Cone beam X-ray luminescence tomography method, system, device and storage medium

Abstract

The invention discloses a cone-beam X-ray luminescence tomography method, system, device and storage medium, which relate to the technical field of optical molecular imaging, and include: obtaining two-dimensional surface measurement images at multiple angles in cone-beam XLCT; establishing a three-dimensional deep encoder-decoder network DeepCB-XLCT; the DeepCB-XLCT network includes a three-dimensional encoder and a three-dimensional decoder with multi-layer spatial convolution and deconvolution operators; inputting the two-dimensional surface measurement images at multiple angles in cone-beam XLCT into DeepCB-XLCT to obtain a three-dimensional image distribution of CB-XLCT nanophosphorus powder to be reconstructed; and obtaining a cone-beam X-ray luminescence tomography image according to the three-dimensional image distribution of the CB-XLCT nanophosphorus powder. The method improves the quality of CB-XLCT reconstruction and can effectively reduce the reconstruction inaccuracy caused by the simplified linear model used in the traditional method.

Description

Technical Field

[0001] The present invention relates to the technical field of optical molecular imaging, and in particular to a cone-beam X-ray luminescence tomography method, system, equipment and storage medium. Background Art

[0002] With the development of X-ray excitable nanoparticles, X-ray luminescence computed tomography (XLCT) has attracted increasing attention due to its high performance. In XLCT, X-ray excitable nanoparticles are used as imaging probes, which emit visible or near-infrared (NIR) light when irradiated with X-rays, which can be detected by an electron multiplying charge-coupled device (EMCCD) camera. By solving an inverse problem using appropriate X-ray and photon transport imaging models, the three-dimensional distribution of nanoparticles in the imaging object can be resolved. Due to the good penetration and collimation of X-rays, XLCT can achieve deeper imaging depths. At the same time, the use of X-ray excited nanoprobes can effectively avoid the interference of autofluorescence and background fluorescence, thereby improving the contrast and resolution of imaging. The above advantages have enabled the rapid development of XLCT, which has broad application prospects in disease diagnosis, new drug research and development.

[0003] After the first demonstration of XLCT, systems with different imaging geometries were proposed to improve spatial and temporal resolution. Compared to narrow-beam and fan-beam XLCT systems, cone-beam XLCT (CB-XLCT) significantly improves the imaging process at the expense of spatial resolution. Due to the high scattering of light propagating in biological tissues, a first-order approximation model of the radiative transfer equation (RTE) is currently used to describe the photon propagation in the CB-XLCT forward model. Considering the complex excitation process of XLCT, the deviation between the approximate linear model of RTE and the true nonlinear photon excitation and propagation cannot be avoided. To improve the reconstruction quality, the reconstructed object function is penalized using structural prior knowledge or sparse regularization, and is mainly solved using statistical iterative algorithms. However, model-based reconstruction may not accurately simulate the complex process and pathology of CB-XLCT imaging, limiting the application of CB-XLCT for in vivo imaging. Summary of the invention

[0004] Aiming at the deficiency that the model-based reconstruction in the prior art may not accurately simulate the complex process and pathology of CB-XLCT imaging, thus limiting the application of CB-XLCT in in vivo imaging, the present invention proposes a cone-beam X-ray luminescence tomography method, system, device and storage medium, and proposes an end-to-end three-dimensional deep codec network DeepCB-XLCT, which directly establishes a nonlinear mapping relationship between the internal X-ray excitable nanoparticle distribution and the boundary fluorescence signal distribution; thereby solving the problems existing in the prior art.

[0005] A cone-beam X-ray luminescence tomography method, comprising:

[0006] Acquire 2D surface measurement images at multiple angles in cone-beam XLCT;

[0007] Establishing a three-dimensional deep encoder-decoder network DeepCB-XLCT; the DeepCB-XLCT network includes a three-dimensional encoder and a three-dimensional decoder with multiple layers of spatial convolution and deconvolution operators;

[0008] The two-dimensional surface measurement images at multiple angles in cone-beam XLCT are input into DeepCB-XLCT, and low-dimensional features are extracted through a three-dimensional encoder. The low-dimensional features are input into a three-dimensional decoder through skip connections and interactively fused with the high-dimensional features in the decoder to obtain the reconstructed three-dimensional image distribution of CB-XLCT nanophosphorus powder.

