D Ruan

Is there an ideal set of prospective scan acquisition phases for fast-helical based 4D-CT?

The article aims to determine if a prospective acquisition algorithm can be used to find the ideal set of free-breathing phases for fast-helical model-based 4D-CT. A retrospective five-patient dataset that consisted of 25 repeated free breathing CT scans per patient was used. The sum of the square root amplitude difference between all the breathing phases was defined as an objective function to determine the optimality of sets of breathing phases. The objective function was intended to determine if a specific set of breathing phases would yield a motion model that could accurately predict the motion in all 25 CT scans. Voxel specific motion models were calculated using all combinations of N scans from 25 breathing trajectories, (3  ⩽  N  ⩽  25), and the minimum number of scans required to absolutely characterize the motion model was analyzed. This analysis suggests that the number of scans could potentially be reduced to as few as five scans. When the objective function was large, the resulting motion model provided an excellent approximation to the motion model created using all 25 scans.

TH‐C‐137‐02: Robotic Radiotherapy Using Intermediate Beam Energies

PURPOSE Intermediate energy (1-2MV) x-rays have steeper depth dose drop-off and sharper penumbra than commonly used 6MV x-rays. Dosimetry benefits of these characteristics are studied on a robotic non-coplanar planning and delivery platform. METHODS Dose of 1MV and 6MV x-rays was calculated using the convolution/superposition algorithm with heterogeneity correction and Monte Carlo calculated dose kernels. The X-ray spectrum was adjusted to match depth dose curves of published data. Thirty noncoplanar beams were selected by a pricing approach from a candidate beam pool, which consisted of 1162 uniformly distributed non-coplanar beams minus beams leading to collision. The collision model was fit to individual treatment sites. Fluence optimization based on 5 mm MLC was performed after adding each beam. Identical objective functions for PTV and organs-at-risk (OARs) were employed in the 3 planning scenarios: 1 MV alone, 6 MV alone and the combination of 1 MV and 6 MV beams (1&6 MV) with the prescription dose covering 95% of the PTV. Four representative cases from the following anatomical sites were included in the study: head and neck, partial breast, lung and liver. RESULTS 1 MV and 1&6 MV plans provided superior OAR sparing for head, liver, partial breast and lung cases while maintaining the same PTV coverage. Compared with 6 MV plan, 1 MV plans reduced the integral dose by 25%, 23%, 19% and 9% for lung, breast, head and liver cases respectively. The plan quality of 1&6 MV plans, which primarily was slightly superior to that of the 1MV only plans. CONCLUSION The dosimetric drawbacks of intermediate energy x-rays are higher skin doses and shallower penetration when few of them are used on a coplanar platform but these drawbacks were effectively overcome on a highly non-coplanar treatment planning platform, where its advantages of normal tissue sparing and sharp penumbra are manifested.

SU-F-R-17: Advancing Glioblastoma Multiforme (GBM) Recurrence Detection with MRI Image Texture Feature Extraction and Machine Learning

PURPOSE To test the potential of early Glioblastoma Multiforme (GBM) recurrence detection utilizing image texture pattern analysis in serial MR images post primary treatment intervention. METHODS MR image-sets of six time points prior to the confirmed recurrence diagnosis of a GBM patient were included in this study, with each time point containing T1 pre-contrast, T1 post-contrast, T2-Flair, and T2-TSE images. Eight Gray-level co-occurrence matrix (GLCM) texture features including Contrast, Correlation, Dissimilarity, Energy, Entropy, Homogeneity, Sum-Average, and Variance were calculated from all images, resulting in a total of 32 features at each time point. A confirmed recurrent volume was contoured, along with an adjacent non-recurrent region-of-interest (ROI) and both volumes were propagated to all prior time points via deformable image registration. A support vector machine (SVM) with radial-basis-function kernels was trained on the latest time point prior to the confirmed recurrence to construct a model for recurrence classification. The SVM model was then applied to all prior time points and the volumes classified as recurrence were obtained. RESULTS An increase in classified volume was observed over time as expected. The size of classified recurrence maintained at a stable level of approximately 0.1 cm3 up to 272 days prior to confirmation. Noticeable volume increase to 0.44 cm3 was demonstrated at 96 days prior, followed by significant increase to 1.57 cm3 at 42 days prior. Visualization of the classified volume shows the merging of recurrence-susceptible region as the volume change became noticeable. CONCLUSION Image texture pattern analysis in serial MR images appears to be sensitive to detecting the recurrent GBM a long time before the recurrence is confirmed by a radiologist. The early detection may improve the efficacy of targeted intervention including radiosurgery. More patient cases will be included to create a generalizable classification model applicable to a larger patient cohort. NIH R43CA183390 and R01CA188300.NSF Graduate Research Fellowship DGE-1144087.

