Lightweight CNN-based microfluidic droplet classification for portable imaging flow cytometry

Classifying microfluidic droplets is an essential step in imaging flow cytometry. While deep learning algorithms can detect and classify such droplets in benchtop laboratory settings, their deployment on portable devices remains challenging because the computational requirements often exceed the capabilities of compact, resource-limited devices. This hinders the transition from stationary lab setups to field-deployable instruments. To tackle this issue, we introduce a customized YoloV4-tiny model deployed on a Raspberry Pi-5 (RPi5) single-board computer. Our neural network is trained using 878 images from a custom dataset of 975 images, derived from two videos captured with real-life microfluidic experimental setup. We evaluate performance based on inference time and mean average precision. Our system successfully classifies three distinct droplet types (no cell, one cell, multiple cells) within 13 ms, achieving a 99.95% mean average precision at an intersection over union threshold of 0.5 ([email protected]). We also compare the classification performance metrics of our customized YoloV4-tiny model against seven other combinations of machine learning models and platforms, including a recent low-cost, highly compact edge device with tensor processing unit capabilities, specifically, the MaixCam board with LicheeRV Nano module (SOPHGO SG2002) running a YoloV5-s model. Compared to this proposed customized YoloV4-tiny on the RPi5, the YoloV5-s on MaixCam achieves a significantly shorter classification time (5.34 ms) owing to its onboard tensor processing unit but suffers from a lower [email protected] of 55.09% due to quantization. Our work shows that carefully designed systems can achieve a balance between speed and accuracy, enabling robust performance even on resource-limited devices and paving the way for microfluidic droplet classification in portable imaging flow cytometry.