Fiducial detection and registration for 3D IMRT QA with organ‐specific dose information

Abstract Purpose Two‐dimensional (2D) IMRT QA has been widely performed in Radiation Oncology clinic. However, concerns regarding its sensitivity in detecting delivery errors and its clinical meaning have been raised in publications. In this study, a robust methodology of three‐dimensional (3D) IMRT QA using fiducial registration and structure‐mapping was proposed to acquire organ‐specific dose information. Methods Computed tomography (CT) markers were placed on the PRESAGE dosimeter as fiducials before CT simulation. Subsequently, the images were transferred to the treatment planning system to create a verification plan for the examined treatment plan. Patient’s CT images were registered to the CT images of the dosimeter for structure mapping according to the positions of the fiducials. After irradiation, the 3D dose distribution was read‐out by an optical‐CT (OCT) scanner with fiducials shown on the OCT dose images. An automatic localization algorithm was developed in MATLAB to register the markers in the OCT images to those in the CT images of the dosimeter. SlicerRT was used to show and analyze the results. Fiducial registration error was acquired by measuring the discrepancies in 20 fiducial registrations, and thus the fiducial localization error and target registration error (TRE) was estimated. Results Dosimetry comparison between the calculated and measured dose distribution in various forms were presented, including 2D isodose lines comparison, 3D isodose surfaces with patient’s anatomical structures, 2D and 3D gamma index, dose volume histogram and 3D view of gamma failing points. From the analysis of 20 fiducial registrations, fiducial registration error was measured to be 0.62 mm and fiducial localization error was calculated to be 0.44 mm. Target registration uncertainty of the proposed methodology was estimated to be within 0.3 mm in the area of dose measurement. Conclusions This study proposed a robust methodology of 3D measurement‐based IMRT QA for organ‐specific dose comparison and demonstrated its clinical feasibility.


