LuminaPath and the Lab Gap in the Image

Why it surfaced

This is an unusually candid AI project: it reports a model scoring 0.647 macro F1 on expert-labelled real cells, while openly documenting an unresolved edge-sharpness gap. Its core argument is that a diagnostic model can learn the laboratory that produced an image instead of the disease in the cell.

LuminaPath generates physics-based synthetic blood-smear images for cancer diagnostic models, varying the microscope, stain, and disease independently.