Source-linked AI summary
GPU-based ultra fast IMRT plan optimization
Chunhua Men, Xuejun Gu, Dongju Choi, Amitava Majumdar, Ziyi Zheng, Klaus Mueller, Steve B. Jiang
TL;DR
Changing patient geometry can make a pre-designed radiotherapy plan less suitable, while online adaptation requires replanning within minutes. This paper implements constrained IMRT fluence-map optimization on GPUs and evaluates it against a CPU implementation. In a prostate case, the GPU achieved 20.1–40.1 speedup over the CPU implementation, with no observable differences in the reported dose displays and curves.
Problem
Online adaptive radiotherapy needs treatment plans redesigned from updated patient geometry within a few minutes, while single-CPU IMRT optimization takes at least several minutes.
Source: §1 · p. 2Method
The paper reoptimizes fluence maps to reproduce the original dose distribution within a tolerance, using gradient projection with Armijo line search and a CUDA GPU implementation.
Source: §2.1 · p. 3, §1 · p. 3, §2.2 · p. 4Results
The GPU implementation achieved speedups of 20.1 to 40.1 over the CPU implementation, with no observable differences in DVHs, dose wash, or isodose curves.
Source: §3.2 · p. 7Takeaways & Limitations
For the tested 9-field prostate IMRT case, the GPU generated an optimal plan in 0.2 to 2.8 seconds.
Source: §20 · p. 7Reproducing the original dose distribution may be too rigid for some new geometries, potentially leaving accumulated dose different from the desired distribution.
Source: §20 · p. 8
Abstract
from arXiv · showhide
The widespread adoption of on-board volumetric imaging in cancer radiotherapy has stimulated research efforts to develop online adaptive radiotherapy techniques to handle the inter-fraction variation of the patient's geometry. Such efforts face major technical challenges to perform treatment planning in real-time. To overcome this challenge, we are developing a supercomputing online re-planning environment (SCORE) at the University of California San Diego (UCSD). As part of the SCORE project, this paper presents our work on the implementation of an intensity modulated radiation therapy (IMRT) optimization algorithm on graphics processing units (GPUs). We adopt a penalty-based quadratic optimization model, which is solved by using a gradient projection method with Armijo's line search rule. Our optimization algorithm has been implemented in CUDA for parallel GPU computing as well as in C for serial CPU computing for comparison purpose. A prostate IMRT case with various beamlet and voxel sizes was used to evaluate our implementation. On an NVIDIA Tesla C1060 GPU card, we have achieved speedup factors of 20-40 without losing accuracy, compared to the results from an Intel Xeon 2.27 GHz CPU. For a specific 9-field prostate IMRT case with 5x5 mm^2 beamlet size and 2.5x2.5x2.5 mm^3 voxel size, our GPU implementation takes only 2.8 seconds to generate an optimal IMRT plan. Our work has therefore solved a major problem in developing online re-planning technologies for adaptive radiotherapy.
1. Introduction
Changing patient anatomy can make a plan based on pre-treatment geometry less suitable during treatment, motivating adaptive radiotherapy. This paper implements GPU-based IMRT optimization as part of UCSD’s SCORE effort to speed online replanning.
1. Introduction
- Plans based on anatomy acquired before treatment may deteriorate as tumor and nearby-organ positions or shapes change during the treatment course.Adaptive radiotherapy responds by imaging changed geometry and modifying the plan.Source: §1 · p. 2
- Online adaptive radiotherapy must replan in a few minutes while a patient waits on the treatment couch, and IMRT optimization on a single CPU takes at least several minutes.SCORE evaluates GPUs as an affordable parallel-computing platform for online replanning.Source: §1 · p. 2
- Prior GPU work in radiotherapy addressed tasks such as image registration and dose calculation, but the authors knew of no GPU-based treatment-plan optimization work.Source: §1 · p. 3
- The paper implements an IMRT optimization algorithm with CUDA on GPUs and evaluates it against a corresponding serial C implementation on a CPU.The work is part of UCSD’s SCORE project for online re-planning.Source: §1 · pp. 2–3
2. Methods and Materials
The methods reproduce an original dose distribution within a tolerance by reoptimizing fluence maps, then solve the resulting constrained quadratic objective with gradient projection and Armijo line search. The GPU implementation distributes major calculations across CUDA kernels and stores the sparse dose matrix in CSR format.
2.1 Fluence map (re-)optimization model
- The re-optimization model seeks to reproduce the original dose distribution within a tolerance, rather than generate the “best plan of the day.”Dose is calculated from beamlet intensities and dose-deposition coefficients; evaluation uses quadratic one-sided voxel-based penalties.Source: §2.1 · p. 3
2.2 The optimization algorithm
- Gradient projection uses the negative gradient while maintaining nonnegative beamlet intensities, and Armijo’s rule decreases the step size until the objective meets its acceptance threshold.The objective is convex quadratic; the model’s only feasibility constraints are nonnegative intensities.Source: §2.2 · p. 4
2.3 GPU implementation
- Four CUDA kernels handle dose, objective, gradient, and fluence-map calculations, while simpler arithmetic runs on the CPU host.The dose-deposition matrix is stored in CSR format for GPU sparse matrix-vector multiplication.Source: Figure 1 · p. 5, §2.3 · p. 6
3. Experiments and Results
The clinical prostate IMRT tests found that GPU optimization was substantially faster than CPU optimization, with no observable differences in plan-quality displays. The results also expose limits in GPU scaling and in the per-fraction re-optimization model.
3.1 Clinical testing case
- The clinical case used nine evenly distributed co-planar beams across three beamlet/voxel-size scenarios, with full-resolution dose evaluation.The PTV prescription was 73.8 Gy; unspecified-tissue voxels were enlarged to reduce the optimization problem size.Source: §3.1 · p. 6
3.2 Test results
- CPU–GPU objective values differed by around 10^-6~10^-5, and no differences were observed in DVHs, dose washes, or isodose curves.Figure 2 shows the Scenario 3 GPU plan’s DVHs and dose color wash on a representative CT slice.Source: §3.2 · p. 7, Figure 2 · p. 7
- 0.2 to 2.8 seconds produced optimal plans, with GPU speedup factors of 20 to 40 over an Intel Xeon 2.27 GHz CPU.The tests used an NVIDIA Tesla C1060 GPU on a 9-field prostate IMRT case.Source: §20 · p. 7
- The GeForce 9500 GT produced 1x-3x speedup, while GPUs with 30 multiprocessors delivered speedups similar to the Tesla C1060.For these test scenarios, the authors found speedup factors depended solely on the number of GPU multiprocessors; the largest dataset required less than 1 GB of memory.Source: §20 · p. 8
- In Scenario 3, dose, objective-value, and gradient kernels achieved speedups of 46, 18, and 12, respectively.Dose and objective calculations were parallelized over voxels, while gradient calculation used bixels; parallel-reduction communication cost limited objective-value speedup relative to dose calculation.Source: §20 · p. 8
- The model reproduces the original plan’s dose distribution for each fraction, which may not preserve the desired cumulative dose when geometry changes.The authors propose future models using dose-volume constraints or setting T_j to the desired cumulative dose minus doses delivered in prior fractions.Source: §20 · p. 8
- Optimized fluence maps require a separate MLC leaf-sequencing stage, which can potentially reduce treatment quality.The authors plan to implement direct aperture optimization on GPUs as an alternative.Source: §20 · p. 8