Accelerating MCNP-based Monte Carlo Simulations for Neutron Transport on GPU

Proceedings of the 53rd Annual ASTRO Meeting Matthew’s correlation coefficient (r). SNP significance was tested using a permutation t test and the Bonferroni multiple comparison correction (81 comparisons). Results: The most statistically significant SNP was rs2032809 (OR = 0.39; Bonferroni-corrected p value = 0.025), within the BBC3 gene, which has been reported as a suppressor of apoptosis. The second most highly correlated SNP did not reach significance after Bonferroni correction, but was also within the BBC3 gene (rs10407441) (OR = 4.01; corrected p value = 0.14). No predictive models performed better than using the top SNP alone in LDA, resulting (on cross validation) in r = 0.37, and AUC = 0.67. Dosimetric variables, including normal tissue complication probability models for rectal and bladder toxicity, did not add predictive power to this model. Conclusions: This is the first analysis of genetic factors related to GI and GU toxicity in a cohort of men treated relatively uniformly with external beam radiotherapy in a randomized trial. The novel bioinformatics strategy helped identify a statistically significant SNP that may help predict toxicity. Author Disclosure: J. Oh: None. J.O. Deasy: B. Research Grant; Varian Corp. R. Stoyanova: None. Z. Saleh: None. M. Buyyounouski: None. H.P. Wong: None. A.P. Apte: None. R.A. Price: None. J. Hu: None. A. Pollack: None.

1003

The IHE-RO Initiative: A Progress Report

R. Rengan1, R. Kapoor2, M. Abdel-Wahab3, A. Ravi4, M. Mietinnen5, C. Field6, B. Curran7, S. Abdul8, J. Palta2, P. Tripuraneni9, et al. 1 University of Pennsylvania, Philadelphia, PA, 2University of Florida, Gainesville, FL, 3IAEA, Vienna, Austria, 4Weill Cornell Medical College, Flushing, NY, 5Varian Medical Systems, Palo Alto, CA, 6Cross Cancer Institute, Calgary, AB, Canada, 7 University of Rhode Island, Providence, RI, 8ASTRO, Fairfax, VA, 9Scripps Clinic, San Diego, CA

Purpose/Objective(s): Integrating the Healthcare Enterprise in Radiation Oncology (IHE-RO) is an ASTRO sponsored initiative, composed of vendors, physicists and clinicians dedicated to improving equipment and software integration issues related to radiation treatment. It seeks to improve the way computer based systems in radiation oncology share information using well-defined data exchange standards (DICOM / HL7). Materials/Methods: At the IHE-RO testing event (Connectathon) in the last 4 years, 11 vendors with 14 different products have successfully developed and tested solutions to connectivity problems identified in treatment planning. These connectivity problems addressed by IHE-RO thus far include (1) multimodality image registration of CT/MR/PET datasets, (2) simple and advanced interoperability of treatment techniques between treatment planning systems (TPS), and (3) seamless flow of standards-based scheduling and delivery information between radiation oncology treatment management systems (TMS) and treatment delivery systems (TDS) such as linear accelerators including robotic linear accelerators. Test tools were developed to verify the exchange of data between TPS, TMS and TDS. Simple and complex treatment techniques such as arc therapy were tested and confirmed to meet IHE-RO specifications. Vendor to vendor testing between TPS, TMS and TDS was performed at the test event. Interfaces between treatment delivery devices and treatment management systems were also developed and tested with the help of these tools. Results: To date, five vendors have successfully tested, passed, and implemented components of the connectivity solutions to multimodality image registration of CT/MR/PET datasets. Four vendors have successfully tested, passed, and implemented components of the simple and advanced interoperability of treatment techniques between treatment planning systems (TPS). Four vendors have successfully tested, passed, and implemented components of the seamless flow of standards based scheduling and delivery information between TMS and TDS. Conclusions: With the help of IHE-RO testing events and test tools, vendors have successfully tested interoperability that can be implemented in the clinic. These solutions will ultimately improve workflow efficiency and serve to improve patient safety in concordance with the ASTRO ‘Target Safely’ initiative. A web-based guide, ‘‘IHE-RO Helper’’ is under development that will allow the end user to determine if TPS ‘‘X’’ can plan a treatment and transfer data to TMS ‘‘Y’’ and then transfer data to TDS ‘‘Z’’ for treatment delivery without any interconnectivity issues. The ‘‘IHE-RO Helper’’ can evolve as a tool that can aid the user community to successfully implement its concepts in the clinic. Author Disclosure: R. Rengan: None. R. Kapoor: None. M. Abdel-Wahab: None. A. Ravi: None. M. Mietinnen: None. C. Field: None. B. Curran: None. S. Abdul: None. J. Palta: None. P. Tripuraneni: None.

