KAHIBARO
42.5. Comparing with Experimental Data
Table of Contents
Detector spectra
When you compare GATE to detector measurements you usually start with spectra, because they are easy to measure and very sensitive to problems in physics, geometry, and digitization. The idea is simple: you simulate exactly the same experimental setup, produce an energy spectrum from the simulation, and overlay it with the measured one. The difficulty is in making “exactly the same” precise enough.
The first step is to define a clear experimental configuration. You need the radionuclide or source type, its nominal activity and reference time, the source geometry (point-like, capsule, extended phantom), the detector model and its main dimensions, and any shielding or collimation. Try to gather manufacturer data sheets for crystal size and material, housing, light guide, and collimator if present, and use these to build the GATE geometry. Small geometry mistakes, such as slightly wrong source–detector distance, can noticeably change count rates and spectral shape.
To make a meaningful comparison, you must reproduce the same acquisition conditions. In the experiment you know the live time, acquisition mode, and any energy windowing or pile-up rejection done by the electronics. In GATE you configure the source activity or number of primaries, the simulation time, and the digitizer chain so that the simulated count rate and dead-time behavior, if modeled, correspond to the measurement. If you only want to compare spectral shape, you can ignore absolute count rate and later normalize the spectra.
The digitizer configuration is often the key point. The measured spectrum already includes detector energy resolution, thresholds, and electronic effects. A spectrum produced directly from hits or deposited energy in the crystal will look too sharp. In GATE you add energy blurring with a resolution that matches the detector at the relevant energy. For example, a NaI(Tl) detector may have 10 percent FWHM at 662 keV. You can model this as a Gaussian resolution, often with a dependence on $\sqrt{E}$:
A common model for Gaussian energy resolution is
$$
\frac{\sigma_E}{E} = \sqrt{\left(\frac{a}{\sqrt{E}}\right)^2 + b^2}
$$
where $E$ is in keV or MeV, $a$ is the stochastic term, and $b$ is a constant term. FWHM is related to $\sigma_E$ by
$$
\text{FWHM} = 2.355 \, \sigma_E.
$$
You tune $a$ and $b$ so that the simulated photopeak width matches the measured one at one or more calibration energies. GATE allows you to implement such formulas in the digitizer.
Once geometry, physics list, and digitizer are in place, you run the simulation and extract the spectrum, typically from singles or hits in the sensitive volume. The spectrum must be processed in the same way as the measured data. Use the same bin width, the same energy range, and, if possible, the same calibration. For a fair comparison, convert both to the same units, for example counts per second per keV, or normalize both spectra to unit area if you are most interested in shape.
When you overlay the spectra, focus on several features. The photopeak position should match the gamma energy after applying the same calibration. If there is a systematic shift, check your energy calibration or the digitizer’s gain. The peak width should match within a few percent; large differences suggest incorrect energy resolution modeling. The Compton continuum should have a similar shape and end point. Differences here often indicate missing or incorrect physics processes, or incorrect detector materials.
You should also compare the relative intensities of different peaks if you have multiple gamma lines. Branching ratios and escape peaks are sensitive to both physics and geometry. For example, single and double escape peaks from pair production will be present only if your energy threshold is low enough and your simulation allows high-energy gamma interactions in the detector and surrounding materials.
Use quantitative metrics to support visual inspection. For photopeaks you can compare peak position, FWHM, and peak-to-Compton ratios. You can calculate a chi-squared or reduced chi-squared between measured and simulated spectra over a given range, after appropriate normalization. Remember that both experimental and Monte Carlo spectra have statistical uncertainties in each bin. To reduce Monte Carlo noise, simulate enough events so that simulated uncertainties are significantly smaller than experimental ones in the energy range of interest.
When spectra do not match, change only one aspect at a time. If the peak is too narrow, adjust resolution parameters. If the Compton edge is in the wrong place, verify energy calibration and physics list choice. If the absolute counts differ, check source activity, solid angle, and acquisition time, and whether dead time is present in the real detector but not modeled in GATE. Careful iteration like this allows you to refine the simulation until spectra agree within your target accuracy, for example within a few percent in key regions.
Dose measurements
Dose comparisons follow the same philosophy as spectra, but you work with spatial distributions of absorbed dose instead of one-dimensional energy spectra. Your goal is to simulate the same irradiation and measurement conditions and to show that the calculated dose agrees with measurements within acceptable clinical or research tolerances.
