Research On Fluorescence Encoded Microsphere Image Detection Method Based On Deep Learning

In recent years,the rapid development of technologies such as genetic testing has placed higher demands on the accuracy and speed of biomolecular detection technology.Fluorescence encoded microspheres are stable biomarkers and have been widely used in high-throughput detection and analysis of nucleic acid sequences,proteins and other biomolecules.Compared with the fluorescence encoded microsphere analysis platform with suspension array technology,the fluorescence encoded microsphere image analysis technology has the characteristics of low cost,simple process and high speed.In order to realize high-accuracy fluorescence encoded microsphere image analysis,this thesis combines the outstanding ability of deep learning technology in the field of computer vision to research on the fluorescence encoded microsphere image detection technology based on deep learning.Moreover,we design a fluorescence encoded microsphere image detection prototype system.In this thesis,the whole architecture of the fluorescence coded microsphere image detection prototype system is researched.The system is divided into hardware subsystem and software subsystem according to their functions.The hardware subsystem is designed with Xilinx Spartan-6 series FPGA as the core.The microsphere images are captured,encoded and transmitted to the software subsystem via USB;the software subsystem receives the microsphere image and implements a deep learning based fluorescence image detection algorithm application.This thesis introduces the system architecture,workflow and hardware and software subsystem implementation in detail.A fluorescence encoded microsphere image analysis algorithm is designed based on the Mask R-CNN target detection network.The pixel-level semantic segmentation of the fluorescence encoded microsphere qualitative image is realized,and the statistical analysis of the corresponding quantitative image is completed.Aiming at the problem that the performance of the trained model is too weak because of the small training set,method of generating the fluorescence encoded microsphere image based on the Cycle-Consistent Adversarial Networks is proposed.The experimental results show that the accuracy of the network training on the extended training set is 94.17% and 95.96%,which is 17.9% and 10.4% higher than that of the synthetic training set.The deep learning based fluorescence encoded microsphere image detection method studied in this thesis combines the deep learning technology with FPGA technology to realize the complete process from acquisition to detection of fluorescence encoded microsphere images.It is an innovative application of fluorescence encoded microsphere,which can achieve high-accuracy biomolecular information detection while reducing application cost and increasing detection speed.