Researchers with interest in localizing para- and diamagnetic objects in MR images, metal artifacts and/or MR simulation.Post-implant dosimetry is an important tool for quality assurance of low-dose rate (LDR) brachytherapy of the prostate¹, where small radiation sources (brachytherapy seeds) are permanently implanted in the prostate to locally deliver a planned dose. In current clinical practice, X-ray based techniques are often used to locate the seeds to calculate the dose distribution, while MRI is used to delineate the prostate and organs at risk. As this requires both CT and MRI scanning, followed by an image registration procedure, this workflow increases patient burden and costs, and renders the procedure labor intensive and error-prone when compared to an MRI only approach. In this study we present a novel method for automatically localizing brachytherapy seeds using only MRI, based on simulation of the metal artifacts originating from the implanted seeds. Contrary to methods that only visualize seeds, such as methods based on susceptibility² or co-RASOR³, this method exactly determines the seed locations and rotations, which are needed for dosimetry.

Acquisition:

We acquired fast 3D SSFP scans (matrix 376x292x75, resolution 1.2 mm isotropic, TE/TR 2.7/4.6 ms, flip angle 10°, scan time 130 seconds) at 3 Tesla (Ingenia, Philips, Best, The Netherlands) for 5 prostate cancer patients. One month prior to the scans the patients were treated with I-125 brachytherapy seeds (selectSeed model 130.002, Isotron, Berlin, Germany), with 40 to 85 seeds implanted per patient.

Simulation:

We used MRI simulation to predict the artifacts around a seed placed in a uniform background. The induced field perturbation was calculated⁴ at a high resolution using an exact model of the brachytherapy seed (Fig 1b). We used a fast simulation method which simulated only the effect of field perturbations on the spatial encoding process (i.e. signal shift, voids and dephasing) (Fig 1c). The simulation time was approximately 300 seconds per 3D scan on a Xeon E5-1607 CPU.

Template matching:

The artifacts were simulated for 321 different orientations of the brachytherapy seed with respect to B0. Phase Correlation template matching⁵ was applied to the acquired in-vivo images using each simulation individually as a template. Per voxel the orientation with the highest match intensity was selected and candidate detections were found using a threshold on this match intensity. To obtain the final detections, the following procedure was used (assuming that the number of implanted seeds is known a priori):

1. For each candidate detection linear regression was used to locally match the acquired image intensity and phase to the corresponding template.

2. The candidate with the lowest regression error is accepted as a final detection.

3. The phase effect of the detected seed is cancelled in the image by subtracting the phase predicted by the simulation.

4. Repeat from 1 until the number of detections is equal to the number of implanted seeds.

The final detections were registered to CT and manually classified as true or false detections, and the distance to the center of the seeds on CT was measured.

Table 1 shows the detection results for each of our patients. The mean sensitivity of the method was 0.97 and the worst case sensitivity 0.95 (subject 5). Figure 2 shows a 3D rendering of the detected seeds superimposed on the MRI scan and an isosurface of the seeds in the registered CT scan.

As opposed to existing MR-based methods that visualize brachytherapy seeds, we presented a method that automatically localizes seeds with high sensitivity. We have shown consistent performance over multiple subjects, which indicates the method is robust against anatomical variations and variations in the configuration of the seeds. Missed and false detections were primarily caused by seeds that were in close vicinity to each other. In these cases the MR artifacts interfere and are not accurately predicted by single seed simulations. However, similar cases may also cause missed detections in CT-based methods.

The locations resulting from our method are intrinsically registered to other MRI scans from the same session, on which the prostate and organs at risk are delineated. Therefore no image registration is needed to perform dosimetry, which should increase the accuracy of the calculated dose distributions. Furthermore, the exact orientation of the seeds is known, which may improve the calculation of dose distributions by taking into account the anisotropy of the dose resulting from rotated seeds. In future research we will compare the proposed MRI-only dosimetry to the current CT-MRI approach.