UC San Diego

Creating efficient, accurate approaches to cytometry is an important problem for clinical diagnostics, biological research, and drug discovery. Cytometry identifies cell types or cell status, separates mature cells from immature ones, detects cancerous cells from healthy normal cells, classifies stem cells during differentiation, and screens drugs based upon how they affect cellular architecture.

Imaging flow cytometry is especially promising in cytometry research area since this image-based cell analysis system is capable of capturing highly sophisticated contents while achieving high-throughput analysis. Analyzing cellular images quickly and accurately is a non-trivial problem. These images are commonly obtained at the microscopic level and therefore are very sensitive to light, are often plagued by visual ‘noise’, and blur easily. Processing these images is highly data-intensive and computationally demanding. Therefore, even state-of-art approaches can achieve either high-throughput or profile the cell contents, but not both. There have consequently been significant demands for a properly designed algorithmic approach, as well as specialized hardware support for it.

This work presents a hardware-accelerated system design for a real-time imaging flow cytometry technique. The main algorithmic approaches in this work are two-folds: 1) morphological feature analysis to describe cellular features and 2) an image segmentation method to classify irregular cell shapes and separate the cellular membrane and nucleus. It first describes a high-throughput and low-latency system design solution for extracting cellular properties from a high frame-rate video. Our system analyzes cell images to understand their mechanical properties, such as shape, size, circularity, or deformability. This work suggests hardware-friendly algorithms and carefully optimized hardware accelerated systems using a reconfigurable hardware, i.e. Field Programmable Gate Arrays (FPGA). Secondly, it describes a streaming data clustering method for image segmentation. Data clustering is commonly used for data analysis but is also a demanding process, even in hardware. The segmentation approach in this work achieves a highly streaming and scalable data clustering solution that runs in the highest throughput in an FPGA while handling high-dimensional data. We evaluate this method and conclude that it outperforms other prior state-of-the-art systems. We generalize our streaming data clustering approach for other clustering problems in various data analysis application domains.