Research On FPGA-based Diabetic Retinopathy Detection System

Diabetic retinopathy(DR)is an ocular manifestation of diabetes and the main cause of visual impairment and blindness in the world today.DR has the problem of a explosive increase in medical imaging data and the lack of professional doctors,which delays the treatment of patient.At the same time,DR detection is mostly implemented by software methods such as SVM classifiers and neural networks.There are still a lot of gaps in the embedded deployment of hardware equipment for DR detection algorithms.So this thesis has carried out a research on a FPGA-based diabetic retinopathy detection system to explore new effective methods for the detection and diagnosis of DR.The main research work includes:1.Research on the simultaneous location and detection algorithm and system of optic disc and macula.Compared with using mathematical morphology to realize the sequential location of the optic disc and the macula,this method is based on the YOLOv4-tiny algorithm,which realizes the simultaneous location and detection of the optic disc and the macula,and solves the dependence of the macular location.Through the self-designed convolutional layer and pooling layer IP core,the entire algorithm is transplanted on the Field Programmable Gate Array(FPGA)platform.The experiment uses the recognized COCO data set and the Kaggle-Diabetic Retinopathy Detection(Kaggle-DR)competition data set to train and test the algorithm,and the final average accuracy rate is 96.11%,and it only takes 150.445 ms to detect a picture.This system is the first time to realize a 38-layer medium-sized neural network by using High Level Synthesis(HLS)language.At the same time,it provides a theoretical research foundation for the realization of a handheld device of DR.2.Research on the two-class algorithm and detection system for diabetic retinopathy.First,we build a two-class neural network algorithm model based on Lenet-5 optimization,and then use the encapsulated IP core and time-division multiplexing technology to build the entire diabetic retinopathy detection and diagnosis system.The experiment uses the MNIST binary format data set converted from the Kaggle-DR data set for the verification.The 10-layer convolutional neural network is deployed on the ZYNQ-7000 xc7z010clg400-1 FPGA chip of Xilinx.The average recognition accuracy after 1000 iterations is 95.1 %.This system is conducive to the auxiliary diagnosis of DR,and has great research significance for realizing the neural network on FPGA to carry out the image processing quickly.