CT image reconstruction method based on pulse neural network

Abstract

A CT image reconstruction method based on a spiking neural network comprises the steps that the spiking neural network of a plurality of neurons with a pulse discharge function is constructed, and an external input current vector at the initial moment of the neurons is a product obtained by sequentially multiplying a CT sampling signal with a sampling matrix and a Dobezi wavelet basis matrix; a leakage integration distribution (LIF) model spiking neural network is adopted for simulation evolution, that is, potential changes of biological neurons are simulated through an RC circuit, a discharge rate vector is obtained when iteration is carried out until the average receiving current of the spiking neural network converges, and then a CT image reconstruction result is obtained through multi-Bezi wavelet basis matrix reconstruction. According to the invention, by setting excitatory or inhibitory neuron connection, the flexibility and adaptability of the network are increased, different optimization problems can be adapted by adjusting the external input current of the neuron, and along with the expansion of the scale of solving the optimization problems, better advantages in energy consumption and processing time are shown.

Description

Technical Field

[0001] The present invention relates to a technology in the field of neural networks, in particular to a CT image reconstruction method based on a pulse neural network. Background Art

[0002] Computed tomography (CT) is a widely used imaging technology in medical diagnosis and treatment, which uses X-rays to generate detailed images of the internal structures of the human body. Currently, the optimization algorithms used in the field of CT image reconstruction, such as gradient descent and greedy algorithms, are limited in efficiency when dealing with large-scale problems, thereby increasing computational overhead and time consumption, while the spike local competition algorithm relies on fully inhibitory connections between neurons, which limits the range of optimization problems it can effectively solve. In addition, the neural network architecture needs to be adjusted for different optimization problems, which means that when implemented on neuromorphic chips, the network architecture needs to be modified for different problems, reducing the versatility of the algorithm and the adaptability of the hardware. Summary of the invention

[0003] In view of the above-mentioned deficiencies in the prior art, the present invention proposes a CT image reconstruction method based on a spiking neural network. According to the dynamic characteristics of biological neurons, a spiking neuron model is designed based on the mathematical characteristics of biological neurons including the LIF model and the Hodgkin-Huxley model. By setting excitatory or inhibitory neuron connections, the flexibility and adaptability of the network are increased. The method can adapt to different optimization problems by adjusting the external input current of the neurons. As the scale of solving the optimization problem increases, the method shows better advantages in energy consumption and processing time.

[0004] The present invention is achieved through the following technical solutions:

[0005] The present invention relates to a CT image reconstruction method based on a pulse neural network. A pulse neural network of several neurons with pulse discharge function is constructed. The external input current vector at the initial moment of the neuron is the product obtained by sequentially multiplying a CT sampling signal with a sampling matrix and a Daubechie wavelet basis matrix. A leakage integrate-and-fire (LIF) model pulse neural network is used for simulation evolution, that is, an RC circuit is used to simulate the potential change of biological neurons, and a discharge rate vector is obtained when the average receiving current of the pulse neural network converges through iteration, and then a CT image is obtained through reconstruction of the Dobechi wavelet basis matrix. Technical Effects

[0006] The present invention designs a generalized spiking local competition network (Generalized Spiking LCA) model based on the spiking neuron model and the brain network connection structure, and adds different external input currents during network evolution to make the network's discharge rate converge to the solution of multiple optimization problems. Compared with the prior art, the present invention significantly improves the computational efficiency of processing large-scale optimization problems, reduces computational overhead and time consumption. Increase the adaptability of algorithms and hardware: The algorithm can adapt to a variety of optimization problems without modifying the network architecture for different problems, thereby improving the adaptability and economy of hardware implementation. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 Iterate the flowchart for the neuron model;

[0008] Figure 2 It is the structure diagram of the pulse neural network;

[0009] Figure 3 It is a flow chart of the present invention;

[0010] Figure 4 It is a schematic diagram of the effect of the embodiment;

[0011] Figure 5 Schematic diagram of the implementation effect for various network dimensions. DETAILED DESCRIPTION

[0012] like Figure 1 As shown, this embodiment relates to a CT image reconstruction method based on a pulse neural network, comprising:

[0013] Step 1: Construct Figure 2 The pulse neural network with N neurons shown in the figure has a connection strength matrix W = Φ between each neuron T Φ-E, where: E is the identity matrix, w ij is the i-th row and j-th column element of the connection strength matrix W, which represents the connection strength between neuron i and neuron j. The external input current vector b = Φ at the initial moment of the neuron T s, where: b = [b1, b2, ..., b N ], vector s=[s1,s2,…,s M ] is the CT sampling signal, M is the dimension of the CT sampling signal, and the matrix Φ=[φ1,φ2,…,φ N ], can be expressed as the multiplication of two matrices, Φ = AΨ. Matrix represents the Dobesi wavelet basis matrix, matrix is a discrete random measurement matrix, where M<<N (in the embodiment, the size ratio ).

