Steganalysis of Medical Radiographs for Radiographic Machine Identification

Large online databases with radiographs from different institutions are being increasingly used in biomedical research consortiums. Radiographs from one machine at a single site may be suboptimal or corrupted. Steganalysis can be used to quality assurance of radiographs. Here, we use a deep learning framework for source identification of radiographs by predicting the exact radiographic machine (manufacturer and model) using a single rich model. A convolutional neural network architecture is applied to extract high-level content-free features, and three fully connected neural network layers are used to predict the radiographic machine source of the radiographs. Potential change in pixel information in medical images can be detected using steganalysis. Steganalysis contributes to maintaining trust in medical systems and ensures the accurate diagnosis and treatment of patients. Patients with cervical (n=2028 patients; 3905 radiographs) and chest (n=1499 patients; 2725 radiographs) radiographs obtained at Mayo Clinic from 01/01/2010 to 12/31/2021 were analyzed. Data was randomly split by patient into training/validation (n=80%) and test (n=20%) datasets respectively. The accuracy in the test dataset was 99.50% (AUC=99.72%) and 96.86% (AUC=98.23%) for cervical and chest radiographic machine identification respectively.

I. Introduction

Deep learning (DL) is utilized more commonly in biomedical and radiology imaging research studies today. These deep learning studies typically require a training/validation dataset, an internal test dataset, and an external dataset [1]. The external dataset is typically a rich and diverse dataset with images collected using different devices (manufacturers and models), from different geographic sites / institutions and from patients with many different sociodemographic factors to ensure that the models created are externally valid and will perform in clinical practice. Imaging repositories including: the Radiological Society of North America (RSNA), National Institutes of Health (NIH) Image Gallery, European Radiology Repository/the European Society of Medical Imaging Informatics (EuSoMII), Medical Imaging Data Resource Center (MIDRC), and the Cancer Imaging Archive (TCIA) have large numbers of radiographs collected on different patients from different institutions across different sites. The original source of the radiograph is often removed, although the institution where the radiograph originated is known. However, in the quality assessment of radiographs for analysis, radiographs from one radiographic machine may be systematically different from the other radiographic machine, and have to be removed from the analysis. Several factors impact the quality and integrity of radiographs, and can be classified into external and internal facts. External factors include variations in input voltage and configuration settings. Internal causes encompass device hardware malfunctions, and software errors in preand post-processing calculations. Both internal and external factors influence the type and quality of radiographic images. Source identification of images is becoming more important since contemporary technology makes it extremely simple and quick to generate convincingly corrupted, low-quality images due to some changes in configuration, hardware inconsistency, tools dependencies, and outdated / deprecated libraries. The importance of source identification of digital radiographs for quality assurance and integrity is known [2]. Corrupted / low quality radiographs have significant implications for research, medical diagnosis and prognosis. Developing a method to identify the source of radiographs will provide a variety of significant benefits [2], [3]. Identifying the source of the medical images might enable researchers to eliminate radiographs that are systematically different from other radiographs which would change the results of the study [4].