KAHIBARO
13.4. Accuracy vs Performance
Table of Contents
Simulation precision
Production cuts and transport parameters directly control how finely GATE follows particles and their secondaries in matter. High precision means that the simulation tracks more of the real physics details, such as low energy secondary particles and short ranges, but this always increases computational cost. When you choose cuts that are too strict, you may obtain beautiful physics detail that you do not actually need for your application while spending far too much CPU time.
In Geant4 and GATE, production cuts are typically expressed as a range, for example a cut of 0.1 mm for electrons. Internally Geant4 converts this range cut into an energy threshold that depends on the material, then decides whether to create a secondary particle or to keep its energy locally deposited. A smaller range cut corresponds to a lower production energy threshold, which allows more secondary particles to be produced and transported. This improves spatial and sometimes spectral accuracy of energy deposition and dose, especially in regions with fine structures such as small detector layers, thin air gaps or interfaces between soft tissue and bone.
For medical physics, sensitivity to cuts depends on the observable that you care about. In many PET or SPECT imaging studies, the main interest is in detecting photons in the hundreds of keV range, and relatively loose electron and positron cuts can already give reliable detection spectra. In contrast, detailed dosimetry in small voxels, proton therapy depth dose assessment or micro-dosimetry for small detectors require tighter cuts to model local energy deposition and scattering accurately. You should also be careful around interfaces because too coarse cuts can smear out sharp dose gradients or change backscatter.
Transport parameters also affect precision. Step limitations, multiple scattering models and step size constraints define how smoothly particles are transported through geometry. Smaller maximum step sizes better resolve boundaries between materials and small volumes, but at the expense of more tracking steps. This can be important in problems where geometric details at the submillimeter scale affect the result, such as narrow collimator holes in SPECT or small detector crystals.
There is no single universal set of cuts that is correct for all problems. A good strategy is to start from recommended defaults for your physics list, then perform a small sensitivity study on a simplified geometry that resembles your application. Vary the cuts and possibly step size limits and observe whether your key observables, such as depth dose, energy spectra or detector count rates, change within your acceptable tolerance. If the observable is stable when relaxing the cuts, you can safely save computation time in the full simulation.
For accurate dosimetry, the absorbed dose $D$ is defined as
$$D = \frac{dE}{dm},$$
where $dE$ is the energy deposited and $dm$ is the mass of the scoring volume. If production cuts are too large, low energy secondaries will be discarded and their energy will be deposited locally in a less realistic way, which can bias $dE$ in small or heterogeneous volumes.
Execution time
Execution time in GATE is dominated by how many particles are simulated and how much work is done per particle. Tighter production cuts and stricter transport parameters increase the number of secondaries and the number of tracking steps. This leads to more geometry boundary checks, more physics process calls and more data for actors and digitizers to handle. Balancing accuracy and performance means adjusting these settings so that you obtain results within your target physical uncertainty while keeping the simulation duration manageable.
From a practical point of view, you can think of performance as a function of both per event cost and number of events. If your cuts are very tight, each event will be slow, but you might need fewer events because each event is more detailed. If your cuts are too loose, events will be fast, but you may introduce systematic bias that does not decrease by adding more events. For most medical physics applications, the useful compromise is moderate cuts that resolve the spatial and energy scales that matter for your observable but ignore unnecessary fine details.
To control execution time, you can gradually relax production cuts, especially for particles or regions that have minor influence on your main results. For example, in a PET simulation, you might keep quite detailed gamma tracking in the detectors and phantom region, while using coarser cuts in the surrounding air or shielding. Region based cuts are designed exactly for this purpose, so that you can keep fine precision where it matters and gain speed elsewhere.
Another practical approach is to monitor event throughput, for example the number of primary events simulated per second, for different cut configurations. By testing a few reasonable settings on a short run, you can quantify how much time each configuration will require for a full simulation and whether the gain in accuracy justifies the extra CPU time. Parallel execution and multithreading help, but even on many cores, overly strict cuts can still make a simulation impractical for large statistics.
For performance tuning, remember two key rules:
- Tighter cuts and smaller step sizes always increase execution time, because they create more secondaries and more tracking steps.
- Relax cuts only after verifying that your main observables do not change beyond your required accuracy.
Balancing these two rules is the core of the accuracy versus performance trade off.
Views: 8