Deep-RSI: Deep learning for Radiographs Source Identification

In forensics, the authenticity of digital images is of the utmost importance, considering that modern technology makes it incredibly simple and quick to alter and generate fake but convincing images. As a result, digital image credibility has decreased, making it difficult to demonstrate the source of images. Prior studies have shown that magnetic resonance imaging (MRI) scans can be traced back to their sources, but radiographs have not. In this paper, we propose the Deep-RSI algorithm, an algorithm that identifies the source (manufacturer and model) of the device used to create radiographs. This is the first time that a medical forensics investigation of this kind will be accomplished to declare and confirm radiograph sources. Researchers in information forensics, security, and medical imaging can use this data to determine scientific fraud, like fake radiographs made from unreliable sources or cut-and-paste fakes. This proposed solution describes how non-content pixels in images enable us to discover the manufacturer and model of a radiographic machine. Since radiographs are obtained from different sites of the body, source recognition has to be sensitive and free of any content-specific information. This will prevent the convolutional neural networks (CNN) from detecting content-specific details and instead identify fingerprints that are unique to the source. CNNs start with low-level features and, in the convolutional blocks, generate high-level features to identify the radiographic machine sources. This proposed solution reports the source (manufacturer and model) of each image. We obtain the highest AUC of 0.97 and a prediction accuracy of 98.54% for radiographic machine manufacturer detection. Our results show that forensic assessments of radiographs can be done with a high level of certainty.