[The Monte Carlo method and parallel estimation in the drawing up of radiosurgery treatment plans]

We investigated the practical application of a calculation algorithm based on the Monte Carlo method to stereotactic radiosurgery treatment planning. In radiosurgery, high dose gradients and the lack of electronic disequilibrium make high resolution matrices and high computing power and speed necessary to obtain accurate dose distribution. To date, the main obstacle to the wider-spread use of the Monte Carlo method has been the huge computing time necessary to obtain a dose distribution on current hardware.

In this project, developed within the ESPRIT program, funded by the European Union, a Parsytec CC (Cognitive Computing) computer was used with 9 processors (Power PC 604, 133 Mhz, RAM 64 Mb) with IBM AIX/EPX OS and availability for Fortran parallel codes compilation, connected to a PC for data input, results rendering, and dose distribution calculation with a conventional algorithm for comparison with the Monte Carlo code (an EGS4 user code). The module named Rapt Region Extractor performs data compression with an octree method without decreasing resolution, for RAM and computing time requirements to remain acceptable. A model of the 6 MV photon beam from Clinac 2100C Varian linear accelerator was devised, based on incident photon energy spectrum and, for each collimator dimension, on bidimensional dose distribution orthogonal to beam direction measured at SSd = SAD = 100 cm.

Parallelization was carried out on event numbers, allowing a simulation speed to number of processor ratio close to unity. A new random number generator was used, capable of correctly running on the parallel architecture. The simulation procedure includes: 1) CT acquisition in DICOM 3.0 format, Analyze or with scanner; 2) Target delineation, treatment arc definition. 3) Dose calculation, with both conventional and Monte Carlo methods. 4) Dose distribution rendering on every transverse, sagittal or coronal planes overlapped in color wash on anatomical representation. Comparison between conventional and Monte Carlo algorithms were carried out on an anthropomorphic phantom and 10 real patients, with 2.5 mm anatomical resolution and standard deviation never exceeding 2%. A simulation with 10,000,000 events and 1% maximum variance can be run in 43'. When PTV is an homogeneous areas the differences between the two methods are around 5%, while when PTV is localized in dishomogeneous areas discrepancies reach 20% in the bone.

In conclusion, the feasibility of direct simulation with the Monte Carlo method in radiosurgery has been demonstrated within time and hardware costs compatible with clinical practice.