One of the many challenges on the way to developing a reliable, reproducible, and efficient solution for the automated diagnosis and treatment of chronic wounds is their accurate detection in images. In this work, we approach this problem using a deep learning object detection architecture (RetinaNet), being an established single-stage object detector known for working well with relatively small and densely positioned objects. We build upon the MMDetection framework and exploit the pre-trained weights of the ResNeXt-101 backbone, for effectively fine-tuning the model to target the chronic wound detection task. In our experimental study, we utilized a publicly available dataset, following a rigorous multi-fold cross-validation procedure, to investigate the generalization abilities of the wound detectors. The obtained average precision of nearly 0.7 and an average recall of 0.6 show the potential utility of deep learners trained from relatively small yet extremely heterogeneous chronic wound images, capturing not only different characteristics of the wounds, but also a variety of background objects. We believe that our flexible processing pipeline represents an important component in the application of machine learning for patient-oriented medical purposes.