U-Net based image segmentation of mesenchymal stem cells

Cell-based therapy is an attractive strategy for the long-term management of various chronic diseases. Mesenchymal stem cells (MSCs) are a heterogeneous group of cells that have demonstrated clinically relevant therapeutic effects. The proliferative and therapeutic potential of MSCs can be characterized by the culture quality, which is reflected by their morphological phenotype. Morphological analysis has been a robust method for monitoring culture quality, but visual inspection is subjective and time-consuming. Our goal is to develop an automated algorithm to segment MSCs for an objective, non-invasive, and rapid cell assessment. We have built an algorithm to segment MSCs using U-Net architecture trained with 71 phase-contrast micro- graphs having 472 cells. MSC culture images are pre-processed and given as inputs to the trained U-Net model that provides a prediction map for cell segmentation. The U-Net output is then post-processed using morphological operations to get rid of false positive cell detections. Results were validated using visual inspection from MSC experts. Our independent test dataset of 36 images consisted of 186 cells. We obtained a sensitivity of 0.742 and a precision of 0.789 for cell detection and a Dice-Sorensen score of 0.823 ± 0.051 for segmentation. The proposed algorithm shows the potential to segment MSCs with accuracy and robustness higher than conventional U-Net. Automated cell segmentation would enable rapid quantification of cytomorphological features and may also drive stem cell quality control processes.