Deep architectures for the segmentation of frontal sinuses in X-ray images: Towards an automatic forensic identification system in comparative radiography

Comparative radiography (CR) is the forensic anthropology technique in which ante-mortem (AM) and post-mortem (PM) radiographic materials (e.g., X-ray images or CTs) are compared in order to determine the identity of a human being (either by providing a positive identification or by reducing the number of candidates). One of the most commonly used anatomical structures in CR are the frontal sinuses, which are used in forensic human identification tasks due to their singularity and high identification power. In order to automate the comparison of frontal sinuses in AM and PM materials, we need to perform their segmentation and registration. However, these tasks are time-consuming, subjective, and complex. This paper presents the first CR-based identification system for the comparison of frontal sinuses and supporting the forensic expert's decision. The proposed system comprises two building blocks. First, a frontal sinus segmentation method in X-ray images using deep convolutional neural networks. Different convolutional architectures are compared in solving the segmentation problem, obtaining an average Dice Similarity Coefficient of the frontal sinuses of 0.823 on a dataset composed of 234 skull radiographs. Second, an evolutionary-based 2D-3D IR method, that searches for the best alignment of segmented AM and PM images using a real-coded evolutionary algorithm. The proposed system is evaluated on a real multiple-comparison identification scenario including 10 X-ray images and 10 CTs, where manual and automatic segmentation approaches are compared. The global results shows that the proposed system is able to filter 50% of the sample. These preliminary results suggest that our system can reliably keep the true positive identity in the first half of the sample, allowing for a significant reduction of forensic experts' workload and shortening identification times. (C) 2021 Elsevier B.V. All rights reserved.