Evaluation of sub-pixel fiducial tracking using image processing

Introduction. Resonant micro-coil fiducial markers are increasingly being used to track devices within an MRI scanner [1,2]. Good results have been obtained with previously described methods of tracking such as multiple application of 1-D gradients [3,4]. Image processing has also been applied to track devices with 3D reconstruction [5]. In this report, we assess various image-processing algorithms in order to increase the precision with which a fiducial position can be determined. We show that sub-voxel sensitivity of inductively coupled fiducials can be achieved, and hence precise locations are obtained from relatively low-resolution rapidly acquired images. We then compare the performance of the techniques for both post-processing and real-time imaging. Methods. Resonant fiducial micro-coils constructed by hand-winding 6 turns of 6 gauge enamelled wire around a 2mm former and connecting a 100pF capacitance were tuned to 63.8MHz. To generate a signal, the coils are filled with approximately 8mm of a water-rich gel. A whole fiducial is around 3 x 3 x 5mm in size. To image the fiducials, we used a Siemens Vision 1.5T scanner running a 2D FLASH sequence with 256 matrix, 200mm square FOV (pixel size 0.78mm x 0.78mm), and 1 flip angle to reduce signal from the phantom. Ten slices of thickness 5mm were taken with fiducials mounted on a software-controlled manipulator arm. Images were processed using MatLab. The manipulator translations ranged from 0.1 to 40mm and real displacements were measured using high precision optical encoders. Series of images were taken with various displacements of the fiducials. An automatic error detection routine was first applied to identify and remove erroneous pixels due to artefacts. The image data were then analysed using five detection algorithms, (1) detection of the Maximum Intensity Pixel (MIP), (2) detection of the MIP from a Subtracted image (SIP), (3) Linear Interpolation between pixels (LI), (4) quadratic Curve Fitting (CF) to pixel values, and (5) quadratic Correlation-Coefficient Fitting (CCF) using a submatrix around the MIP [6]. In Figure 1 it can be seen how higher precision is achieved by reconstruction of the curves in CF and CCF. Figure 2 shows the correlation between calculated and measured distances at large and sub-pixel distances, as well as fiducial SNR variations for rotations of the manipulator. Results.