Bayesian segmentation of human facial tissue using 3D MR-CT information fusion, resolution enhancement and partial volume modelling

• Bayesian method provides resolution enhanced segmentation of human head.
• Segmentation classes incllude muscle, bone, fat, air and skin.
• Tests were performed on 3D MR and CT images, as well as registered MR-CT images.
• The most successful results were obtained by the information fusion of MR and CT.
• Free parameters of the algorithm can be adjusted in a more systematic way.

BackgroundAccurate segmentation of human head on medical images is an important process in a wide array of applications such as diagnosis, facial surgery planning, prosthesis design, and forensic identification.ObjectivesIn this study, a Bayesian method for segmentation of facial tissues is presented. Segmentation classes include muscle, bone, fat, air and skin.MethodsThe method presented incorporates information fusion from multiple modalities, modelling of image resolution (measurement blurring), image noise, two priors helping to reduce noise and partial volume. Image resolution modelling employed facilitates resolution enhancement and superresolution capabilities during image segmentation. Regularization based on isotropic and directional Markov Random Field priors is integrated. The Bayesian model is solved iteratively yielding tissue class labels at every voxel of the image. Sub-methods as variations of the main method are generated by using a combination of the models.ResultsTesting of the sub-methods is performed on two patients using single modality three-dimensional (3D) image (magnetic resonance, MR or computerized tomography, CT) as well as registered MR-CT images with information fusion. Numerical, visual and statistical analyses of the methods are conducted. High segmentation accuracy values are obtained by the use of image resolution and partial volume models as well as information fusion from MR and CT images. The methods are also compared with our Bayesian segmentation method proposed in a previous study. The performance is found to be similar to our previous Bayesian approach, but the presented methods here eliminates ad hoc parameter tuning needed by the previous approach which is system and data acquisition setting dependent.ConclusionsThe Bayesian approach presented provides resolution enhanced segmentation of very thin structures of the human head. Meanwhile, free parameters of the algorithm can be adjusted for different imaging systems and data acquisition settings in a more systematic way as compared with our previous study.