"FAST, PRECISE, AND EASY-TO-USE METHOD FOR DETECTING CELLS ON CHIP USING FASTER R-CNN"

Cell detection is ubiquitous in the analysis of microfluidic cell assays, in cell biology, immunology, oncology, and toxicology research [1].One major challenge to achieving reliable cell detection in image-based methods is the microfluidic channel network and crowded cell environment, which render difficult the detection of individual cells.Manual annotation is possible; however, analyzing a large amount of data would require hundreds of hours of tedious labor and is subject to user bias.Therefore, automated image-processing methods are of high interest to speed up the analysis, facilitate the work, and avoid user bias.Numerous image-processing methods have been developed to automate cell detection [2]; however, such techniques, often based on intensity thresholding, struggle to discriminate the cells from the complex microfluidic environment and distinguish individual cells in clusters.Machine-learning techniques have emerged that can handle complex scenes in bright-field imaging [3], but they require some expertise to implement and tedious annotation sessions for training.Here, we present a facile method for cell detection in microfluidic arrays using Faster R-CNN.We have built a package based on Faster R-CNN [4], a region-based convolutional neural network that returns the objects' locations as rectangles (Fig. 1).Annotating the image training set is simple and fast (~5-30 min) because annotating only takes two clicks per label, circumventing long and tedious sessions of drawing cell contours.In addition, the models generated can achieve high precision (> 95 % for K562 suspension cells).Finally, our straightforward guide and ready-to-use model enable even those without expertise to install and run a Faster R-CNN model in less than an hour.We trained a neural network using images of suspension cells in a microfluidic array and evaluated the model on new images.The model performs very well for a variety of conditions (Fig. 2): slightly out-of-focus imaging planes, various noise levels, and different spatial resolutions.Moreover, achieving a high precision rate of 94% requires only a small number of annotations (~40), and the performances of the model do not depend on the annotator (Fig. 3).Finally, we demonstrate the use of Faster R-CNN cell detection in a cell killing assay, where cell death is quantified using an apoptosis fluorescent reporter (Fig. 4).The combination of bright field images for cell detection and fluorescence reporters highlights the method's broad potential for applications in cell tracking and gene expression monitoring.We have demonstrated an image-processing method based on Faster R-CNN for cell detection in microfluidic assays.In minutes, one can generate high-precision models (> 95 % precision) and quickly infer cell position on chip (~ 2 s/image).We provide an online resource and guidelines for anyone to set up, train their own model, and run their model.We also made available our pre-trained model, which may already achieve good accuracy without further modification.Due to the ease of implementation and high performance of the method, we envision that Faster R-CNN cell detection will have a broad use in microfluidic cell-based assays.