A scatter correction method for cone-beam CT scatter projection ratio estimation

Abstract

The present invention provides a scatter correction method for cone-beam CT scatter projection ratio estimation. This method uses a neural network model to estimate the scatter projection ratio distribution of test objects with different materials and structures, adaptively acquires scatter field information, and optimizes the projection image by eliminating blurring artifacts caused by scattering through subtraction. The method can quickly output scatter projection ratio information through a pre-trained network, improving the efficiency of projection optimization while also addressing issues such as poor consistency in scatter suppression. The method boasts excellent reliability, stability, and versatility, significantly reducing local artifacts and image blur, and significantly improving cone-beam CT projection quality.

Description

Technical Field

[0001] The invention relates to a scatter correction method for cone-beam CT scatter projection ratio estimation, and belongs to the technical field of medical CT imaging and industrial CT non-destructive testing. Background Art

[0002] As an advanced medical imaging and industrial nondestructive testing technology, cone-beam computed tomography (CBCT) can clearly and accurately display the internal structure of the object being inspected in the form of two-dimensional or three-dimensional tomographic images without destroying the object, thereby detecting its internal defects or measuring its internal dimensions. Its intuitive imaging and high resolution have unique advantages in the nondestructive testing of complex components, including medical diagnosis and treatment, safety inspections, and product quality inspection and control.

[0003] The actual X-ray imaging process is affected by various factors. The coupling of scattering information and projection information results in low-resolution images, which is a major factor limiting imaging quality. In industrial nondestructive testing, in particular, scattering information from high-density objects severely affects and obscures image detail, resulting in low contrast in reconstructed slices and making it difficult to discern object outlines and boundaries. This poses challenges for defect identification in cone-beam CT nondestructive testing. Therefore, optimizing cone-beam CT projection quality and improving the quality of reconstructed slices are particularly important.

[0004] Scatter information is typically coupled into the projection information as a low-frequency component, which can be suppressed using both hardware and software methods. Hardware methods obtain scatter information through physical manipulation and computation, primarily involving collimators, air gaps, filters, scanning slits, and radiopaque lead strips. Software methods, based on the projection image, utilize digital image processing techniques to analyze the image and estimate the illuminated object, thereby determining the scatter distribution pattern. These methods primarily include convolution, deconvolution, and Monte Carlo simulation. Neither hardware nor software correction methods alone can achieve ideal results, leading to the emergence of hybrid correction methods that combine hardware and software. The beam stop array (BSA) method is a combined hardware and software projection correction method. This method places a scatter plate in front of the detector and assumes that the light intensity at the corresponding lead strip positions on the detector is generated by scattered photons. Therefore, the output values at the corresponding lead block positions are two-dimensionally interpolated and expanded to the entire plane to obtain a scatter intensity distribution map, thereby achieving projection scatter correction. The scatter correction method based on complementary gratings uses Gaussian filtering to fill in the missing data of the grating scattering image, then splices the partial scattering obtained by interlaced slit scanning, and uses the complementary method of two gratings to obtain the overall scattering field, thereby completing the scattering correction. This method has certain effects, but the accuracy of the splicing is difficult to grasp. The hybrid scattering model uses the scattered point interpolation method combined with the scattered point convolution model to achieve projection scattering estimation on the projection boundary area, but the projection scattering correction is not thorough, the original image information is lost, and the image resolution is reduced. The main technical shortcomings of scattering suppression in practical applications include:

[0005] (1) The scattered field is affected by many uncertain factors, and the scattered field distribution based on the model has poor accuracy;

[0006] (2) The scattering hardware correction process is complex and requires multiple scanning and projection operations, resulting in high scanning costs;

[0007] (3) Scanned objects with different structures exhibit non-uniform scattering distributions, and a single model leads to the loss of local image details.

[0008] In summary, the current scatter correction method does not fully reflect the influence of the structure and material of the detection object on the scattered field, resulting in insufficient projection correction and inconsistent blur artifact information. It cannot meet the practical application requirements of CT high-precision medical imaging and the needs of industrial intra-visual non-destructive testing. Summary of the Invention

[0009] To address practical issues such as low cone-beam CT projection image correction accuracy, large scatter field estimation deviations, and inconsistent reconstructed blur artifact information, the present invention provides a scatter correction method for cone-beam CT scatter projection ratio estimation. This method utilizes a neural network model to estimate the scatter projection ratio distribution of objects with different materials and structures, adaptively acquires scatter field information, and optimizes the projection image by eliminating blur artifacts caused by scattering through subtraction. This method can rapidly output scatter projection ratio information through a pre-trained network, improving the efficiency of projection optimization while also addressing issues such as poor scatter suppression consistency.