1004

Accelerating MCNP-based Monte Carlo Simulations for Neutron Transport on GPU

1

C. Gong , J. Liu1, B. Yang1, L. Deng2, G. Li2, X. Li1, Q. Hu1, Z. Gong1 National University of Defense Technology, 410073, China, 2Institute of Applied Physics and Computational Mathematics, 100088, China 1

Purpose/Objective(s): Neutron transport simulation has found uses in many problem areas including medical physics, nuclear reactors and radiation shielding. MCNP is a general-purpose, continuous-energy, generalized-geometry, time-dependent, coupled neutron/photon/electron Monte Carlo transport code. This work aims at porting the computation-intensive kernel of MCNP-4C on a single instruction multiple data (SIMD) architecture. Materials/Methods: The implementation is based on the scalable parallel programming model CUDA, which uses C as a highlevel programming language, and the highly parallel, multithreaded, many core Graphics Processing Unit (GPU). In order to efficiently use the computing capability of GPU, it needs to improve the memory access efficiency and use the massive GPU threads. An appropriate data structure named Structure of Array (SOA) is applied on the data representations of neutrons histories to satisfy the coalesced global memory access. The most frequently accessed data such as the random numbers is stored in the onchip shared memory to improve performance. The tracks of neutron histories involve sampling the velocity, energy, scattering angles, particle weight and collisions and are distributed to the massive thin GPU threads on average. Each thread block has 64 threads and the total threads are issued as many as possible. The total threads in each kernel funcation call are limited by both memory amount and neutron numbers. The synchronizations are performed by the kernel function call. The simulation was executed on an NVIDIA M2050 GPU. M2050 is specially designed for high performance computing which contains 448 CUDA cores at 1.15 GHz and 2.6 GB memory.

S157

I. J. Radiation Oncology d Biology d Physics

S158

Volume 81, Number 2, Supplement, 2011

Results: A speed-up of a factor 16.3 to 23.67 compared to a single core CPU has been observed in our present implementation using M2050 with full double-precision floating-point operations. The results of neutron transport with 10 MeV in three-dimensional sphere geometry produced from our program showed an exactly agreement with those from the MCNP on CPU. Conclusions: This study shows that MCNP-based MC simulation of neutron transport is accurate and feasible on GPU. As the future works, the neutron transport technology on GPU will be applied to a more practical neutron dose calculation and more MCNP-based particle transport problems will be accelerated on GPU. Acknowledgement: This study was supported by National Natural Science Foundation of China (No. 60673150, No. 60970033) and National Basic Research Program of China (No. 61312701001). Author Disclosure: C. Gong: None. J. Liu: None. B. Yang: None. L. Deng: None. G. Li: None. X. Li: None. Q. Hu: None. Z. Gong: None.

1005

Optimal Classification to Characterize Tumor Heterogeneity and At-risk Tumor Voxels by DCE MRI for Early Prediction of Therapy Outcome

W. Yuh1, Z. Huang2, G. Jia1, S. S. Lo3, J. Zhang1, J. Z. Wang1, D. Zhang1, Z. K. Shah1, N. A. Mayr1 Ohio State University, Columbus, OH, 2East Carolina University, Greenville, NC, 3Case Western Reserve University, Cleveland, OH 1

