Manipulation of biomedical images to misrepresent experimental results has plagued the biomedical community for a while. Recent interest in the problem led to the curation of a dataset and associated tasks to promote the development of biomedical forensic methods. Of these, the largest manipulation detection task focuses on the detection of duplicated regions between images. Traditional computer-vision based forensic models trained on natural images are not designed to overcome the challenges presented by biomedical images. We propose a multi-scale overlap detection model to detect duplicated image regions. Our model is structured to find duplication hierarchically, so as to reduce the number of patch operations. It achieves state-of-the-art performance overall and on multiple biomedical image categories.

MONet: Multi-scale Overlap Network for Duplication Detection in Biomedical Images