Accuracy of low‐dose proton CT image registration for pretreatment alignment verification in reference to planning proton CT

Abstract Purpose Proton CT (pCT) has the ability to reduce inherent uncertainties in proton treatment by directly measuring the relative proton stopping power with respect to water, thereby avoiding the uncertain conversion of X‐ray CT Hounsfield unit to relative stopping power and the deleterious effect of X‐ ray CT artifacts. The purpose of this work was to further evaluate the potential of pCT for pretreatment positioning using experimental pCT data of a head phantom. Methods The performance of a 3D image registration algorithm was tested with pCT reconstructions of a pediatric head phantom. A planning pCT simulation scan of the phantom was obtained with 200 MeV protons and reconstructed with a 3D filtered back projection (FBP) algorithm followed by iterative reconstruction and a representative pretreatment pCT scan was reconstructed with FBP only to save reconstruction time. The pretreatment pCT scan was rigidly transformed by prescribing random errors with six degrees of freedom or deformed by the deformation field derived from a head and neck cancer patient to the pretreatment pCT reconstruction, respectively. After applying the rigid or deformable image registration algorithm to retrieve the original pCT image before transformation, the accuracy of the registration was assessed. To simulate very low‐dose imaging for patient setup, the proton CT images were reconstructed with 100%, 50%, 25%, and 12.5% of the total number of histories of the original planning pCT simulation scan, respectively. Results The residual errors in image registration were lower than 1 mm and 1° of magnitude regardless of the anatomic directions and imaging dose. The mean residual errors ranges found for rigid image registration were from −0.29 ± 0.09 to 0.51 ± 0.50 mm for translations and from −0.05 ± 0.13 to 0.08 ± 0.08 degrees for rotations. The percentages of sub‐millimetric errors found, for deformable image registration, were between 63.5% and 100%. Conclusion This experimental head phantom study demonstrated the potential of low‐dose pCT imaging for 3D image registration. Further work is needed to confirm the value pCT for pretreatment image‐guided proton therapy.


| INTRODUCTION
Proton therapy provides superior dose distributions in the low to intermediate dose range compared to photon therapy, which may lead to improved outcomes for some types of cancer and reduced side effects. [1][2][3] Uncertainties in patient positioning and beam range as well as internal changes of tumor and patient anatomy could, however, compromise treatment effectiveness. 4 Therefore, efforts to develop and improve treatment planning accuracy and image guidance for proton therapy are ongoing. 5,6 Currently, for treatment planning in proton therapy, an X-ray CT dataset of the patient is acquired and Hounsfield units of the scan are converted to relative stopping power (RSP). This conversion is one important source for range uncertainties, which are typically estimated on the order of 3-5% of the planned proton range. 7 Replacing X-ray planning CT with proton CT (pCT) planning CT simulations with individual proton tracking during the scan has been proposed as a low-dose method to reduce this planning uncertainty; pretreatment pCT would also provide a method for pretreatment verification of correct patient setup and RSP distribution. This method is currently in the preclinical stage of its development. [8][9][10] The potential advantages of pCT for image guidance in the treatment room are several-fold: (a) There is a dose advantage compared to X-ray cone-beam CT (CBCT) and (b) there is absence of artifacts often present in X-ray CT based reconstructions; (c) using the same radiation source would allow imaging the patient immediately before treatment in the treatment position; (D) finally, the largest advantage of pCT would be that it could detect range errors before treatment in addition to serving as a low-dose alignment technique that could replace CBCT. Therefore, daily 3D verification of patient alignment relative to the proton beam and confirmation that the RSP distribution on the beam path has not changed from the original treatment plan could be a valuable development for proton therapy, as it would allow better treatment accuracy and narrower margins, especially for hypofractionated treatment schedules.
Proton CT based on individual particle tracking utilizes position and direction information of the protons before and after the patient and measures the energy deposited by protons that traversed the object in a scintillator. Using this information from many protons, typically of the order of 100 protons per cm 2 , coming in from many discrete or continuous directions, one can reconstruct the distribution of the RSP with sufficient spatial resolution. 10 One of the challenges in proton imaging is the degraded spatial resolution due to multiple Coulomb scattering (MCS) inside the imaged object. To improve the resolution, several most likely path (MLP) formulations have been proposed and are used in pCT image reconstruction. [11][12][13] Iterative algorithms can then be used to reconstruct 3D pCT images from radiological projections. With these developments, including fast parallel processing of the acquired pCT data, a clinical setting for pCT appears feasible.
The purpose of this study was to evaluate the performance of pCT for pretreatment image guidance using rigid and deformable image registration algorithms. A high-quality planning CT simulation scan was created by experimentally scanning a head phantom and a reconstruction algorithm using all available proton histories and FBP as initial iterate followed by an iterative reconstruction algorithm. In addition, pretreatment pCT scans were generated for different imaging doses by selecting different number of proton histories entering the reconstruction and using only fast FBP as the reconstruction method. These pretreatment scans were then rigidly transformed by prescribing random 3D errors (rotations and translations) to simulate random alignment errors. The study endpoint was the accuracy of the image registration algorithm in recovering the original planning pCT simulation scan as a function of the different imaging dose levels. In the second part of the study, a deformation field derived from a real patient was applied (a) to the original planning pCT study to simulate a deformed pretreatment pCT using all histories and FPB plus iterative reconstruction that could be used for replanning and (b) to the FBPonly reconstructed preplanning pCT scans to simulate the accuracy of registration in the presence of deformation and at different doses.

