13 Computational Biomechanics of the Brain in the Operating Theatre
329
13.3.1 Before Surgery: Model Generation
As presented in Sect. 13.2.1, the model must be built from pre-operative MRI images
acquired a few hours before surgery. This procedure must then be as automatic as
possible to be compatible with a clinical workflow.
13.3.1.1 Geometries from MRI Images
Several atlas-based automatic algorithms such as [23–25] have provided detailed
solutions, with regard to the resolution of the meshes, but their implementation
is difficult. In our research context, we have chosen a semiautomatic coarse
segmentation method.
Soft tissues are segmented from pre-operative T2-FLAIR (fluid-attenuated inversion recovery) MRI. This sequence reduces the signal from the cerebrospinal fluid,
thus highlighting brain lesions. The complete organ is first extracted using the BET
(brain extraction tool) algorithm proposed by [9]. It is fully automatic and runs
in just a few seconds. The tumour is then segmented, after manually initialising
some seeds, via a growing region algorithm implemented in the ITK-SNAP
software [26]. For a more comprehensive study using our model in clinic, more
advanced segmentation methods, potentially publicly available, should obviously
be investigated.
Soft tissues are meshed homogeneously with linear tetrahedral elements, using
the CGAL library [27], leading to a coarse mesh of 2000 nodes for a single
hemisphere. Its surface is extracted to form the collision mesh that will be used
to impose constraints with the skull and the probe footprint. This collision mesh is
decimated (i.e. about 300 vertices) in order to reduce the number of constraints.
13.3.1.2 Segmentation of the Vascular Tree
The cerebrovascular tree is segmented from pre-operative angiographic MRI
(ARM). While this modality provides poor contrasts for soft tissues, vessels appear
with very high intensities.
A review of vessel segmentation methods from medical images can be found
in [28]. In our work, we reconstructed the cerebrovascular tree using the maximum
intensity projection (MIP) technique proposed by [29]. The histogram of this image
reveals two peaks corresponding to the background of the image (black) and the soft
tissues of the brain (dark grey), respectively. The vessels are then segmented using a
threshold value chosen just after the second peak of the histogram. This is illustrated
in Fig. 13.2a.
The shape of this segmentation, and more particularly its outer envelope, is very
sensitive to the image quality and segmentation parameters. For example, vessels
often appear thicker in US than in ARM. In order to minimise the segmentation
329
13.3.1 Before Surgery: Model Generation
As presented in Sect. 13.2.1, the model must be built from pre-operative MRI images
acquired a few hours before surgery. This procedure must then be as automatic as
possible to be compatible with a clinical workflow.
13.3.1.1 Geometries from MRI Images
Several atlas-based automatic algorithms such as [23–25] have provided detailed
solutions, with regard to the resolution of the meshes, but their implementation
is difficult. In our research context, we have chosen a semiautomatic coarse
segmentation method.
Soft tissues are segmented from pre-operative T2-FLAIR (fluid-attenuated inversion recovery) MRI. This sequence reduces the signal from the cerebrospinal fluid,
thus highlighting brain lesions. The complete organ is first extracted using the BET
(brain extraction tool) algorithm proposed by [9]. It is fully automatic and runs
in just a few seconds. The tumour is then segmented, after manually initialising
some seeds, via a growing region algorithm implemented in the ITK-SNAP
software [26]. For a more comprehensive study using our model in clinic, more
advanced segmentation methods, potentially publicly available, should obviously
be investigated.
Soft tissues are meshed homogeneously with linear tetrahedral elements, using
the CGAL library [27], leading to a coarse mesh of 2000 nodes for a single
hemisphere. Its surface is extracted to form the collision mesh that will be used
to impose constraints with the skull and the probe footprint. This collision mesh is
decimated (i.e. about 300 vertices) in order to reduce the number of constraints.
13.3.1.2 Segmentation of the Vascular Tree
The cerebrovascular tree is segmented from pre-operative angiographic MRI
(ARM). While this modality provides poor contrasts for soft tissues, vessels appear
with very high intensities.
A review of vessel segmentation methods from medical images can be found
in [28]. In our work, we reconstructed the cerebrovascular tree using the maximum
intensity projection (MIP) technique proposed by [29]. The histogram of this image
reveals two peaks corresponding to the background of the image (black) and the soft
tissues of the brain (dark grey), respectively. The vessels are then segmented using a
threshold value chosen just after the second peak of the histogram. This is illustrated
in Fig. 13.2a.
The shape of this segmentation, and more particularly its outer envelope, is very
sensitive to the image quality and segmentation parameters. For example, vessels
often appear thicker in US than in ARM. In order to minimise the segmentation