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.