Calibration and deskew

Camera and illumination

Camera calibration converts ADU to nonnegative float32 intensity:

intensity = max((camera_counts - camera_offset) * camera_conversion, 0)

The stage-scan path can additionally divide by the measured detector-X response. Cropping retains the original detector-column offset so that this correction uses the matching calibration columns. Optional illumination correction divides the calibrated data by a CYX illumination field. Empty-channel decisions occur after camera calibration and before illumination division.

These calculations are separate from deconvolution and interpolation. See camera_correct for parameter units and detector-gain behavior.

Instrument and laboratory coordinates

The raw stack is ordered scan, camera Y, camera X. For a lateral scan, plane position s, detector coordinates u, v, and oblique angle theta map as:

X = u
Y = s + v * cos(theta)
Z = v * sin(theta)

The orthogonal interpolator projects each laboratory output point onto adjacent camera planes and uses four source samples. X remains the detector-column axis. Output shape, padding, origin, and scan orientation are determined from the acquisition geometry. See the geometry diagram for the interpolation construction.

Sampling and intensity

Before Z averaging, laboratory spacing is the camera pixel size along each spatial axis. --z-downsample-level N averages N adjacent laboratory Z planes, giving ZYX spacing (N * pixel_size, pixel_size, pixel_size).

Interpolation includes the existing camera-pixel/scan-step intensity factor, pixel_size / scan_step. This convention must be included when comparing absolute intensities or integrated fluorescence with a simulated object. Unsupported regions and padding remain zero. A partly covered Z bin retains the full averaging divisor; its geometric coverage is accounted for during fusion.

The default output casts clipped values to uint16. --save-float32 preserves fractional intensity. Changing output dtype does not change the geometric model.

Execution

CPU kernels parallelize independent detector or output rows. Coordinates retain float64 precision, with float32 interpolation weights and intensities.