Label-free AI-driven drug screening in 3D cancer models using longitudinal brightfield imaging

Functional drug screening in three-dimensional (3D) cancer models provides biologically relevant treatment-response evidence. However, viability assessment typically relies on destructive endpoint assays with limited temporal resolution. Label-free brightfield (BF) imaging enables longitudinal non-destructive monitoring of the same 3D cultures while preserving volumetric information, yet extracting reliable drug-response readouts from BF Z-stacks is challenging due to imaging artefacts, densely packed and overlapping cells, and treatment-induced changes in cell morphology. We introduce an automated, label-free, AI-driven workflow combining Cellpose-SAM 3D segmentation with a new BF-specific post-processing module to correct imaging artefacts and merged cell errors. The resulting foreground fraction provides a viability readout at every imaging timepoint rather than only at assay endpoints. Across two biologically distinct 3D cancer models, BR4 (trametinib-treated, BRAF-inhibitor-resistant melanoma) and A-431 (cisplatin-treated epidermoid carcinoma), BF-derived viability agreed strongly with independent endpoint references (CellTiter-Glo 3D: Pearson r = 0.959-0.971; Airyscan confocal live-cell counts: Pearson r = 0.948-0.990). These results support longitudinal label-free BF imaging as a practical AI-enabled approach for generating time-resolved drug-response readouts that complement endpoint assays and help bridge the gap between computational drug screening and functional experimental validation.