Steganalysis of Medical Radiographs: a Deep Learning Approach Comparing the Importance of Using Content Pixels and Content-Free Pixels

Content-free pixels, which do not contain actual information, have been used in research studies for passive watermarking of radiographs. However, no studies have compared the effectiveness of content-free pixels to content pixels for watermarking. In this study, we propose a radiograph steganalysis solution that can be used to identify the source of a radiograph and be potentially used to identify whether a radiograph is fake. The solution uses a deep-learning architecture for automating computer-generated fake radiograph detection and compares the performance of passive watermarking using the content-free pixels to that using the content pixels. We use patients who had radiographs of the abdomen, pelvis, and lumbar spine at Mayo Clinic (01/01/2010 - 12/31/2021). The patients (n = 4722, radiographs = 10937) were randomly split into training/validation (80%, n = 3778, radiographs = 8998) and test (20%, n = 944, radiographs = 1939) datasets by patient. We evaluate and obtain the highest source identification for patient level model classification of pelvis (accuracy (ACC) = 98.6%, area under curve (AUC) = 95.34%, precision = 99.11%, recall = 98.6%) in content pixel analysis and for patient level model classification of lumbar spine (ACC = 97.2%, AUC = 97.34%, precision = 97.18%, recall = 97.2%) in content-free pixel analysis. This research confirms that the steganalysis can be performed on content-free and content pixels from radiographs. These results will be valuable for medical forensic and legal communities.