SU‐F‐J‐181: An Alternative Patient Alignment Tool On TomoTherapy: The First In‐ Human Megavoltage‐Topogram Acquisition

PURPOSE To show the first in-human Megavoltage (MV)-Topogram acquisition for the evaluation of the potential for MV-Topogram-based alignment as an alternative to MVCT for reducing dose and imaging time. METHODS A lung cancer patient was enrolled in an ongoing IRB-approved clinical trial at our institute. The patient was set up using the clinical protocol employing positioning lasers. 3.2mm diameter tungsten spheres were placed on the patients skin at their alignment tattoos to check surface-based marker concordance between topograms and MVCT. Anterior-Posterior (AP) and lateral (LAT) MV-Topograms were acquired using gantry angles of 0°/90° with a 1mm collimator opening, all MLC leafs open, 4cm/s couch speed, and 12.5s scanning time. The topogram acquisition was immediately followed by the normal MVCT scan acquisition. MV-Topograms were reconstructed from the detector exit-data using in-house developed software. The topograms were also enhanced using contrast-limited adaptive histogram equalization (CLAHE). The MV-Topograms were registered to reference kV-based digitally reconstructed topograms. The localization results were compared against results obtained comparing the clinical MVCT to the kVCT simulation. RESULTS The shifts using the unenhanced Topograms, enhanced Topograms, and MVCT were (LAT, LONG, VERT, ROLL) (5.8mm, 2.6mm, -5.6mm, 0.34°), (3.9mm, 2.5mm, -2.2mm, 0.65°) and (2.4mm, 1.5mm, -3.0mm, 0.5°), respectively. The magnitude alignment differences between the enhanced Topograms and MVCT were within 1.5 mm and 0.15°. The average MVCT and total Topogram acquisition times were 272.9s ± 31.5s and 46s, respectively. CONCLUSION MV-Topograms have the potential for providing equivalent performance with less dose and acquisition time than the traditional MVCT technique. We are evaluating other sites as well as adding patients to develop statistically significant analyses regarding the alignment quality differences. MV-Topograms are likely to be most clinically useful for bony anatomy and radiopaque marker-based alignments. The study was supported by an Accuray Grant.

TH‐CD‐202‐06: A Method for Characterizing and Validating Dynamic Lung Density Change During Quiet Respiration

PURPOSE To obtain a functional relationship that calibrates the lung tissue density change under free breathing conditions through correlating Jacobian values to the Hounsfield units. METHODS Free-breathing lung computed tomography images were acquired using a fast helical CT protocol, where 25 scans were acquired per patient. Using a state-of-the-art deformable registration algorithm, a set of the deformation vector fields (DVF) was generated to provide spatial mapping from the reference image geometry to the other free-breathing scans. These DVFs were used to generate Jacobian maps, which estimate voxelwise volume change. Subsequently, the set of 25 corresponding Jacobian and voxel intensity in Hounsfield units (HU) were collected and linear regression was performed based on the mass conservation relationship to correlate the volume change to density change. Based on the resulting fitting coefficients, the tissues were classified into parenchymal (Type I), vascular (Type II), and soft tissue (Type III) types. These coefficients modeled the voxelwise density variation during quiet breathing. The accuracy of the proposed method was assessed using mean absolute difference in HU between the CT scan intensities and the model predicted values. In addition, validation experiments employing a leave-five-out method were performed to evaluate the model accuracy. RESULTS The computed mean model errors were 23.30±9.54 HU, 29.31±10.67 HU, and 35.56±20.56 HU, respectively, for regions I, II, and III, respectively. The cross validation experiments averaged over 100 trials had mean errors of 30.02 ± 1.67 HU over the entire lung. These mean values were comparable with the estimated CT image background noise. CONCLUSION The reported validation experiment statistics confirmed the lung density modeling during free breathing. The proposed technique was general and could be applied to a wide range of problem scenarios where accurate dynamic lung density information is needed. This work was supported in part by NIH R01 CA0096679.

