Detectors¶

The detector turns the optical image (photons per second at each pixel) into the digital image (integer gray values). This is where the exposure time enters, and where most of the noise comes from.

sim = ms.Simulation(..., detector=ms.CameraCMOS(qe=0.8, read_noise=1.5))

In dictionary form, camera_type selects the class: detector={"camera_type": "EMCCD", "em_gain": 300}.

Exposure time¶

Simulation.exposure_ms is the integration time per pixel, in milliseconds:

• for a camera, the exposure time of each frame;
• for a point scanner, the pixel dwell time (typically 0.0005–0.01 ms).

A channel can override it with OpticalConfig.exposure_ms, and digital_image(exposure_ms=...) overrides both, which is a quick way to compare exposure times on one optical image:

optical = sim.optical_image()
short = sim.digital_image(optical, exposure_ms=10)
long = sim.digital_image(optical, exposure_ms=1000)

The detector simulation pipeline¶

All detectors run the same sequence of steps. They differ in their parameters, and in which steps do nothing for them.

1. Expected signal. Photons per second × integration time gives the mean number of photoelectrons per pixel. Dark current (and clock-induced charge, for cameras) is added.
2. Shot noise. The number of photoelectrons is drawn from a Poisson distribution with that mean.
3. Full well. A camera pixel cannot hold more than full_well electrons.
4. Multiplication. EMCCDs and PMTs multiply each electron by a random gain before readout. The randomness adds noise.
5. Readout. The amplifier adds Gaussian read noise, and the ADC converts electrons to gray values, rounded and clipped to the top of the bit_depth range. Then any digital gain and the offset are applied, and the result is clipped to the range again.

The parameters shared by all detectors:

Give either electrons_per_adu or relative_gain, not both. For cameras the default is relative_gain=1, which maps a full pixel to the top of the ADC range.

SNR against signal¶

Cameras¶

CCD¶

CameraCCD: all pixels share one output amplifier. Typical values: QE 0.5–0.7, read noise 5–10 e⁻, full well 15,000–20,000 e⁻.

EMCCD¶

CameraEMCCD adds an electron-multiplying register before the amplifier. Each of its em_stages cells (604 by default) multiplies the charge by a small random amount, for a mean gain em_gain (typically 100–1000). The signal is lifted far above the read noise, so single photons become detectable, but the excess noise factor rises toward 2: at high gain an EMCCD has the SNR of a noise-free detector with half its QE.

ms.CameraEMCCD(qe=0.95, em_gain=300, read_noise=50, electrons_per_adu=20)

With an EMCCD, electrons_per_adu is usually easier to reason about than relative_gain, because relative_gain refers to the pixel full well, before the EM gain.

CMOS and sCMOS¶

CameraCMOS: each pixel has its own amplifier. Typical values for a scientific CMOS: QE 0.7–0.95, read noise 1–2 e⁻, electrons_per_adu around 0.5 at 16 bits.

Point detectors¶

A point scanner records one pixel at a time, so its detectors have no pixels to hold charge, no full well, and no relative_gain. Set their gain with electrons_per_adu, or the parameters below. exposure_ms is the pixel dwell time.

PMT¶

PMT models a photomultiplier tube read out in analog mode. A photoelectron from the photocathode is multiplied by a chain of dynodes, and the resulting current is integrated over the dwell time.

• Gain. Set the high voltage hv as on the microscope. The gain follows a power law, ref_gain × (hv / ref_hv) ** hv_exponent. The defaults model a Hamamatsu GaAsP photosensor module controlled by a 0.5–0.9 V signal, with a gain of 5 × 10⁵ at 0.8 V. To set the gain directly, use hv_gain instead.
• Gain noise. The excess noise factor follows from the gain per dynode, and increases as the HV is lowered (about 1.3 at 0.9 V, 1.5 at 0.5 V for the default tube).
• Digital gain. digital_gain multiplies the digitized signal, like the digital gain setting in some acquisition software. It amplifies signal and noise equally.
• Saturation. If max_output_current_ua is set, the gain drops as the average anode current approaches that limit. The current is averaged over divider_time_constant_ms along the scan, so a bright region dims itself and the lines that follow it.
• Amplifier bandwidth. If time_constant_us > 0, the signal of each pixel leaks into the next few pixels along X, producing a streak at short dwell times.

HyD¶

HyD models a hybrid photodetector (as in Leica HyD detectors) in photon-counting mode. Every photon gives nearly the same pulse, so pulses are counted: no gain noise and no read noise. With the default electrons_per_adu=1, the output is the photon count.

After each pulse the detector is blind for dead_time_ns (1.5 ns by default), so at high count rates some photons are missed: a true rate \(n\) is recorded as \(n / (1 + n\tau)\), about 6% low at 40 million counts per second. Dead time also makes the counts more regular than Poisson.

Detector library¶

microsim.schema.detectors.lib has detectors with published or measured parameters:

The source code of detectors/lib.py lists where each value comes from.

The detector docstrings in the API reference describe each model in more detail, including what is not modeled.