Purpose: To implement and validate the accuracy of an intensity‐based 2D to 3D rigid registration algorithm for proton treatment systems. Methods: A 2D‐3D rigid registration algorithm (REG23) was previously validated for linac machines. We adapted REG23 for use in both gantry‐based and Stereotactic Alignment in Radiosurgery (STAR) proton treatment systems. REG23 registration was run on NVIDIA Quadra600 GPU card using NCC metrics and AMOEBA optimizer and validated using dual orthogonal kV images acquired during cranial target treatments: 66 fractions to 7 patients in gantry system and 135 fractions to 9 patients in STAR. Rectangle ROIs covering the whole region superior to the base‐of‐skull were used for REG23 registration and the accuracy was evaluated using clinical utilized fiducial‐based 2D‐3D ray back‐projection rigid registration as the baseline. Results: For the gantry system, the differences between REG23 and baseline were ‐ 0.01±0.63 mm, 0.43±0.61 mm, 0.12±0.60 mm in the left‐right, superior‐inferior, and anterior‐posterior directions; 0.15±0.44°, 0.11±0.23°, and 0.19±0.37° in pitch, roll, and yaw, respectively. The vector difference was 1.04±0.48 mm. For STAR system, the REG23 results were − 0.34±0.45 mm, −0.10±0.36 mm, −0.38±0.75 mm, −0.14±0.54°, 0.04±0.18°, and 0.03±0.35° different than the baseline in the left‐right, superior‐inferior, and anterior‐posterior directions, pitch, roll, and yaw. The vector difference was 0.93±0.53 mm. The time for REG23 automatic registration was 29.7±9.5 seconds. Conclusion: We demonstrated that the intensity based 2D‐3D rigid registration algorithm REG23 provided sub‐millimeter accuracy and better than 0.5° for both gantry‐based and STAR proton treatment systems for cranial patients. This accuracy is within patient setup tolerance for current fractionated proton treatment systems. Given its accuracy and efficiency, we believe REG23 has great potential for clinical utilization in proton radiation therapy by replacing the implanted fiducial or anatomic feature based patient setup method.The project was supported by the Federal Share of program income earned by Massachusetts General Hospital on C06 CA059267, Proton Therapy Research and Treatment Center.

S1142 nd ESTRO Forum 2013 CT may hamper contouring and dose calculation accuracy.Although not in widespread routine use, deformable image registration (DIR), preferable unsupervised, is used for auto-segmentation of organs/volumes based on those delineated on a reference planning CT.Using the daily pre-treatment registrations between planning CT and CBCT the delivered dose at each fraction can be recalculated on the CBCT anatomy with proper calibration of the systems.Based on DIR, accumulation of the dose can be performed from day to day, and delivered dose can be compared directly with the planned dose, and dose-volume histograms for OAR can be evaluated.Alternatively, the consecutive DIRs can be combined to represent a mean deformation vector field which can be applied to the planning CT with planned dose distribution to obtain a measure of delivered dose.For doseresponse relations, the accumulated delivered dose needs to be combined with validated clinical outcome measures.

Purpose: To determine the robustness and accuracy of an open source 2D/3D GPU accelerated image registration algorithm in the context of cranial image guided radiotherapy. Methods: The open source 2D/3D image registration algorithm, Reg23, has been released under the GNU license. The algorithm utilizes an iterative digitally reconstructed radiograph (DRR) approach to the image registration problem. The DRR generator is accelerated on a GPU to rapidly iterate the optimization process. Multiple cost functions are supported and were analyzed. Robustness was determined by comparing a baseline set of orthogonal kV images of a cranial phantom with a predetermined isocenter to the planned isocenter in the CT image set and introducing more than 6000 combinations of rotations and translation to the position of the isocenter in the CT. Accuracy and time efficiency of various cost function were analyzed for the virtual patient shifts. Furthermore, a set of 43 experimental orthogonal images were acquired with a linac mounted kV imaging system of predetermined physical shifts which were compared to the results of the Reg23 algorithm. Results: The Reg23 algorithm was found to be accurate to 0.04±0.02mm for the virtual isocenter shifts and 0.23±0.40mm for real images compared to the CBCT registration results. Time to solution could be reduced from >70 s to < 40 s without a significant change in the algorithm accuracy depending upon the cost function employed. Conclusions: The Reg23 algorithm is robust and sensitive to sub‐mm variations of virtual shifts of the isocenter position. The Normalized Cross Correlation (NCC) cost function was determined to be most accurate and fastest for cranial image registration. For real experimental data, the Gradent Difference (GD) cost function was most accurate and both GD and NCC delivered results accurate to within 0.5 mm and 0.4° when compared to CBCT/CT registrations.

Image-guided alignment procedures in radiotherapy aim at minimizing discrepancies between the planned and the real patient setup. For that purpose, we developed a 2D/3D approach which rigidly registers a computed tomography (CT) with two x-rays by maximizing the agreement in pixel intensity between the x-rays and the corresponding reconstructed radiographs from the CT. Moreover, the algorithm selects regions of interest (masks) in the x-rays based on 3D segmentations from the pre-planning stage. For validation, orthogonal x-ray pairs from different viewing directions of 80 pelvic cone-beam CT (CBCT) raw data sets were used. The 2D/3D results were compared to corresponding standard 3D/3D CBCT-to-CT alignments. Outcome over 8400 2D/3D experiments showed that parametric errors in root mean square were <0.18° (rotations) and <0.73 mm (translations), respectively, using rank correlation as intensity metric. This corresponds to a mean target registration error, related to the voxels of the lesser pelvis, of <2 mm in 94.1% of the cases. From the results we conclude that 2D/3D registration based on sequentially acquired orthogonal x-rays of the pelvis is a viable alternative to CBCT-based approaches if rigid alignment on bony anatomy is sufficient, no volumetric intra-interventional data set is required and the expected error range fits the individual treatment prescription.

We present a novel view on 2D/3D image registration by introducing a generic algorithmic framework that is based on supervised machine learning (SML). First and foremost, this class of algorithms, referred to as texture model registration (TMR), aims at making 2D/3D registration applicable for time-critical image guided medical procedures. TMR methods are two-stage. In a first offline pre-computational stage, a prediction rule is derived from a pre-interventional 3D image and according geometric constraints. This is achieved by computing digitally reconstructed radiographs, pre-processing them, extracting their texture, and applying SML methods. In a second online stage, the inferred rule is used for predicting the spatial rigid transformation of unseen intrainterventional 2D images. A first simple concrete TMR implementation, referred to as TMR-PCR, is introduced. This approach involves principal component regression (PCR) and simple intermediate pre-processing steps. Using TMR-PCR, first experimental results on five clinical IGRT 3D data sets and synthetic intra-interventional images are presented. The implementation showed an average registration rate of 48 Hz over 40000 registrations, and succeeded in the majority of cases with a mean target registration error smaller than 2 mm. Finally, the potential and characteristics of the proposed methodical framework are discussed.

Recently, Birkfellner et al. proposed a novel image-to-image merit function (stochastic rank correlation, SRC) for robust intensity-based 2D/3D image registration. In this work, we summarize the basic idea of SRC, and present a generic ITK-based implementation of this image-to-image metric including tests for software verification. Moreover, we provide two simple examples that demonstrate the usage of this metric: a) within the native ITK 2D/3D image registration method, and b) within a recently published extended ITK-based 2D/3D registration framework. It is, however, important to note, that this paper neither covers a comprehensive evaluation of SRC, nor a comparison with other metrics. It rather shows that SRC appears to succeed on a femoral and a porcine data set in the course of ITK-based 2D/3D image registration.