Ke Nie

Site‐specific deformable imaging registration algorithm selection using patient‐based simulated deformations

PURPOSE The accuracy of deformable image registration could have a significant dosimetric impact in radiation treatment planning. Various image registration algorithms have been developed for clinical application. However, validation of these algorithms in the current clinical setting remains subjective, relying on visual assessment and lacking a comparison to the ground-truth deformation. In this study, the authors propose a framework to quantitatively validate various image registration solutions by using patient-based synthetic quality assurance (QA) phantoms, which can be applied on a site-by-site basis. METHODS The computer-simulated deformation was first generated with virtual deformation QA software and further benchmarked using a physical pelvic phantom that was modeled after real patient CT images. After the validity of the virtual deformation was confirmed, a set of synthetic deformable images was produced to simulate various anatomical movements during radiotherapy based on real patient CT images. Three patients with head-and-neck, prostate, and spine cancer were included. The transformations included bladder filling, soft tissue deformation, mandible, and vertebral body movement, etc., which provided various ground-truth images to validate deformable registration. Several clinically available deformable registration algorithms were tested on these images with multiple registration setups, such as global or regional and single-pass or multipass optimization. The generated deformation fields and the ground-truth deformation are compared using voxel-by-voxel based analysis as well as regional based analysis. RESULTS Performance of registration algorithms varies with clinical sites. The voxel-by-voxel analysis showed the intensity-based free-form deformation by MIM generated the greatest accuracy for low-contrast small regions that underwent significant deformation, such as bladder expansion for prostate. However, for large field deformations with strong contrast, this approach may increase errors, which is especially evident in the cranial spinal irradiation (CSI) case. Both single-pass and multipass B-spline registrations performed well for the head-and-neck patient and CSI patients. CONCLUSIONS QA for deformable image registration is essential to verify the cumulated dose for accurate adaptive radiotherapy. In this study, the authors develop a workflow that can validate image registration techniques for several different clinical sites and for various types of deformations using patient-based simulated deformations. This work could provide a reference for clinicians and radiation physicists on how to choose appropriate image registration algorithms for different situations.

Rectal Cancer: Assessment of Neoadjuvant Chemoradiation Outcome based on Radiomics of Multiparametric MRI.

Purpose: To evaluate multiparametric MRI features in predicting pathologic response after preoperative chemoradiation therapy (CRT) for locally advanced rectal cancer (LARC). Experimental Design: Forty-eight consecutive patients (January 2012–November 2014) receiving neoadjuvant CRT were enrolled. All underwent anatomical T1/T2, diffusion-weighted MRI (DWI) and dynamic contrast-enhanced (DCE) MRI before CRT. A total of 103 imaging features, analyzed using both volume-averaged and voxelized methods, were extracted for each patient. Univariate analyses were performed to evaluate the capability of each individual parameter in predicting pathologic complete response (pCR) or good response (GR) evaluated based on tumor regression grade. Artificial neural network with 4-fold validation technique was further utilized to select the best predictor sets to classify different response groups and the predictive performance was calculated using receiver operating characteristic (ROC) curves. Results: The conventional volume-averaged analysis could provide an area under ROC curve (AUC) ranging from 0.54 to 0.73 in predicting pCR. While if the models were replaced by voxelized heterogeneity analysis, the prediction accuracy measured by AUC could be improved to 0.71–0.79. Similar results were found for GR prediction. In addition, each subcategory images could generate moderate power in predicting the response, which if combining all information together, the AUC could be further improved to 0.84 for pCR and 0.89 for GR prediction, respectively. Conclusions: Through a systematic analysis of multiparametric MR imaging features, we are able to build models with improved predictive value over conventional imaging metrics. The results are encouraging, suggesting the wealth of imaging radiomics should be further explored to help tailoring the treatment into the era of personalized medicine. Clin Cancer Res; 22(21); 5256–64. ©2016 AACR.

An automated deformable image registration evaluation of confidence tool.

