Automatic segmentation of facial soft tissue in MRI data based on non-rigid normalization in application to soft tissue thickness measurement
Introduction
Knowledge about facial soft tissue thickness is important for the forensic facial reconstruction, head anthropometry, head models, surgery planning, skull stripping, defacing, soft tissue thickness measurements and for academic purposes. Currently, the most popular methods for measuring the thickness of soft tissue are ultrasonography, manual measurement in computed tomography (CT) and MRI images based on expert knowledge [1] or opaque markers placed on the surface of the face, visible in MRI imaging [2]. The most promising are MRI based methods due to the lack of direct contact, high spatial resolution and good contrast in soft tissue. It is the preferred modality for anatomical imaging of head structures because it does not involve exposure to radiation. This allows to build large soft tissue database. The main disadvantage of MRI is a weak signal from bone structures and difficulty during separation of the bone and the air segments.
Segmentation of tissues in the human head has been extensively researched over the last decade. The most important issue is the correct identification of head tissues such as brain, skull, soft tissue and air cavities. Other requirements for head segmentation method are full automation, low consumption of time and resource. Accurate segmentation of the head tissues is important for the construction of patient-specific models of complete head [3]. Such models are used for numerical simulation of the behavior of electromagnetic fields in tissues, for the study of electrical activity in the cerebral cortex in electroencephalography and transcranial electric stimulation, as well as their magnetic equivalents – magnetoencephalography (MEG) and transcranial magnetic stimulation [4,5].
There are several approaches to segmentation of the human head in MRI data that often were a by-product of classification techniques designed to categorize head tissues:
- -
Markov random field used to classify MRI data into gray matter (GM), white matter (WM), cerebrospinal fluid (CSF), skull, and background [6];
- -
thresholding and Laplacian of Gaussian operations on successive transverse slices to detect the skin/skull and skull/brain boundaries [7];
- -
neural networks-based methods to segment skull and brain from successive T1-weighted MRI slices [8,9];
- -
thresholding and region growing to segment skull in MRI volumes [10];
- -
simple thresholding and manual segmentation aimed at detecting the boundaries of skull and scalp in order to create an average head model for forward modeling in MEG [11];
- -
deformable model to segment skull from MRI images [5];
- -
thresholds and sequences of morphological operations for skull and scalp segmentation [3];
- -
using CT information for segmentation of skull in coregistered MRI image of the same subject [12,13].
Multimodal approaches are also implemented where information from both CT and MRI scans is used for accurate segmentation of both, the skull and soft tissue. However, these approaches are not practical since simultaneous CT and MRI examinations are rarely recommended. Multiple MRI acquisitions can be used for head segmentation. Proton-density-weighted (PD-weighted) MRI was found more suitable for identification of CSF-skull boundary than T1-weighted MRI [14].
Automatic methods of head segmentation based on deformable models (DM) and tissue probability maps (TPM) are the most promising. The use of 3D deformable surface model can potentially improve efficiency and robustness and ensures global smoothness and coherence of the rendered surface. The main problem of the DM technique is the result being highly sensitive to the initialization of the deformable model [15]. A hybrid 3D image segmentation method which combines region growing methods and DM to obtain accurate and topologically preserving surface structures of anatomical objects was described [16].
Another promising approach is the probabilistic method, where the automatic head segmentation is based on fitting the TPM to an individual MRI image, so the similarity of both images being observed is maximized [4]. However, the results of segmentations acquired by the TPM method still have morphological errors (segments are discontinuous, and some tissue voxels erroneously fall into other segments, e.g. in skull and air segments) and the tissue boundaries are unnecessarily rough making subsequent meshing and forward modelling intractable [4]. Apart from that, some regions of the skull (e.g. ocular globes, nasal bones, portions of the upper region of the skull), which are thin compared to the resolution of MRI and other bones could be poorly segmented, with holes and “disconnected” voxels. Currently, the most popular and globally accessible is the probabilistic head atlas developed for the imaging data software toolkit Statistical Parametric Mapping (SPM) [17]. The atlas consists of six average-shaped tissue probability maps representing brain, skull, soft tissue and background. Unfortunately, this atlas covers a limited field of view (FOV) excluding the neck and jaw. A very detailed anatomical model of the human head and neck based on T1, T2, PD-weighted MRI data, magnetic resonance angiography data and diffusion tensor imaging was also presented [18]. However, no probabilistic atlas based on this model was developed.
The majority of the currently available algorithms are focused on segmentation in standard FOV which does not cover the whole head (only the brain area) [3,16]. The authors of [4,19] proposed fully automatic whole head segmentation. Huang et al. [4,19] improved the TPM and the Unified Segmentation algorithm [20] implemented in the SPM and proposed a rigorous Bayesian inference framework combining image intensity model, anatomical prior and morphological constraints using Markov random field. However, according to the visual assessment of results obtained using these methods [4,5] the accuracy of skull and soft tissue segmentation may be insufficient for our purposes. The segments obtained may be excessively smoothed, resulting in a loss of important anatomical features, e.g. nasal bone, nasal spine, maxilla, mandible usually may be deformed (Fig. 5 in [4], Fig. 6 in [19]).
Therefore, we propose to use patient-specific models of a complete head for the measurement of the thickness of the soft tissue. Our main goal was to develop universal method for soft tissue segmentation of complete head using MRI images. Currently most attention is paid to the segmentation and use of the brain or skull segments, whereas the soft tissue segment is usually obtained as an extra result of secondary importance. A very important feature from the point of view of collection of complete data on the facial soft tissue is that the approach proposed here is applied to the entire head image, thus does not limit the FOV, as is in the case of majority of other applications. The method is expected to enable a new application of patient-specific models and elastic matching to the measurement of the soft tissue thickness, with a perspective of collection of massive experimental data and creation of database of facial tissue thickness.
