BACKGROUND
In vitro human cell line models have been widely used for biomedical research to predict clinical response, identify novel mechanisms and drug response. However, one-fifth to one-third of cell lines have been cross-contaminated, which can seriously result in invalidated experimental results, unusable therapeutic products and waste of research funding. Cell line misidentification and cross-contamination may occur at any time, but authenticating cell lines is infrequent performed because the recommended genetic approaches are usually require extensive expertise and may take a few days. Conversely, the observation of live-cell morphology is a direct and real-time technique.
OBJECTIVE
The purpose of this study was to construct a novel computer vision technology based on deep convolutional neural networks (CNN) for “cell face” recognition. This was aimed to improve cell identification efficiency and reduce the occurrence of cell-line cross contamination.
METHODS
Unstained optical microscopy images of cell lines were obtained for model training (about 334 thousand patch images), and testing (about 153 thousand patch images). The AI system first trained to recognize the pure cell morphology. In order to find the most appropriate CNN model，we explored the key image features in cell morphology classification tasks using the classical CNN model-Alexnet. After that, a preferred fine-grained recognition model BCNN was used for the cell type identification (seven classifications). Next, we simulated the situation of cell cross-contamination and mixed the cells in pairs at different ratios. The detection of the cross-contamination was divided into two levels, whether the cells are mixed and what the contaminating cell is. The specificity, sensitivity, and accuracy of the model were tested separately by external validation. Finally, the segmentation model DialedNet was used to present the classification results at the single cell level.
RESULTS
The cell texture and density were the influencing factors that can be better recognized by the bilinear convolutional neural network (BCNN) comparing to AlexNet. The BCNN achieved 99.5% accuracy in identifying seven pure cell lines and 86.3% accuracy for detecting cross-contamination (mixing two of the seven cell lines). DilatedNet was applied to the semantic segment for analyzing in single-cell level and achieved an accuracy of 98.2%.
CONCLUSIONS
This study successfully demonstrated that cell lines can be morphologically identified using deep learning models. Only light-microscopy images and no reagents are required, enabling most labs to routinely perform cell identification tests.
Development of a shear stress-free microfluidic gradient generator capable of quantitatively analyzing single-cell morphology