Proceedings of the 29th Annual International Conference of the IEEE EMBS Cité Internationale, Lyon, France August 23-26, 2007.
SaC04.6
3D Segmentation with an Application of Level Set-Method using MRI Volumes for Image Guided Surgery A. Bosnjak, G. Montilla, R. Villegas, I. Jara
Abstract— This paper proposes an innovation in the application for image guided surgery using a comparative study of three different method of segmentation. This segmentation method is faster than the manual segmentation of images, with the advantage that it allows to use the same patient as anatomical reference, which has more precision than a generic atlas. This new methodology for 3D information extraction is based on a processing chain structured of the following modules: 1) 3D Filtering: the purpose is to preserve the contours of the structures and to smooth the homogeneous areas; several filters were tested and finally an anisotropic diffusion filter was used. 2) 3D Segmentation. This module compares three different methods: Region growing Algorithm, Cubic spline hand assisted, and Level Set Method. It then proposes a Level Set-based on the front propagation method that allows the making of the reconstruction of the internal walls of the anatomical structures of the brain. 3) 3D visualization. The new contribution of this work consists on the visualization of the segmented model and its use in the presurgery planning.
usages. The following sections propose different modules and especially the 3D segmentation methods. Finally, figures depict the 3D reconstruction and segmentation of white matter of the brain, ventricles and cerebellum based on the 3D level set segmentation method. A conclusion summarizes the originality and efficiency of the methods. II. METHODOLOGY The proposed 3D segmentations are included in a processing chain from acquisition of MRI volumes to the 3D reconstruction and visualization for planning a guided surgery (figure 1).
I. INTRODUCTION The modeling, reconstruction, and visualization of organs and internal structures of the human body, issued from MRI volumes, or CT scan volumes are important to improve the medical diagnosis and the therapy. Two-dimensional (2D) and three-dimensional (3D) medical imagery assists largely in medical therapy such as image guided surgery applications. Shape recovery of medical organs may be more complex compared to other computer vision. This is primarily due to the large shape variability, structure complexity, several kinds of artifacts, and restrictive body scanning methods [1]. For example: recovery of the white matter and gray matter boundaries in the human brain slices is a challenge due to its highly convoluted structure. On the other hand, Vega et al. [2], proposed another method for 3D visualization of intracranial aneurysm using a multidimensional transfer functions, because a clear identification of vascular tissues is vital for the planning of surgical treatment of intracranial aneurysms. This paper presents a comparative study of three different segmentation methods with a purpose of clinical practice Manuscript received April 1, 2007. A. Bosnjak, G. Montilla, R. Villegas, “Centro de Procesamiento de Imágenes”. Facultad de Ingeniería. Universidad de Carabobo. Valencia. Venezuela. (e-mail:
[email protected] ) . I. Jara, “Hospital Metropolitano del Norte”, Valencia. Venezuela.
1-4244-0788-5/07/$20.00 ©2007 IEEE
Figure 1. Processing chain of MRI volumes.
A. DICOM and DICOMDIR Reader DICOMDIR directory files are useful in medical applications because they allow organized access to images and information sets. For this reason we design and implemented software for reading DICOM Directory Files [3]. This software was developed using the DCMTK 2005 library [4], which has several years of evolution and continuous use in medical applications. This tool has been successfully integrated into an application for neurosurgery preoperative planning [5]. It can also be attached to any other software under development that requires the handling of DICOM images and DICOMDIR directory files. Figure 2 depict the results of the integration of the DICOMDIR reading tool into a medical application for neurosurgery planning. This tool is coupled with a graphical interface for browsing studies, series and images. B. 3D Filtering, Anisotropic diffusion filter The pre-processing is an important module in this chain, since the segmentation will be better if we have no noise images. With this objective, we evaluated two types of
5263
filters, an anisotropic filter proposed by Perona and Malik [6], and anisotropic filter based on min / max flow scheme, proposed by Malladi and Sethian [7]. The second filter is a scheme for image enhancement and noise removal based on the level set theory. This filter was modified and extended to the third dimension in order to eliminate the speckle noise and to keep the most significant edges of the volumes.
are noisy and the gray level cannot stop the connected region.
Figure 3. Segmentation using the Region Growing Algorithm
Figure 2. DICOMDIR reader integrated to a graphical interface for the navigation inside the volume.
