US 12,336,827 B2
Cardiac transmembrane potential reconstruction method based on graph convolutional neural network and iterative threshold contraction algorithm
Huafeng Liu, Hangzhou (CN); and Lide Mu, Hangzhou (CN)
Assigned to ZHEJIANG UNIVERSITY, Hangzhou (CN)
Appl. No. 17/625,766
Filed by ZHEJIANG UNIVERSITY, Zhejiang (CN)
PCT Filed Aug. 19, 2021, PCT No. PCT/CN2021/113447
§ 371(c)(1), (2) Date Jan. 9, 2022,
PCT Pub. No. WO2022/262109, PCT Pub. Date Dec. 22, 2022.
Claims priority of application No. 202110682912.1 (CN), filed on Jun. 18, 2021.
Prior Publication US 2023/0263450 A1, Aug. 24, 2023
Int. Cl. A61B 5/318 (2021.01); G16H 50/50 (2018.01)
CPC A61B 5/318 (2021.01) [G16H 50/50 (2018.01); A61B 2576/023 (2013.01)] 7 Claims
OG exemplary drawing
 
1. A cardiac transmembrane potential reconstruction method based on graph convolutional neural network and iterative threshold contraction algorithm, comprising the following steps:
(1) performing enhanced CT plain scan on the patient's torso, and acquiring enhanced CT slice images of the patient's thoracic cavity;
(2) establishing a finite element model of a heart surface and a torso surface according to the enhanced CT slice images, and finding the positive relationship between a body surface potential and a cardiac transmembrane potential, that is, Φ=HU, where Φ is the body surface potential and U is the cardiac transmembrane potential, H represents the forward conversion matrix between H and Φ;
(3) collecting the patient's multi-lead body surface ECG signal Φ, and filtering the ECG signal;
(4) according to the obtained Φ and H, initializing the cardiac transmembrane potential signal;
(5) according to an initial value of the cardiac transmembrane potential signal, the cardiac transmembrane potential is reconstructed by an iterative threshold contraction algorithm of embedded a graph convolutional neural network;
wherein, in the step (3), the patient is made to wear a body surface potential detection device with 64-lead electrodes distributed for measurement, and the patient's 64-lead body surface ECG signal is collected and obtained, and the 64-lead body surface ECG signal is processed by wavelet transform to flatten and denoise the signal;
wherein, a formula for initializing the cardiac transmembrane potential signal in the step (4) is as follows:

OG Complex Work Unit Math
wherein, u(0) is a initial value of the cardiac transmembrane potential signal, ∥ ∥F represents a F norm, T represents a transposition, and Q represents a reverse conversion matrix between Φ and U,
wherein, the specific implementation of the step (5) is as follows:
(5.1) mapping update steps in ISTA to a graph convolutional neural network model composed of a fixed number of block cascades, and each block corresponds to an iteration in ISTA;
(5.2) replacing a regularization matrix in a L1 norm regularization with a nonlinear transformation function corresponding to a block, the following objective reconstruction equation is established:

OG Complex Work Unit Math
wherein, ∥ ∥2 represents a 2 norm, ∥ ∥1 represents a 1 norm, λ is a regularization coefficient, and Γ( ) is the nonlinear transformation function corresponding to a block;
(5.3) optimizing the above-mentioned objective reconstruction equation by using ISTA, each iteration in ISTA is divided into two steps: the first step is gradient descent, and the second step is to solve a near-end operator;
(5.4) training the graph convolutional neural network model based on ISTA by using existing data samples, and finally using the trained network model to solve a ECG inverse problem, so as to complete the reconstruction of the cardiac transmembrane potential.