| INTRODUCTION
Intensity modulated radiation therapy (IMRT) quality assurance (QA) or patient-specific QA has been widely performed in the clinic to verify treatment planning dose calculation as well as the delivery system of a linear accelerator (LINAC) with multileaf collimators (MLCs). [1][2][3][4] However, its sensitivity in detecting errors and its relevance to clinical judgment has been extensively discussed by physicists. [5][6][7][8] In 2018, AAPM task group 218 was published in order to address the issues of IMRT QA and review the existing measurement-based methods and computer reconstruction methods. 9 It was concluded that the conventional gamma test should be reviewed on a structure by structure basis if the QA method allows for it. Purely using the passing rate for evaluation could underestimate the clinical consequences because the passing rate only summarizes the gamma test in aggregate. In addition, computed and measured DVH comparisons can provide more clinically relevant information. The study also addressed that the dose difference criterion would ideally be customized for different anatomical structures and the predicted dose in the structures. For example, the dose criterion in the spinal cord for a predicted cord dose of 45 Gy should be tighter than the tolerance in the cord with a predicted dose of 20 Gy. A recent study evaluated current measurement-based QA at multiple institutions using the IROC head and neck IMRT phantom. 10 The results showed that traditional IMRT QA methods performed consistently poorly in searching for a large error or a moderate error regardless of whether a 3%/3 mm or a 2%/2 mm criteria was used.
This work aims to resolve the issues regarding IMRT QA by demonstrating a measurement-based methodology using fiducial registration and structure-mapping to acquire organ-specific dose information. PRESAGE three-dimensional (3D) dosimeters (Heuris Inc.) have been recognized as true 3D dosimeters because dose deposition in the 3D space is readout using an optical scanner with no computer modeling involved. 11,12 The dosimeter consists of an optically clear polyurethane matrix, containing a leuco dye and free radical initiators that exhibits a radiochromic response when exposed to ionizing radiation. In 2012, the first comprehensive application of 3D dosimetry to verify a complex radiation treatment was proposed. 13 The novelty of this work was to transform measured 3D dose distribution in the phantom back to the patient CT data, and thus enabling DVHs comparison. However, the study addressed that the methodology was limited to the accuracy of the 3D dose measurement, as well as the dose transformation between the phantom CT and the patient CT since the dose deposition at the two different geometries cannot be adequately described by a simple transformation matrix. Also, it was not clear how the correlation between the coordinates of the evaluation space and the reference space was established.
Furthermore, several publications have shown 3D dose measurement of IMRT fields using different types of 3D dosimeters (Gel, PRESAGE etc) and dose read-out tools. 11,[14][15][16][17][18][19][20][21][22][23] One of the most significant source of errors remains in the 3D registration between the measured dose and planned dose, which requires fiducial markers to be shown with sufficient contrast in two different image modalities, simulation CT and optical CT. The registration error is important for 3D measurement-based QA because the dosimeters were read-out by an optical CT scanner with different orientations than the CT scan. No previous research has analyzed the effect of the registration errors in 3D dose comparison, or have reported the accuracy of the registration. Previous studies have addressed that the result of registration errors in the manual alignment of the measured and calculated dose distributions leads to the gamma failing points at the sharp dose gradient regions. 15,18 A robust and accurate registration between the treatment planning coordinates and the dosimeter coordinates is therefore one of the key components to true 3D dose comparison. One could find the 'best match' through the extended use of manual registration. However, a rigorous and fair dose distribution comparison cannot be established when exclusive manual registration is used to align the dose distributions and the results are operator-dependent. This study aims to resolve these concerns by proposing a methodology using automatic fiducial registration algorithm and commercially available structure-mapping application in clinical TPS. First of all, fiducial-based registration was employed to register the optical CT dose images to the simulation CT dosimeter images in order to correlate the two coordinate systems. Second, using the coordinates of the fiducials, patients' anatomical structures were mapped to the dosimeter coordinates for structure-by-structure 3D dose comparison using Eclipse structure mapping application  In this study, the dosimeter received only one fraction of dose while in the Results section, the measured dose was scaled to the prescribed total dose for the presentation.
After irradiation, the 3D dose distribution of the irradiated dosimeter was readout by a single laser beam optical-CT scanner (OCT) modified from the OCTOPUS TM scanner 11 at our institution. Four hundred projections were generated for one slice with slice thickness of 1 mm. For each projection (13.5 cm), 5000 data points were acquired. 3D dose images with submillimeter resolution were reconstructed using filtered back-projection algorithm. An in-house MATLAB (MathWorks, Inc) code was developed to perform the reconstruction algorithm and an automatic fiducial localization algorithm to register the markers in the OCT dose images to the CT simulation images. Figure 4 shows each step of the algorithm. Before localizing the markers, three image sets, CT simulation images, calculated dose images, OCT dose images were resampled to have the same size and resolution (1 mm). A region of interests (ROI) was selected to reduce the image size and pixels with image intensities higher or lower than a specific range were filtered out. In the marker localization phase, the prominence, of each pixel was calculated for both the CT and OCT image sets. The prominence measures how one pixel stands out from the surrounding pixels. Four pixels with the highest prominence values were selected in both image sets representing the fiducial points. Using singular-value decomposition, rotation (R), and translation matrix (t) for the point-based registration were found 24 : where X, Y are the matrices, consisting of three rows and four col-

2.B | Registration error estimation
The overall dose comparison errors between calculated dose on the phantom and measured dose include real delivery errors to be detected, dosimetry uncertainties and registration errors between where N is the number of the fiducials. squared TRE at a point r: where  ing points relative to the structures. In the first case, a VMAT plan for cerebellar metastasis was selected for the demonstration. Figure 6 shows the results of the measured dose distribution and its relative location to the structures from 3D Slicer. In this case, the target is close to the brainstem and thus, sparing of the OAR is critical. With the proposed method, dose fall-off in the high dose gradient region between the target and the OAR can be evaluated.
In Fig. 7  In addition to the DVH comparison, 3D gamma analysis was performed on the measured and calculated dose matrices. The passing rates were 99.2% and 96% using 3%, 3 mm and 3%, 2 mm criteria (with a 30% threshold). However, merely looking at the passing rate is challenging to make a clinical judgment. Using 3D Slicer, pixels that fail the 3%, 3 mm gamma test can be shown in 3D space ( Fig. 9). The failing pixels are mostly in the region of a steep dose fall-off outside the PTV, where the coverage of the PTV is influenced.
In the second case, three malignant lesions in brain, PTV at the frontal lobe (PTV frontal), PTV near thalamus (PTV thalamus) and PTV near globus pallidus (PTV GP) were irradiated using three noncoplanar arcs with a single isocenter. Figure 10 shows the measured dose distribution in 3D using 3D Slicer. In this case, high gradient F I G . 7. Three orthogonal views of the measured (red) and TPS-calculated (blue) dose distribution comparison.
WANG ET AL.
| 29 dose regions were scattered at different places to cover three targets. Both OAR, chiasm and brain stem were in the low dose region.
The proposed method not only assessed the dose coverage of individual lesions but also the dose fall-off outside the targets and low dose spill into normal brain.