C. Segmentation by dynamical models For medical images, the segmentation is about the localization of a closed interface between the inside and the outside of the considered object, e.g. an organ. Among the segmentation methods, there are efficient ones in which the interface corresponds to the dynamical boundary between two media [8]. The interface's shape is modified with constraints depending on data. It stops and fits the boundary between two media. C.1. Region Growing Algorithm This segmentation technique, segments objects inside an image that contains many bifurcations and protuberances. It is not necessary to know a priori the topology of the object. This method is initialized using a vicinity of pixels or a pixel seed, placed by the operator. Once placed the vicinity or seed inside the object to be identified on the image, begins the study of the neighboring pixels, comparing them with the vicinity using a threshold condition that is adjusted by the operator. It is continuously searching neighbors that fulfill this condition. It marks the pixels that belong to the new vicinity and stops the growth at the border of the object, with the objective of finding the contour. The region growing algorithm was expanded to 3D volumes, and we used a connected topology of the six neighbor voxels. Figure 3 depicts the results of segmentation using this algorithm, when we made the reconstruction using a marching cubes method. The main drawback are the many non-identified structures inside of head because MRI image
C.2. Cubic Spline Hand Assisted. When the physician chooses this method of segmentation, the software provides the tools to place points inside the image. These points should be placed in the interface between the tumor or the organ to be segmented. The software has the algorithm to transform this array of points into a cubic spline, in addition to the 3D visualization of the segmented area. Figure 4 depicts the segmentation of the brain ventricle using this manual method. The top image shows the tools for manual segmentation, the bottom image shows the different curves for the 3D reconstruction of the brain ventricle. C.3. 3D Level Set Methods. A physical shaped model was introduced recently by Malladi et al. [9]. They have developed a propagation model based on a non-intersecting closed curve or surface, where the propagation speed depends on the curvature. Here also, the level set splits the space into two regions (inside and outside), it is called “interface” in the Sethian's literature [8]. The front propagates itself inside the image adapting and fitting the walls of the structure or 3D object. This technique solves two problems of the snakes [10] a) It allows the segmentation of objects with many bifurcations and protuberances, and recovers complex shapes inside the image. b) It is not necessary to know a priori the topology of an object for its recovering. We consider the interface as the final localization of a closed surface Γ(t ) , propagating along its normal direction with speed V depending on the mean curvature. The level set methodology of Sethian, is about seeing the propagating interface as the zero level set of a higher dimensional r function ψ . The initial function ψ ( s , t = 0 ) , verifies r r Γ(t = 0) = {s ( x, y, z ) / ψ ( s , t = 0) = 0} (1)
5264
The fact that the zero level set of ψ
matches the
propagating surface Γ ( t ) can be expressed by r ψ ( s ( x, y , z ) , t ) = 0
k ( x, y , z ) =
(2)
In order to obtain the equation of the motion of the level set function; we derive the previous equation with regard to time, using the chain rule: r r ψ t + ∇ψ ( s ( x, y, z ) , t ) .s ′ ( x, y, z ) = 0 (3) Simplifying the equation, we obtain: ψ t + V ∇ψ = 0 (4) r r r where V = s ′ ( x, y, z ) ⋅ n ( x, y, z ) and n is normal to the surface defined by: r ∇ψ n= ∇ψ
borders of the image, we use the parameter k ( x, y, z ) as: 1 1 + Grad _ Mod
Grad _ Mod = exp
( ∇G*I ( x , y , z ) −α ) / γ
ψ in, +j ,1k = ψ in, j , k − Δt V k ( i, j , k ) ∇i , j , kψ in, j , k
(8) (9)
(10)
where the speed V depends on the c mean curvature. In order to accelerate the convergence, it is possible to initialize the level set with a fast marching method. Figure 5 gives the reconstruction of the brain structures with the front propagation segmentation. III. RESULTS
(5)
Finally, we carried out the comparative analysis of the three segmentation methods using a qualitative method. We have different conclusions for each one of these methods depending on the application. The first method, of regions growing, is a quick method that makes a quick segmentation of the brain and the skin. However, it cannot discriminate the internal structures of the brain since there will always be a voxel or a pixel for continues growing the region. The second is a manual method. It adapts perfectly for the localization of tumors inside of the brain. However, the segmentation is done manually. It is a repetitive task that consumes too much of a physician’s time. The third segmentation method uses a level set. The figure 5 shows the results of the front propagation model. To accelerate the convergence and the computing time, the distance function was calculated only on a thin layer around the surface of the volume that propagates. By the same motivation, the level set is initialized with the result of a fast marching method [8]. The only disadvantage is the computing time. The computing time of the front propagation method is around 4 minutes, without mesh computing (using marching cubes method) or 3D visualization. In future works, we consider the fusion of the second method with the level set method. It would provide more useful results in cases of tumors detection into the brain. IV. RESULTS
Figure 4. Segmentation using Cubic Spline Hand Assisted.
