Clarification on GetMu coefficient definition and detector signal estimation

[EslamMTaha]

Hello XCIST Team,

I have a question regarding the physics model used for detector efficiency calculations.

When GetMu is called to estimate detection probability, does it return the Total Linear Attenuation Coefficient (mu_total) or the Energy Absorption Coefficient (mu_en)?

I am asking because Rayleigh (coherent) scattering contributes to the total interaction probability, but since it deposits negligible energy, it typically does not contribute to the signal in an energy-integrating or photon-counting detector.

If GetMu returns mu_total, does XCIST assume that Rayleigh events contribute to the signal, or is there an internal mechanism to handle these non-depositing interactions?

[MingyeWu]

Hi EslamMTaha,
Yes, that's correct. Not only Rayleigh scattering, Compton scattering also does not deposit the full photon energy in the detector.
The most rigorous way to model detector behavior is through a spectral response matrix, i.e. for each incident or interaction energy, the matrix provides the resulting detected spectrum. The challenge with this approach is that generating an accurate response matrix usually requires Monte Carlo simulation. In XCIST, the PCCT model already adopts this method, but it's only for the certain material and pixel dimension.
The current energy‑integrating (EI) CT detector model uses the total linear attenuation coefficients, which is the simplest approach and only calculates the interaction ratio. Even so, this approach works reasonably well for high‑Z detector materials because scattering contributes only a small fraction of the interactions, and it can be easily adapted to other detector designs. That said, we are preparing to release a more accurate EICT model that explicitly removes Rayleigh scattering.

[EslamMTaha]

Hi MingyeWu,

Thank you for the quick confirmation and the explanation regarding the spectral response matrix.

To share how I came across this. I was comparing XCIST results against GATE Monte Carlo simulations using monochromatic 120 keV photons.

I observed that the XCIST signal was consistently about 4% higher than the signal produced by GATE. Given that the Rayleigh cross-section at 120 keV constitutes approximately 6% of the total attenuation for this material, the ~4% signal discrepancy aligns very well with the inclusion of Rayleigh events in the total interaction ratio.

[MingyeWu]

Thanks for sharing!
GATE or GEANT4 is an excellent choice when model accuracy and full physics fidelity are essential. In contrast, XCIST (CatSim) is designed to model the entire imaging chain with a good computational efficiency.
Another key difference compared to Monte Carlo simulation is the scatter modeling. XCIST (CatSim) does not include highly accurate scatter physics, so if scatter plays an important role in your study, I would recommend using Monte Carlo‑based tools.

[EslamMTaha]

That trade-off is exactly why I am exploring XCIST. My research focuses on radiomics, which requires generating large datasets to analyze feature stability. Doing this with GATE is prohibitively slow for the volume of data I need, even when employing variance reduction techniques.

Because of this, I was concerned about whether XCIST's simplified scatter modeling would affect the texture features. However, my initial validation has been very promising.

I simulated a water phantom with a Cone Beam CT geometry in both GATE and XCIST. While there is that ~4% signal difference we discussed, the resulting images look quite similar visually. More importantly, the first-order statistics and GLCM features were highly consistent between the two simulations.

[MingyeWu]

That looks promising!
I recommend that you create a more accurate detection model in XCIST, you may refer the original one "Detection_EI.py".
And change the cfg parameter cfg.scanner.detectionCallback = "Detection_EI" to point to the new model.