SU-D-BRB-01: A Comparison of Learning Methods for Knowledge Based Dose Prediction for Coplanar and Non-Coplanar Liver Radiotherapy

PURPOSE The predictive power of knowledge based planning (KBP) has considerable potential in the development of automated treatment planning. Here, we examine the predictive capabilities and accuracy of previously reported KBP methods, as well as an artificial neural networks (ANN) method. Furthermore, we compare the predictive accuracy of these methods on coplanar volumetric-modulated arc therapy (VMAT) and non-coplanar 4π radiotherapy. METHODS 30 liver SBRT patients previously treated using coplanar VMAT were selected for this study. The patients were re-planned using 4π radiotherapy, which involves 20 optimally selected non-coplanar IMRT fields. ANNs were used to incorporate enhanced geometric information including liver and PTV size, prescription dose, patient girth, and proximity to beams. The performance of ANN was compared to three methods from statistical voxel dose learning (SVDL), wherein the doses of voxels sharing the same distance to the PTV are approximated by either taking the median of the distribution, non-parametric fitting, or skew-normal fitting. These three methods were shown to be capable of predicting DVH, but only median approximation can predict 3D dose. Prediction methods were tested using leave-one-out cross-validation tests and evaluated using residual sum of squares (RSS) for DVH and 3D dose predictions. RESULTS DVH prediction using non-parametric fitting had the lowest average RSS with 0.1176(4π) and 0.1633(VMAT), compared to 0.4879(4π) and 1.8744(VMAT) RSS for ANN. 3D dose prediction with median approximation had lower RSS with 12.02(4π) and 29.22(VMAT), compared to 27.95(4π) and 130.9(VMAT) for ANN. CONCLUSION Paradoxically, although the ANNs included geometric features in addition to the distances to the PTV, it did not perform better in predicting DVH or 3D dose compared to simpler, faster methods based on the distances alone. The study further confirms that the prediction of 4π non-coplanar plans were more accurate than VMAT. NIH R43CA183390 and R01CA188300.

TH-CD-207A-07: Prediction of High Dimensional State Subject to Respiratory Motion: A Manifold Learning Approach

PURPOSE The development of high dimensional imaging systems (e.g. volumetric MRI, CBCT, photogrammetry systems) in image-guided radiotherapy provides important pathways to the ultimate goal of real-time volumetric/surface motion monitoring. This study aims to develop a prediction method for the high dimensional state subject to respiratory motion. Compared to conventional linear dimension reduction based approaches, our method utilizes manifold learning to construct a descriptive feature submanifold, where more efficient and accurate prediction can be performed. METHODS We developed a prediction framework for high-dimensional state subject to respiratory motion. The proposed method performs dimension reduction in a nonlinear setting to permit more descriptive features compared to its linear counterparts (e.g., classic PCA). Specifically, a kernel PCA is used to construct a proper low-dimensional feature manifold, where low-dimensional prediction is performed. A fixed-point iterative pre-image estimation method is applied subsequently to recover the predicted value in the original state space. We evaluated and compared the proposed method with PCA-based method on 200 level-set surfaces reconstructed from surface point clouds captured by the VisionRT system. The prediction accuracy was evaluated with respect to root-mean-squared-error (RMSE) for both 200ms and 600ms lookahead lengths. RESULTS The proposed method outperformed PCA-based approach with statistically higher prediction accuracy. In one-dimensional feature subspace, our method achieved mean prediction accuracy of 0.86mm and 0.89mm for 200ms and 600ms lookahead lengths respectively, compared to 0.95mm and 1.04mm from PCA-based method. The paired t-tests further demonstrated the statistical significance of the superiority of our method, with p-values of 6.33e-3 and 5.78e-5, respectively. CONCLUSION The proposed approach benefits from the descriptiveness of a nonlinear manifold and the prediction reliability in such low dimensional manifold. The fixed-point iterative approach turns out to work well practically for the pre-image recovery. Our approach is particularly suitable to facilitate managing respiratory motion in image-guide radiotherapy. This work is supported in part by NIH grant R01 CA169102-02.

