Multi-GPU configuration of 4D intensity modulated radiation therapy inverse planning using global optimization

We report on the design, implementation and characterization of a multi-graphic processing unit (GPU) computational platform for higher-order optimization in radiotherapy treatment planning. In collaboration with a commercial vendor (Varian Medical Systems, Palo Alto, CA), a research prototype GPU-enabled Eclipse (V13.6) workstation was configured. The hardware consisted of dual 8-core Xeon processors, 256 GB RAM and four NVIDIA Tesla K80 general purpose GPUs. We demonstrate the utility of this platform for large radiotherapy optimization problems through the development and characterization of a parallelized particle swarm optimization (PSO) four dimensional (4D) intensity modulated radiation therapy (IMRT) technique. The PSO engine was coupled to the Eclipse treatment planning system via a vendor-provided scripting interface. Specific challenges addressed in this implementation were (i) data management and (ii) non-uniform memory access (NUMA). For the former, we alternated between parameters over which the computation process was parallelized. For the latter, we reduced the amount of data required to be transferred over the NUMA bridge. The datasets examined in this study were approximately 300 GB in size, including 4D computed tomography images, anatomical structure contours and dose deposition matrices. For evaluation, we created a 4D-IMRT treatment plan for one lung cancer patient and analyzed computation speed while varying several parameters (number of respiratory phases, GPUs, PSO particles, and data matrix sizes). The optimized 4D-IMRT plan enhanced sparing of organs at risk by an average reduction of [Formula: see text] in maximum dose, compared to the clinical optimized IMRT plan, where the internal target volume was used. We validated our computation time analyses in two additional cases. The computation speed in our implementation did not monotonically increase with the number of GPUs. The optimal number of GPUs (five, in our study) is directly related to the hardware specifications. The optimization process took 35 min using 50 PSO particles, 25 iterations and 5 GPUs.

Prognostic factors associated with the accuracy of deformable image registration in lung cancer patients treated with stereotactic body radiotherapy

We evaluated the accuracy of an in-house program in lung stereotactic body radiation therapy (SBRT) cancer patients, and explored the prognostic factors associated with the accuracy of deformable image registrations (DIRs). The accuracy of the 3 programs which implement the free-form deformation and the B-spline algorithm was compared regarding the structures on 4-dimensional computed tomography (4DCT) image datasets between the peak-inhale and peak-exhale phases. The dice similarity coefficient (DSC) and normalized DSC (NDSC) were measured for the gross tumor volumes from 19 lung SBRT patients. We evaluated the accuracy of DIR using gross tumor volume, magnitude of displacement from 0% phase to 50% phase, whole lung volume in the 50% phase image, and status of tumor pleural attachment. The median NDSC values using the NiftyReg, MIM Maestro and Velocity AI programs were 1.027, 1.005, and 0.946, respectively, indicating that NiftyReg and MIM Maestro programs had similar accuracy with an uncertainty of < 1 mm. Larger uncertainty of 1 to 2 mm was observed using the Velocity AI program. The NiftyReg and the MIM programs provided higher NDSC values than the median values when the gross tumor volume was attached to the pleura (p <0.05). However, it showed a different trend in using the Velocity AI program. All software programs provided unexpected results, and there is a possibility that such results would reduce the accuracy of 4D treatment planning and adaptive radiotherapy. The unexpected results may be because the tumors are surrounded by other tissues, and there are differences regarding the region of interest for rigid and nonrigid registration. Furthermore, our results indicated that the pleural attachment status might be an important predictor of DIR accuracy for thoracic images, indicating that there is a potentially large dose distribution discrepancy concerning 4D treatment planning and adaptive radiotherapy.

Cavernous Nerve Injury by Radiation Therapy May Potentiate Erectile Dysfunction in Rats

PURPOSE/OBJECTIVES

Radiation-induced erectile-dysfunction (RiED) is one of the most common side effects of radiation therapy (RT) and significantly reduces the quality of life (QoL) of cancer patients. Approximately 50% of prostate cancer patients experience RiED within 3 to 5 years after completion of RT. A series of vascular, muscular, and neurogenic injuries after prostate RT lead to RiED; however, the precise role of RT-induced neurogenic injury in RiED has not been fully established. The cavernous nerves (CN) are postganglionic parasympathetic nerves located beside the prostate gland that assist in penile erection. This study was designed to investigate the role of CN injury, tissue damage, and altered signaling pathways in an RiED rat model.

