Brightfield correction pilot

Six separate wells from the same experiment and field index form a normalized median reference. The target field is excluded. First-folder D-row targets use D-row donors to avoid mixing the earlier A-row acquisition regime. Donors are populated images, not confirmed empty controls; median aggregation suppresses variable cells but may retain recurring sample structures.

The target/reference ratio is fitted with an object-masked smooth function (10 µm Gaussian sigma). A shared fitted background is used across z. Saved output is signed float64 relative contrast: (image − fitted background) / fitted background, with a denominator floor. Camera offset is unknown, so this is empirical contrast normalization, not a calibrated transmission measurement. It cannot restore missed focus or information obscured by dust.

Provisional artifact masks exclude reference intensity below 0.65 or above 1.5 times the donor-normalized baseline, plus unsafe denominators and local deviations exceeding 4% in the donor reference, dilated by about 1 µm. The mask may miss artifacts or exclude usable regions and requires review. Cell contours remain to be checked against originals. Each overview uses its own contrast limits; crops use their parent overview limits. Fine-structure correlation is a diagnostic, not a segmentation accuracy score. Pilot measurements and reference provenance.