TH-AB-202-08: A Robust Real-Time Surface Reconstruction Method On Point Clouds Captured From a 3D Surface Photogrammetry System

PURPOSE Surface photogrammetry (e.g. VisionRT, C-Rad) provides a noninvasive way to obtain high-frequency measurement for patient motion monitoring in radiotherapy. This work aims to develop a real-time surface reconstruction method on the acquired point clouds, whose acquisitions are subject to noise and missing measurements. In contrast to existing surface reconstruction methods that are usually computationally expensive, the proposed method reconstructs continuous surfaces with comparable accuracy in real-time. METHODS The key idea in our method is to solve and propagate a sparse linear relationship from the point cloud (measurement) manifold to the surface (reconstruction) manifold, taking advantage of the similarity in local geometric topology in both manifolds. With consistent point cloud acquisition, we propose a sparse regression (SR) model to directly approximate the target point cloud as a sparse linear combination from the training set, building the point correspondences by the iterative closest point (ICP) method. To accommodate changing noise levels and/or presence of inconsistent occlusions, we further propose a modified sparse regression (MSR) model to account for the large and sparse error built by ICP, with a Laplacian prior. We evaluated our method on both clinical acquired point clouds under consistent conditions and simulated point clouds with inconsistent occlusions. The reconstruction accuracy was evaluated w.r.t. root-mean-squared-error, by comparing the reconstructed surfaces against those from the variational reconstruction method. RESULTS On clinical point clouds, both the SR and MSR models achieved sub-millimeter accuracy, with mean reconstruction time reduced from 82.23 seconds to 0.52 seconds and 0.94 seconds, respectively. On simulated point cloud with inconsistent occlusions, the MSR model has demonstrated its advantage in achieving consistent performance despite the introduced occlusions. CONCLUSION We have developed a real-time and robust surface reconstruction method on point clouds acquired by photogrammetry systems. It serves an important enabling step for real-time motion tracking in radiotherapy. This work is supported in part by NIH grant R01 CA169102-02.

TH-CD-206-04: Learning Relevance Criterion for Multi-Atlas Based Image Segmentation

PURPOSE It is important to effectively identify the subset of relevant training atlases in learning based segmentation. Since segmentation based geometry relevance between atlases and the target is inaccessible due to the unknown target segmentation, image feature based relevance criteria have to be used instead. This study aims to learn an image based criterion that best reflects the underlying geometry agreement. METHODS We learn the image-based relevance criterion in the form of a Mahalanobis distance on linearly transformed image feature, which is specialized to image intensity in this work, and optimize such linear transformation. Given a set of atlases, i.e., image/segmentation pairs, a linear transformation is optimized to make the Mahalanobis distance small between the geometrically relevant atlases and large between the irrelevant pairs. Performance assessment and comparison with the commonly-used mean square distance (MSD), a special case of the Mahalanobis distance with a trivial linear transform, are performed based on clinical brain MR images. RESULTS The proposed surrogate learning approach was validated with multi-atlas based corpus callosum segmentation. Compared to MSD, our learned surrogate demonstrated superiority in selecting the geometrically relevant atlases. The learned surrogate yielded an improvement on the average pairwise Dice similarity coefficient (DSC) of the selected atlases from .004 to .007, and a corresponding improvement on the ultimate segmentation accuracy from .003 to .005, over an atlas subset size from 1 to 20. CONCLUSION This work provides a systematic methodology to learn task-specific criterion to select atlases for image segmentation, and demonstrates effectiveness in identifying the most relevant atlases. We are working on investigating various image features and extending to nonlinear transformations for further improvement.

SU‐D‐202‐06: Prospective Free‐Breathing CT Scan Selection for 5DCT

PURPOSE 5DCT employs 25 fast helical scans and breathing surrogate monitoring to sample the respiratory cycle. Deformable image registration is used to fit a correspondence model between tissue motion and breathing amplitude and rate. The number of scans was chosen to ensure a high probability that tissues were imaged at sufficiently distinct breathing phases for accurate modeling of the entire breathing cycle. This work describes a method to prospectively select scan start times and reduce the protocols number of scans from 25 to 6. METHODS Breathing traces from 7 patients imaged with 5DCT were used to simulate acquisition of 6 scans. Breathing phase was estimated using only observations from previous time points. Cross-correlation between the representative breath and the most recent half period was continuously computed. If phase and cross-correlation criteria were met, scans were triggered with a 2 second delay before acquisition. Blind acquisition, 6 scans separated by a fixed delay, was modeled at staggered start times. The spread of prospectively and blindly sampled breathing waveforms was characterized using a previously published objective function. RESULTS Prospectively selected scans ranked on average in the 84th percentile of objective function values obtained by blind acquisition at staggered start times for 7 patient breathing traces. CONCLUSION A method to prospectively determine scan start times for 5DCT was developed and tested by simulating acquisition on patient breathing traces. The method is computationally inexpensive enough for real-time implementation and would result in an imaging dose of less than one quarter of the current 5DCT protocol.