EveryWere Lab of University of Milan. The goal is to detect the presence of effusion in ultrasound images of knees in individuals affected by hemophilia, a genetic disorder related to impairment of blood clotting mechanisms. Patients cannot always receive medical attention when suffering a contusion, so that medication is prescribed based on how the patient feels, possibly resulting in patients not receiving medication when needed, or in undue administration. Thus, automating the way in which a pathological knee is identified would enhance the process of treatment prescription, both providing better healthcare to patients suffering from this disorder, and allowing for a more efficient allocation of resources in the sanitary system. Here, we address the issue of anomaly detection by adopting an approach based on image reconstruction by impainting. An autoencoder-based convolutional network is trained to reconstruct anomaly-free images, where some pixels have been masked. Then, the trained network –which has learnt how to properly reconstruct healthy knees – is fed with ultrasound images of both normal and anomalous knees, and the output image is compared with the original one. The idea is that the network will use the context information from pixels around the impainting mask to recreate an image without anomalies, so that the reconstruction error will be particularly high for anomalous instances, allowing for their identification.