Region-Based Steganalysis of Medical Radiographs for Radiographic Machine Identification
Farid Ghareh Mohammadi, Ronnie Sebro · 2023
New advances in artificial intelligence (AI) allow us to fake digital images that are difficult for humans to distinguish from real images, including images used in health care like radiographs. Malware creating forged digital radiographs have the potential to have severe negative repercussions for patients' diagnosis and treatment, therefore there is a need to validate the radiographs' source used in health-care. We address this challenge and propose a region-based steganalysis algorithm using a deep learning framework that identified the region of radiographs which has the most informative pixels and patterns for determining the radiographs' source. The deep learning algorithm uses a convolutional neural network (CNN) with four convolutional layers with different filters followed by three layers of fully connected convolutional neural network (FCNN). We used radiographs of the knees (n = 1418), legs (n = 616), ankles (n = 1290) and feet (n = 1074) of patients at Mayo Clinic (01/01/2010 - 12/31/2021) and identified the radiographs' source (manufacturer). The dataset was randomly split by patient into training/validation (n = 3635, 80%) and test (n = 763, 20%), and after tuning evaluated using only one radiograph for each patient in the test dataset. The algorithm yields a model prediction performance for a region of feet radiographs with 98.06% accuracy (Area Under the Curve (AUC) = 98.56%). This novel research is the first in medical forensic imaging that identifies the content-free region of radiographs that is most informative to determine the radiographs' source. These results will be invaluable for detection of fake radiographs and scientific fraud.