The algorithm of the soft tissue measurement was described in our previous work [21]. For accurate measurement of the soft tissue thickness, we need to define craniometric and anthropometric landmarks located on the surface of skull and face. During the tests, we noticed that after automatic definition, some landmarks were erroneously located inside the tissue or in the air. We therefore proposed a correction requiring knowledge of the surface of the skull and face. Our intention is to use the brain, skull and air segments to obtain a reliable inner border of the soft tissue segment by masking it using these segments. The skull segment is also needed for the verification of the positions of the craniometric landmarks. Thus, segmentation of the brain, skull and air cavities was performed. The segmentation algorithm is based on nonlinear normalization of the MRI template to MRI data and application of the transform matrix to the head model.
Section snippets
Materials and methods
MRI images from the public IXI database containing T1, T2 and PD-weighted images from healthy human subjects [22] were used in this study. Each record is annotated with race, sex, age, height and weight. Only T1-weighted MRI images were used, because only this modality contains complete head data. Spatial resolution of each record is 0.9375 mm × 0.9375 mm × 1.2 mm and corresponds to 256 × 256 × 150 matrix.
The proposed segmentation algorithm (see Subsection 2.3) uses preprocessed MRI data and
Results
The metrics presented in Table 1 show that the accuracy of the soft tissue segmentation is similar in all spatial planes and that the results are repeatable (see Fig. 2).
The average specificity and sensitivity values calculated for the soft tissue segment are 0.9312 ± 0.0236 and 0.8783 ± 0.0428, respectively (Table 5). Less than 7% of voxels from other structures were erroneously included in the soft tissue segment. Around 13% of voxels were erroneously included elsewhere. The soft tissue
Discussion
The best results of segmentation were obtained for the brain and soft tissue. The errors in the brain segmentation may result from imprecisely defined meninges that envelope the brain, which for the purpose of this study were included into the brain segment. In T1-weighted MRI images the border of the dura mater and the skull is not very sharp, therefore sometimes the dura mater can be erroneously classified as a part of the skull segment. The errors in the soft tissue segmentation could be
Conclusions
The proposed method of head segmentation is automatic, consuming little time and resources with limited influence of the operator on segmentation results, as only the head model is developed with the intervention of the operator. The skull, brain and air segments are only needed to obtain the inner border of the soft tissue segment and for verification of acquired soft tissue segment. The segmented head structures do not contain disconnected voxels. Thus, the method provides complete and smooth
Acknowledgements
This work has been supported with statutory funds of the Institute of Metrology and Biomedical Engineering, Warsaw University of Technology.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (30)
- et al.
Segmentation of the skull in MRI volumes using deformable model and taking the partial volume effect into account
Med. Image Anal.
(2000) - et al.
Segmentation of T1 MR scans for reconstruction of resistive head models
Comput. Methods Programs Biomed.
(1997) - et al.
Deformable models in medical image analysis: a survey
Med. Image Anal.
(1996) - et al.
A combined region growing and deformable model method for extraction of closed surfaces in 3D CT and MRI scans
Comput. Med. Imaging Graph.
(2009) - et al.
Unified segmentation
Neuroimage
(2005) - et al.
Midline facial soft tissue thickness database of Turkish population: MRI study
Forensic Sci. Int.
(2012) - et al.
Use of magnetic resonance imaging to measure facial soft tissue depth
Cleft Palate Craniofac. J.
(2007) - et al.
Segmentation of skull and scalp in 3-D human MRI using mathematical morphology
Hum. Brain Mapp.
(2005) - et al.
Fully automated whole-head segmentation with improved smoothness and continuity, with theory reviewed
PLoS One
(2015) - et al.
Markov random field segmentation of brain MR images
IEEE Trans. Medical Imaging
(1997)
3-dimensional rendering of MR (magnetic resonance) images
IEEE WESCANEX 93 Communications, Computers and Power in the Modern Environment — Conference Proceedings
Tissue segmentation of MRI of the head by means of a Kohonen map
Proceedings of 18th Annual International Conference of the IEEE EMBS
Applying a visual segmentation algorithm to brain structures MR images
First International IEEE EMBS Conference on Neural Engineering
Numerical estimation of skull and cortex boundaries from scalp geometry
Proceedings of the 20th Annual International Conference of the IEEE EMBS
A multi-resolution approach for intracranial volume segmentation from brain images
Med. Phys.
Cited by (2)
Quantitative analysis of facial soft tissue using weighted cascade regression model applicable for facial plastic surgery
2024, Signal Processing: Image CommunicationInter-Subject Variability of Skull Conductivity and Thickness in Calibrated Realistic Head Models
2020, NeuroImageCitation Excerpt :While accurate skull thickness can best be determined using CT (Lillie, Urban, Lynch, Weaver, & Stitzel, 2016), the non-invasive procedure proposed here is based on MRI. Using T1-weighted (T1w) MRI, Gorbenko, Mikołajczyk, Jasionowska, Narloch, and Kałużyński (2020) accurately segmented soft tissues, but with only 67% specificity and 83% sensitivity, the results concerning the skull were rather limited. This is because the low contrast between CSF and cranial tissue makes it difficult to estimate the inner skull boundary from T1w MRI alone.