[0010] The technical solution adopted by the present invention to solve the technical problem includes the following steps:

[0011] Step 1: Collect several different types of scanning projections to obtain the scattered field;

[0012] Step 2: Calculate the scattered projection ratio image and use the scanned projection and scattered projection ratio images to construct a sample training set;

[0013] Step 3: Complete the training of the cone-beam CT projection-scatter projection ratio neural network model by building a neural network model and setting network parameters;

[0014] Step 4: Use the trained neural network model to estimate the scattering projection ratio of the detection object projection and obtain the scattered field distribution;

[0015] Step 5: Perform scatter correction on the actual sampled projection and complete high-quality reconstruction of the corrected projection through the reconstruction algorithm.

[0016] In the above step 1: several different types of scanning projections are obtained, including: single grating projection A, object projection B, single grating + object projection C.

[0017] In the above step 1, the specific steps of obtaining the scattered field include:

[0018] (1) Acquire a grating projection image A, perform statistics on the grating projection information, and use the two peaks appearing in the statistics as the thresholds for dividing the grating grid area and the slit area, denoted as τ1 and τ2;

[0019] (2) The grid area and slit area of the grating + object projection C are extracted according to the threshold value. The area with a value less than τ1 is the grid area, and the area with a value greater than τ2 is the slit area. The grid area image and the slit area image are interpolated and Gaussian filtered to obtain the grating scattering image of the grid area (called image C1) and the object + grating scattering image of the slit area (called image C2).

[0020] (3) Subtract image C1 from image C2 to obtain a pure projection image, called image D. Subtract image D from the scanned projection B to obtain the scattered field E.

[0021] In the above step 2: the scattered projection ratio image is calculated, and the specific steps of using the scanning projection and scattered projection ratio images to construct a sample training set include:

[0022] (1) Ratio the scattered field E obtained in step 1 to the scanned projection B, and record it as the scattered projection ratio image R, where

[0023] (2) The scanning projection B and the scattered projection ratio image R are grouped together, the sample set training block size is set to L×L, the sliding step is recorded as T, and the data set is expanded by a combination of rotation, flipping, and superposition to obtain the training sample set.

[0024] In step 4 above, the specific steps of estimating the scattered projection ratio of the detection object projection using the trained neural network model and obtaining the scattered field distribution include:

[0025] 1) The projection of the detected object is recorded as B n As the network input, use the neural network model trained in step 3 to obtain the scattering projection ratio image corresponding to the projection, which is recorded as R g As network output;

[0026] 2) The scattering projection ratio image obtained by the network output is converted into the image according to the model E g =B n ×R g Calculations are performed to obtain an estimate of the scattered field.

[0027] In the above step 5: the actual sample projection is scatter corrected by the actual projection B n and the scattered field E g Perform subtraction processing to complete the projection correction, that is, B c =B n -E g .

[0028] The beneficial effects of the present invention are as follows: the cone-beam CT projection quality optimization method based on neural network scattering projection ratio prediction provided by the present invention is applicable to the scattering field estimation of objects with arbitrary complex structures and different densities. The method has good reliability, stability and versatility, greatly reduces local artifacts and image blur, and significantly improves the cone-beam CT projection quality.

[0029] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a flow chart of the algorithm of the present invention.

[0031] Figure 2Comparison of linear grayscale at the same position in the reconstructed slices before and after scatter correction for cone-beam CT. DETAILED DESCRIPTION

[0032] A projection scan of a titanium alloy object was performed using an existing industrial cone-beam CT device (the X-ray source was Comet's MXR-451HP / 11, the flat-panel detector was PerkinElmer's XRD 1621AN15 ES, and it was equipped with a scanning mechanism, a system control, and a computer for calculation). The cone-beam CT scatter correction method for scatter projection ratio prediction using the method of the present invention was performed by performing the following steps:

[0033] Step 1: Using a multi-energy spectral X-ray source industrial cone-beam CT device, select a X-ray source voltage of 430 kV and a current of 0.15 mA. Scanning geometry parameters are: source-to-detector distance of 1202.8 mm, and source-to-rotation center distance of 952.3 mm. Reconstruction resolution is 512 × 512. Circular scanning is performed to obtain a single grating projection A at 0°, object projections B from 0° to 360°, and a single grating + object projection C. The specific steps include:

[0034] (1) Select the grating projection image A at the 0° angle position, perform statistics on the grating projection information, and use the two higher peaks in the statistics as the thresholds for dividing the grating grid area and the slit area, where τ1 = 2600 and τ2 = 39000;

[0035] (2) The grid area and slit area of the grating + object projection C at the 0° angle position are extracted according to the threshold. The area with C ≤ 2600 is the grid area, and the area with C ≥ 39000 is the slit area. The grid area image and the slit area image are interpolated and Gaussian filtered to obtain the grating scattering image of the grid area (called image C1) and the object + grating scattering image of the slit area (called image C2).

[0036] (3) Subtract image C1 from image C2 to obtain the projection image, called image D. Subtract image D from the scanning projection B at the 0° angle position to obtain the object scattering field E.

[0037] Step 2: Calculate the scattered projection ratio image. The specific steps of using the scanned projection and scattered projection ratio images to construct a sample training set include:

[0038] (1) Ratio the scattered field E obtained in step 1 to the scanned projection B, and record it as the scattered projection ratio image R, where

[0039] (2) The scanning projection B and the scattered projection ratio image R are grouped together, the sample set training block size is set to 32×32, the sliding step is 16, and the data set is expanded by a combination of rotation, flipping, and superposition to obtain approximately 550,000 training sample blocks.

[0040] Step 3: By constructing a U-net neural network framework, designing 15 convolutional layers + nonlinear activation layers, setting the network learning rate to 0.01, the training batch size to 200, and training for 100 generations, the cone-beam CT projection-scattering projection ratio model training was completed;

[0041] Step 4: Use the trained U-net neural network model to estimate the scattering projection ratio of the titanium alloy object projection. The specific steps to obtain the scattered field distribution include:

[0042] (1) The projection of the titanium alloy object is denoted as B n As the network input, use the U-net neural network model trained in step 3 to obtain the scattering projection ratio image corresponding to the projection, which is recorded as R g As network output;

[0043] (2) The scattering projection ratio image obtained by the U-net neural network output is converted into the image according to the model E g =B n ×R g Calculations are performed to obtain an estimate of the scattered field.

[0044] Step 5: Perform scatter correction on the projection of the titanium alloy object, according to the projection B of the titanium alloy object sampled in step 4 n and the scattered field E estimated by the U-net neural network g Perform subtraction processing to obtain the corrected projection B c =B n -E g And complete the high-quality reconstruction of the corrected projection through the reconstruction algorithm.

Claims

1. A scatter correction method for cone-beam CT scatter projection ratio estimation, characterized in that The steps include: Step 1: Collect several different types of scanning projections to obtain the scattered field; Step 2: Calculate the scattered projection ratio image and use the scanned projection and scattered projection ratio images to construct a sample training set; Step 3: Complete the training of the cone-beam CT projection-scatter projection ratio neural network model by building a neural network model and setting network parameters; Step 4: Use the trained neural network model to estimate the scattering projection ratio of the detection object projection and obtain the scattered field distribution; Step 5: Perform scatter correction on the actual sampled projection and complete high-quality reconstruction of the corrected projection through the reconstruction algorithm; In step 2, the scattered field E obtained in step 1 is compared with the scanning projection B, and the result is recorded as the scattered projection ratio image R, where Then, the scanning projection B and the scattered projection ratio image R are grouped together, the sample set training block size is set to L×L, the sliding step is recorded as T, and the data set is expanded by a combination of rotation, flipping, and superposition to obtain the training sample set; In step 4, the scattering projection ratio image R obtained by outputting the trained neural network model is g According to Model E g =B n ×R g Calculate the scattered field estimate, B n Indicates a scan projection.

2. The scatter correction method for cone-beam CT scatter projection ratio estimation according to claim 1, characterized in that: A neural network is used to estimate the scattering projection ratio of the detection object projection. The scattering projection ratio is defined as the ratio of the scattering field obtained by physical hardware to the corresponding projection, which is used as a means to indirectly obtain the scattering field distribution of the sampled projection.

Images