2.A | Proton CT scanner and study design
The prototype pCT scanner, built by the pCT collaboration was used for this work (Fig. 1). It consists of a front and rear tracker system used to extrapolate the proton path before and after the object and a ton gantry, the current prototype pCT scanner is this limited to registering about 1 million protons per second. In a future implementation, the pCT scanner acquisition rate will be increased by a factor 2-3, making it possible to acquire the scan in 2-3 rotations at 1 rpm. The tracker and MSS data of individual protons were read out by a custom high-speed data acquisition (DAQ) system, capable of handling data rates on the order of 1 million protons/sec. 8,10 To determine WEPL, the MSS detector response was calibrated using a step-phantom of known water-equivalent thickness. 14 For high-fidelity treatment planning pCT simulations, a 3D filtered back projection (FBP) algorithm was employed initially to determine the object boundaries; subsequently it was used as the first iterate for the subsequent iterative image reconstruction. The reconstruction for the planning pCT simulation was achieved in under 7 min with high-performance computing. 15 The FBP without further refinements of RSP values by iterative reconstruction was obtained in under 1 min, and was used for image registration in a pretreatment situation (pretreatment pCT).
Image registration (IR) of the pretreatment pCT scan to the original planning pCT simulation was used to determine the spatial transform for the alignment of the head phantom after the study had been intentionally been transformed by a random 3D vector and three random rotations about the cardinal axes. A rigid IR procedure was used for finding three translations and rotation angles that realigned the pretreatment pCT to the original planning pCT simulation.

2.B | Experimental pCT data
For the planning pCT simulation scan, 90 projections of the pediatric head phantom (model 715-HN, CIRS) were obtained with the prototype proton CT scanner. 10 The pCT data processing and image reconstruction steps are as follows. The acquired pCT data (histories) are checked for completeness and consistency and then converted to tracker coordinates and MSS response values. A pre-scan WEPL calibration scan with a calibration object is used to construct a calibrated relationship to convert MSS responses to WEPL values. Since the active tracker area is 9 cm in cranio-caudal direction, two successive scans of the head phantom were obtained with a longitudinal shift of the phantom of about 8 cm between the two scans. For each scan, a total number of about 200 M protons entered the reconstruction process. For the planning pCT simulation scan the 3D FBP algorithm was used as the initial step producing an initial approximate solution followed by five iterations of the total-variation superiorization diagonally relaxed projections (TVS-DROP) algorithm described and used for pCT reconstruction previously 16 (Fig. 3).   17 Mattes mutual information, 18 often applied for multimodality images, was used as the similarity metric. The intrinsic advantage of this method is image rescaling when the discrete density function is built. 19 This metric tends to map homogeneous regions from the moving image into homogeneous regions of the

