Escolar Documentos
Profissional Documentos
Cultura Documentos
1 Introduction
True 3D imaging modalities, like MRI and CT, are extremely potent with regard to their rendering capabilities, but are dicult to use for intra-operative procedures, mainly due to obstructive hardware and the latent images. US,however, has been becoming known as a widely popular image guidance modality,since it is real-time,convenient to use in the operation room and inexpensive compared to CT and MRT. In order to provide the physicians with a 3D real-time visualization of the internal anatomy, individual 2D US images must be assembled into 3D volumes and then the position of surgical tools are related with respect to the reconstructed US volume.The goal is to preserve true anatomical shape by means of a consistent reconstruction of 3D volumes Using US as a guidance modality for surgical procedures would require tracking the imaging probe with a magnetic or optical tracking device.A xed transformation between the US beam and the tracking device needs to be determined, so that arbitrary image pixels can be referenced in a global frame.Obtaining this xed transformation is referred to as ultrasound calibration.After calibration, a 3D volume is reconstructed by some surface- or voxel-based method ,and the data is visualized with some appropriate combination of surface extraction, volume rendering, re-slicing, panoramic viewing, or multi-planar techniques.Clearly, the accuracy of calibration is the most signicant factor ,which greatly inuences the quality of the reconstructed volume and visualization. With regard to the currently known calibration processes, an object of known geometrical properties(phantom) is scanned by the tracked US probe and then various mathematical procedures are applied to determine the unknown transformation that maximize the similarity between the US images and the actual phantom.Geometrical model base phantoms like points, plane exist and some studies have compared their accuracy and performance.The cross-wire and three-wire
phantoms require long time of acquisition and are hard to automate, while the single-wall as in Cambridge phantom is automatic repeatable method. Hopkins phantom which is based on using a recent closed-form formulation allows for oine image processing leading to immediate calibration if the simple docking guide mechanism is involved.The basic concept and reconstruction model of ultrasound calibration will be introduced in the section 2.In section 3 and 4,traditional calibration phantoms and Hopkins phantom will be discussed respectively and the experiment results and the comparison between the dierent phantoms will be summarized in section 5.
2 Reconstruction[3]
2.1 Processing
A 3D free-hand examination consists of three stages:scanning, reconstruction and visualization, which is illustrated by gure 1. Before scanning, some sort of position sensor is attached to the probe,which is typically the receiver of an electro-magnetic position sensor.Measurements from the position sensor are used to determine the positions and orientations of the B-scans with respect to the xed transmitter.In the reconstruction-stage, the set of acquired B-scans and their relative positions are used to build a regular voxel array.Finally, the voxel array is visualized using ,for example, volume rendering or surface rendering.
to the transmitter Tand nally to the reconstruction volume C.The process is illustrated in the gure 3.
2.2.2 Mathematical formulation The overall transformation can be expressed as the multiplication of homogeneous transformation matrices: Cx =C TT T TR R TP Px where Sx u S v Px = y 0 1 The standard notation J TI is adopted, which means the transformation from coordinate system I to coordinate systemJ.u and v are the column and row indices of the pixel in the cropped image,
and Sx and Sy are scale factors with units of mm/pixel.Cx is the pixels location in the coordinate system. A transformation between two coordinate systems has six degrees of freedom:three rotation(, , ) and three translation(x, y, z).The rotation between two coordinate systems is eected by rst rotating through around the x-axis, then through around the y-axis, and nally through around the z-axis.Using this convention,the homogeneous matrix describing the transformation is illustrated in the gure 4.
Each of the transformation matrices plays a dierent role in reconstruction.T TR is derived directly from the position sensor readings.C TT is included largely as a matter of convenience. R TP needs to be determined by calibration. The scale factors Sx and Sy could be derived from the axis markings on the B-scan. Once Cx has been found for every pixel, the voxels of C can be set according to the intensities of the pixels they intersect.
It is proved that the detecting a line in an image is easier than a point in an image and it is possible to locate a straight line even when portions of the line are corrupted or missing.On the contrary, the same cannot be said of dots and crosses.In addition, practically speaking, calibration with wire-based phantoms seems to be done by means of locating the dots by hand.In contrast, automatic line detection algorithm can be applied to the single-wall calibration so that time can be saved to collect the images. One signicant problem of single-wall calibration is caused by the width of the ultrasound beam and the nature of specular reection.When the beam is not normal to the wall, the rst echo to return to the probe comes from the edge of the beam closest to the wall, as shown in the gure 8(b).In gure 8, point B is encountered by the ultrasound pulse before point A on the centerline.The echo from B produces a response in the image which does not reect the true position of the wall.
Figure 8: Minimal sequence of motions and beam thickness problem in single-wall calibration[3]
Insert ultrasound probe into clamp and tighten bolts so that the slots through the two sides of the clamp are aligned with the scan plane of the probe. Immerse the phantom in water,slot the clamp over the bar,and scan the bar with the probe from all possible angles.A clear image of the bar will always be visible in the B-scan Because of the clarity of the images and the fact that the phantom produces a line in the B-scan,it is possible to automatically detect the line in each B-scan. Once these lines have been located, calibration proceeds is exactly the same as that for single-wall phantom. Accurate calibration requires a minimal sequence of motions,as shown in the gure 9.The procedure is fully automated and takes less than ve minutes to complete,including scanning, line detection and optimization.It is not possible for the wired-based techniques to do the calibration so fast, since the dots need to be located by hand in each image. The B-scan images from the traditional calibration phantoms are shown in the gure 10 together.
