Digital holographic microscopy(DHM) offers label-free, full-field imaging of live-cell
samples bycapturing optical path differences to produce quantitative phaseimages.
Accurate cell segmentation from phase images is crucial forlong-term quantitative
analysis. However, complicated cellular states(e.g., cell adhesion, proliferation,
and apoptosis) and imagingconditions (e.g., noise and magnification) pose significantchallenge
to the accuracy of cell segmentation. Here, we introduceDL-CSPF, a deep-learning-based
cell segmentation method with aphysical framework designed for high-precision live-cell
analysis.DL-CSPF utilizes two neural networks for foreground-backgroundsegmentation
and cell detection, generating foreground edges and “seedpoints.” These features serve
as input for a marker-controlledwatershed algorithm to segment cells. By focusing
on foreground edgesand “seed points”, which have lower information entropy than completecell
contours, DL-CSPF achieves accurate segmentation with a reduceddataset and without
manual parameter tuning. We validated thefeasibility and generalization of DL-CSPF
using various open-sourceand DHM-collected datasets, including HeLa, pollen, and COS-7
cells.Long-term live-cell imaging results further demonstrate that DL-CSPFreliably
characterized and quantitatively analyzed the morphologicalmetrics across the cellular
lifecycle, rendering it a promising toolfor biomedical research.