METHODS AND MATERIALS

Male rats were exposed to a single dose of 25 Gy prostate-confined RT. Erectile function was evaluated by intracavernous pressure (ICP) measurements conducted both 9 and 14 weeks after RT. Neuronal injury was evaluated in the CN using quantitative polymerase chain reaction, conduction studies, transmission electron microscopy, and immunoblotting. Masson trichrome staining was performed to elucidate fibrosis level in penile tissues.

RESULTS

There were significant alterations in the ICP (P<.0001) of RT rats versus non-RT rats. TEM analysis showed decreased myelination, increased microvascular damage, and progressive axonal atrophy of the CN fibers after RT. Electrophysiologic analysis showed significant impairment of the CN conduction velocity after RT. RT also significantly increased RhoA/Rho-associated protein kinase 1 (ROCK1) mRNA and protein expression. In addition, penile tissue showed increased apoptosis and fibrosis 14 weeks after RT.

CONCLUSIONS

RT-induced CN injury may contribute to RiED; this is therefore a rationale for developing novel therapeutic strategies to mitigate CN and tissue damage. Moreover, further investigation of the RhoA/ROCK pathway's role in mitigating RiED is necessary.

Inversed-Planned Respiratory Phase Gating in Lung Conformal Radiation Therapy

PURPOSE

To assess whether the optimal gating window for each beam during lung radiation therapy with respiratory gating will be dependent on a variety of patient-specific factors, such as tumor size and location and the extent of relative tumor and organ motion.

METHODS AND MATERIALS

To create optimal gating treatment plans, we started from an optimized clinical plan, created a plan per respiratory phase using the same beam arrangements, and used an inverse planning optimization approach to determine the optimal gating window for each beam and optimal beam weights (ie, monitor units). Two pieces of information were used for optimization: (1) the state of the anatomy at each phase, extracted from 4-dimensional computed tomography scans; and (2) the time spent in each state, estimated from a 2-minute monitoring of the patient's breathing motion. We retrospectively studied 15 lung cancer patients clinically treated by hypofractionated conformal radiation therapy, for whom 45 to 60 Gy was administered over 3 to 15 fractions using 7 to 13 beams. Mean gross tumor volume and respiratory-induced tumor motion were 82.5 cm³ and 1.0 cm, respectively.

RESULTS

Although patients spent most of their respiratory cycle in end-exhalation (EE), our optimal gating plans used EE for only 34% of the beams. Using optimal gating, maximum and mean doses to the esophagus, heart, and spinal cord were reduced by an average of 15% to 26%, and the beam-on times were reduced by an average of 23% compared with equivalent single-phase EE gated plans (P<.034, paired 2-tailed t test).

CONCLUSIONS

We introduce a personalized respiratory-gating technique in which inverse planning optimization is used to determine patient- and beam-specific gating phases toward enhancing dosimetric quality of radiation therapy treatment plans.

Characterizing spatiotemporal information loss in sparse-sampling-based dynamic MRI for monitoring respiration-induced tumor motion in radiotherapy

PURPOSE

Sparse-sampling and reconstruction techniques represent an attractive strategy to achieve faster image acquisition speeds, while maintaining adequate spatial resolution and signal-to-noise ratio in rapid magnetic resonance imaging (MRI). The authors investigate the use of one such sequence, broad-use linear acquisition speed-up technique (k-t BLAST) in monitoring tumor motion for thoracic and abdominal radiotherapy and examine the potential trade-off between increased sparsification (to increase imaging speed) and the potential loss of "true" information due to greater reliance on a priori information.

METHODS

Lung tumor motion trajectories in the superior-inferior direction, previously recorded from ten lung cancer patients, were replayed using a motion phantom module driven by an MRI-compatible motion platform. Eppendorf test tubes filled with water which serve as fiducial markers were placed in the phantom. The modeled rigid and deformable motions were collected in a coronal image slice using balanced fast field echo in conjunction with k-t BLAST. Root mean square (RMS) error was used as a metric of spatial accuracy as measured trajectories were compared to input data. The loss of spatial information was characterized for progressively increasing acceleration factor from 1 to 16; the resultant sampling frequency was increased approximately from 2.5 to 19 Hz when the principal direction of the motion was set along frequency encoding direction. In addition to the phantom study, respiration-induced tumor motions were captured from two patients (kidney tumor and lung tumor) at 13 Hz over 49 s to demonstrate the impact of high speed motion monitoring over multiple breathing cycles. For each subject, the authors compared the tumor centroid trajectory as well as the deformable motion during free breathing.

