Find in Library

Leukocyte quantification from whole blood can aid in detecting and managing infections, cardiovascular diseases, and immune system responses. In addition to traditional leukocyte quantification devices such as flow cytometers and benchtop fluorescent microscopes, smartphone-based particle quantifiers are becoming popular because they provide results at a fraction of the cost. One major limitation of these smartphone-based devices is their dependence on desktop computers for data processing, which keeps them from reaching their true translation potential as point of care (POC) devices. In this paper, we present a computer vision and machine learning-enabled methodology to count particles imaged from our 3D printed smartphone-coupled fluorescent microscope. Multiple convolution neural networks (CNN) using different filter sizes were implemented and trained with various learning rates (0.001, 0.0001) and batch sizes (8,16,32). The performance of these trained networks was then tested on green fluorescent microparticles and leukocytes and compared against the ground truth obtained using ImageJ. An R² value of 0.99 was observed. Next, when cross-validation was done to validate the efficacy of the designed CNN architecture, and the predicted results showed a good correlation (R² = 0.99) when compared against the ground truth. The performance of the trained model was also evaluated on particles conjugated with a different fluorophore and an R² value of 0.99 was observed, showcasing its efficacy and versatility. This trained model was then integrated into an Application Programming Interface (API) and is available online for the broader community usage.