CN108230277B - Dual-energy CT image decomposition method based on convolutional neural network - Google Patents

Abstract

The invention relates to the technical field of medical dual-energy image decomposition and image processing, in particular to a dual-energy CT image decomposition method, and particularly relates to a dual-energy CT image decomposition method based on a convolutional neural network. A dual-energy CT image decomposition method based on a convolutional neural network comprises the following steps: designing a convolutional neural network model as a mapping function D (mu) in a dual energy decomposition model_H,L(ii) a Θ); training the convolutional neural network through a convolutional neural network model and a training data set, and effectively estimating parameters theta of the convolutional neural network; and (3) carrying out base material efficient decomposition on the dual-energy CT image by using the trained convolutional neural network and the convolutional neural network parameter theta obtained in the step (2). According to the method, through the establishment of a double-input and double-output convolution neural network model and cross convolution, reasonable shunting of different base materials in the high-energy CT image and the low-energy CT image is realized, so that the quality of base material decomposition of the double-energy CT image is effectively improved.

Description

A Decomposition Method of Dual-Energy CT Image Based on Convolutional Neural Network

technical field

The invention relates to the technical field of medical dual-energy image decomposition and image processing, in particular to a dual-energy CT image decomposition method, in particular to a dual-energy CT image decomposition method based on a convolutional neural network.

Background technique

Dual-energy CT image reconstruction has been increasingly used in medical imaging, security inspection, non-destructive testing and other fields. Compared with the traditional single-energy spectrum CT imaging technology, dual-energy CT can use image attenuation information under different energy spectra to realize Identification of different physical materials. Dual-energy CT technology breaks the physical limitations of traditional single-energy spectral CT, and has become a hot and difficult issue in the field of CT imaging.

The core theory of dual-energy CT imaging technology is dual-energy CT image reconstruction algorithm, in which the decoupling of material and energy cross information is one of the key issues. Compared with the traditional CT image reconstruction theory, the difficulty of the dual-energy CT image reconstruction algorithm lies in the nonlinearity, multi-solution and ill-posedness of the problem, and high dimensionality. Due to the different application requirements of dual-energy CT imaging, different imaging modes used, and different acquired data information, the corresponding dual-energy CT image reconstruction algorithms are also different. At present, dual-energy CT image reconstruction algorithms can be mainly divided into three categories: iterative direct reconstruction algorithms, image reconstruction algorithms based on projection domain preprocessing, and image reconstruction algorithms based on image domain postprocessing.

The iterative dual-energy CT image reconstruction algorithm is suitable for both high and low energy projection sets that are geometrically consistent or inconsistent, and the reconstructed image has a high signal-to-noise ratio. But these methods usually require huge computational overhead and slow computational speed, which seriously reduces the practicability of the algorithm. The image reconstruction algorithm based on projection domain preprocessing makes full use of the high and low energy spectral information and the multi-color projection generation model to convert the nonlinear solution problem into a linear solution problem, so that the conventional CT image reconstruction algorithm can be used to achieve dual-energy CT image reconstruction. In theory, the influence of hardening artifacts can be effectively eliminated, and accurate physical parameter distribution information can be obtained. The calculation is simple and the efficiency is high. It is the mainstream reconstruction method of dual-energy CT image reconstruction technology at present. However, such methods are highly dependent on the calibration process and require the high and low energy projection datasets to be spatially geometrically consistent, i.e. each pair of high and low energy projection measurements needs to follow the same ray path.

The image reconstruction algorithm based on image domain post-processing first reconstructs high- and low-energy CT images from high- and low-energy projection data using traditional CT image reconstruction algorithms, and then decomposes high- and low-energy CT reconstructed images in the image domain to obtain the physical properties of the object tomography. Parametric distribution image. Realizing the decomposition of different materials under high and low energy images is the key to image reconstruction algorithms based on image post-processing. The dual-energy CT image reconstruction algorithm based on image domain post-processing does not require high spatial geometric consistency of high-energy and low-energy projection data, and the calculation is simple, which can suppress hardening artifacts to a certain extent. Moreover, the method can be directly applied to the existing imaging system without additional hardware equipment, thus saving cost. Therefore, this method is widely used in existing dual-energy CT imaging systems. Based on this, this patent designs a dual-energy CT image decomposition method based on convolutional neural network.

SUMMARY OF THE INVENTION

Aiming at the problem that the base material image obtained by the existing dual-energy CT image decomposition method contains a lot of noise and has a low signal-to-noise ratio, the present invention provides a dual-energy CT image decomposition method based on a convolutional neural network. The output of the convolutional neural network model and the establishment of cross-convolution can realize the reasonable shunting of different base materials in high-energy CT images and low-energy CT images, thereby effectively improving the quality of dual-energy CT image base material decomposition.