4 Hopkins phantom[1][2][4][5]
With regard to Hopkins phantom,a position sensor is attached to the probe for tagging each image/volume with its position and orientation in space.The traditional US calibration framework in gure 3 is integrated into the AX = XB framework as in gure 11, using a recent closed form solution for the AX = XB problem.It uses minimal US imaging allowing for an immediate result. In this section, the original idea will be rst discussed, and then the mathematical formulation of this model.Finally ,the experiment,calibration setup and protocol, will be described and summarized.
4.1 Idea
The closed form solution comes originally from a basic issue, which is to determine the spatial relationship between a camera mounted into a robot end-eector.This spatial relationship is a rigid transformation, that is, a rotation and a translation, known as the hand-eye transformation. A classical approach (Tsai and Lenz 1989;Chen 1991;Daniilidis and Bayro-Corrochano 1996; Horaud and Dornaika 1995; Shiu and Ahmad 1989; Chou and Kamel 1991; Wang 1992) states that when the camera undergoes a motion A = (Ra , ta ) and the corresponding end-eector motion is B = (Rb , tb ), then they are conjugated by the hand-eye transformation X = (Rx , tx ) (Fig. 2). This yields the following homogeneous matrix equation: AX = XB,where A is estimated, B is assumed to be known, and X is the unknown.This is illustrated in the gure 11(b).
Figure 11: AX=XB method in the Hopkins Calibration and hand-eye calibration[1][5]
The estimated US image frame motion is given by A (),as described in the gure 12,where Ra is the rotation of US image frame between pose 1 and 2 and is the unknown scale factor vector that relates the translation vector ua in voxel space to US image frame translation vector ta (in mm). Using AX = XB and the equations from the gure 12,the result that is computed in the gure 12 is obtained.The gure 13 illustrates only the left side of AX = XB,and the right side is computed by means of the same calculation.According the result of grue 13, the rotation equation and translation equation are obtained.
Figure 13: Mathematical formulation(2) What we will do next step is to reduce the nonlinear model to linear formulation. A new formulation,Sylvester equation: UX + XW = T, that is very similar to the homogeneous equation AX = XB is inspired and Hussein et al. proved that Sylvester equation is usually formulated as a linear system by means of the Kronecker product: (U I + I W ) vec (V ) = vec (T ) (1)
The computation process how the property of Kronecker product applies to our model is described in the gure 14.Finally, Andre et. al[5] proved that two independent motions with non-parallel axes is sucient to recover a unique solution for AX = XB.
10
4.3.2 Images acquisition For the images acquisition, an optical pointer pivoted to obtain an accurate estimate of the desired 3D point is used to collect 3D points of each of the plates for oine processing. Figure 16 presents that 3D points are rst registered to provide a local coordinate system for each thin plates,and then the relative transformations between each pair of plates can be calculate from the local coordinate systems.What important is that the poses of the three plates should be arranged carefully to give the optimal results for the two motions required by AX = XB,based on previous experiments.
11
4.3.3 US probe calibration By calibrating, A1 , A2 have to be estimated by means of the method discussed in the paragraph Images acquisition.That is,A1 and A2 transformation matrices are the relative transformations between the plates of the calibration phantom with a positional oset based on the pixel coordinates of the phantom in the acquired US image.Two protocols are tested for computing A1 and A2 matrices: Move the US tracked probe such that the probe is parallel to the thin plate and the image plane of US images shows the middle of the thin plate. Apply the same protocol, but collect multiple tracking data from one end of the thin plate to the other end and using an averaging technique to nd the center of the thin plate. The result of scanning can be seen in the gure 15 (b).It yields sharp US images that can even be processed automatically.Three sets of tracking and US image data are sucient to solve the mathematical formulation,and an additional 3 sets of data(48 datasets) will ensure a wellconditioned problem and produce comparable results to previous calibration.
12
Figure 18: Docking station[1] the line detection algorithm. It speeds up the calibration process greatly.The performance of the Cambridge phantom was signicantly better than any of the other methods, while calibration was performed in a matter of minutes (less than ve minutes). Compared to the traditional calibration phantoms, Hopkins phantom indicate signicant potential in using a simple calibration phantom in conjunction with the AX = XB closed form formulation. It uses optical digitization with a calibrated pointer to replace with a great extent the traditional segmentation of points/planes in US images.The tracked pointer appeared to introduce signicantly less error than the resolution of the US image caused in the traditional approaches.It also provided very accurate calibration results using signicantly fewer US images and requires only minimal image segmentation.
13
References
[1] Emad Boctor, Anand Viswanathan, Michael Choti, Russell H. Taylor, Gabor Fichtinger, Gregory Hager, A novel closed form solution for ultrasound calibration, IEEE Int Symp. On Biomedical Imaging, 2004 [2] Emad Boctor, Anand Viswanathan, Michael Choti, Russell H. Taylor, Gabor Fichtinger, and Gregory Hager,A Novel Closed Form Solution for Ultrasound Calibration, IEEE Int Symp. On Biomedical Imaging, 2004 [3] R. W. Prager, Rohling R. N., Gee A. H., and Berman L.,Rapid Calibration for 3-D Freehand Ultrasound, US in Med. Biol., 24(6):855-869, 1998. [4] John W. Brewer, Kronecker Products and Matrix Calculus in System Theory, IEEE Trans. Circuits and systems, 25(9) Sep.1978. [5] Nicolas Andre and Radu Horaud and Bernard Espiau,Robot Hand-Eye Calibration Using Structure from Motion, International J. of Robotics Research, 20(3), pp 228-248, 2001. [6] Laurence Mercier, Thomas Lang, Frank Linkseth,and D.Louis Collins,A Review Of Calibration Techniques For Freehand 3-D Ultrasound Systems, Ultrasound in Med. & Biol., Vol. 31, No. 4, pp. 449 471, 2005.
14