RESULTS

In the rigid and deformable phantom studies, the RMS error of target tracking at the acquisition speed of 19 Hz was approximately 0.3-0.4 mm, which was smaller than the reconstructed pixel resolution of 0.67 mm. In the patient study, the dynamic 2D MRI enabled the monitoring of cycle-to-cycle respiratory variability present in the tumor position. It was seen that the range of centroid motion as well as the area covered due to target motion during each individual respiratory cycle was underestimated compared to the entire motion range observed over multiple breathing cycles.

CONCLUSIONS

The authors' initial results demonstrate that sparse-sampling- and reconstruction-based dynamic MRI can be used to achieve adequate image acquisition speeds without significant information loss for the task of radiotherapy guidance. Such monitoring can yield spatial and temporal information superior to conventional offline and online motion capture methods used in thoracic and abdominal radiotherapy.

A robust real-time surface reconstruction method on point clouds captured from a 3D surface photogrammetry system

PURPOSE

To develop a robust and real-time surface reconstruction method on point clouds captured from a 3D surface photogrammetry system.

METHODS

The authors have developed a robust and fast surface reconstruction method on point clouds acquired by the photogrammetry system, without explicitly solving the partial differential equation required by a typical variational approach. Taking advantage of the overcomplete nature of the acquired point clouds, their method solves and propagates a sparse linear relationship from the point cloud manifold to the surface manifold, assuming both manifolds share similar local geometry. With relatively consistent point cloud acquisitions, the authors propose a sparse regression (SR) model to directly approximate the target point cloud as a sparse linear combination from the training set, assuming that the point correspondences built by the iterative closest point (ICP) is reasonably accurate and have residual errors following a Gaussian distribution. To accommodate changing noise levels and/or presence of inconsistent occlusions during the acquisition, the authors further propose a modified sparse regression (MSR) model to model the potentially large and sparse error built by ICP with a Laplacian prior. The authors evaluated the proposed method on both clinical point clouds acquired under consistent acquisition conditions and on point clouds with inconsistent occlusions. The authors quantitatively evaluated the reconstruction performance with respect to root-mean-squared-error, by comparing its reconstruction results against that from the variational method.

RESULTS

On clinical point clouds, both the SR and MSR models have achieved sub-millimeter reconstruction accuracy and reduced the reconstruction time by two orders of magnitude to a subsecond reconstruction time. On point clouds with inconsistent occlusions, the MSR model has demonstrated its advantage in achieving consistent and robust performance despite the introduced occlusions.

CONCLUSIONS

The authors have developed a fast and robust surface reconstruction method on point clouds captured from a 3D surface photogrammetry system, with demonstrated sub-millimeter reconstruction accuracy and subsecond reconstruction time. It is suitable for real-time motion tracking in radiotherapy, with clear surface structures for better quantifications.

Predictive modeling of respiratory tumor motion for real-time prediction of baseline shifts

Baseline shifts in respiratory patterns can result in significant spatiotemporal changes in patient anatomy (compared to that captured during simulation), in turn, causing geometric and dosimetric errors in the administration of thoracic and abdominal radiotherapy. We propose predictive modeling of the tumor motion trajectories for predicting a baseline shift ahead of its occurrence. The key idea is to use the features of the tumor motion trajectory over a 1 min window, and predict the occurrence of a baseline shift in the 5 s that immediately follow (lookahead window). In this study, we explored a preliminary trend-based analysis with multi-class annotations as well as a more focused binary classification analysis. In both analyses, a number of different inter-fraction and intra-fraction training strategies were studied, both offline as well as online, along with data sufficiency and skew compensation for class imbalances. The performance of different training strategies were compared across multiple machine learning classification algorithms, including nearest neighbor, Naïve Bayes, linear discriminant and ensemble Adaboost. The prediction performance is evaluated using metrics such as accuracy, precision, recall and the area under the curve (AUC) for repeater operating characteristics curve. The key results of the trend-based analysis indicate that (i) intra-fraction training strategies achieve highest prediction accuracies (90.5-91.4%); (ii) the predictive modeling yields lowest accuracies (50-60%) when the training data does not include any information from the test patient; (iii) the prediction latencies are as low as a few hundred milliseconds, and thus conducive for real-time prediction. The binary classification performance is promising, indicated by high AUCs (0.96-0.98). It also confirms the utility of prior data from previous patients, and also the necessity of training the classifier on some initial data from the new patient for reasonable prediction performance. The ability to predict a baseline shift with a sufficient look-ahead window will enable clinical systems or even human users to hold the treatment beam in such situations, thereby reducing the probability of serious geometric and dosimetric errors.