Deformable image registration (DIR) is a powerful tool for radiation oncology, but it can produce errors. Beyond this, DIR accuracy is not a fixed quantity and varies on a case-by-case basis. The purpose of this study is to explore the possibility of an automated program to create a patient- and voxel-specific evaluation of DIR accuracy. AUTODIRECT is a software tool that was developed to perform this evaluation for the application of a clinical DIR algorithm to a set of patient images. In brief, AUTODIRECT uses algorithms to generate deformations and applies them to these images (along with processing) to generate sets of test images, with known deformations that are similar to the actual ones and with realistic noise properties. The clinical DIR algorithm is applied to these test image sets (currently 4). From these tests, AUTODIRECT generates spatial and dose uncertainty estimates for each image voxel based on a Students t distribution. In this study, four commercially available DIR algorithms were used to deform a dose distribution associated with a virtual pelvic phantom image set, and AUTODIRECT was used to generate dose uncertainty estimates for each deformation. The virtual phantom image set has a known ground-truth deformation, so the true dose-warping errors of the DIR algorithms were also known. AUTODIRECT predicted error patterns that closely matched the actual error spatial distribution. On average AUTODIRECT overestimated the magnitude of the dose errors, but tuning the AUTODIRECT algorithms should improve agreement. This proof-of-principle test demonstrates the potential for the AUTODIRECT algorithm as an empirical method to predict DIR errors.

Performance variations among clinically available deformable image registration tools in adaptive radiotherapy — how should we evaluate and interpret the result?

The purpose of this study is to evaluate the performance variations in commercial deformable image registration (DIR) tools for adaptive radiation therapy and further to interpret the differences using clinically available terms. Three clinical examples (prostate, head and neck (HN), and cranial spinal irradiation (CSI) with L‐spine boost) were evaluated in this study. Firstly, computerized deformed CT images were generated using simulation QA software with virtual deformations of bladder filling (prostate), neck flexion/bite‐block repositioning/tumor shrinkage (HN), and vertebral body rotation (CSI). The corresponding transformation matrices served as a “reference” for the following comparisons. Three commercialized DIR algorithms: the free‐form deformation from MIMVista 5.5 and the RegRefine from MIMMaestro 6.0, the multipass B‐spline from VelocityAI v3.0.1, and the adaptive demons from OnQ rts 2.1.15, were applied between the initial images and the deformed CT sets. The generated adaptive contours and dose distributions were compared with the “reference” and among each other. The performance in transferring contours was comparable among all three tools with an average Dice similarity coefficient of 0.81 for all the organs. However, the dose warping accuracy appeared to rely on the evaluation end points and methodologies. Point‐dose differences could show a difference of up to 23.3 Gy inside the PTVs and to overestimate up to 13.2 Gy for OARs, which was substantial for a 72 Gy prescription dose. Dosevolume histogram‐based evaluation might not be sensitive enough to illustrate all the detailed variations, while isodose assessment on a slice‐by‐slice basis could be tedious. We further explored the possibility of using 3D gamma index analysis for warping dose variation assessment, and observed differences in dose warping using different DIR tools. Overall, our results demonstrated that evaluation based only on the performance of contour transformation could not guarantee the accuracy in dose warping, while dose‐transferring validation strongly relied on the evaluation endpoint. As dose‐transferring errors could cause misinterpretations when attempting to accumulate dose for adaptive radiation therapy and more DIR tools are available for clinical use, a standard and clinically meaningful quality assurance criterion should be established for DIR QA in the near future. PACS number(s): 87.57

SU-E-J-263: Dosimetric Analysis On Breast Brachytherapy Based On Deformable Image Registration

PURPOSE To quantitatively compare and evaluate the dosimetry difference between breast brachytherapy protocols with different fractionation using deformable image registration. METHODS The accumulative dose distribution for multiple breast brachytherapy patients using four different applicators: Contura, Mammosite, Savi, and interstitial catheters, under two treatment protocols: 340cGy by 10 fractions in 5 days and 825cGy by 3 fractions in 2days has been reconstructed using a two stage deformable image registration approach. For all patients, daily CT was acquired with the same slice thickness (2.5mm). In the first stage, the daily CT images were rigidly registered to the initial planning CT using the registration module in Eclipse (Varian) to align the applicators. In the second stage, the tissues surrounding the applicator in the rigidly registered daily CT image were non-rigidly registered to the initial CT using a combination of image force and the local constraint that enforce zero normal motion on the surface of the applicator, using a software developed in house. We calculated the dose distribution in the daily CTs and deformed them using the final registration to convert into the image domain of the initial planning CT. The accumulative dose distributions were evaluated by dosimetry parameters including D90, V150 and V200, as well as DVH. RESULTS Dose reconstruction results showed that the two day treatment has a significant dosimetry improvement over the five day protocols. An average daily drop of D90 at 1.3% of the prescription dose has been observed on multiple brachytherapy patients. There is no significant difference on V150 and V200 between those two protocols. CONCLUSION Brachytherapy with higher fractional dose and less fractions has an improved performance on being conformal to the dose distribution in the initial plan. Elongated brachytherapy treatments need to consider the dose uncertainty caused by the temporal changes of the soft tissue.