Changing the continuous equations into the discrete ones, we obtain: ψ in, +j ,1k − ψ in, j , k (6) + V ∇i , j , kψ in, j , k = 0 Δt and finally: ψ in, +j ,1k = ψ in, j , k − Δt V ∇i , j , kψ in, j , k (7) With the purpose to stop the front propagation on the
(CLINICAL PRACTICE)
Here we present the use in the preoperatory planning of a treatment surgery, for a 65 years old female patient with a sellar arachnoid cyst or arachnoidocele. The cyst had a 1.8 cm diameter and it caused a compressive effect on the hipophisis, therefore the removal by a stereotactic frame guided transphenoidal surgical approach was indicated [11] (figure 6). Only DICOM images coming from CT were used for the planning. The precision of the stereotactic coordinates provided by the software was contrasted in parallel with the ones calculated by the traditional planning procedure, verifying the approximation error was acceptable and less than 1 mm. The figure 6 shows the use of the software in the afore mentioned case planning. This
5265
intervention was successfully done in the Hospital Metropolitano del Norte (Valencia-Venezuela) and the neurosurgeon stated his satisfaction with the support provided by the software in the planning and segmentation of this case.
V. CONCLUSION We presented a semi-automatic segmentation method for the cerebral structures in order to locate precisely the left and right ventricles of heads, the gray matter and white matter, in addition to the cerebellum and other important structures for the medical diagnosis in MRI volumes. The obtained results are promising as they have been visually evaluated by a neurosurgeon. Furthermore, our application works on conventional PCs platforms equipped with commercial graphic cards containing NVIDIA GPU’s. It permits the successful planning and simulation of neurosurgeries with stereotactic frame in a precise, fast and easy interactive way. REFERENCES
Figure 5. 3D Segmentation of Skin, white matter of the brain, ventricles and cerebellum
Presently, we have carried out three successful cases of surgical pre-planning using this software jointed with an stereotactic frame. The first two segmentation methods are integrated in the software of surgical planning with the purpose of calculating the stereotactic coordinates, and the spatial position of the tumor. The results of Level Set method illustrated on figure 5, remains a prospect that is yet to be included on the software of surgical planning for clinical practice at this time.
Figure 6. Using the software for the sellar arachnoid cyst surgery planning. The left superior corner shows a 3D reconstruction of the patient’s head with the simulated probe (top left) and the widget for tissues classification (top right).
[1] Suri Jasjit S., “Two-Dimensional Fast Magnetic Resonance Brain Segmentation”. IEEE Engineering in Medicine and Biology. July-August 2001, pp. 84-95. [2] Vega F., Hastreiter P., Tomandl B., Nimsky C., Greiner G. “3D Visualization of Intracranial Aneurysms with Multidimensional Transfer Functions”. Bildverarbeitung für die Medizin, 2003, pp. 46-50. [3] Villegas R., Montilla G., Villegas H. “A Software Tool for Reading DICOM Directory Files”. International Journal of Healthcare Information Systems and Informatics. Vol. 2, Nº 1. January-March 2007, pp. 54-70. [4] DCMTK 2005, “Digital imaging and communications in medicine tool kit”. DICOM toolkit software documentation. Oldenburger Forschungs und Entwicklungsinstitut für Informatik-Werkzeuge und Systeme (OFFIS). Retrived on September 29, 2005, from http://dicom.offis.de/dcmtk.php.en [5] Montilla G., Bosnjak A., Jara I., Villegas H. “Computer Assisted Planning using dependent Texture Mapping and Multiple Rendering projections in Medical Applications”. Proceedings of 3rd European Medical & Biological Engineering Conference. Czech Republic, 2005, pp. 44204425. [6] Perona P., Malik J. “Scale-space and Edge Detection Using Anisotropic Diffusion”. IEEE Transactions on Pattern analysis and Machine Intelligence. Vol 12., Nº 7. July 1990, pp. 629-639. [7] Malladi, R., and Sethian, J.A., “Image Processing: Flows under Min/Max Curvature and Mean Curvature”. Graphical Models and Image Processing. Vol 58., Nº 2. 1996, pp. 127141. [8] Sethian, J. “Level Set Methods and Fast Marching Methods. Evolving Interfaces in Computational Geometry, Fluid Mechanics, Computer Vision, and Materials Science”. Cambridge University Press. 1999. [9] Malladi R, Sethian J, Vemuri B. “Shape Modelling with Front Propagation: A Level Set Approach”. IEEE Trans. on Pattern Analysis and Machine Intelligence. Vol. 17, N° 2. February 1995, pp. 158-175. [10] Kass M., Witkin A., Terzopoulos D. “Snakes : Active Contour Models” Int’l Journal of Computer Vision, 1988, pp. 321-331. [11] Kouri J., Chen M., Watson J., Oldfield E. “Resection of suprasellar tumors by using a modified transphenoidal approach”. Journal of Neurosurgery. Vol. 92., 2000, pp. 10281035.
5266