[0009] According to the reconstructed three-dimensional image distribution of the CB-XLCT nanophosphorus powder, a cone-beam X-ray luminescence tomography image is obtained.

[0010] Furthermore, the 3D encoder includes a plurality of convolutional layers and pooling layers, the 3D decoder includes a plurality of non-sampling layers, and a connection layer is provided between the 3D encoder and the 3D decoder.

[0011] Furthermore, the training of the DeepCB-XLCT network specifically includes the following steps:

[0012] acquiring a plurality of projection images;

[0013] Resize multiple projected images to 128×128;

[0014] The x- and y-axis slice sizes of the 3D images output by the DeepCB-XLCT network were set to 128 × 128, and the z-axis reconstruction resolution was set to 1 mm;

[0015] The adjusted projection images are divided into a training set and a validation set, and the DeepCB-XLCT is trained using the training set.

[0016] Furthermore, the loss function of the DeepCB-XLCT network includes the mean square error MSE between the reconstruction result and the true result and the result similarity loss SSIM, as well as the loss ROI of the target area, and its loss function loss is expressed as:

[0017] loss=MSE(x)+MSE(ROI)+2*(SSIM(x)+SSIM(ROI)).

[0018] Furthermore, the optimizer used in the DeepCB-XLCT network is the Adam algorithm implemented on TensorFlow, with epoch=50 and batch size=64.

[0019] Furthermore, the quality of the reconstructed 3D image of the CB-XLCT nanophosphorus powder was quantitatively evaluated by the dice similarity coefficient DICE and the contrast-to-noise ratio CNR; wherein the DICE reflects the similarity between the real target and the reconstructed target, which is expressed as:

[0020]

[0021] Among them, ROI t and ROI r represent the regions of the true target and the reconstructed target, respectively, and || defines the number of voxels in a region;

[0022] The CNR is used for quantitative evaluation of noise and artifacts in the reconstructed image and is expressed as:

[0023]

[0024] Among them, ROI and BCK represent the target and background regions of the imaging object respectively, and w ROI 、w BCK is a weighting factor determined by the relative volume of the target and background, μ ROI , μ BCK is the average intensity value of ROI and BCK, represent the variance of ROI and BCK respectively.

[0025] Furthermore, a cone-beam X-ray luminescence tomography system comprises:

[0026] An acquisition module, used for acquiring two-dimensional surface measurement images at multiple angles in cone-beam XLCT;

[0027] A network building module for building a three-dimensional deep encoder-decoder network DeepCB-XLCT; the DeepCB-XLCT network includes a three-dimensional encoder and a three-dimensional decoder with multiple layers of spatial convolution and deconvolution operators;

[0028] A calculation module is used to input the two-dimensional surface measurement images at multiple angles in cone-beam XLCT into DeepCB-XLCT to obtain the reconstructed three-dimensional image distribution of CB-XLCT nanophosphorus powder;

[0029] The imaging module is used to obtain a cone beam X-ray luminescence tomography image according to the reconstructed three-dimensional image distribution of the CB-XLCT nano-phosphorus powder.

[0030] Furthermore, a cone-beam X-ray luminescence tomography computer device includes: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the cone-beam X-ray luminescence tomography method when executing the computer program.

[0031] Furthermore, a readable storage medium is provided, wherein the readable storage medium stores a computer program, wherein the computer program includes program instructions, and when the program instructions are executed by a processor, they are used to execute the steps of the cone-beam X-ray luminescence tomography method.

[0032] The present invention provides a cone-beam X-ray luminescence tomography method, system, device and storage medium, which have the following beneficial effects:

[0033] The present invention proposes an end-to-end three-dimensional deep encoder-decoder network DeepCB-XLCT for CB-XLCT reconstruction, which directly establishes a nonlinear mapping relationship between the internal X-ray excitable nanoparticle distribution and the boundary fluorescence signal distribution to improve the quality of CB-XLCT reconstruction; it can effectively reduce the reconstruction inaccuracy caused by the simplified linear model used in traditional methods; compared with traditional iterative methods, the proposed DeepCB-XLCT method can improve spatial resolution and reconstruction accuracy, thereby promoting the widespread application of CB-XLCT in vivo. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 Schematic diagram of DeepCB-XLCT network architecture in an embodiment of the present invention;

