Development of a robust MRI fiducial system for automated fusion of MR‐US abdominal images

Abstract We present the development of a two‐component magnetic resonance (MR) fiducial system, that is, a fiducial marker device combined with an auto‐segmentation algorithm, designed to be paired with existing ultrasound probe tracking and image fusion technology to automatically fuse MR and ultrasound (US) images. The fiducial device consisted of four ~6.4 mL cylindrical wells filled with 1 g/L copper sulfate solution. The algorithm was designed to automatically segment the device in clinical abdominal MR images. The algorithm's detection rate and repeatability were investigated through a phantom study and in human volunteers. The detection rate was 100% in all phantom and human images. The center‐of‐mass of the fiducial device was robustly identified with maximum variations of 2.9 mm in position and 0.9° in angular orientation. In volunteer images, average differences between algorithm‐measured inter‐marker spacings and actual separation distances were 0.53 ± 0.36 mm. “Proof‐of‐concept” automatic MR‐US fusions were conducted with sets of images from both a phantom and volunteer using a commercial prototype system, which was built based on the above findings. Image fusion accuracy was measured to be within 5 mm for breath‐hold scanning. These results demonstrate the capability of this approach to automatically fuse US and MR images acquired across a wide range of clinical abdominal pulse sequences.

Unfortunately, lesions may have low conspicuity with US imaging, particularly in obese patients, in patients with prior chemotherapy treatment or in patients with diffuse parenchymal disease. [1][2][3] Thus, fusion or co-registration of real-time US with previously obtained MRI has been advocated for guidance of percutaneous interventions on challenging lesions. MR-US image fusion pairs the advantages of MRI, namely high-contrast resolution and lesion conspicuity, with the real-time capabilities of US guidance, and has been shown to be clinically beneficial. [1][2][3][4][5][6][7][8] Most current commercial MR-US image fusion applications rely on manual co-registration of images. (One example is General Electric's LOGIQ E9 ultrasound system with Volume Navigation 9 [10][11][12][13] Active fiducial marker devices are limited in that they are complex, are susceptible to radio-frequency heating during MRI, and rely on precise tuning and calibration. 10 On the other hand, recently investigated passive marker systems 10-13 required use of custom or specific MRI pulse sequences with some relying on spatial frequency images 10,11 or only being limited to specific clinical application (i.e. fixed in headrest for intracranial imaging only). 13 Additional approaches to MR-US fusion without the use of external fiducial markers have also been reported. 14,15 These methods, however, require either acquisitions of 3D US images combined with extensive computation time and initial manual three-point rigid registration, 14 or have been demonstrated to be successful only with a single MRI pulse sequence. 14,15 In a proposed system, automated MR-US image fusion works as follows: an MRI fiducial device attached to a patient is imaged in an MR scanner, and subsequently, automatically segmented within the acquired image set. When the patient undergoes a US-guided interventional procedure, an electromagnetic sensor is attached to the fiducial and placed on the patient in the same location as it was during the prior MRI examination. The position of the US transducer is then known relative to the fixed fiducial and, consequently, known relative to the MR image set (assuming the MRI fiducial device is accurately located). The real-time US images are then directly fused with the MRI images without the need for manual identification of shared anatomical landmarks. This image fusion concept has been successfully developed for CT-US image fusion applications and is commercially available. 1 In this article, we present development of a MRI fiducial system, comprising a passive MRI fiducial marker device and a corresponding autosegmentation algorithm. The individual fiducial markers were designed to yield high signal-to-noise ratio (SNR) values across a wide range of MR pulse sequences, patient sizes, and acquisition geometries. Likewise, the segmentation algorithm was designed to be robust and function well for different image acquisition parameters. The initial application of this work is to couple the fiducial device-algorithm system with a commercial US scanner capable of manual MR-US image fusion, such as General Electric's LOGIQ E9 ultrasound and Volume Navigation systems, 9 allowing automated MR-US image fusion to be achieved. However, the fiducial device and auto-segmentation algorithm should be generally applicable to other clinical problems in which automatic registration of MR data sets is beneficial.

2.A | Fiducial device prototype
The fiducial marker device was designed with the goal of being detectable in images acquired with the wide array of possible pulse sequences used in our institution's clinical abdominal MRI protocol (Table 1). The prototype device, shown in Fig. 1, consisted of three cylindrical reservoirs with 12.7 mm inner diameter and depth, arranged to form a scalene triangle with the following side lengths: 50.7, 69.2, and 88.9 mm. A fourth reservoir was positioned 12.7 mm above the centroid of the triangle as shown in Fig. 1(a). Actual marker separation distances were within an estimated tolerance of 1 mm. The reservoirs were filled with 1 g/L (6.265 mM) copper sulfate solution. The location of the device was defined as the centerof-mass of the group of four markers (derived from their individual, intensity-weighted, center-of-mass coordinates), shown in Fig. 1(a).
The orientation of the device was defined as the cross product of the vectors from marker A to marker B and marker A to marker C, see Fig. 1(c).
The choice of 1 g/L copper sulfate solution as a fiducial marker material was motivated by published reports indicating this material as suitable for preparation of high-contrast MRI markers. 16,17 In addition, liquid copper sulfate solutions are readily available and convenient for fabricating custom fiducial markers. The particular selection of the marker size of 12.7 mm took into account the average slice thickness and slice gap of the RF pulse sequences listed in Table 1: 5-6 mm and 1-2 mm, respectively (Fig. 2).

