Microfluidic Droplet Classification through Tuned Convolutional Neural Network on a Resource Constrained Platform

Microfluidic droplet classification is an essential research step in imaging flow cytometry. Although deep learning algorithms have proven effective for detecting and classifying microfluidic droplets in benchtop setups, their implementation in portable devices presents a significant challenge. The computational complexity of deep learning needed to meet the speed and accuracy requirements of imaging flow cytometer often exceed the capabilities of portable resource-constrained devices, making it difficult to transition from benchtop setups to field-deployable devices. To address this challenge, we present a tuned YoloV4-tiny model mapped onto a Raspberry Pi-5 single board computer. The neural network is trained on 878 images from our own dataset consisting of 975 images from two videos recorded on our microfluidic experimental setup. The performance is measured in terms of inference time and mean average precision. Our implementation is able to classify three different types of droplets within 13 ms while maintaining an accuracy rate of more than 98% accuracy.