[0035] Figure 2 Schematic diagram of the edge-to-edge distances between two targets studied using a cylinder model in the simulation of an embodiment of the present invention, where (a) 1 mm, (b) 1.5 mm, and (c) 2 mm;

[0036] Figure 3 is a schematic diagram of a CB-XLCT system in an embodiment of the present invention;

[0037] Figure 4 Schematic diagram of a dual-target phantom experiment in an embodiment of the present invention;

[0038] Figure 5 Schematic diagram of a tomographic image reconstructed by using the DeepCB-XLCT network in an embodiment of the present invention;

[0039] Figure 6 Schematic diagram of the reconstruction results obtained by the appendix (a), MAP (b), T-FISTA (c), ADMLEM (d) and DeepCB-XLCT network (e) in the embodiments of the present invention;

[0040] Figure 7 A schematic diagram of tomographic images of targets at different distances and different noise levels reconstructed based on the DeepCB-XLCT network in an embodiment of the present invention;

[0041] Figure 8 It is a schematic diagram of a tomographic image of a target located at different distances reconstructed in an ablation experiment in an embodiment of the present invention;

[0042] Fig. 9 Schematic diagram of tomographic images of three targets reconstructed using different methods in an embodiment of the present invention;

[0043] Fig.10 It is a schematic diagram of reconstructing tomographic images of targets at different distances based on different phantom experiment algorithms in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0045] The present invention proposes an end-to-end three-dimensional deep encoder-decoder network (DeepCB-XLCT) for CB-XLCT reconstruction. The DeepCB-XLCT network consists of a three-dimensional encoder network and a three-dimensional decoder network with multi-layer spatial convolution and deconvolution operators. The input is a two-dimensional surface measurement image of CB-XLCT at 24 angles, and the output is the reconstructed three-dimensional distribution of CB-XLCT nanophosphorus powder. After the two-dimensional surface measurement images at 24 angles are input into the encoder, sparse optical features can be obtained. The extracted low-dimensional features are input into the decoder through jump connections, interactively fused with the high-dimensional features of the decoder, and finally the distribution of nanophosphorus powder is reconstructed.

[0046] The network structure details proposed by the present invention are as follows Figure 1 As shown in the figure; the basic structure of the 3D-en-decoder network consists of two parts: a 3D encoder network that captures the distribution over optical and other feature codes, and a 3D decoder network that learns the conditional distribution of y to produce a certain x. The 3D encoder consists of several convolutional layers followed by batch norm, ReLU activation function, and pooling layers. The 3D decoder has several unsampling layers followed by convolution, batch norm, and ReLU activation functions. There is a full connection layer between the 3D encoder and the 3D decoder. While training this input-to-output mapping, a low-dimensional joint manifold of data in the image and sensor domains is implicitly learned, capturing highly expressive representations that are robust to noise and other input perturbations.

[0047] To support network training, a large number of training data sets are required. In CB-XLCT, simulation is usually performed before phantom and in vivo experiments because it is easy to control and evaluate, which means that a large number of simulation training sets can be generated through the simulation program. 6000 simulation samples are used for training and 2000 simulation samples are used as validation sets to determine the optimal model.

[0048] In addition, all projection images are resized to 128 × 128 before being fed into the network. The output of this network is a 3D image. The size of the xy slices is also 128 × 128, and the reconstruction resolution of the z axis is 1 mm. In summary, the shapes of the input and output are 128 × 128 × 24 and 128 × 128 × 7, respectively.

[0049] The loss function of the DeepCB-XLCT network consists of two parts: the mean square error (MSE) between the reconstruction result and the true result and the structural similarity loss (SSIM). Figure 2 and Figure 4 The target region shown is smaller than the entire imaging area, so the loss of the target region (ROI) is also considered. Therefore, the loss function of the DeepCB-XLCT network can be calculated as:

[0050] loss=MSE(x)+MSE(ROI)+2*(SSIM(x)+SSIM(ROI)) (1)

[0051] The optimizer used in this network is the Adam algorithm implemented on tensorflow. It uses epoch=50 and batch size=64.

[0052] The performance of the trained 3D-decoder network in CB-XLCT reconstruction was evaluated through numerical simulations and phantom experiments. Four conventional methods, Adaptive FISTA, MAP, T-FISTA, and ADMLEM, were compared and also implemented to solve equation (9). Their hyperparameters and maximum number of iterations were set according to published references to ensure the convergence of the reconstruction.

