A method for fusing and reconstructing anisotropic 3D images from multiple perspectives

Abstract

This invention relates to a method for fusing and reconstructing multi-view anisotropic three-dimensional images, comprising: acquiring three orthogonal MRI images of a small mammalian organ; aligning the spatial proportions of the three orthogonal MRI images to obtain images I1, I2, and I3; extracting feature points on a reference image R1; calculating the affine transformation matrix of the reference image and the floating image using the correspondence between two sets of feature point pairs corresponding to the obtained reference image and the floating image; obtaining two generated images R2 and R3; obtaining three images R1', R2', and R3' after brightness adjustment; and obtaining a three-dimensional isotropic image, i.e., the final output image I. o This invention precisely stitches and aligns three images from different perspectives. By fusing and reconstructing multiple orthogonal low-resolution interlayer images, the clarity of the originally low-resolution images is greatly improved, providing high-definition image data support for the study of organ function in small mammals.

Description

Technical Field

[0001] This invention relates to the field of 3D biomedical image processing technology, and in particular to a method for fusing and reconstructing multi-view anisotropic three-dimensional images. Background Technology

[0002] Magnetic resonance imaging (MRI) is a non-invasive imaging technique that produces detailed three-dimensional anatomical images, commonly used for disease detection, diagnosis, and treatment monitoring. Due to limitations in imaging technology, current MRI imaging is primarily used for imaging human organs. Typical medical equipment can achieve slice resolution at the millimeter level; improving slice resolution requires a very strong imaging magnetic field.

[0003] Currently, MRI imaging of small mammals remains challenging. For example, the brain of a tree shrew is less than two centimeters in size, requiring very high-resolution imaging equipment to acquire images suitable for analysis. This results in demanding imaging conditions and a limited number of image samples. Current methods for improving image resolution through slices often require pre-defined conditions and extensive training with large amounts of high-quality data, which contradicts the reality of the difficulty in obtaining high-quality image data in practice. Summary of the Invention

[0004] To address the limitation of existing imaging equipment in obtaining high-resolution inter-slice images for MRI imaging of small organs in small animals, the present invention aims to provide a method for fusing and reconstructing multi-view anisotropic three-dimensional images by acquiring approximately orthogonal low-resolution inter-slice images through three scans from different perspectives, followed by a series of processing steps to fuse and reconstruct isotropic high-resolution isotropic three-dimensional images. This method provides high-definition image data support for organ function research in small mammals, and is highly operable and has good fusion and reconstruction results.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for fusing and reconstructing multi-view anisotropic three-dimensional images, the method comprising the following sequential steps:

[0006] (1) Obtain three orthogonal MRI images of organs from small mammals;

[0007] (2) Convert the three orthogonal MRI images from nii format to v3draw format. By calculating the ratio of intra-slice resolution to inter-slice resolution, use linear interpolation to interpolate the inter-slice pixels so that the spatial ratio of the three orthogonal MRI images is consistent, and obtain images I1, I2 and I3.

[0008] (3) After linear interpolation, select the image with the best quality among the three images I1, I2 and I3 as the reference image R1, and the other two images I2 and I3 as floating images. Select a small number of feature points representing the image direction on the three images R1, I2 and I3 respectively, and use the 2.5D Harris corner detection algorithm to extract the feature points on the reference image R1.

[0009] (4) Use the CLM coherent marker mapping matching algorithm to search for feature points on the two floating images I2 and I3 that match the feature points on the reference image R1, and use the correspondence between the two sets of feature point pairs corresponding to the reference image and the floating image to calculate the affine transformation matrix of the reference image and the floating image.

[0010] (5) Use the affine transformation matrix to perform an affine transformation on the two floating images, so that the two floating images I2 and I3 are completely aligned with the reference image R1, and two generated images R2 and R3 are obtained.

[0011] (6) Using the brightness distribution of the reference image R1 as a reference, adjust the brightness of the two generated images R2 and R3 to make the brightness distribution of the three consistent, and obtain the three images R1', R2' and R3' after brightness adjustment;

[0012] (7) For the three images R1', R2' and R3', the three images R1', R2' and R3' are fused and reconstructed according to the 3D gradient enhancement fusion method to obtain a 3D isotropic image, which is the final output image I. o .

[0013] Step (2) specifically includes the following steps:

[0014] (2a) Convert an anisotropic 3D nii format image to v3draw format;

[0015] (2b) Calculate the upsampling factor between image slices. Based on the ratio of intra-slice resolution to inter-slice resolution, obtain the intra-slice resolution of the image as n*n micrometers and the inter-slice resolution as m micrometers, where m>n. Calculate the inter-slice upsampling factor as α=m / n.

[0016] (2c) Calculate the image size after upsampling based on the upsampling factor: If the initial anisotropic image size is X*Y*Z, then the output image size after interpolation is X*Y*(Z*α);

[0017] (2d) Perform interlayer upsampling on the three-dimensional anisotropic image according to the output image size. The interpolation method is bilinear interpolation. Perform the same interpolation method on the three images to obtain three images I1, I2 and I3 with the same spatial ratio.

[0018] Step (3) specifically includes the following steps:

[0019] (3a) Compare the three images after linear interpolation, select the image with the best quality I1 as the reference image and name it R1, and the two floating images are I2 and I3; the best quality means the largest field of view and uniform brightness distribution;

[0020] (3b) Select 8 feature points on each of the three images R1, I2 and I3 that can represent the approximate directional distribution of the three images;

[0021] (3c) Using the 2.5D Harris corner detection algorithm, non-maximum suppression windows are set respectively. The radii of the three-dimensional non-maximum suppression window and the two-dimensional non-maximum suppression window are set to 25 and 20 respectively. 700 feature points are extracted on the reference image R1.

