Method for reconstructing incomplete data of X-ray absorption contrast computed tomography based on deep learning

Abstract

The present invention discloses a method for reconstructing incomplete data of X-ray absorption contrast computed tomography (CT) based on deep learning (DL). The method includes the following steps: using a filtered back projection (FBP) algorithm to obtain an initial reconstructed image; forward projecting the initial reconstructed image to obtain artifact-contaminated complete projection sequences; using a DL technique to process the artifact-contaminated projection sequences to obtain artifact-free projection sequences; using the FBP algorithm to reconstruct the artifact-free projection sequences to obtain a final reconstructed image. Compared with the traditional incomplete data reconstruction methods, the examples of the present invention feature a simpler calculation process, fewer parameters to be manually set, a faster calculation speed and higher image quality.

Claims

1. A method for reconstructing incomplete data of X-ray absorption contrast computed tomography (CT) based on deep learning (DL), the method comprising: a step 1 of using a filtered back projection (FBP) algorithm to reconstruct incomplete absorption contrast projection sequences obtained by an X-ray absorption contrast CT system to obtain an initial reconstructed image, wherein the initial reconstruction image in an initial reconstruction result of FBP comprises artifacts and noise due to the incomplete projection sequences; a step 2 of using a forward projection operator to forward project the initial reconstructed image to obtain artifact-contaminated complete projection sequences, wherein the artifact-contaminated complete projection sequences are obtained by forward projecting image structure information and the artifacts into the projection sequences; a number of the sequences satisfies a complete condition, that is, a Nyquist sampling theorem; a step 3 of using a DL technique to process the artifact-contaminated complete projection sequences to obtain artifact-free complete projection sequences, wherein the complete projection sequences obtained by processing the artifact-contaminated complete projection sequences by using the DL technique do not comprise the artifacts; a number of the sequences satisfies a complete condition; and a step 4 of using the FBP algorithm to reconstruct the artifact-free complete projection sequences to obtain a final reconstructed image.

2. The method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to claim 1, wherein the incomplete projection sequences are incomplete data generated as a result of various imaging condition restrictions or special needs, and comprise sparse-view data and limited-view data.

3. The method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to claim 1, wherein in step 1, the FBP algorithm that is used to generate the initial reconstructed image is expressed by Formula (1): $\begin{array}{cc}\beta \left(r,\theta \right)=\frac{1}{2}{\int }_0^{2\pi }\frac{1}{U^2}P\left(\omega ,\varphi \right)\frac{D}{\sqrt{D^2+{\omega }^2}}*h\left(\omega \right)d\varphi & \left(1\right)\end{array}$ wherein, β(r,θ) represents a reconstruction result, (r,θ) represents polar coordinates, U represents a weight matrix of the imaging system, P(ω,ϕ) represents projection sequences, D represents a distance from a ray source to a rotation center of the imaging system, h represents inverse Fourier transform (IFT) of a filter, ω represents a position of a detection element on a detector, and ϕ represents a rotation angle of the imaging.

4. The method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to claim 1, wherein in step 2, the forward operator that is used to process the initial reconstructed image to produce a projection is expressed by Formula (2):
P(ω,ϕ)=∫.sub.−∞.sup.+∞β(r,θ)dl  (2) wherein, P(ω,ϕ) represents the artifact-contaminated complete projection sequences comprising image structure information and artifacts, the number of the sequences being the same as the complete data; β(r,θ) represents the initial reconstructed image; l represents a projection path.