[0053] Numerical simulation device: First, numerical simulation was performed using a cylindrical phantom. Figure 2 As shown, the cylinder model used in the simulation studies the edge-to-edge distance between two targets: (a), 1 mm, (b), 1.5 mm, (c) 2 mm) consisting of a large glass cylinder (3.0 cm in diameter and 2.3 cm in height) and two small tubes (4 mm in diameter and 4 mm in height) filled with Y2O3:Eu 3+ .(50mg / ml), 1% fat emulsion was added to the glass cylinder, and the absorption coefficient and reduced scattering coefficient were set to 0.02 and 10cm-1 To evaluate the reconstruction results for targets at two different distances, numerical simulations were performed for targets at distances of 2.0, 1.5, and 1.0 mm (the edge distance between the two targets).

[0054] For the numerical simulation, the imaging model was discretized into 2695 nodes and 12285 tetrahedral elements located in a three-dimensional area of ​​3.0×3.0×2.3cm. To make the results comparable to the phantom experiments, for the numerical simulation, the distance of the x-ray source from the center of rotation of the imaging system was set to 26.3cm and the EMCCD camera was positioned perpendicular to the x-ray source detector axis, with a distance of 45.0cm between the EMCCD and the center of rotation. The voltage and current of the cone-beam x-ray source were set to 50kVp and 1mA, respectively. In a 360° scan, 24 projections were obtained for simulation every 15°. After obtaining the optical luminescence at different angles, a Gaussian white noise with zero mean and a signal-to-noise ratio (SNR) of 30dB was added to all the projections to simulate the noise measurement. In order to further verify the performance of the method at different noise levels, the present invention also added Gaussian white noise of 25dB, 20dB, and 15dB to the projections, respectively.

[0055] Phantom imaging experiment setup:

[0056] To further verify the performance of the proposed method and the performance of real luminescence measurement, a phantom experiment was performed using a customized CB-XLCT system, which mainly includes a microfocused X-ray source (Oxford Instruments, UK), a rotating platform, a flat-panel X-ray detector (2923, Dexela, UK), and an electron multiplying charge-coupled device (EMCCD) camera (iXon DU-897, Andor, UK), as shown in Figure 3 Schematic diagram of the CB-XLCT system shown; the maximum voltage of the x-ray source is 80 kVp and the maximum power is 80 W. The distances from the x-ray source to the rotation center of the imaging system and to the flat panel detector are 26.3 cm and 86.3 cm, respectively. The EMCCD camera with a Nikon 50 mm f / 1.8D lens is oriented 90° toward the x-ray source-detector axis and 45.0 cm from the rotation center. To effectively reduce dark noise, the minimum cooling temperature of the EMCCD camera is -80°.

[0057] The structure of the physical phantom used in the imaging experiment is as follows Figure 4 Figure 2 shows the setup of the dual-target phantom experiment. The concentrations of the two targets were 50 mg / ml and 50 mg / ml, respectively, and the distances between the edges of the two targets were 2.3 mm, 1.7 mm, and 1.0 mm, respectively, based on the observations from the simulation study. A glass cylinder (4.0 cm diameter, 4.0 cm height) containing a mixture of water and lipids was fixed on a rotating stage. Two small glass tubes (3 mm diameter) were filled with Y2O3:Eu 3+(50 mg / ml) were placed symmetrically in the cylinder to simulate two targets. The edge-to-edge distances (EEDs) between the two tubes were 2.3 mm, 1.7 mm, and 1.0 mm, respectively.

[0058] In the imaging experiments, the voltage and current of the X-ray source were set to 40 kVp and 1 mA, respectively. The phantom was rotated from 0° to 360°, and the EMCCD camera acquired optical images every 15°. The exposure time of the EMCCD camera was set to 0.5 s, and the EM gain and binning time were set to 260 and 1 × 1. For CT imaging, the angular increment was 1° on a 360° circular orbit, and the acquisition time for each projection was 150 ms. The proposed (FDK) algorithm was used for CT reconstruction.

[0059] Quantitative evaluation: The quality of the reconstructed CB-XLCT images was quantitatively evaluated using indicators such as the Dice Similarity Coefficient (DICE) and the Contrast to Noise Ratio (CNR). DICE reflects the similarity between the real target and the reconstructed target and can be calculated using the following method:

[0060]

[0061] ROI t and ROI r represent the regions of the true target and the reconstructed target, respectively, and || defines the number of voxels in a region.

