UC-IEG-VIII

From EGI Knowledge Base

Use Case title: e-imrt simulation and diagnostic tool

Short description: eIMRT is a framework for helping radiotherapists to select the best cancer treatment using High Performance Computing which comprises several tools as:

• Monte Carlo methods for treatment verification. They accurately model the interaction of radiation with matter and are the dominant standard in dose calculation techniques. They achieve more accurate results than convolution/superposition methods, at a higher computational cost. Although forthcoming achievements may render Monte Carlo a valid near-real-time treatment planning alternative, it currently plays an outstanding role as a validation technique. Therefore, the eIMRT platform will use it to verify the results of commercial treatment planning systems for any kind of radiotherapy plan of external photon beams.
• Optimization algorithms, based on mixed Monte Carlo C/S dose computation techniques, will produce optimized treatment plans of a quality (in terms of dose conformation and organ sparing) hardly achievable with commercial TPS.

Actors involved:

---

Nowadays, the verification runs in five phases:

• Phase 1: Accelerator simulation. It takes the data describing the geometry of the linear accelerator, its radiation source and the treatment to be verified and produce the input files for the next. Due to the reduced CPU time, this step can be executed locally at the server side. It produces from few files for CRT treatments (about one per angle) to several thousands in case of IMRT.
• Phase 2: Accelerator treatment head simulation. The whole accelerator treatment head is simulated for each input file. This step executes the BEAMnrc code for each input file. It needs about 20 hours of CPU (in a Pentium IV at 3.0GHz) for each treatment. It generates one output file for each input. It executed on Grid.
• Phase 3: Patient simulation. It collects and groups together the output of previous step and converts the CT data describing a patient to densities for further calculation, taking into account the data of the characterization of the computerized tomography (CT). The output of this phase produces the input files for simulating the dose deposition inside the patient using also Monte Carlo. It is executed locally.
• Phase 4: Dose delivered to the patient. The dose-inside-the-patient is calculated using the DOSXYZnrc Monte Carlo code. Since this task is highly parallelizable, it can be divided in many different independent jobs. Currently, the calculation is divided in one job for each incident angle, although divisions with finer granularity are possible. It is executed on Grid and, cumulatively consumes about 35 hours of a single CPU.
• Phase 5: Dose collection and end of process. The results are merged into a final single dose distribution, the process is done locally.

The optimization process is still under development. It comprises three phases:

• Preprocessing.
• Delivered dose calculation to the patient from the different angles and beam energies. It is highly distributed, because each combination of angle-energy can be calculated independently.
• Optimization. A short process which calculates the best treatments. Produce a set of solutions selected by the best DVH. The output shows only these DVHs.

The full CPU-time for optimization depends on the number of angles and energies to be selected. It can be from 15 minutes to several hours, but it should be as much interactive as possible.

The e-IMRT application will benefit mainly from:

Interactivity: In optimization, if some of the results (jobs) are available to the final user before all of them end, the user could interact with the simulation in order to visualize the treatment impact on the human body and changes the requirements for optimization.

Parallelism: There are two levels of parallelism in this application: Coarse grain parallelism, so that the problem has tens to hundreds of independent jobs in verification. Fine grain parallelism, because each MC job in verification can be split in several independent jobs. In optimization is possible to use MPI in a master-slave mode to improve interactivity and time-to-solution (this is work still pending to be done, but plans to have it at 2008Q1)

Visualization: There is a user interface already built that is based on web services and Java and could be integrated in the migrating desktop.