[0014] Step 2: Use the leaky integrate-fire (LIF) model to simulate the evolution of the spiking neural network constructed in step 1, that is, simulate the potential change of biological neurons through the RC circuit, iteratively calculate the discharge rate of each neuron, obtain the sparse representation of the CT image under the Dobesi wavelet basis, and then obtain the CT image reconstruction result through the matrix Ψ, which specifically includes:

[0015] 2.1) According to the LIF model, the membrane potential V of the i-th neuron in the spiking neural network i (t)Satisfy: like Figure 1 As shown, the neuron membrane potential V at time t i (t) and the external input current I i (t) Get the value of neuron membrane potential V at time t+1 i (t+1), when V i (t+1) is greater than the potential threshold V th , then the neuron is judged to be discharged at this time and the kth discharge time t of neuron i is recorded i,k = t+1, at the same time neuron i sends a pulse signal to all other connected neurons and resets to the resting potential V rest =0, where: R m is the resistance value of the neuron, C m is the capacitance of the capacitor in the RC circuit, I i (t) is the external input current.

[0016] In the LIF model, when the external input current is a constant value, that is, I i (t) = I con , I con ∈[0,+∞), the neuron discharge rate a spike The function relationship between the current and the external input current is g(·), that is, a spike =g(I con ).

[0017] 2.2) If Figure 3 As shown, simulate evolution, including:

[0018] ① At the initial time t0, for each neuron in the pulse neural network, the external input current I i (t0) = b i , neuronal membrane potential V i (t0)=0,i=1,2,…,N,b i is the i-th component of the external input current vector b;

[0019] ②At time t, through the neuron membrane potential V i(t) and the external input current I i (t), and simultaneously update the membrane potential of all neurons in the spiking neural network to obtain V i (t+1), i=1,2,…,N. When V i (t+1) is greater than the potential threshold V th , then the neuron is judged to be discharged at this time and the kth discharge time t of neuron i is recorded i,k =t+1;

[0020] ③Calculate the receiving current μ of each neuron i (t+1)=b i -∑ j≠i w ij α(t+1)*σ j (t+1), i=1,2,…,N, where: α(t+1)=e -t-1 is the decreasing function of synaptic current, σ j (t+1)=∑ k δ(t+1-t j,k ) is the discharge event sum of neuron j, δ(·) is the impulse function of the pulse signal, t j,k is the time of the kth discharge of the jth neuron, and * represents the convolution operation.

[0021] ④ At time t1, calculate the average current of neuron i at time t=t1: The average discharge rate is Where: i (t) = ∑ k δ(tt i,k ) is the sum of discharge events of neuron i at time t, i = 1, 2, ..., N. Here t0 refers to the initial time referred to in ①, in this embodiment, t0 = 0, t1 refers to the time value of each iteration, in this embodiment, t1 = 1, 2, ..., 2000.

[0022] ⑤ Update the external input current at time t+1 to I i (t+1)=g -1 (T λ (u i (t+1))), where: λ is the regularization parameter, so the voltage membrane potential V at time t+1 is obtained i (t+1) and the external input current I i (t+1);

[0023] ⑥ Execute steps ② to ⑤ several times in a loop. When the average current u i(t) When convergence, the discharge rate vector a = [a1, a2, ..., a N ].

[0024] The convergence mentioned here means: satisfying |u i (t+1)-u i (t)|<0.01 for all i=1,2,…,N.

[0025] 2.3) Using the discharge rate vector as the sparse representation of the Dobesi wavelet basis matrix Ψ, the CT image reconstruction result s is calculated. rec =Ψa. It should be noted that at this time s rec is a column vector, and needs to be rec Perform rearrangement operation and convert it into a two-dimensional matrix form to obtain the final CT image s img .

[0026] After specific experiments, the following Figure 2 The spiking neural network of the present invention has a total of 65536 neurons, and after iterative simulation 2000 times using the CT sampling signal on the medical image data set BrainWeb, the discharge rate vector a of the spiking neural network at time t=2000 is calculated to obtain the final CT image reconstruction result.

[0027] like Figure 4 As shown, the peak signal-to-noise ratio PSNR = 20log 10 (255 / ||s ori -s rec ||2) to evaluate the CT image reconstruction effect, where: s ori is the original CT image signal, ||·||2 represents the vector norm, and the existing FISTA algorithm Figure 4 (b) Compared with the present invention Figure 4 The PSNR of (c) reaches 34.6dB.

[0028] Compared with the prior art, the present invention uses a generalized pulse local competition network model designed based on the pulse neuron model and the brain network connection structure. It can achieve a significantly faster convergence speed than the traditional optimization algorithm FISTA in the field of signal recovery. In dealing with CT image reconstruction problems, as the scale of solving the optimization problem increases, the performance of this method is more significantly improved compared with the FISTA optimization algorithm, showing better advantages in energy consumption and processing time. Figure 5 As shown in Figure 3, when the number of network neurons reaches 32,000, the processing time can be improved by two orders of magnitude.

