Label-free live cell analysis in transmitted light using cross-channel supervised deep learning

Label-free transmitted-light (TL) imaging offers a non-invasive route to monitor live-cell responses in high-content screening, but robust segmentation and viability classification remain difficult because TL images have low intrinsic contrast, poorly defined cellular boundaries, and substantial morphological variability after compound treatment. We developed an AI-enabled workflow in IN Carta Image Analysis Software (Molecular Devices) that uses cross-channel supervision to generate fluorescence-derived ground-truth masks from Hoechst- and Ethidium Homodimer-1-positive populations. These masks are then used to train TL-based segmentation models for label-free inference. The resulting masks were used with TL-derived morphology, intensity, spatial, and texture features to enumerate live cells and support quantitative compound-response analysis. Across U2OS, HeLa, HEK, and HCT116 cells, TL segmentation achieved precision above 0.9 for most compound and concentration conditions, with many conditions exceeding 0.95. F1 scores were above 0.9 for the majority of treatments, with modest reductions in highly perturbed or fragmented samples. Performance was also retained when trained models were applied to additional experimental plates, supporting the consistency of the workflow across experiments. Live-cell detection was the most stable readout, whereas dead-cell classification was more variable in conditions with heterogeneous staining, fragmented cells, or morphological ambiguity. These results demonstrate that fluorescence channels can be used during model development to overcome the TL annotation bottleneck, while the final analysis can be performed from transmitted-light images alone. The workflow provides a scalable framework for label-free cell counting and viability assessment in high-content screening, reducing assay complexity and sample perturbation while maintaining quantitative performance across multiple cell models.