Purpose/Objective(s): The ability to quantify the tumor subvolume with unfavorable functional/biological imaging properties within the heterogeneous tumor has recently been reported to improve the early prediction of tumor control and survival in cervix cancer. However, the optimal threshold of the imaging biomarker’s values to functionally differentiate unfavorable vs. favorable tumor voxels have not been determined at different treatment time points. This study applied dynamic contrast-enhanced (DCE) MRI to classify at-risk tumor voxels, likely contributing to treatment failure, with various DCE threshold values and correlate the at-risk tumor subvolume component within the heterogeneous tumor with clinical therapy outcome for early outcome prediction. Materials/Methods: In 104 patients with Stage IB2-IVA cervix cancer, DCE-MRI scans were obtained at three different treatment time points: before and during treatment at 2 to 2.5 weeks (20 – 25 Gy) and at 4 to 5 weeks (40 – 50 Gy). For each tumor voxel, the plateau signal intensity (SI) was derived from the dynamic time-SI curve of the DCE-MRI. For each imaging time point, SI threshold values ranging from 1 to 3 were explored to define at-risk tumor voxels. The component of at-risk tumor voxels within the entire tumor voxel population was correlated with clinical outcome. The optimal SI thresholds to differentiate local control vs. failure and disease-specific survival vs. cancer death were determined for each imaging time point using ROC and Mann-Whitney analysis. Mean follow-up was 6.8 years. Results: The optimal SI threshold values, before treatment and at 2 to 2.5 and 4 to 5 weeks into treatment, are summarized in Table 1 (rows 3, 4, and 6). These optimal thresholds showed minor differences among different time points throughout the radiation therapy course. When a uniform of SI = 2.1 was applied to the different time points to classify at-risk voxels, the differentiation of favorable vs. unfavorable outcome continued to be highly significant for all three MRI imaging time points (row 7). Conclusions: These results support the concept that at-risk tumor voxels, adversely influencing treatment outcome, can be defined by DCE MRI with SI threshold values. Although optimal SI thresholds vary slightly with imaging time during the therapy course, a universal threshold of SI = 2.1 remains efficacious for all three time points to quantify at-risk tumor voxels and predict therapy outcome. Table: Optimal SI threshold value to classify unfavorable tumor voxels at different imaging times Local tumor control MRI time point

Pre-RT

Maximum area under ROC curve (AUC) First derivative of AUC Optimal SI p value with time-specific SI p value with same SI = 2.1 for all times

0.66 0 2.1 0.03 0.03

20–25 Gy 0.78 0 2.0 8.0E 1.2E

5 4

Disease-specific survival 40–50 Gy 0.80 0 2.1 3.7E 3.7E

5 5

Pre-RT

20–25 Gy

40–50 Gy

0.66 0 1.8 0.009 0.015

0.71 0 2.1 0.001 0.001

0.69 0 1.8 0.003 0.008

Author Disclosure: W. Yuh: B. Research Grant; NIH R01 CA 71906. Z. Huang: None. G. Jia: None. S.S. Lo: None. J. Zhang: None. J.Z. Wang: None. D. Zhang: None. Z.K. Shah: None. N.A. Mayr: B. Research Grant; NIH R01 CA 71906.

1006

Gross Tumor Volume (GTV) Target Delineation: Comparison of Expert and Non-expert Contouring with Quantitative PET Parameters

J. Kalpathy-Cramer1, J. Duppen2, J. Nijkamp2, C. R. N. Rasch2, C. R. Thomas1, C. D. Fuller3 Knight Cancer Institute, Oregon Health & Science University, Portland, OR, 2Netherlands Cancer Institute-Antoni van Leeuwenhoek Hospital, Amsterdam, Netherlands, 3University of Texas Health Science Center at San Antonio, San Antonio, TX

1

Purpose/Objective(s): Target delineation for lung radiotherapy is a highly operator dependent process. PET-CT imaging has been demonstrated to improve inter-observer region of interest (ROI) variation in target delineation. The specific aims of this analysis were to determine whether quantitative PET segmentation parameters for expert user-derived ROIs were substantially distinct from non-expert ROIs using quantitative analysis. Materials/Methods: Seventeen radiation oncologists (4 experts, 13 non-experts) were provided a coregistered PET-CT dataset of a standardized lung radiotherapy case, allowing visual display of PET intensity map (e.g., without SUV values) and matching CT dataset. Users contoured the case twice (once with a mouse, once with a pen-tablet device) using a custom delineation analysis software (Big Brother, NKI-AVL). GTV volumes were collected for central analysis. ROIs were converted to DICOM-RT format and exported to an analytic software (TaCTICS), http://skynet.ohsu.edu/tactics/, allowing calculation of a composite expert and non-expert ROIs using Warfield’s simultaneous truth and performance level estimation (STAPLE) algorithm. The Planning PET