This document describes a process for segmenting medical images to create 3D models for finite element analysis. Specifically, it focuses on segmenting an intervertebral disc into the nucleus pulposus and annulus fibrosus regions using micro-MRI images. A series of image processing algorithms are applied, including noise filtering, region growing, level set filtering and smoothing. Parameters for generating a finite element mesh from the segmented images are also described. The results demonstrate an automatic procedure for segmenting the disc regions and generating a FE mesh with 0.25mm grid size suitable for biomechanical modeling.
Abstract from medical image to 3 d entities creation
1. 5º CONGRESSO NACIONAL DE BIOMECÂNICA
R.M. Natal Jorge, J.M.R.S. Tavares, J. Belinha, MPL Parente, PALS Martins (Eds)
Espinho, Portugal, 8 e 9 de Fevereiro, 2013
FROM MEDICAL IMAGE TO 3D ENTITIES CREATION
Diogo, S; Claro, J C P
CT2M – Centre of Mechanical & Materials Technologies, University of Minho
KEYWORDS: Segmentation, FE mesh generation, Intervertebral Disc
1. INTRODUCTION
2.2 CONFIDENCE CONNECTED REGION
In order to obtain a 3D entity, for use in a
GROWING
Finite Element (FE) simulation biomodel a
This algorithm involves the selection of
set of procedures need to take place and,
seed points, which will control the
among them, one of the most important is
formation of regions into which the image
the segmentation of the medical images.
will be segmented. It examines the
Segmentation can be defined as the process
neighbouring pixels of the initial seed point
of decomposition of a 3D domain,
and determines whether the pixel
reconstructed from 2D images, into
neighbours should be added to the region.
different segments for subsequent
In this work we used this algorithm twice to
generation of a suitable for the generation
produce the results seen in Figures 1(c) and
of a FE mesh.
1(d).
The aim of this study is to built up an
automatic procedure for the segmentation 2.3 LEVEL SET FILTER – SHAPE DETECTION
of an Intervertebral Disc (IVD), taking into IMAGE FILTER
consideration its two main regions: the This filter represents the evolving contour
Nucleus Pulposus (NP) and Annulus using a signed function, where its zero level
Fibrosus (AF) departing from micro-MRI. corresponds to the actual contour. Then, a
similar flow can be deriving for the implicit
2. METHODOLOGY
surface, according to the motion equation of
In this work a commercial software was
the contour, that when applied to the zero-
used and the algorithms employed for
level will reflect the propagation of the
image segmentation will be described.
contour. [1]
Firstly, we will describe the procedures for
The results are shown in Figure 1(e).
the NP. Figure 1(a) represents the original
image. For the AF, the same procedures 2.4 SMOOTHING – RECURSIVE GAUSSIAN
were applied. Results for the AF FILTER
segmentation are shown in figure 2. The filter smoothes an image by convolving
it with a Gaussian kernel. It is a good way
2.1 NOISE FILTERING - CURVATURE
to remove noise. The key parameter is the
ANISOTROPIC DIFFUSION FILTER
Gaussian sigma, which controls the width
This filter preserves edges. This filter of the Gaussian kernel. The larger the value
requires three parameters: the number of of sigma, the larger is the kernel support,
iterations to be performed, the time step, which tends to smooth the image.
and the conductance parameter used in the Results are shown in Figure 1(f).
computation of the level set evolution.
Typical values for 2D were used and the 3. RESULTS AND DISCUSSION
results can be seen in Figure 1(b). The resolution of the medical images was
0.12x0.12x0.59 mm.
2. 5º CONGRESSO NACIONAL DE BIOMECÂNICA
R.M. Natal Jorge, J.M.R.S. Tavares, J. Belinha, MPL Parente, PALS Martins (Eds)
Espinho, Portugal, 8 e 9 de Fevereiro, 2013
The results of the image segmentation can
be seen in the images below, been Figure 1
for the NP and Figure 2 for the AF.
(a)
(a) (b)
(c) (d)
(b)
(e) (f)
Figure 1 – Image segmentation for NP.
(a) Original.
(b) Noise filtering – Curvature anisotropic diffusion
filter.
(c) Confidence connected region growing. (c)
(d) Confidence connected region growing. Figure 2 – Image segmentation for AF.
(e) Level Set filter – Shape detection image filter. (a) Confidence connected region growing.
(f) Smoothing – Recursive Gaussian filter. (b) Level set filter – Shape detection image filter.
(c) Smoothing – Recursive Gaussian filter.
The geometrical accuracy of the FE mesh
depends on the size of the grid dimensions.
In this study the grid size used was 0.25
mm and the various parameters used in the
treatment are listed in Table 1.
3. 5º CONGRESSO NACIONAL DE BIOMECÂNICA
R.M. Natal Jorge, J.M.R.S. Tavares, J. Belinha, MPL Parente, PALS Martins (Eds)
Espinho, Portugal, 8 e 9 de Fevereiro, 2013
Table 1 - Different parameters used in FE mesh
generation
Voxel
dimension 0.117188 0.117188 0.593415
[mm]
Voxel 0.62
Diagonal Figure 4 - Sagittal cross section of the obtained FE
Grid Size 0.25 mesh
N 0.70
Ϭ1 (0.5 xN) 0.35
ACKNOWLEDGEMENTS
Ϭ2 (1 x N) 0.70
This work was performed within the
Ϭ3 (2 x N) 1.41
European Project: NP Mimetic –
Ϭ4 (4 x N) 2.80 Biomimetic Nano-Fibre Based Nucleus
Ϭ5 (6 x N) 4.20 Pulposus Regeneration for the Treatment
off Degenerative Disc Disease, funded by
the European Commission under FP7 (grant
After analysis and selection of the
NMP-2009-SMALL-3-CP-FP 246351).
parameters’ values the FE final entities
mesh is shown in Figures 3 and 4. REFERENCES
1. Osher, S. and N. Paragios, Geometric Level Set
Methods in Imaging, Vision, and Graphics. 2003: Springer.
Figure 3 - FE mesh obtained for N = 0.7' and Ϭ = 4.2