Accelerating intelligent microfluidic image processing with transfer deep learning: A microchannel droplet/bubble breakup case study

Abstract

Deep learning has recently been leveraged in the microfluidic research for intelligent image processing. However, it usually requires massive human efforts to label sufficient images for obtaining a precise deep learning model. We herein propose an integrated microfluidic image processing method based on transfer deep learning for the rapid deployment of Mask R-CNN, which is capable of reutilizing models trained in previous microfluidic domains to accelerate the analysis of droplets and bubbles in the images of the target microfluidic domain. The transfer deep learning is validated in the microfluidic research with a constrictive microchannel, which is commonly used to break large droplets or bubbles to enhance emulsification. With only 40 labeled images and about 10 min training, the integrated image processing method achieves precise recognition results for both droplets and bubbles in the microfluidic experiments. Applying the trained Mask R-CNN to unlabeled images, the breakup of droplets is found to be highly dependent on the low flow rate, but the bubble breakup is much easier to happen in the same microfluidic equipment. Critical conditions of droplet breakup and size laws of droplet and bubble are precisely determined based on the recognition results of the trained Mask R-CNN.