Deep learning-enabled inertial microfluidic biosensor for high-efficiency automated point-of-care blood cell counting.

While rapid quantification of blood cells is fundamental to modern hematological diagnosis and treatment, traditional hematology analyzers are constrained by bulky, costly hardware and the need for skilled operators, restricting their practical deployment in point-of-care testing (POCT) and primary healthcare settings. This work successfully develops a high-precision, low-cost microfluidic automated sensing platform for point-of-care cell quantitative counting by integrating inertial microfluidic technology with an image recognition algorithm. By introducing a novel loss function for small targets, integrating a multi-scale feature fusion network, and incorporating a hybrid attention mechanism, the recognition and counting accuracy of the algorithm for cells in complex microfluidic images is improved. The cell capture performance of the platform is systematically characterized to elucidate the influence of the Reynolds number (Re) and capture time on the capture yield, thereby establishing the optimal conditions for stable operation. On this basis, a stable linear model between cell-capture images of varying quality and cell concentrations within a specific range is established, and successfully used for Red Blood Cell (RBC) count prediction. Ultimately, this work provides a novel automated sensing platform that is reliable, efficient, and cost-effective for early screening and the monitoring of critical indicators in blood-related diseases.