The ideal scenario would be to measure any changes in patient anatomy and recalculate the plan immediately prior to delivering each treatment fraction, while the patient is lying on the couch. Implementing such real-time adaptive radiotherapy in clinical practice is challenging, however, as the replanning must be completed within a few minutes. This is almost impossible to achieve with the central processing unit (CPU)-based computational frameworks used in current clinical settings.
Replanning of intensity-modulated radiotherapy (IMRT) involves three tasks: patient modelling, dose calculation and plan optimization. While parallel processing could accelerate these calculations, traditional supercomputers are too costly and inconvenient for general clinical use. Instead, researchers at the University of California, San Diego (UCSD), are looking to another option: graphics processing units (GPUs). With up to hundreds of processing cores, GPUs make high-performance computation affordable.
The research team – headed up by Steve Jiang, co-director of the University's Center for Advanced Radiotherapy Technologies – has now demonstrated the potential of GPUs to overcome the computational bottleneck of real-time online replanning. The group's initial results are published in two Physics in Medicine & Biology research papers.
Determining dose
The first of the team's publications, co-authored by Xuejun Gu, Jiang and colleagues, describes a GPU-based framework for ultrafast IMRT dose calculation based on a finite-size pencil-beam (FSPB) model, in which the dose deposition coefficient is calculated independently for individual beamlets and voxels (Phys. Med. Biol. 54 6287).
The dose calculation was performed on a Tesla C1060 card from NVIDIA (Santa Clara, CA) – a GPU designed specifically for scientific computation – and using NVIDIA's CUDA development platform. The Tesla C1060 has 30 multiprocessors and 4 GB memory, costs around $1500 and can be readily inserted into a workstation PC. For comparison, the researchers also ran the calculations on a 2.27–GHz Intel Xeon CPU.
The team initially evaluated the code using a 30 × 30 × 30 cm water phantom irradiated by five co-planar 10 × 10 cm, 6 MV beams. The computational time on the CPU ranged from 21 to 124 s. This was reduced to less than 0.5 s when using the GPU – equivalent to a speed-up factor of around 400.
The researchers also tested the FSPB implementation on a clinical prostate-cancer case, calculating the dose deposition coefficients for a nine-field IMRT plan. Here, the sequential CPU computation took about 4.8 min, while the parallel GPU implementation took 0.7 s.
Optimized approach
The second paper in the series, co-authored by Chunhua Men, Steve Jiang and colleagues, details the implementation of IMRT plan optimization on a GPU framework. Again, Jiang and his team used CUDA to implement the optimization algorithm on the NVIDIA Tesla C1060 GPU, and also ran the algorithm in sequential C code on the Intel Xeon 2.27 GHz CPU (Phys. Med. Biol. 54 6565).
The researchers examined a prostate-cancer case, with three combinations of beamlet/voxel size. For each scenario, nine co-planar beams were evenly distributed around the patient and the target dose was 73.8 Gy. The voxel size was 4 × 4 × 4 mm for scenarios 1 and 2, and 2.5 × 2.5 × 2.5 mm for scenario 3. The beamlet size was 10 x 10 mm for scenario 1 (2055 beamlets) and 5 x 5 mm (6453 beamlets) for scenarios 2 and 3.
In all cases, plan optimization was much faster on the GPU than the CPU, with running times reduced from 3.81 to 0.19 s, 16.4 to 0.49 s, and 111.8 to 2.79 s, for scenarios 1, 2 and 3, respectively. This corresponds to speed-up factors of 20.1, 33.5 and 40.1 when using the GPU. The gain achieved by switching from CPU to GPU increased with the size of the optimization problem.
Jiang says that the team has also developed GPU-based algorithms for performing the third task – automated organ segmentation – details of which that will appear in a future publication. However, he notes that there's much work still to be done before treatment replanning on a GPU framework can be widely implemented at clinical sites.
"After developing the GPU-based computational tools required for real-time online replanning (which is where we are now), we need to integrate these tools into clinical treatment-planning systems, and then evaluate and improve their accuracy, efficiency and robustness under clinically realistic conditions," he explained.
"Our next step will be to integrate, test and improve the developed computational tools," Jiang told medicalphysicsweb. "In the meanwhile, a computational infrastructure will be developed to facilitate a streamlined clinical workflow, which differs dramatically from the current clinical workflow."
• Related articles in PMB
GPU-based ultra-fast dose calculation using a finite size pencil beam model
Xuejun Gu et al Phys. Med. Biol. 54 6287
GPU-based ultrafast IMRT plan optimization
Chunhua Men et al Phys. Med. Biol. 54 6565
On-line re-optimization of prostate IMRT plans for adaptive radiation therapy
Q Jackie Wu et al Phys. Med. Biol. 53 673
GPU-based streaming architectures for fast cone-beam CT image reconstruction and demons deformable registration
G C Sharp et al Phys. Med. Biol. 52 5771
Formulating adaptive radiation therapy (ART) treatment planning into a closed-loop control framework
Adam de la Zerda et al Phys. Med. Biol. 52 4137
Real-time 3D computed tomographic reconstruction using commodity graphics hardware
Fang Xu et al Phys. Med. Biol. 52 3405