Parameter inversion for cardiovascular hemodynamics based on deep learning

Considering that the existing hemodynamic parameter inversion methods have some problems in practical application, such as a large amount of calculation and easy divergence of iteration, a new cardiovascular hemodynamic parameter inversion method based on deep learning was proposed. Firstly, a multi-scale hemodynamic model coupling one-dimensional model and zero-dimensional model was established. Then, based on convolution neural network and fully connected neural network, a hybrid multi-source-input deep network model for parameter inversion was proposed. Aiming at the problem of noise interference in measurement waveforms, an integrated network model using multiple depth networks was proposed to improve the inversion accuracy. Based on the parameter sensitivity analysis, the parameter inversion experiments of the proposed method were carried out to study the inversion accuracy under different noise levels, and they are compared with those obtained by Kalman filter method. The results show that the prediction errors of blood pressure and blood flow waveform are significantly lower than those of the existing methods. The proposed method can accurately and efficiently realize the parameter inversion of the cardiovascular model, and shows a good application prospect. © 2022, Editorial Department of Journal of Southeast University. All right reserved.