Machine-learning Enhanced Biomedical Data Analytics for Point-of-care Diagnosis System
With the recent advance of microfluidic-based lab-on-a-chip integration, lensless microfluidic imaging system with super-resolution (SR) algorithm has become a promising solution to miniaturize the conventional bulky optical-lens-based flow cytometer for portable and high-throughput cell detection. The previous lensless microfluidic imaging system however requires a multi-frame-SR based image processing with limited throughput to be realized on hardware. This talk presents two single-frame super-resolution algorithms using online machine-learning for lensless microfluidic imaging. One is based on extreme-learning machine and the other one is based convolution-neuron-network. Both requires only one frame of image to correct the resolution of the lensless images and can be realized with compact hardware (ASIC chip and FPGA) to perform a real-time image processing. A corresponding contact-imaging based microfluidic cytometer prototype is demonstrated for cell recognition and counting. Compared with the commercial flow cytometer, less than 8% error is observed for absolute number of microbeads; and 0.10 coefficient of variation is observed for cell-ratio of mixed RBC and HepG2 cells in solution.