In order to achieve the above object, the present invention adopts the following technical solutions:

A dual-energy CT image decomposition method based on convolutional neural network, comprising the following steps:

Step 1: design the convolutional neural network model as the mapping function D (μ _{H, L} ; Θ) in the dual-energy decomposition model;

Step 2: Train the convolutional neural network through the convolutional neural network model and the training data set, and effectively estimate the parameter Θ of the convolutional neural network;

Step 3: Use the trained convolutional neural network and the convolutional neural network parameters Θ obtained in step 2 to efficiently decompose the base material of the dual-energy CT image.

Preferably, the convolutional neural network model is designed as a dual-input and dual-output network structure model to realize the direct input of high-energy CT images and low-energy CT images and the direct output of images of different materials.

Preferably, the dual-input, dual-output network structure model establishes a cross convolution to achieve reasonable shunting of information of different base materials in high-energy CT images and low-energy CT images.

Preferably, short links are established in the convolutional neural network model.

Preferably, the training data set includes input data and output data of a convolutional neural network model, the output data is a base material image, and the input data is a dual-energy CT image obtained according to the base material image and corresponding energy information, The dual-energy CT images include high-energy CT images and low-energy CT images; the output data is used as label data.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention decomposes the dual-energy CT image based on the dual-input and dual-output convolutional neural network model, which can effectively avoid the information crosstalk between the input and output ends of different energies and different material information.

2. The cross-convolution-based convolutional neural network model of the present invention can reasonably divide the information of different materials in high-energy CT images and low-energy CT images to reach different output ends along the cross network structure.

3. The present invention is based on the network residual design idea, and establishing short links in the convolutional neural network model can effectively improve the training efficiency of the convolutional neural network, which is beneficial to the subsequent design of deeper networks.

Description of drawings

FIG. 1 is a schematic diagram of the basic flow of a method for decomposing a dual-energy CT image based on a convolutional neural network according to the present invention.

2 is a dual-input and dual-output network structure model of a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

3 is a dual-input and dual-output cross-network structure model of a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

4 is a dual-energy CT image convolutional neural network model of a dual-energy CT image decomposition method based on a convolutional neural network of the present invention;

FIG. 5 is a digital simulation test phantom filled with bone and tissue material according to a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

FIG. 6 is an X-ray high- and low-energy spectrum information diagram of a SpekCalc software simulation of a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

FIG. 7 is a high-energy CT image generated by using the test phantom of FIG. 5 and the energy spectrum information of FIG. 6 according to a method for decomposing a dual-energy CT image based on a convolutional neural network of the present invention.

FIG. 8 is a low-energy CT image generated by using the test phantom of FIG. 5 and the energy spectrum information of FIG. 6 according to a method for decomposing a dual-energy CT image based on a convolutional neural network of the present invention.

FIG. 9 is a bone image obtained by using simulation data in a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

FIG. 10 is a tissue image obtained by a method of decomposing a dual-energy CT image based on a convolutional neural network of the present invention through simulation data.

FIG. 11 is an actual high-energy CT image of a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

FIG. 12 is an actual low-energy CT image of a dual-energy CT image decomposition method based on a convolutional neural network of the present invention.

FIG. 13 is a skeleton image obtained by a method of decomposing a dual-energy CT image based on a convolutional neural network of the present invention through actual data.

FIG. 14 is a tissue image obtained by a method of decomposing a dual-energy CT image based on a convolutional neural network of the present invention through actual data.

Detailed ways

For ease of understanding, the following explanations are made to some terms that appear in the specific embodiments of the present invention:

BP algorithm: Error back propagation algorithm.

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments:

Example 1:

As shown in Figure 1, a method for decomposing a dual-energy CT image based on a convolutional neural network of the present invention includes the following steps:

Step S101: design the convolutional neural network model of dual input and dual output as the mapping function D (μ _{H, L} ; Θ) in the dual energy decomposition model, wherein μ _{H, L} are dual energy CT images, and the convolutional neural network The network model is designed as a dual-input and dual-output network structure model, as shown in Figure 2. In the figure, A and B are the input high-energy CT images and low-energy CT images, and M1 and M2 are the output base material 1 and base material 2. "→" represents two different network structures; and a cross-convolution is established in the network structure model of double input and double output, as shown in Figure 3; short links are established in the convolutional neural network model. The specific convolutional neural network model is shown in Figure 4: Among them, the convolutional neural network model contains twelve convolutional layers. Except for the final output convolution (two in total, corresponding to two label data), after each convolution Following a Batch Normalization (BN) and a rectified linear unit (ReLU), each block in Figure 4 represents a combination of convolution, BN and ReLU, and Data_A, Data_B, Label_A, and Label_B in turn represent the input low-energy CT Image, input high-energy CT image, output base material 1 image, output base material 2 image. The entire convolutional neural network model is divided into three layers, of which the first layer (Stage 1) is the feature extraction layer, the second layer (Stage 2) is the shunt layer, and the third layer (Stage 3) is the output layer. The feature extraction layer is composed of 1×7×7×64 convolutions and outputs 64-dimensional feature images; the shunt layer is composed of 4 cross convolutions and 2 residual blocks, and the convolution dimensions are 64×3×3× 64; the output layer consists of convolutions with dimensions of 64 × 5 × 5 × 1, combining 64-dimensional features into the output image.