5. The method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to claim 1, wherein in step 3, the DL technique used to process the artifact-contaminated complete projection sequences to obtain artifact-free complete projection sequences is expressed by Formulas (3) to (6): $\begin{array}{cc}\hat{P}\left(\omega ,\varphi \right)=F\left(\Lambda \left(f\left(P\left(\omega ,\varphi \right)\right)\right)\right)+P\left(\omega ,\varphi \right)& \left(3\right)\\ f\left(P\left(\omega ,\varphi \right)\right)=W^T.Math.P\left(\omega ,\varphi \right)+\mathrm{Bias}& \left(4\right)\\ \mathrm{Error}=\frac{1}{2m}{\left(P\left(\omega ,\varphi \right)-\hat{P}\left(\omega ,\varphi \right)\right)}^2& \left(5\right)\\ {\omega }_j^{t+1}={\omega }_j^t-\eta .Math.\frac{\partial \mathrm{Error}}{{\omega }_j^t}& \left(6\right)\end{array}$ wherein, {circumflex over (P)}(ω,ϕ) represents the artifact-free complete projection sequences; f represents an encoding network, which uses a convolutional neural network to extract features from the artifact-contaminated complete projection sequences {circumflex over (P)}(ω,ϕ); Λ represents a nonlinear mapping function; F represents a decoding network, which uses the convolutional neural network to analyze the artifact information from high-level features obtained from the encoding; Error represents a learning target of the DL technique in this step to measure a difference between an output value and a true value; W and Bias represent parameters that need to be learned in the convolutional neural network, wherein the parameters are updated by using a gradient descent algorithm by finding a partial derivative of the learning target to the parameters; η represents a learning rate; ωj.sup.t+1 represents a learned network parameter.

6. The method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to claim 1, wherein in step 3, the DL technique is used to process the artifact-contaminated complete projection sequences instead of the initial reconstruction result.

Description

BRIEF DESCRIPTION OF DRAWINGS

(1) FIG. 1 is a flowchart of a method for reconstructing incomplete data of X-ray absorption contrast computed tomography (CT) based on deep learning (DL) according to an example of the present invention.

(2) FIG. 2 is a structural diagram of an example of a DL technique of the method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to an example of the present invention.

(3) FIGS. 3A-3D show projection sequences of sparse-view data, artifact-contaminated projection sequences and artifact-free projection sequences of X-ray incomplete data and complete projection sequences according to an example of the present invention.

(4) FIGS. 4A-4C show an initial reconstruction result and a final reconstruction result of the sparse-view data of the X-ray incomplete data and a reconstructed image of the complete projection sequences according to an example of the present invention.

(5) FIGS. 5A-5D show projection sequences of limited-view data, artifact-contaminated projection sequences and artifact-free projection sequences of the X-ray incomplete data and complete projection sequences according to an example of the present invention.

(6) FIGS. 6A-6C show an initial reconstruction result and a final reconstruction result of the limited-view data of the X-ray incomplete data and a reconstructed image of the complete projection sequences according to an example of the present invention.

REFERENCE NUMERALS

(7) 20. artifact-contaminated projection sequence; 21. first-level encoding layer; 22. second-level encoding layer; 23. third-level encoding layer; 24. fourth-level encoding layer; 25. fifth-level encoding layer; 26. first-level decoding layer; 27. second-level decoding layer; 28. third-level decoding layer; 29. fourth-level decoding layer; and 210. artifact-free projection sequence.

DETAILED DESCRIPTION

(8) The present invention is described in further detail below with reference to the accompanying drawings and specific implementations.

(9) FIG. 1 is a flowchart of a method for reconstructing incomplete data of X-ray absorption contrast computed tomography (CT) based on deep learning (DL) according to an example of the present invention. The example of the present invention provides a method for reconstructing common incomplete data (sparse-view and limited-view) of X-ray absorption contrast CT based on DL. The method specifically includes the following steps:

(10) S101: use a filtered back projection (FBP) algorithm to reconstruct incomplete projection sequences obtained by a CT system to obtain an initial reconstructed image, where the initial reconstruction image in an initial reconstruction result of FBP includes artifacts and noise due to the incomplete projection sequences.

(11) S102: use a forward projection operator to forward project the initial reconstructed image to obtain artifact-contaminated complete projection sequences, where the artifact-contaminated complete projection sequences are obtained by projecting image structure information and the artifacts into the projection sequences; a number of the sequences satisfies a complete condition.

(12) S103: use a DL technique to process the artifact-contaminated complete projection sequences to obtain artifact-free complete projection sequences, where the complete projection sequences obtained by processing the artifact-contaminated complete projection sequences by using the DL technique do not include the artifacts; a number of the sequences satisfies a complete condition.

