Morphology-Based Label-Free Cell Separation System Utilizing a Cell-friendly Photoresist and Deep Learning

Abstract Cell separation technology is essential for cell-based analysis and therapy. However, high-fidelity isolation of specific cell subsets from limited specimens remains a significant challenge, as conventional methods such as fluorescence-activated cell sorting (FACS) and magnetic-activated cell sorting (MACS) often require large starting populations of labeled cells and disruptive suspension states. To address this challenge, a label-free, morphology-based cell separation platform is developed to selectively isolate adherent cells while preserving their native state. Three representative cell types found in tumor tissues—cancer cells, fibroblasts, and macrophages—are used as model systems. Cell mixtures are seeded onto a cell-friendly photoresist surface, which dissolves in cell culture media upon light exposure without damaging cells, and differential interference contrast (DIC) images are acquired. A deep learning algorithm based on a Mask R-CNN model is employed to recognize, classify, and generate segmentation masks for cells in real time. The classification model achieve high accuracy, with all diagonal values exceeding 0.8 at an IoU threshold of 0.5. By projecting light through a dynamic photomask generated via a digital micromirror device (DMD), target cells attached to the photoresist surface are selectively detached with detachment efficiencies over 90% for all cell types. This platform enables automated, label-free, and spatially selective cell isolation from small samples and demonstrates scalability through multi-position operation using a motorized stage.