2.E | Performance evaluation
To evaluate the accuracy of rigid registration using pCT scans, 10 random 6-degree-of-freedom (DOF) transformations (translation and rotation) were created using orthogonal sampling 21 and applied to each set of images to be registered. The images were then resampled using the Lanczos filter in the Amira 3D software platform (ver-  The parameters can be changed; therefore, it is possible to decide how much similarity is enough to stop the IR process. In our case, the minimum step length of 0.001 (as suggested on ITK documentation examples) was maintained and was found to be sufficient to reach clinical accuracy of the procedure in an acceptable time.
To evaluate the accuracy of deformable registration using pCT scans, a realistic deformation field was obtained from the planning X-ray CT and subsequent cone beam CT of a real patient treated with radiotherapy for head and neck cancer. The deformation field was then applied (a) to the original planning CT simulation study to represent a high quality pCT study of a realistically deformed phantom at the time of treatment, (Fig. 4), and (b) to the pCT FBP images reconstructed at 100%, 50%, 25%, and 12.5% of the total dose of the planning pCT scan to simulate fast/low dose pretreatment pCT image reconstructions. The individual image sets were then deformably registered to the original planning CT simulation study.
After registering each pair of images, the scale invariant feature transform (SIFT) 22 was used to extract features and to calculate the residual 3D distance between corresponding landmarks to numerically assess the quality of the registration. 23 The deformable image registration (DIR) procedures were carried out on the same notebook as rigid transformation procedures: the

3.A | Rigid registration
After the registration procedure, the differences between imposed errors and suggested corrections were calculated. The mean and standard deviation values of the residual distance for the 10 different simulated shifts for each IR modality are summarized in Table 3 for translation and in Table 4

3.B | Deformable registration
After the deformable registration procedure, on average, 44 corresponding markers between the fixed and the transformed image were identified using SIFT (Fig. 7) for the pCT images. The percentage of sub-millimetric errors of the residual distance between landmarks calculated for each case after DIR are presented in Table 5. An example of images before and after DIR is presented in Fig. 8.

| DISCUSSION
Image registration is an important aspect of image-guided radiotherapy, and is particularly important for accurate proton therapy. In this work, we explored in an initial, admittedly limited experimental study, the use of a preclinical prototype pCT scanner for pretreatment alignment with a head phantom. Proton CT requires highenergy protons to traverse the patients for imaging. At this point, the pCT method is limited to head and neck applications but is expected to also work for most patients in the thorax region; remaining body regions (pelvis and abdomen) would require energies in excess of 250 MeV, which are currently not clinically available, but should become available soon. For body scans, the use of helium ions would be more advantageous since it is less effected by MCS.
F I G . 4. Sagittal mid-plane reconstruction after a patient-specific deformation field was applied to the planning pCT simulation study.  the pCT scanner can be used for obtaining proton 2D projections to be used on a 2D-3D registration. This a procedure analogous to the currently used method of registering X-ray DRRs from the planning CT to two in-room orthogonal X-ray projections.
In summary, this was the first study of using pCT for planning and pretreatment patient alignment. Our study was limited to a single pCT study that was mathematically modified. The next step in this research will be to perform a more realistic study with an actually modified head phantom position, deformation, and changes in RSP values registered to an original pCT planning simulation scan.

| CONCLUSION
This work demonstrated the potential of 3D head image registration based on proton CT for in-room pretreatment verification. The developed algorithms for image registration can be accurate even at very low proton imaging doses. Nevertheless, the alignment could be influenced by image artifacts that were introduced by the fast filtered back projection reconstruction.

ACKNOWLEDG MENT
The development of proton CT technology has been made possible by the contribution of many individuals. For this work, we particu-