Classifying cell viability using a label-free approach: Integration of phase-contrast imaging, Raman spectroscopy, and deep learning

Cell viability assays have been widely used to discover and evaluate a compound in drug development and cancer research. The conventional cell viability assays usually rely on colorimetric and fluorescence techniques to quantify cellular metabolism and viability. However, these techniques may potentially damage cells, affecting the accuracy of cell viability measurements. To address these concerns, we developed an automatic cell viability classification system based on a simple biological microscope to predict cell viability from label-free phase-contrast images. The system automatically obtains single-cell positions and images, which are then analyzed by a trained deep-learning model to classify cell viability. After image acquisition, the single-cell viability was used to Raman spectroscopy to quantify the MTT formazan, which correlated to cellular metabolic activity. For classifying cell viability, three CNN-based models (VGG-16, DenseNet-121, and Xception) were employed. According to the results, the VGG-16 model achieved the highest performance, with an average accuracy and sensitivity of 89 % in 3-fold cross-validation to classify cell viability between non- and minor-damaged conditions. The results therefore demonstrated that the developed system provides a simple and efficient solution for classifying cell viability, with promising applications in biomedical research and drug screening.