First clarify the experimental setup. For fundamental tests you might use a simple water phantom with ionization chamber measurements along a central axis. For imaging or therapy equipment tests you may use slab phantoms, anthropomorphic phantoms, or patient-specific measurements. You need detailed information about phantom dimensions and materials, detector type and position, irradiation field size and shape, beam energy or spectrum, and any collimation or modulation devices such as multileaf collimators or range shifters.
In GATE you reproduce the phantom geometry and define materials that approximate those in the experiment, often water or tissue-equivalent plastics. Use a dose actor to record absorbed dose in a voxelized grid. The grid resolution must be chosen with care. If your detector has a small sensitive volume, for example an ionization chamber cavity of a few millimeters, a very coarse dose grid will smear out local dose gradients and make direct comparisons unreliable. Typically you choose voxel sizes comparable to or smaller than the detector dimensions.
Next you configure the radiation source to match the experimental beam. For external beams this includes beam energy, field size, source-to-surface distance, and beam modifiers. For internal dosimetry you define the activity distribution used in the experiment. The physics configuration must be appropriate for the energy range and medium. For megavoltage photon beams you include the recommended electromagnetic and, if relevant, hadronic processes. For proton beams you ensure that hadronic and nuclear interactions governing the Bragg peak are correctly modeled.
Dose comparisons require correct normalization. The dose actor usually reports dose in gray per primary particle or per history. The experiment measures absolute or relative dose for a certain number of monitor units or a certain beam-on time at a specified dose rate. To relate the two, you can either scale the simulated dose to match the measured dose at a reference point, or you can use information about machine output (for example, calibration of cGy per monitor unit) to translate delivered monitor units into an equivalent number of simulated primaries. For many validation studies, relative comparisons normalized at a reference depth are sufficient.
You then extract dose profiles and depth-dose curves from the simulation that correspond to the available measurements. For example, for a photon beam in water you compare simulated percentage depth dose along the central axis and lateral dose profiles at several depths. For proton therapy you compare depth-dose distributions to verify Bragg peak position, width, and distal falloff.
To judge agreement you use quantitative metrics. Two very common ones are relative dose difference and gamma analysis.
If $D_{\text{meas}}$ is the measured dose and $D_{\text{sim}}$ is the simulated dose, the point-wise relative dose difference is
$$
\Delta D_{\text{rel}} = \frac{D_{\text{sim}} - D_{\text{meas}}}{D_{\text{meas}}} \times 100 \,\%.
$$
Gamma analysis combines dose difference and distance-to-agreement. For each point in the measured distribution you find the minimum value of
$$
\gamma = \min \sqrt{\left(\frac{\Delta D}{\Delta D_{\text{crit}}}\right)^2 + \left(\frac{\Delta r}{\Delta r_{\text{crit}}}\right)^2},
$$
where $\Delta D_{\text{crit}}$ is the dose difference criterion (for example 3 percent) and $\Delta r_{\text{crit}}$ is the distance criterion (for example 3 mm). Passing means $\gamma \le 1$.
In simple one-dimensional cases, plotting simulated and measured profiles on the same axes allows you to see differences in field size, penumbra width, and depth behavior. Compute key parameters such as depth of maximum dose, half-maximum field width, and, for proton beams, Bragg peak depth. Agreement within one or two millimeters in position and a few percent in dose is often considered good, depending on the application.
Dose measurements themselves have uncertainties. Ion chambers have calibration uncertainties, positioning errors, and correction factors for temperature, pressure, and recombination. Film and array detectors have calibration and response non-uniformity. When you compare to simulation, consider these experimental uncertainties and, where possible, include them in your interpretation. On the simulation side, Monte Carlo statistical uncertainty decreases as you increase the number of histories. For reliable comparisons, aim for simulation uncertainties smaller than experimental ones in regions of interest, often below 1 to 2 percent.
Discrepancies can come from multiple sources. If the simulated depth-dose curve is too shallow or too deep, check beam energy or source-to-surface distance. If lateral profiles are wider or narrower, verify field size, jaw or MLC positions, and phantom geometry. For proton beams, incorrect range suggests problems with material composition or physics settings. For internal dosimetry, mismatches may be related to activity distribution, CT-to-material conversion, or voxel size. As with spectra, change one aspect at a time, run new simulations, and document each modification.
By systematically comparing GATE dose calculations to well characterized experimental measurements, and by understanding acceptable levels of agreement for your purpose, you build confidence in your simulation models. Once validated in simple benchmark scenarios, these models can be extended to more complex clinical or research applications with a clear understanding of their accuracy and limitations.
Views: 9