Fast leaf-fitting with generalized underdose/overdose constraints for real-time MLC tracking

PURPOSE

Real-time multileaf collimator (MLC) tracking is a promising approach to the management of intrafractional tumor motion during thoracic and abdominal radiotherapy. MLC tracking is typically performed in two steps: transforming a planned MLC aperture in response to patient motion and refitting the leaves to the newly generated aperture. One of the challenges of this approach is the inability to faithfully reproduce the desired motion-adapted aperture. This work presents an optimization-based framework with which to solve this leaf-fitting problem in real-time.

METHODS

This optimization framework is designed to facilitate the determination of leaf positions in real-time while accounting for the trade-off between coverage of the PTV and avoidance of organs at risk (OARs). Derived within this framework, an algorithm is presented that can account for general linear transformations of the planned MLC aperture, particularly 3D translations and in-plane rotations. This algorithm, together with algorithms presented in Sawant et al. ["Management of three-dimensional intrafraction motion through real-time DMLC tracking," Med. Phys. 35, 2050-2061 (2008)] and Ruan and Keall [Presented at the 2011 IEEE Power Engineering and Automation Conference (PEAM) (2011) (unpublished)], was applied to apertures derived from eight lung intensity modulated radiotherapy plans subjected to six-degree-of-freedom motion traces acquired from lung cancer patients using the kilovoltage intrafraction monitoring system developed at the University of Sydney. A quality-of-fit metric was defined, and each algorithm was evaluated in terms of quality-of-fit and computation time.

RESULTS

This algorithm is shown to perform leaf-fittings of apertures, each with 80 leaf pairs, in 0.226 ms on average as compared to 0.082 and 64.2 ms for the algorithms of Sawant et al., Ruan, and Keall, respectively. The algorithm shows approximately 12% improvement in quality-of-fit over the Sawant et al. approach, while performing comparably to Ruan and Keall.

CONCLUSIONS

This work improves upon the quality of the Sawant et al. approach, but does so without sacrificing run-time performance. In addition, using this framework allows for complex leaf-fitting strategies that can be used to account for PTV/OAR trade-off during real-time MLC tracking.

A continuous surface reconstruction method on point cloud captured from a 3D surface photogrammetry system

PURPOSE

To accurately and efficiently reconstruct a continuous surface from noisy point clouds captured by a surface photogrammetry system (VisionRT).

METHODS

The authors have developed a level-set based surface reconstruction method on point clouds captured by a surface photogrammetry system (VisionRT). The proposed method reconstructs an implicit and continuous representation of the underlying patient surface by optimizing a regularized fitting energy, offering extra robustness to noise and missing measurements. By contrast to explicit/discrete meshing-type schemes, their continuous representation is particularly advantageous for subsequent surface registration and motion tracking by eliminating the need for maintaining explicit point correspondences as in discrete models. The authors solve the proposed method with an efficient narrowband evolving scheme. The authors evaluated the proposed method on both phantom and human subject data with two sets of complementary experiments. In the first set of experiment, the authors generated a series of surfaces each with different black patches placed on one chest phantom. The resulting VisionRT measurements from the patched area had different degree of noise and missing levels, since VisionRT has difficulties in detecting dark surfaces. The authors applied the proposed method to point clouds acquired under these different configurations, and quantitatively evaluated reconstructed surfaces by comparing against a high-quality reference surface with respect to root mean squared error (RMSE). In the second set of experiment, the authors applied their method to 100 clinical point clouds acquired from one human subject. In the absence of ground-truth, the authors qualitatively validated reconstructed surfaces by comparing the local geometry, specifically mean curvature distributions, against that of the surface extracted from a high-quality CT obtained from the same patient.

RESULTS

On phantom point clouds, their method achieved submillimeter reconstruction RMSE under different configurations, demonstrating quantitatively the faith of the proposed method in preserving local structural properties of the underlying surface in the presence of noise and missing measurements, and its robustness toward variations of such characteristics. On point clouds from the human subject, the proposed method successfully reconstructed all patient surfaces, filling regions where raw point coordinate readings were missing. Within two comparable regions of interest in the chest area, similar mean curvature distributions were acquired from both their reconstructed surface and CT surface, with mean and standard deviation of (μrecon=-2.7×10(-3) mm(-1), σrecon=7.0×10(-3) mm(-1)) and (μCT=-2.5×10(-3) mm(-1), σCT=5.3×10(-3) mm(-1)), respectively. The agreement of local geometry properties between the reconstructed surfaces and the CT surface demonstrated the ability of the proposed method in faithfully representing the underlying patient surface.