(13) FIG. 2 is a structural diagram of an example of a DL technique of the method for reconstructing incomplete data of X-ray absorption contrast CT based on DL according to an example of the present invention. As shown in FIG. 2, in the example of the present invention, a convolutional neural network is composed of 5-level encoding layers and 4-level decoding layers. The height and width of feature maps in each level of encoding layer are reduced by one time, and the number of corresponding feature maps is doubled. The height and width of feature maps in each level of decoding layer are doubled, and the number of corresponding feature maps is reduced by one time. The feature maps with the same height and width in an encoding network and a decoding network are splicedtogether to serve as the input feature map of the next-level decoding layer.

(14) S104: use the FBP algorithm to reconstruct the artifact-free complete projection sequences to obtain a final reconstructed image.

(15) Compared with the existing methods, the method for reconstructing incomplete data of X-ray absorption contrast CT in the examples of the present invention makes full use of all the information obtained by the CT system and retains the tiny details in the reconstructed image, thereby achieving higher quality of the reconstructed image.

(16) In order to prove the effects of the above examples, experiments were carried out in the following example of the present invention, including:

(17) (1) Set experimental conditions. The experiments included a sparse-view experiment and a limited-view experiment. The sparse-view incomplete data were composed of projections of 90 angles obtained by 360° circumferential scanning, and the limit-view incomplete data were composed of projections of 180 angles obtained by scanning within [0-90°].

(18) (2) Use an FBP algorithm to obtain an initial reconstruction result.

(19) (3) Use a forward projection operator to obtain artifact-contaminated projection sequences.

(20) (4) Process the artifact-contaminated complete projection sequences according to FIG. 2 and Formulas (3) to (6) to obtain artifact-free complete projection sequences.

(21) (5) Use the FBP algorithm to obtain a final reconstruction result.

(22) FIGS. 3A-3D show projection sequences of sparse-view data, artifact-contaminated projection sequences and artifact-free projection sequences of X-ray incomplete data and complete projection sequences according to an example of the present invention. FIGS. 4A-4C show an initial reconstruction result and a final reconstruction result of the sparse-view data of the X-ray incomplete data and a reconstructed result/image of the complete projection sequences according to an example of the present invention. FIGS. 3 and 4 indicate that the method for reconstructing incomplete data of X-ray absorption contrast CT based on DL can effectively process the incomplete sparse-view data.

(23) FIGS. 5A-5D show projection sequences of limited-view data, artifact-contaminated projection sequences and artifact-free projection sequences of the X-ray incomplete data and complete projection sequences according to an example of the present invention. FIGS. 6A-6C show an initial reconstruction result and a final reconstruction result of the limited-view data of the X-ray incomplete data and a reconstructed result/image of the complete projection sequences according to an example of the present invention. FIGS. 5 and 6 indicate that the method for reconstructing incomplete data of X-ray absorption contrast CT based on DL can effectively process the incomplete limited-view data.

(24) Compared with the traditional reconstruction methods of incomplete data, the examples of the present invention feature a simpler calculation process, fewer parameters to be manually set and a faster calculation speed. Compared with the traditional reconstruction methods of incomplete data based on DL, the examples of the present invention use the DL technique to process the artifact-contaminated projection sequences instead of the initial reconstructed image, so that the weighting coefficient in the reconstruction method can be used again to obtain more image details, thereby improving the image quality.

(25) It should be understood that the device and method disclosed by the examples of the present invention can be implemented in other manners. For example, the example of the DL network structure described above is only illustrative. For example, the encoding network and decoding network described only present a logical function division, and there may be other division methods in actual implementation. For example, the learning target and parameter update method in the example of the DL network structure are only intended for simple functional description, and there may be other methods in actual implementation.

(26) Finally, it should be noted that the above examples are merely intended for describing the technical solutions of the present invention, but not for limiting the present invention. Although the present invention is described in detail with reference to the above examples, persons of ordinary skill in the art should understand that they may still make modifications to the technical solutions described in the above examples or make equivalent replacements to some or all technical features thereof, without departing from the scope of the technical solutions of the examples of the present invention.