[0062] CNR is used to quantitatively evaluate the noise and artifacts in the reconstructed image, as shown in formula (3):

[0063]

[0064] Among them, ROI and BCK represent the target and background regions of the imaging object, w ROI 、w BCK is a weighting factor determined by the relative volume of the target and background, μ ROI , μ BCK is the average intensity value of ROI and BCK, represent the variance of ROI and BCK respectively.

[0065] Numerical simulation:

[0066] (1) Resolution experiment: First, different algorithms are used to reconstruct XLCT tomographic images of targets located at different distances, such as Figure 5As shown in the figure, tomographic images of targets positioned at different distances are reconstructed, and the first to fifth columns are the results obtained using the Attachment, MAP, T-FISTA, ADMLEM and the proposed DeepCB-XLCT network, respectively. The first to third rows are the reconstruction results with edge-to-edge distances of 2.0mm, 1.5mm and 1.0mm, respectively. All reconstruction results are normalized according to their maximum values. For algorithms such as Attachment, MAP and T-FISTA, the distributions of the two targets are difficult to distinguish. Based on the ADMLEM method, the distributions of the two targets can be distinguished, and the position error is relatively large. In contrast, compared with the other four algorithms, the reconstruction algorithm using the DeepCB-XLCT network has better shape similarity with the ground truth and less noise, as shown in Figure 5 Table 1 summarizes the quantitative evaluation of the reconstruction results using different methods. It can be seen from the proposed DeepCB-XLCT network that among the five methods, the reconstruction results also achieved the highest reconstruction quality in terms of CNR and shape similarity, further confirming that Figure 5 Observation results in . Figure 6 For along Figure 5 The contour drawn by the green dashed line in the middle, with an edge-to-edge distance of 1.0 mm, (a)-(e) are the contours of the reconstruction results obtained using the attachment, MAP, T-FISTA, ADMLEM, and the proposed DeepCB-XLCT network, respectively. The results again show that the values ​​of the images reconstructed by the proposed DeepCB-XLCT network are closer to the ground truth images.

[0067] Table 1 CNR and DICE of the reconstruction results of two targets under different ed in numerical simulation

[0068]

[0069] (2) Robustness experiment and ablation experiment at different noise levels: In order to further verify the performance of the method at different noise levels, the present invention reconstructed the XLCT tomographic images of dual-target simulation with signal-to-noise ratios of 25dB, 20dB, 10dB, and 15dB, such as Figure 6 The quantitative evaluation of the reconstruction results is summarized in Table 2. It can be seen from the proposed DeepCB-XLCT network that even if the signal-to-noise ratio drops to 15dB, the reconstruction results still achieve acceptable reconstruction quality in terms of CNR and shape similarity, which further verifies the improvement of the proposed algorithm.

[0070] like Figure 7The figure shows the reconstructed tomographic images of targets at different distances and noise levels based on the proposed DeepCB-XLCT network. The signal-to-noise ratios of the first to fourth columns are 30dB, 25dB, 20dB, and 15dB, respectively. The first to third rows are the reconstruction results with edge distances of 2.0mm, 1.5mm, and 1.0mm, respectively.

[0071] Table 2 CNR and DICE of reconstruction results of two targets with different ed at different noise levels

[0072]

[0073] In order to study the impact of the skip connection module and ROILoss on the reconstruction results, the present invention conducted ablation experiments based on targets located at different distances, such as Figure 8 As shown, in the ablation experiment, the tomographic images of targets located at different distances are reconstructed. The first to fourth rows are the results obtained based on the proposed DeepCB-XLCT network without skip connections and ROI loss network, the proposed DeepCB-XLCT network without skip connections, the proposed DeepCB-XLCT network without ROI loss, and the proposed DeepCB-XLCT network, respectively. The first to third columns are the reconstruction results with edge-to-edge distances of 2.0 mm, 1.5 mm, and 1.0 mm, respectively.

[0074] Table 3 summarizes the quantitative evaluation of the reconstruction results in the ablation experiments. Without the skip connection module and ROILoss, the quality of the reconstruction results is significantly degraded. Compared with the skip connection module, ROILoss has a greater impact on the reconstruction results.

