Towards Deep Learning Approaches for Quantitative Analysis of High-Throughput DLD

Abstract Microfluidics has shown great promise for the sorting or separation of biological cells such as circulating tumor cells since the first studies came out a few decades ago. With recent advances in high-throughput microfluidics, analysis of massive amounts of data needs to be completed in an iterative, timely manner. However, the majority of analysis is either performed manually or through the use of superimposing multiple images to define the flow of the particles, taking a significant amount of time to complete. The objective of the work is to increase the efficiency and repeatability of particle detection in a high-throughput deterministic lateral displacement (DLD) device. The program proposed is in the early stages of development, but shows promise. The average time it takes to analyze over a gigabyte of video is 24.21 seconds. The average percent error of the total particle detection was 21.42%. The assumptions made for the initial version of the program affect the accuracy of the particle in wall detection, so new techniques that do not follow the assumptions will need to be investigated. More work will be completed to implement machine learning or deep learning to assist in the development of DLD devices.