HPC Boosts Medical Physics
When it comes to employing physics in medicine, there are two major fields in terms of their relevance in clinical practice: medical imaging and radiation therapy. A recent paper from an Argentinian research duo addresses how these domains can benefit from high-performance computing techniques.
Medical imaging and radiation therapy both rely heavily on computational resources. Ideally, computational work can be performed in real-time or near-real-time to benefit patient outcome as much as possible, the researchers note.
While execution times have dropped significantly with the advent of faster CPUs, wait times are still problematic. In tomographic image reconstruction, internal dosimetry calculation and radiotherapy planning, accelerating these processes is enormously important, “not only for the patient – whose quality of life improvement is the ultimate goal-, but also for optimizing professional work in a busy hospital environment.”
Over the last several years, the rise of multicore and GPU-based computing has boosted many technical computing domains, including the field of medical physics. The research paper explores the ways that medical physics has benefited from advances in HPC and specifically GPU computing.
The authors describe two typical lines of research in medical image processing, image segmentation and registration, that are good candidates for parallel computing on GPU cores. Image segmentation, which falls under general image processing, involves the identification and further classification of different constituents or textures depicted in a given dataset. In the case of biomedical images, this discovery process is crucial to both diagnosis and therapy. The authors found that implementing an image segmentation algorithm on GPU delivered impressive results, a 15x speedup in comparison to the optimized code running on a CPU-only setup.
The second medical imaging process, known as registration, involves bringing two or more datasets into spatiotemporal alignment. There are many reasons this is done, including diagnostic power enhancement after comparing different modalities, disease follow-up, and assistance in radiotherapy planning. It’s a complex process and the algorithm designed by the researchers requires 30-40 minutes of CPU to register two 512x512x50 voxel datasets. Because the algorithm uses a hierarchical subdivision scheme, the authors are confident that it will benefit from acceleration using parallel computing.
Radiotherapy is the second main area examined in the paper. “In Radiation Therapy, the calculation of the dose delivered by ionizing radiation and the use of optimization algorithms on advanced methods of treatment, are the main areas where GPU programming has its greatest impact,” write the authors. There are different ways of computing this dose. There is a 2D solution, known as the pencil beam algorithm, and a 3D algorithm known as convolution/superposition. The authors note that other research groups have developed reformulated pencil beam and convolution/superposition algorithms for GPU-based processing, with speedups of 200-400x.
At the authors’ home institution, Fundación Escuela de Medicina Nuclear de Mendoza, they are working to refine these techniques using the accelerative power of the GPU when it’s feasible to do so. It’s worth noting that even when an algorithm, e.g. Monte Carlo, is ideal for parallel computation, the complexity of the method can limit the acceleration potential.
The clinical value of this work is the development of treatment plan that strikes the best compromise between the dose of radiation delivered to the tumor and dose received by healthy organs located around it.