Step S102: train the convolutional neural network through the convolutional neural network model and the training data set, and effectively estimate the convolutional neural network parameter Θ; The convolutional neural network is trained by the convolutional neural network model, so that the parameters Θ of the convolutional neural network can be effectively estimated, and the trained convolutional neural network is obtained. In the supervised training process, the initial learning rate, step size and weight value are Set to 10e-6, 0.95, 0.0005. The training data set includes input data and output data of the convolutional neural network model, the output data is a base material image, the input data is a dual-energy CT image obtained by decomposing the base material image and corresponding energy information, and the Dual-energy CT images include high-energy CT images and low-energy CT images; the output data is used as label data. As an embodiment, dual-energy CT equipment is used to obtain high-energy and low-energy CT images of the human body, and according to the clinical experience of doctors, the high- and low-energy CT images of the human body are segmented to obtain bone images (base material 1) and tissue images under different energies. (Base material 2), take it as output data, namely label data, use SpekCalc software to obtain X-ray high and low energy spectrum information, generate projections and reconstruct high and low energy CT images according to the high and low energy spectrum information, and obtain high and low energy CT images. High and low energy CT images are used as input data to build a training dataset. As an embodiment, a total of 1300 pairs of 512×512 pixel bone images and tissue images were obtained, and 1300 pairs of 512×512 pixel high-energy and low-energy CT images were generated. Segment 1300 pairs of 512 × 512 pixel bone images and tissue images with a stride of 48 to obtain 105300 pairs of image blocks with a size of 128 × 128 pixels as the output data of the convolutional neural network model, that is, label data; High and low-energy CT images of 512 × 512 pixels were segmented with a stride of 48, and 105,300 pairs of image patches of size 128 × 128 pixels were obtained as the input data of the convolutional neural network model.

Step S103: Use the trained convolutional neural network and the convolutional neural network parameters Θ obtained in step S102 to efficiently decompose the base material of the dual-energy CT image.

As an embodiment, the high-energy CT images and low-energy CT images shown in FIGS. 7 and 8 are generated by simulation using FIG. 5 and FIG. 6 , and the trained convolutional neural network is used to process FIGS. 7 and 8 , and the obtained The output decomposition image results are shown in Figure 9 and Figure 10.

As an embodiment, for the high-energy CT images and low-energy CT images shown in FIGS. 11 and 12 , the network decomposition results obtained by the method of this embodiment are shown in FIGS. 13 and 14 , respectively.

The above are only the preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, without departing from the principles of the present invention, several improvements and modifications can be made. It should be regarded as the protection scope of the present invention.

Claims (3)

1. a dual-energy CT image decomposition method based on convolutional neural network, is characterized in that, comprises the following steps: Step 1: design the convolutional neural network model as the mapping function D in the dual-energy decomposition model (μ _{H, L} ; Θ), wherein μ _{H, L} are dual-energy CT images; The convolutional neural network model is designed as a double input , The network structure model of double output is used to realize the direct input of high-energy CT image and low-energy CT image and the direct output of different material images; Reasonable shunting of different base material information in CT images; the convolutional neural network model contains twelve convolutional layers, except for the final output convolution, each convolution is followed by a batch and linear correction unit, the entire convolutional neural network The model is divided into three layers, of which the first layer is the feature extraction layer, the second layer is the shunt layer, and the third layer is the output layer. Feature image, the shunt layer consists of 4 cross convolutions and 2 residual blocks, the convolution dimensions are 64 × 3 × 3 × 64, the output layer is composed of 64 × 5 × 5 × 1 convolution, the The 64-dimensional feature combination is the output image; Step 2: Train the convolutional neural network through the convolutional neural network model and the training data set, and effectively estimate the parameter Θ of the convolutional neural network; Step 3: Use the trained convolutional neural network and the convolutional neural network parameters Θ obtained in step 2 to efficiently decompose the base material of the dual-energy CT image. 2 . The method for decomposing a dual-energy CT image based on a convolutional neural network according to claim 1 , wherein a short link is established in the convolutional neural network model. 3 . 3. a kind of dual-energy CT image decomposition method based on convolutional neural network according to claim 1, is characterized in that, described training data set comprises the input data and output data of convolutional neural network model, and described output data is a base material image, the input data is a dual-energy CT image obtained according to the base material image and corresponding energy information, and the dual-energy CT image includes a high-energy CT image and a low-energy CT image; the output data is used as label data.

Images