2.B | Segmentation algorithm
Operational steps of the MATLAB â (The MathWorks Inc., Natick, MA, 2000) algorithm used to automatically segment markers from MRI datasets are described below and in Fig. 2. The empirically determined parameters associated with each step are listed in Table 2. 2.B.1 | Step 1: Determination of "noise threshold,"

N T
The purpose of this step is to remove the majority of voxels containing background noise only. A histogram of voxel intensity values (bin size set to single integer and number bins set to the maximum voxel signal), h I ð Þ, is expected to contain a major peak at a low-intensity value, I noise , which is associated with image noise. Additional peaks at higher intensity values, I signal , are associated with MRI signal. As a result, the derivative of the histogram will change sign at some intermediate intensity value, I 0 , between the low-intensity and high-intensity peaks. We define the noise threshold as the voxel value corresponding to this zero-crossing of the histogram derivative:  retained, while all other objects are removed. The reference signal threshold, R T , is defined as the average intensity value of all voxels in connected voxel sets. Connected voxel sets were determined using the standard MATLAB function "bwareaopen."

2.B.3 | Step 3: Size and signal discrimination
Connected voxel sets are assessed for signal intensity and size using an iterative segmentation process using progressively decreasing signal thresholds defined as S n T ¼ n = 4 Â R T ; wheren ¼ 7; 6; 5; 4; 3; 2 and R T is the reference signal threshold. With each iteration, threshold S n T is applied to the image and the sets such that 0.35V Marker < V 0 < 1.3V Marker are segmented. Results of each iterative step are added to results from the previous step forming a composite three-dimensional image of the individual marker candidates. If the addition of a new marker candidate yields an object with a volume greater than 1.3V Marker , then that added marker candidate is discarded.

2.B.4 | Step 4: Shape discrimination
The marker candidates are evaluated based on their shape. Only those with their longest dimension not exceeding 26.9 mm (1.5 times the known longest dimension of original cylindrical marker) are retained; all other candidates are removed.

2.B.5 | Step 5: Inter-marker distance discrimination
In this last step, the algorithm looks for a group of four marker candidates whose intensity-weighted centers-of-mass are separated by distances equal to the physical inter-marker separations of the fiducial device. If four candidates are found fitting that condition within the tolerance of AE1 mm, they are consequently identified as the segmented markers of the fiducial device. Otherwise, the separation tolerance is incrementally increased by 1 mm, and the process is repeated until the four markers are found.

2.C.2. | Volunteer trials
The fiducial device was imaged on five volunteers (three males, two   the position of the fiducial device was marked on the volunteer's skin with ink prior to the MRI scans, and a fiducial device was placed on the volunteer at the marked location for US imaging. For both the phantom and volunteer experiments, an electromagnetic tracking sensor, which tracks the position and orientation of the ultrasound probe, was affixed to the MRI fiducial device as depicted in Fig. 3(b). This tracking device is one element of the Volume Navigation component that is integrated with the US scanner.
For each experiment, the MR image set was loaded onto the US scanner and then processed by the auto-segmentation algorithm. T A B L E 3 Device location in the initial, baseline acquisition, and differences in measured device location between different acquisitions and the baseline acquisition. The device centers-of-mass are given in the LPS patient coordinate system and in units of millimeters.  Table 3 shows the center-of-mass coordinates calculated based on the baseline acquisition and compares them with those calculated based on acquisitions using oblique scanning planes and translated device, as described in Section 2.C.1. The coordinate differences shown in Table 3 Table 5.

3.C. | Proof-of-concept
Using the integrated, commercial system, the fiducial device was located with the auto-segmentation algorithm in both the phantom and volunteer experiments. MR-US fused images were immediately displayed following auto-registration. Example screen capture images are shown in Fig. 5. In the fused phantom images, user-identified T A B L E 4 Orientation angles for the initial, baseline acquisition, and differences in measured orientation angles between the different acquisitions and the baseline acquisition. The device was placed on a level base, so the anticipated orientation angles were (90#-90û). The direction cosines are listed for each unique acquisition.     | 269 patient, and general workflow-specific to a given clinical practicewould greatly aid the successful implementation of this system.
Although tests of this system were successful, the following future developments of this system could improve its robustness and clinical applicability. The algorithm can be modified to find the center-of-mass of the device if only three markers are successfully segmented. Moreover, an option for user input to identify the location of the markers would mitigate a complete failure to segment the markers. In addition, different device shapes and marker sizes could be tailored for specific applications. The device in this investigation was designed with abdominal applications in mind. For instance, smaller markers (and a smaller device footprint) could be used for MRI examinations that reconstruct thinner slices with thinner slice gaps. Also, all image acquisitions from the entire examination could be determined and averaged with appropriate (resolution-based) weighting factors applied for each sequence, which could possibly provide more accurate localization and further reduce the likelihood of localization failure. Lastly, this system could be combined with deformable image registration techniques that require an initial three-point registration of the images, 14 in which case this system would provide the necessary initial global rigid registration.

| CONCLUSION
The work presented in this report demonstrates that the two-component MRI