Automatic segmentation of facial soft tissue in MRI data based on non-rigid normalization in application to soft tissue thickness measurement

https://doi.org/10.1016/j.bspc.2019.101698Get rights and content

Highlights

  • Segmentation provides complete segments of the soft tissue, skull, brain and air.

  • Results are repeatable as the accuracy metrics have similar values in all planes.

  • Specificity and specificity of the soft tissue segmentation equal 93% and 87%.

Abstract

For measuring the thickness of soft tissue in magnetic resonance (MRI) images, precise borders between skull and face surfaces should be known. We present an algorithm for segmentation of the human head in T1-weighted MRI images that generates smooth, complete segments of head tissues for further landmarks definition and measurements of the soft tissue thickness of the human head. As a segmentation tool we use an algorithm based on nonlinear normalization of the MRI template to MRI data and application of transform matrix to the head model. The algorithm uses preprocessed subject MRI data and a head model with separate tissue segments. The head model is obtained using a hybrid algorithm and consists of four segments: soft tissue, skull, brain and air. To assess the precision of segmentation, specificity, sensitivity, Dice and Jaccard Similarity Coefficients were computed. The algorithm was tested on MRI images from 10 Caucasian adults from free public database IXI. Specificity of 93% and 98% and sensitivity of 87% and 93% was achieved for soft tissue and brain segment, respectively. Specificity of 67% and 72% and sensitivity of 83% and 62% was achieved for the skull and air segments, respectively.

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)

  • C. Chu et al.

    3-dimensional rendering of MR (magnetic resonance) images

    IEEE WESCANEX 93 Communications, Computers and Power in the Modern Environment — Conference Proceedings

    (1993)
  • S. Congorto et al.

    Tissue segmentation of MRI of the head by means of a Kohonen map

    Proceedings of 18th Annual International Conference of the IEEE EMBS

    (1996)
  • P. Belardinelli et al.

    Applying a visual segmentation algorithm to brain structures MR images

    First International IEEE EMBS Conference on Neural Engineering

    (2003)
  • H.A. Haque et al.

    Numerical estimation of skull and cortex boundaries from scalp geometry

    Proceedings of the 20th Annual International Conference of the IEEE EMBS

    (1998)
  • H. Soltanian-Zadeh et al.

    A multi-resolution approach for intracranial volume segmentation from brain images

    Med. Phys.

    (1997)
  • Cited by (2)

    • Inter-Subject Variability of Skull Conductivity and Thickness in Calibrated Realistic Head Models

      2020, NeuroImage
      Citation 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.

    View full text