Automatic Comparative Chest Radiography Using Deep Neural Networks

Comparative medical radiography is a scientific identification method that involves directly comparing antemortem (AM) radiographs (e.g., X-rays) to postmortem (PM) radiographs taken of a deceased individual. Forensic anthropologists use chest radiographs for comparative radiography due to their availability through health screening. In this paper, we leverage the power of deep neural networks and expert domain knowledge to create a radiographic identification system based on AM and PM chest radiographs. We compiled a dataset of 5,165 anonymized chest radiographs representing 760 individuals from two databases: NIH Chest X-Ray Database and case files from the Michigan State University Forensic Anthropology Laboratory (MSUFAL). We first manually annotated 3 different regions of interest (ROIs) on the radiographs based on expert domain knowledge, viz., thoracic vertebrae from T1-T5; clavicles; and complete vertebral column. We then explored three families of deep neural networks, each selected for its unique strengths in addressing specific challenges in deep learning for processing these ROIs as well as the entire radiograph. Our experiments reveal several compelling findings: (a) the thoracic vertebrae from T1-T5 results in better recognition performance compared to other regions; (b) Efficient Nets result in better recognition accuracy compared to ResNets and DenseNets; and (c) an ensemble of models based on different networks and different ROIs further improves recognition accuracy. We release the expert annotated MSUFAL dataset and our codebase to advance research in comparative medical radiography for deceased identification. This research marks a significant step forward in forensic radiography, being the first to systematically assess the impact of ROIs for robust human identification.