SU-C-18A-07: The Importance of Image Processing for Simulated Deformations.

PURPOSE Deformations can be digitally applied to patient images to evaluate the accuracy of deformable image registration (DIR). However, this can also deform image noise and artifacts, leaving a fingerprint of the underlying deformation that could skew accuracy determination. Image processing can be used to erase this fingerprint and create a more realistic DIR test scenario. The importance of image processing to simulated deformations is tested here. METHODS These tests utilize a virtual pelvic phantom, made from a patient CT image, and a pelvic-shaped water phantom to acquire noise signatures. Two image-filtering techniques are tested here: a spatial convolution with a Gaussian (SCG) and an edgepreserving filter (EPF) that preferentially removes the Fourier components associated with noise. Four different processing scenarios are evaluated here. The first is no processing (NP). For the second, noise from the water phantom is added to the set of test images without filtering (NF). The third and fourth scenarios add noise after applying the SCG and EPF filtering methods. EPF provides the most realistic test scenario. These processing scenarios are tested for their effect on the spatial accuracy of the DIR algorithms from MIM Software and Velocity Medical Solution. RESULTS For NP, NF, SCG, and EPF, the mean errors from MIM were 0.78, 1.11, 1.44, and 1.24 mm, respectively. The corresponding maximum errors for MIM were 21.7, 18.8, 26.1, and 21.1 mm, respectively. Velocity was relatively insensitive to these different processing scenarios, where the mean errors ranged from 1.66 to 1.74 mm and the maximum errors from 12.1 to 12.4 mm. CONCLUSION Velocity creates globally smooth deformations, whereas MIM exhibits much more local pliability. This local pliability makes MIM sensitive to differences in image processing. Thus, an objective evaluation of its accuracy with simulated deformations must utilize realistic noise scenarios, with carefully balanced image processing.

TU-C-141-09: An Automated Workflow for Patient-Specific Verification of Deformable Image Registration

Purpose: To develop an automated workflow for patient‐specific verification of deformable image registration (DIR). Methods: The central concept behind this verification process is digitally applying test deformations to patient images and comparing to those returned from clinical DIR algorithms. The patient‐specific nature of these test deformations is the novel feature of this workflow. First, a set of different test DIR algorithms is applied to the original clinical images. The deformations produced by these algorithms represent plausible deformations and are applied to the clinical moving image, producing a set of test fixed images. A combination of computed‐tomography images of a water phantom with frequency‐domain image filtering is utilized to first remove image noise and then reintroduce distinct noise fluctuations on the moving and fixed test images, which would otherwise skew DIR error evaluation. The clinical algorithm is then applied to the set of test images, which yields a set of theoretical deformations. For each image voxel, the set of theoretical deformations are compared to those actually applied. The maximum discrepancy in each direction becomes the theoretical uncertainty for that voxel on the clinical image. This creates an uncertainty vector field, which can be applied to patient contours and dose to evaluate potential clinical errors from DIR uncertainty. The results of this process are benchmarked against an anthropomorphic pelvic phantom, which possesses a measured ground‐truth deformation. Results: The theoretical DIR uncertainties from the workflow gave an accurate indication for the spatial location of the actual errors, but tended to overestimate the error magnitude. The maximum theoretical uncertainties in the anterior and lateral directions were 15.4 and 11.5 mm, respectively, compared to the maximum measured errors of 12.6 and 3.0 mm, respectively. Conclusion: This workflow demonstrates the ability to perform automated patient‐specific verification for DIR. Further work to develop the method is ongoing.

WE‐E‐108‐10: Validating a 192Ir‐Based Small Animal Irradiation Apparatus Using a 3D‐Printed Applicator: Comparison Between TG‐43, Monte Carlo and Films Dosimetry

