Automatic segmentation of skeletal structures in X-ray images using deep learning for comparative radiography

Abstract

• Comparative radiography is a forensic human identification technique. • Comparative radiography requires the segmentation of radiographs. • We use deep learning to automate the segmentation of radiographs. • We validate our method using chest and head radiographs. • Our proposal serves as computer-aided decision support system to filter candidates. INTRODUCTION: Comparative radiography is a forensic identification technique based on the comparison of skeletal structures in ante-mortem and post-mortem radiographic data to determine the identity of a deceased person. Several works have tackled its automation using different approaches but all of them require the manual segmentation of the skeletal structure’s silhouette. MATERIALS AND METHODS: The radiograph segmentation task has been automated using convolutional neural networks. We have developed a deep network able to accurately segment the skeletal structure of interest within a radiograph. It requires only 200 labelled radiographs to be trained, and has been applied to two problems: (1) the segmentation of clavicles in chest radiographs using the JSRT dataset; and (2) the segmentation of frontal sinuses in head radiographs provided by the Hospital de Castilla-La Mancha (Spain). RESULTS AND DISCUSSION: We achieve human-competitive performance in the segmentation of clavicles in chest radiographs (average Dice Similarity Coefficient 0.939) and high-quality segmentation results in the segmentation of frontal sinuses in head radiographs (0.823). The automatic segmentations of frontal sinuses obtain similar results to manual ones for the decision-making task. Specifically, both manual and automatic segmentation allows 50% of the sample to be filtered. In fact, the positive match is always located among the best first 5 matches provided by our system. CONCLUSIONS: This automatic segmentation framework comprises a first step towards a computer-aided decision support system in comparative radiography, where the resulting segmentation is employed in an image registration pipeline as part of the decision-making process.