The application of parallel processing to radiotherapy dose computation

Abstract

Delivery of the correct dose to patients undergoing high-energy x-ray radiotherapy treatment for cancer requires accurate simulation of photon beams. Current algorithms using equivalent depth or tissue-air ratio (TAR) correction techniques fail to accurately model electron transport, and hence show inaccuracy in inhomogeneous regions of the body (especially at high photon energies). Newer techniques such as superposition and Monte Carlo simulation do model these effects, but require a large computational effort. Convolution in Fourier space is adequately fast using even standard microprocessor technology, but this approach lacks the flexibility offered by a density-scalable superposition kernel, available only in real space. However, in order for the real-space superposition algorithm to be clinically useful it is necessary to reduce planning times to acceptable levels. This has been achieved using a multicomputer system based on Inmos T800 transputer modules, where the superposition is calculated using a master/worker network employing synchronous message passing. The calculation is partitioned by decomposing the terma array into a number of command vectors or “grains.” The master task sends a vector to each worker task as it becomes free, and the worker performs the superposition for those interation voxels in the vector, storing the result in a local copy of the dose array. Upon completion of the entire calculation, the worker dose arrays are collected and summed to form the complete dose distribution. This approach has a very small communication overhead and exhibits near-linear speedup with increasing processor number. The superposition algorithm has been incorporated into the GRATIS Treatment Planning System developed at the University of North Carolina. This system, based on the Uɴɪx and X standards, has been installed on a Sun SPARCstation network, with the transputer network attached to the Sbus port of one SPARCstation. The treatment planning system can access the transputers from anywhere in the network, using Sun’s RPC (Remote Procedure Call) standard. Energy deposition kernels required by the superposition algorithm have been produced using the EGS4 Monte Carlo Code System (Stanford Linear Accelerator Center). The user code RTPEDK has been developed to generate spherically symmetric kernels, which are then interpolated by the treatment planning system to form cartesian kernels for use in superposition. Two other user codes, RTPCYL and RTPCART, have been developed to model more general radiotherapy problems based on cylindrical and cartesian scoring geometries, respectively.