[0075] Table 3 CNR and DICE of the reconstruction results of two different ED target ablation experiments

[0076]

[0077] (3) Subsequently, the XLCT tomographic images of the three targets were reconstructed using different algorithms to verify the performance of the method in multiple targets, e.g. Fig. 9 As shown in the figure, the tomographic images of three targets and the true positions of the targets are reconstructed using different methods. The first column is the true position of the target, and the second to sixth columns are the results obtained using the attachment, MAP, T-FISTA, ADMLEM and the proposed DeepCB-XLCT network. The first to third rows are the reconstruction results of three targets with different positions. Except for the proposed DeepCB-XLCT network, no algorithm can distinguish them, which further verifies the improvement of the algorithm on multi-target XLCT tomographic images.

[0078] Phantom experiment: In order to compare the performance of different algorithms, the present invention reconstructs XLCT tomographic images of phantom experimental targets at different distances based on different algorithms, such as Fig.10 As shown in the figure, based on different phantom experimental algorithms, the tomographic images of the target at different distances are reconstructed. The first to third rows show the EEDs of 2.3 mm, 1.7 mm, and 1.0 mm, respectively. The reconstructions of the AD algorithm and the MAP algorithm are from the first to the fifth column, respectively. For the AD algorithm and the MAP algorithm, it is difficult to distinguish the distributions of the two targets. Compared with the attached method and MAP, based on the T-FISTA and ADMLEM algorithms, the distributions of the two targets can be effectively distinguished when the EED is 2.3 mm and 1.7 mm, but the task becomes difficult to distinguish the distributions of the two targets when the EED is 1.0 mm. In contrast, when the EED is 1.0 mm, no algorithm can distinguish them except the proposed DeepCB-XLCT network, as shown in Fig.10 shown.

[0079] The quantitative evaluation of the reconstruction using different methods in the phantom experiment is summarized in Table 4. The results show that even in the phantom experiment, the proposed DeepCB-XLCT network still has better performance in shape recovery and image contrast compared with the traditional reconstruction methods, which further confirms Fig.10 Observation results in .

[0080] Table 4 CNR and DICE of the reconstruction results of two targets with different eed in the phantom experiment

[0081]

[0082] The present invention proposes a DeepCB-XLCT method to improve the quality of CB-XLCT reconstruction. It directly establishes a nonlinear mapping relationship between the distribution of internal X-ray excitable nanoparticles and the distribution of boundary fluorescence signals, effectively reducing the reconstruction inaccuracy caused by simplified linear models and iterative calculations. Numerical simulation and phantom experimental results show that compared with traditional iterative methods, the proposed DeepCB-XLCT method can improve spatial resolution and reconstruction accuracy, thereby promoting the widespread application of CB-XLCT in vivo.

[0083] Based on the same inventive concept, the present invention also proposes a cone-beam X-ray luminescence tomography system, comprising:

[0084] The acquisition module is used to acquire two-dimensional surface measurement images at 24 angles in cone-beam XLCT.

[0085] A network building module is used to establish a three-dimensional deep encoder-decoder network DeepCB-XLCT; the DeepCB-XLCT network includes a three-dimensional encoder and a three-dimensional decoder with multiple layers of spatial convolution and deconvolution operators.

[0086] The calculation module is used to input the two-dimensional surface measurement images of 24 angles in cone-beam XLCT into DeepCB-XLCT to obtain the reconstructed three-dimensional image distribution of CB-XLCT nanophosphorus powder.

[0087] The imaging module is used to obtain a cone beam X-ray luminescence tomography image according to the reconstructed three-dimensional image distribution of the CB-XLCT nano-phosphorus powder.

[0088] Based on the same inventive concept, the present invention also proposes a cone-beam X-ray luminescence tomography computer device, comprising: a memory, a processor, and a computer program stored in the memory, and the processor implements the steps of the cone-beam X-ray luminescence tomography method when executing the computer program.

[0089] Based on the same inventive concept, the present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the cone-beam X-ray luminescence tomography method.