[0029] The above-mentioned specific implementation can be partially adjusted in different ways by those skilled in the art without departing from the principle and purpose of the present invention. The protection scope of the present invention shall be based on the claims and shall not be limited by the above-mentioned specific implementation. Each implementation scheme within its scope shall be subject to the constraints of the present invention.

Claims

1. A CT image reconstruction method based on a spiking neural network, characterized in that: A spiking neural network with several neurons having pulse discharge function is constructed. The external input current vector at the initial moment of the neuron is the product of the CT sampling signal, the sampling matrix and the Daubechie wavelet basis matrix in sequence. The leaky integrate and fire (LIF) model is used to simulate the evolution of the spiking neural network, that is, the potential change of biological neurons is simulated by an RC circuit, and the discharge rate vector is obtained by iterating until the average receiving current of the spiking neural network converges, and then the CT image reconstruction result is obtained by reconstructing the Doubechie wavelet basis matrix.

2. The CT image reconstruction method based on pulse neural network according to claim 1 is characterized in that: The pulse neural network described above has a connection strength matrix W between neurons = Φ T Φ-E, where: E is the identity matrix, w ij is the i-th row and j-th column element of the connection strength matrix W, which represents the connection strength between neuron i and neuron j. The external input current vector b = Φ at the initial moment of the neuron T s, where: b = [b1, b2, ..., b N ], vector s = [s1, s2, …, s M ] is the CT sampling signal, M is the dimension of the CT sampling signal, and the matrix Φ = [φ1, φ2, …, φ N ].

3. The CT image reconstruction method based on pulse neural network according to claim 2 is characterized in that: The matrix Φ is obtained by multiplying two matrices, Φ = AΨ, where: the matrix represents the multiwavelet basis matrix, and the matrix is the discrete random measurement matrix, where M << N (in the embodiment, the size ratio is ).

4. The CT image reconstruction method based on pulse neural network according to claim 1 is characterized in that: The LIF model refers to: the membrane potential V of the i-th neuron in the spiking neural network i (t)Satisfy: The neuron membrane potential V at time t i (t) and the external input current I i (t) Get the value of neuron membrane potential V at time t+1 i (t+1), when V i (t+1) is greater than the potential threshold V th , then the neuron is judged to be discharged at this time and the kth discharge time t of neuron i is recorded i,k = t+1, at the same time neuron i sends a pulse signal to all other connected neurons and resets to the resting potential V rest =0, where: R m is the resistance value of the neuron, C m is the capacitance of the capacitor in the RC circuit, I i (t) is the external input current.

5. The CT image reconstruction method based on pulse neural network according to claim 1 or 4, characterized in that: In the LIF model, when the external input current is a constant value, that is, I i (t) = I con , I con ∈[0,+∞), the neuron discharge rate a spike The function relationship between the current and the external input current is g(·), that is, a spike =g(I con ).

6. The CT image reconstruction method based on pulse neural network according to claim 1, characterized in that: The simulated evolution specifically includes: ① At the initial time t0, for each neuron in the pulse neural network, the external input current I i (t0) = b i , neuronal membrane potential V i (t0)=0,i=1,2,…,N,b i is the i-th component of the external input current vector b; ②At time t, through the neuron membrane potential V i (t) and the external input current I i (t), and simultaneously update the membrane potential of all neurons in the spiking neural network to obtain V i (t+1), i=1, 2,...,N, when V i (t+1) is greater than the potential threshold V th , then the neuron is judged to be discharged at this time and the kth discharge time t of neuron i is recorded i，k =t+1; ③Calculate the receiving current μ of each neuron i (t+1)=b i -∑ j≠i w ij α(t+1)*σ j (t+1), i=1, 2,...,N, where: α(t+1)=e -t-1 is the decreasing function of synaptic current, σ j (t+1)=∑ k δ(t+1-t j，k ) is the discharge event sum of neuron j, δ(·) is the impulse function of the pulse signal, t j,k is the time of the kth discharge of the jth neuron, * represents the convolution operation, ④ At time t1, calculate the average current of neuron i at time t=t1: The average discharge rate is Where: i (t) = ∑ k δ(tt i，k ) is the sum of discharge events of neuron i at time t, i = 1, 2, ..., N; ⑤ Update the external input current at time t+1 to I i (t+1)=g -1 (T λ (u i (t+1))), where: λ is the regularization parameter, so the voltage membrane potential V at time t+1 is obtained i (t+1) and the external input current I i (t+1); ⑥ Execute steps ② to ⑤ several times in a loop. When the average current u i (t) When convergence, the discharge rate vector a = [a1, a2, ..., a N ].

7. The CT image reconstruction method based on pulse neural network according to claim 1 or 6, characterized in that: The convergence mentioned here means: satisfying |u i (t+1)-u i (t)|<0.01 for all i=1, 2,…, N.

8. The CT image reconstruction method based on pulse neural network according to claim 6 is characterized in that: The CT image reconstruction result is obtained by reconstructing the Dobesi wavelet basis matrix, that is, the discharge rate vector is used as the sparse representation of the Dobesi wavelet basis matrix Ψ, and the CT image reconstruction result s is calculated. rec =Φa.

Images