Application Of CNN Optimization Design Based On APSOC In The Classification Of Congenital Heart Disease

The congenital heart disease,abbreviated as congenital heart disease,is a serious disease that is prone to occur at birth and affects the child’s development and growth.The clinical diagnosis of congenital heart disease is divided into two stages: the initial diagnosis and the diagnosis.The screening of the initial diagnosis mainly depends on cardiac auscultation,which is highly professional and difficult for primary-level doctors to grasp;the diagnosis is mainly through echocardiography.This equipment is expensive and requires professional operation.It is difficult to spread to township health centers.Yunnan is a high incidence area of congenital heart disease,which is limited by the medical conditions in remote mountainous areas of the province.At present,the screening of congenital heart disease mainly depends on the provincial medical team to go to the countryside.The grassroots township hospitals are unable to complete independently.The traditional method of congenital heart disease screening Low efficiency and high cost make it difficult to cover the whole province.Using digital signal processing and artificial intelligence technology to achieve machine-assisted diagnosis is a challenging research topic.This article attempts to use embedded devices,adopt CPU + FPGA heterogeneous deployment and combine deep learning methods to collect and collect heart sound signals in real time.Analysis is expected to achieve the purpose of machine-assisted auscultation.This article first introduces the relevant knowledge of heart sound signals and the physiological and pathological characteristics of electrocardiogram(ECG)signals.Based on the current mainstream all programmable system on chip(APSOC)heterogeneous development platform,a portable heart sound ECG acquisition and Heart sound classification machine assisted initial diagnosis equipment.This research takes APSOC-Zedboard platform as the core,and implements signal acquisition and analysis in a way that software and hardware work together.The waveform drawing of heart sound signals and ECG signals is realized based on CPU architecture;the real-time classification of heart sound signals by convolutional neural network(CNN)is realized based on CPU + FPGA heterogeneous,to achieve the purpose of collecting heart sound ECG signals and real-time classification of heart sound signals.The data acquisition part of this research mainly relies on the CPU resources of the PS part in APSOC-Zedboard to support the Linux system to run the heart sound ECG acquisition program compiled by Qt Creator.The signal is convenient for the user to visualize the related operations of data collection.The heart sound data analysis part of this study uses the CPU + FPGA heterogeneous deployment method to transplant the complete CNN model trained on the PC side to the embedded platform,and evaluates the best deployment plan for maximizing the use of FPGA to achieve CNN real-time Classification heart sound sample data.The experimental results of this study show that the heart sound ECG signals collected by the data collection part in real time and in parallel can be stored in the heart sound database;the accuracy of the CNN classified heart sound sample data realized by the CPU + FPGA best heterogeneous deployment scheme reaches 86% The maximum consumption is 99% of DSP48 E,and the fastest time to classify 200 heart sound samples is 0.77 seconds.It can be optimized for packaging and used in machine-assisted initial diagnosis of congenital heart disease in primary hospitals.