[0090] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A cone-beam X-ray luminescence tomography method, characterized in that: include: Acquire 2D surface measurement images at multiple angles in cone-beam XLCT; Establish a three-dimensional deep encoder-decoder DeepCB-XLCT network; The DeepCB-XLCT network includes a three-dimensional encoder and a three-dimensional decoder with multiple layers of spatial convolution and deconvolution operators; the three-dimensional encoder includes multiple convolution layers and pooling layers, the three-dimensional decoder includes multiple non-sampling layers, and a connection layer is provided between the three-dimensional encoder and the three-dimensional decoder; The two-dimensional surface measurement images at multiple angles in cone-beam XLCT are input into the DeepCB-XLCT network, low-dimensional features are extracted through the three-dimensional encoder, and the low-dimensional features are input into the three-dimensional decoder through jump connections to interact and fuse with the high-dimensional features in the decoder to obtain the reconstructed three-dimensional image distribution of CB-XLCT nanophosphorus powder; the loss function of the DeepCB-XLCT network includes the mean square error MSE and the result similarity loss SSIM between the reconstruction result and the true result, as well as the loss of the target area ROI, and its loss function loss is expressed as: loss=MSE(x)+MSE(ROI)+2*(SSIM(x)+SSIM(ROI)); According to the reconstructed 3D image distribution of the nanophosphorus powders from CB-XLCT, cone beam X-ray luminescence tomography images were obtained.

2. A cone-beam X-ray luminescence tomography method according to claim 1, characterized in that: The training of the DeepCB-XLCT network specifically includes the following steps: acquiring a plurality of projection images; Resize multiple projected images to 128×128; The x- and y-axis slice sizes of the 3D images output by the DeepCB-XLCT network were set to 128 × 128, and the z-axis reconstruction resolution was set to 1 mm; The adjusted projection images are divided into a training set and a validation set, and the DeepCB-XLCT network is trained using the training set.

3. The cone-beam X-ray luminescence tomography method according to claim 1, characterized in that: The optimizer used in the DeepCB-XLCT network is the Adam algorithm implemented on TensorFlow, with epoch=50 and batch size=64.

4. The cone-beam X-ray luminescence tomography method according to claim 1, characterized in that: The quality of the reconstructed 3D image of CB-XLCT nanophosphorus powder is quantitatively evaluated by the dice similarity coefficient DICE and the contrast-to-noise ratio CNR; wherein the DICE reflects the similarity between the real target and the reconstructed target, which is expressed as: Among them, ROI t and ROI r Represent the regions of the real target and the reconstructed target respectively, || defines the number of voxels in a region; The CNR is used for quantitative evaluation of noise and artifacts in the reconstructed image and is expressed as: Among them, ROI and BCK represent the target and background regions of the imaging object respectively, and w ROI 、w BCK is a weighting factor determined by the relative volume of the target and background, μ ROI , μ BCK is the average intensity value of ROI and BCK, represent the variance of ROI and BCK respectively.

5. A cone-beam X-ray luminescence tomography system, characterized in that: include: An acquisition module, used for acquiring two-dimensional surface measurement images at multiple angles in cone-beam XLCT; A network construction module, used for establishing a three-dimensional deep encoder-decoder DeepCB-XLCT network; the DeepCB-XLCT network includes a three-dimensional encoder and a three-dimensional decoder with multiple layers of spatial convolution and deconvolution operators; the three-dimensional encoder includes multiple convolution layers and pooling layers, the three-dimensional decoder includes multiple non-sampling layers, and a connection layer is provided between the three-dimensional encoder and the three-dimensional decoder; The calculation module is used to input the two-dimensional surface measurement images at multiple angles in the cone-beam XLCT into the DeepCB-XLCT network, extract low-dimensional features through the three-dimensional encoder, input the low-dimensional features into the three-dimensional decoder through jump connections, and interactively fuse with the high-dimensional features in the decoder to obtain the reconstructed three-dimensional image distribution of the CB-XLCT nanophosphorus powder; the loss function of the DeepCB-XLCT network includes the mean square error MSE and the result similarity loss SSIM between the reconstruction result and the true result, as well as the loss of the target area ROI, and its loss function loss is expressed as: loss=MSE(x)+MSE(ROI)+2*(SSIM(x)+SSIM(ROI)); The imaging module is used to obtain a cone beam X-ray luminescence tomography image according to the reconstructed three-dimensional image distribution of the CB-XLCT nano-phosphorus powder.

6. A cone beam X-ray luminescence tomography computer device, characterized in that: include: A memory, a processor and a computer program stored in the memory, wherein the processor implements the steps of the cone-beam X-ray luminescence tomography method according to any one of claims 1 to 4 when executing the computer program.

7. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the cone-beam X-ray luminescence tomography method according to any one of claims 1 to 4.

Images