3.B | Registration error estimation
Essential factors affecting FLE and FRE are the image features of the fiducial markers in both image sets (OCT and CT images). In  Figure 16 shows the isovalue lines of TRE in the axial and coronal views. Due to symmetric configuration of the fiducial points, the results in coronal view are the same as those in sagittal view. In the region of measured dose distribution, which is usually at the center of the dosimeter, the estimated TRE is smaller than 0.3 mm. In addition, TRE was estimated by analyzing 10 fiducial markers, previously registered as targets. After registration, all of them are shown to be at the same coordinates in the CT and OCT images. We were unable to measure submillimeter registration errors because the resolution limit of treatment planning exported dose images and CT images is 1 mm.

| DISCUSSION
A robust methodology of 3D IMRT QA using point-based registration and structure mapping was proposed in this study, which aims to improve the correlation between IMRT QA evaluation and the    The sensitivity and specificity of an IMRT QA method to detect planning or delivery errors relates to the uncertainties of the whole QA procedure. Therefore, the source and magnitude of the uncertainties should be estimated. More significant uncertainties than the errors to be detected could result in a high rate of false positives.
The sources of uncertainties of the proposed IMRT QA method include fiducial registration, dose measurements, structures mapping and dosimeter setup. Using the pixel-to-pixel mapping of the Eclipse treatment planning system, the uncertainty from structure mapping is negligible. The dosimeter setup error relates to the laser error and operator error, which is similar to all the measurement-based IMRT QA methods. In this study, errors from fiducial registration were analyzed. FLE, FRE, and TRE were estimated to be less than a millimeter. TRE of pixels in 3D space of the dosimeter was calculated to be smaller than 0.3 mm. The highest resolution of Eclipse treatment planning dose calculation is 1 mm. Therefore, the proposed fiducial markers and configuration can provide sufficient accuracy for dose comparison.
The PRESAGE dosimeter is accurate in terms of relative dose distribution measurement but is not ready for absolute dosimetry. The selection of the normalization point of the measured dose distribution could affect the interpretation of the results. In this study, the normalization point was chosen to be in a uniform high dose region. Moreover there are differences between the inhomogeneity of the real patient and the dosimeter, and thus the magnitude of the discrepancy between the measurement and the calculation evaluated using the phantom could be different than the real discrepancy in the patients. This is the same as all the other measurement-based IMRT QA methods used routinely in clinical practice. To improve the correlation, phantom size and shape should be close to patient's geometry. As 3D printing becomes more common and low-cost, patient-specific phantom could be utilized for radiotherapy dosimetry. 27 This work has provided a clinically feasible methodology utilizing an automatic fiducial registration algorithm and commercially available structure-mapping application in clinical TPS, which is a step toward the implementation of a foolproof 3D dosimetric verification system with organ-specific dose information for routine clinical use.
With the acquired information, organ-specific dose difference criterion could be implemented in the future. Moreover our study adds on to the current methods for 3D dosimetric analysis by reporting the registration error as part of the dose comparison error. More convenient, user-independent and time-efficient optical scanners and programs are being developed at our lab so that 3D dosimetry can become clinically available and easily accessible in the future.

| CONCLUSIONS
In this study, we introduced a robust methodology of 3D measurement-based IMRT QA for organ-specific dose comparison. With accurate point-based registration between measured and calculated image spaces, a precise spatial correlation between the two can be found. In addition, the patient's anatomical structures can be mapped to the CT images of the phantom using the coordinates of the fiducials. This work demonstrates two clinical cases and shows the capability of 3D organ-specific dose comparison. In addition, a comprehensive analysis of the registration uncertainties was performed. This work aims to improve the current 2D measurement based IMRT QA and shows the clinical feasibility of 3D dosimetry for future use.

CONFLI CT OF INTEREST
The authors declare no conflict of interest.

AUTHOR CONTRI BUTIONS
Yi-Fang Wang : Contribution Statement: The author made substantial contribution to conception and design, drafting the article, analysis, interpretation of data, and final approval of the version to be submit-