A computational framework for CT-based craniofacial reconstruction

Abstract

Introduction Estimating the facial outlook from an unidentified skull is a challenging task in forensic investigations. This paper presents the implementation and testing of a fully automatic pipeline for computerized craniofacial reconstruction (CFR). Methods A database of age-, gender- and BMI-annotated full-head 3D cross-sectional CT reference scan images is used. The database is continuously being upgraded with new reference cases. Furthermore, the procedure can be easily extended to include data of different ethnicities when such data become available. Facial and skeletal surfaces are automatically extracted for every image in the database. Reconstruction of the face of an unidentified skull (scanned ‘as is’, not requiring a time-consuming soft tissue removal procedure) is started by warping all skull surfaces in the database to the target skull. The resulting transformations are then applied to the associated facial surfaces resulting in as many warped facial surfaces as individuals in the database. A single consensus facial surface is generated by combining the individual reconstructions taking into account gender, BMI and age matching. By implementing this procedure on a cluster platform, reconstructions can be obtained in less than 10 min, irrespective of the number of reference images, which is considerably shorter than current manual reconstruction procedures. Results Validation was performed using a cross-validation leave-one-out scenario where each image in the database was reconstructed using all other images in the database as reference. The reconstructed and actual facial surfaces are then compared using statistics of intersurface distances. On a database of 72 reference individuals a maximal reconstruction error of 4 mm was obtained.