A Facile Method Based on Faster R-CNN for Cell Detection in Microfluidic Devices

Cell detection is ubiquitous in the analysis of microfluidic cell assays. In cell biology, immunology, oncology, and toxicology research, studying cellular response starts with identifying the cells on chip. The large amount of data generated in such assays requires automating image analysis. While multitudes of image processing tools exist, the microfluidic channel network and crowded cell environment make it difficult to identify and track cells by conventional image processing techniques. In contrast, machine learning-based techniques may overcome this challenge. Two important challenges in implementing these techniques are that it often requires tedious image labeling and coding expertise. Here, we present a facile method for cell detection in microfluidic arrays using Faster region-based convolutional neural network (R-CNN) that addresses both challenges. First, image labeling is fast and easy, because Faster R-CNN only needs bounding boxes as labels to generate training data. Second, we provide a ready-to-use model and a guide for training a Faster R-CNN model that does not require coding expertise. We demonstrate that Faster R-CNN does not need trade-offs between precision and user-friendliness: we created a model that detects cells with an average precision over 98% using a few hundred annotations, which takes less than half an hour. We show that shapes created by the microfluidic structure alone or its interplay with cells are not misidentified as cells. We show for the first time cell detection using Faster R-CNN in microfluidic chips; we envision that this approach will have a broad use in many on-chip fundamental biology and drug-discovery assays.