PURPOSE Accurate assessment of dose delivered is key to early prediction of radiation-induced anatomic and physiologic changes vital in providing the most accurate patient-specific treatments. We are conducting a translational small-animal study using functional Hyperpolarized-13C-Urea with DCE-MRI to evaluate tumor perfusion changes following local targeted Ir-192 irradiation using the Leipzig applicator. The purpose of this study is to present and evaluate a novel Monte Carlo (MC) tool, which provides heterogeneous dose predictions for this specific application, and to compare the results against the TG-43 dose calculation. METHODS CT scans were obtained with a Leipzig applicator mold (fabricated using a 3D-printer to avoid metal artifact) centrally placed on top of the CyberKnife Ball Cube QA phantom . The CT density of the mold was overridden to be that of the applicator material, Tungsten. A Monte Carlo platform, ALGEBRA (ALgorithm for heterogeneous dosimetry based on GEANT4 for BRAchytherapy) was used for simulation. Dose measurements were done using Gafchromic EBT2 film placed orthogonally inside the Ball Cube exposed to Ir-192 with the Leipzig applicator in place. Simple TG-43 calculations for a free source in water was done using the Oncentra treatment planning system. RESULTS The two-dimensional planar dose distributions obtained from MC simulation showed strong agreement (within 4-5%) with dose measurements while TG-43 calculations showed differences up to 20%. Compared to TG-43, the MC Result was attenuated less at the surface because of the Leipzig air cavity, but penetrated less due to the collimation effect. CONCLUSION We have validated a new MC simulation tool using a Leipzig applicator for small animal irradiation. This tool will provide the basis for studies associating tumor response with the actual dose delivered. The tool can be further used to construct dosimetry information for clinical treatments using the Leipzig applicator as an alternative to superficial/orthovoltage radiation treatment.

SU‐D‐211‐04: Sector Intensity Modulated (SIM) Gamma Knife Stereotactic Radiosurgery

PURPOSE The latest Gamma Knife (GK) system, Perfexion, consists of 192 Co-60 sources divided into eight sectors. Treatment delivery includes multiple shots placed at different positions. For every shot, each sector can be either blocked or open with four different aperture sizes. However, the beam-on time is designed to be fixed. We proposed an innovative concept, Sector Intensity Modulated (SIM) Gamma Knife by dynamically varying the beam-on time for each individual sector to improve stereotactic radiosurgery planning quality. METHODS The anatomic structures and dose matrices from each sector for every shot were obtained from the GK workstation. The beam-on time for each sector was decomposed with various discrete levels and brute-force algorithm was used to get the optimal solution. The resulting SIM plan was then re-entered into the GK workstation. Six indices were used to benchmark the plan quality: Coverage, Conformality, Gradient, Maximum Dose(s) to critical structure(s), Volume receiving over 8 and 12 Gy. All the SIM plans in comparison with the original plans were further reviewed by an experienced oncologist. RESULTS The simulations were tested on various pituitary adenoma cases. Results consistently showed that SIM yielded better plans with all quantitative indices improved compared to original plan. It provides better conformality, quicker drop off of the isodose line outside the tumor, lower doses to the critical structures as optical- nerve/chiasm while maintaining at least 99% coverage of the tumor. Results were more favorable according to oncologists view. In particular, up to 20% or 0.6 cc volume decrease in healthy tissue receiving 8 Gy was observed. This may translate into clinically observable reduction in acute/late toxicities. CONCLUSIONS Our preliminary results show that Sector Intensity Modulated Gamma Knife offers superior treatment plans compared to the originally delivered plans. Further works as adding dynamic shot location and dynamic shot shaping will be discussed.

Dosimetric feasibility of 4DCT-ventilation imaging guided proton therapy for locally advanced non-small-cell lung cancer

BackgroundThe principle aim of this study is to incorporate 4DCT ventilation imaging into functional treatment planning that preserves high-functioning lung with both double scattering and scanning beam techniques in proton therapy.MethodsEight patients with locally advanced non-small-cell lung cancer were included in this study. Deformable image registration was performed for each patient on their planning 4DCTs and the resultant displacement vector field with Jacobian analysis was used to identify the high-, medium- and low-functional lung regions. Five plans were designed for each patient: a regular photon IMRT vs. anatomic proton plans without consideration of functional ventilation information using double scattering proton therapy (DSPT) and intensity modulated proton therapy (IMPT) vs. functional proton plans with avoidance of high-functional lung using both DSPT and IMPT. Dosimetric parameters were compared in terms of tumor coverage, plan heterogeneity, and avoidance of normal tissues.ResultsOur results showed that both DSPT and IMPT plans gave superior dose advantage to photon IMRTs in sparing low dose regions of the total lung in terms of V5 (volume receiving 5Gy). The functional DSPT only showed marginal benefit in sparing high-functioning lung in terms of V5 or V20 (volume receiving 20Gy) compared to anatomical plans. Yet, the functional planning in IMPT delivery, can further reduce the low dose in high-functioning lung without degrading the PTV dosimetric coverages, compared to anatomical proton planning. Although the doses to some critical organs might increase during functional planning, the necessary constraints were all met.ConclusionsIncorporating 4DCT ventilation imaging into functional proton therapy is feasible. The functional proton plans, in intensity modulated proton delivery, are effective to further preserve high-functioning lung regions without degrading the PTV coverage.