CONCLUSIONS

The authors have integrated and developed an accurate level-set based continuous surface reconstruction method on point clouds acquired by a 3D surface photogrammetry system. The proposed method has generated a continuous representation of the underlying phantom and patient surfaces with good robustness against noise and missing measurements. It serves as an important first step for further development of motion tracking methods during radiotherapy.

Inverse 4D conformal planning for lung SBRT using particle swarm optimization

A critical aspect of highly potent regimens such as lung stereotactic body radiation therapy (SBRT) is to avoid collateral toxicity while achieving planning target volume (PTV) coverage. In this work, we describe four dimensional conformal radiotherapy using a highly parallelizable swarm intelligence-based stochastic optimization technique. Conventional lung CRT-SBRT uses a 4DCT to create an internal target volume and then, using forward-planning, generates a 3D conformal plan. In contrast, we investigate an inverse-planning strategy that uses 4DCT data to create a 4D conformal plan, which is optimized across the three spatial dimensions (3D) as well as time, as represented by the respiratory phase. The key idea is to use respiratory motion as an additional degree of freedom. We iteratively adjust fluence weights for all beam apertures across all respiratory phases considering OAR sparing, PTV coverage and delivery efficiency. To demonstrate proof-of-concept, five non-small-cell lung cancer SBRT patients were retrospectively studied. The 4D optimized plans achieved PTV coverage comparable to the corresponding clinically delivered plans while showing significantly superior OAR sparing ranging from 26% to 83% for D max heart, 10%-41% for D max esophagus, 31%-68% for D max spinal cord and 7%-32% for V 13 lung.

Radiotherapy Planning Using an Improved Search Strategy in Particle Swarm Optimization

OBJECTIVE

Evolutionary stochastic global optimization algorithms are widely used in large-scale, nonconvex problems. However, enhancing the search efficiency and repeatability of these techniques often requires well-customized approaches. This study investigates one such approach.

METHODS

We use particle swarm optimization (PSO) algorithm to solve a 4D radiation therapy (RT) inverse planning problem, where the key idea is to use respiratory motion as an additional degree of freedom in lung cancer RT. The primary goal is to administer a lethal dose to the tumor target while sparing surrounding healthy tissue. Our optimization iteratively adjusts radiation fluence-weights for all beam apertures across all respiratory phases. We implement three PSO-based approaches: conventionally used unconstrained, hard-constrained, and our proposed virtual search. As proof of concept, five lung cancer patient cases are optimized over ten runs using each PSO approach. For comparison, a dynamically penalized likelihood (DPL) algorithm-a popular RT optimization technique is also implemented and used.

RESULTS

The proposed technique significantly improves the robustness to random initialization while requiring fewer iteration cycles to converge across all cases. DPL manages to find the global optimum in 2 out of 5 RT cases over significantly more iterations.

CONCLUSION

The proposed virtual search approach boosts the swarm search efficiency, and consequently, improves the optimization convergence rate and robustness for PSO.

SIGNIFICANCE

RT planning is a large-scale, nonconvex optimization problem, where finding optimal solutions in a clinically practical time is critical. Our proposed approach can potentially improve the optimization efficiency in similar time-sensitive problems.

Prediction of high-dimensional states subject to respiratory motion: a manifold learning approach

The development of high-dimensional imaging systems in image-guided radiotherapy provides important pathways to the ultimate goal of real-time full volumetric motion monitoring. Effective motion management during radiation treatment usually requires prediction to account for system latency and extra signal/image processing time. It is challenging to predict high-dimensional respiratory motion due to the complexity of the motion pattern combined with the curse of dimensionality. Linear dimension reduction methods such as PCA have been used to construct a linear subspace from the high-dimensional data, followed by efficient predictions on the lower-dimensional subspace. In this study, we extend such rationale to a more general manifold and propose a framework for high-dimensional motion prediction with manifold learning, which allows one to learn more descriptive features compared to linear methods with comparable dimensions. Specifically, a kernel PCA is used to construct a proper low-dimensional feature manifold, where accurate and efficient prediction can be performed. A fixed-point iterative pre-image estimation method is used to recover the predicted value in the original state space. We evaluated and compared the proposed method with a PCA-based approach on level-set surfaces reconstructed from point clouds captured by a 3D photogrammetry system. The prediction accuracy was evaluated in terms of root-mean-squared-error. Our proposed method achieved consistent higher prediction accuracy (sub-millimeter) for both 200 ms and 600 ms lookahead lengths compared to the PCA-based approach, and the performance gain was statistically significant.