[0022] (3d) Filter the 700 generated feature points and finally retain 500 feature points.

[0023] Step (4) specifically includes the following steps:

[0024] (4a) Using the selected feature points, calculate the coarse-precision three-dimensional affine transformation matrix from floating image I2 to reference image R1, and from floating image I3 to reference image R1. Use the coarse-precision three-dimensional affine transformation matrix to perform a preliminary affine transformation on the two floating images I2 and I3 to reference image R1 to obtain floating image I2' and floating image I3'. The basic directions of the three images R1, I2' and I3' are consistent.

[0025] (4b) Input the reference image R1 and extract the feature points on the reference image R1 using the 2.5D Harris corner detection algorithm, as well as the floating images I2' and I3' after orientation adjustment. Use the CLM coherent marker mapping matching algorithm, set the search radius to 10, and find the feature points on the floating images I2' and floating I3' that match the feature points on the reference image R1 through iterative search.

[0026] (4c) Use the inverse matrix of the coarse-precision three-dimensional affine transformation matrix to inversely transform the feature points on floating images I2' and I3 to floating images I2 and I3, and obtain the feature points on floating images I2 and I3 that accurately match the reference image R1. Using these two sets of accurately matched feature point pairs, calculate the affine transformation matrix of floating image I2 to reference image R1 and floating image I3 to reference image R1.

[0027] Step (5) specifically includes the following steps:

[0028] (5a) Generate two empty images with the same size as the reference image R1 and a pixel value of 0. Obtain the pixel value of each pixel in the two empty images and calculate the position of each pixel in the two empty images in the original floating images I2 and I3 according to the affine transformation matrix.

[0029] (5b) Linear interpolate the pixel values ​​of the generated image using the pixel values ​​of the eight neighboring locations around this position in the original image; if the position of the target image pixel point in the original image exceeds the boundary, the pixel value of the part exceeding the boundary is taken as the pixel value of the reference image R1 at that position, the generated image R2 is obtained according to the floating image I2, and the generated image R3 is obtained according to the floating image I3.

[0030] Step (6) specifically includes the following steps:

[0031] (6a) Select a sliding one-dimensional window with a radius of 5, and perform mean smoothing filtering on the X, Y and Z views of the three images R1, R2 and R3 respectively to obtain three smoothed images S1, S2 and S3.

[0032] (6b) Traverse the three images R1, R2, and R3 pixel by pixel, adjust the brightness values ​​of the generated images R2 and R3, and after brightness matching, the pixel values ​​at the corresponding positions are R2(x, y, z) = R2(x, y, z) * (S1(x, y, z) / S2(x, y, z)) and R3(x, y, z) = R3(x, y, z) * (S1(x, y, z) / S3(x, y, z)), thus obtaining the three images R1', R2', and R3' with consistent brightness after brightness adjustment.

[0033] Step (7) specifically includes the following steps:

[0034] (7a) Generate an empty image with the same size as the three images R1', R2' and R3' and a pixel value of 0;

[0035] (7b) For the completely empty image in step (7a), iterate through each pixel position to obtain the final output image I. o For each pixel position (x,y,z) in the image, extract the corresponding pixel values ​​R'(x,y,z), R'(x+1,y,z), R'(x-1,y,z), R'(x,y+1,z), R'(x,y-1,z), R'(x,y,z+1), and R'(x,y,z-1) from the three images R1', R2', and R3' respectively.

[0036] (7c) The three images are fused using a three-dimensional gradient enhancement fusion method, as follows: For the final output image I o Assigning pixel values ​​at each pixel location in the data, and taking... Where a = 4b, and b ranges from 0 to 1, adjusted according to the fusion effect. In the above formula, if I o (x,y,z)<0 or I o If (x,y,z)>255, then take By iterating through the images one by one, we can obtain the final output image I. o .

[0037] As can be seen from the above technical solution, the beneficial effects of the present invention are as follows: First, by accurately stitching and aligning three images from different perspectives, the images with different field of view ranges obtained from multiple perspectives can be complementary, and the effective information contained in the three images can be integrated; Second, by fusing and reconstructing multiple orthogonal low-resolution interlayer images, the clarity of the originally low-resolution images is greatly improved, providing high-definition image data support for the study of organ function in small mammals; Third, the present invention proposes a new image acquisition method that can perform fusion and reconstruction based on a small number of images from different perspectives, which is highly operable and has good fusion and reconstruction effect. Attached Figure Description

[0038] Figure 1 This is a flowchart of the method of the present invention;

[0039] Figure 2 These are three low-resolution anisotropic images acquired between layers;

[0040] Figure 3 This is the result of interlayer interpolation on a low-resolution anisotropic image;

[0041] Figure 4 An illustration showing the effect of accurately registering two floating images to a reference image;

[0042] Figure 5 This is a diagram showing the effect of brightness matching for three images;

[0043] Figure 6 This is the result of fusing and reconstructing three images. Detailed Implementation

[0044] like Figure 1 As shown, a method for fusing and reconstructing multi-view anisotropic 3D images includes the following sequential steps:

[0045] (1) Obtain three orthogonal MRI images of organs of small mammals; due to limitations in imaging technology, the inter-slice resolution of the obtained images is lower than the intra-slice resolution, and it is difficult to achieve a strictly orthogonal effect when operating the imaging.

[0046] (2) Convert the three orthogonal MRI images from nii format to v3draw format. By calculating the ratio of intra-slice resolution to inter-slice resolution, use linear interpolation to interpolate the inter-slice pixels so that the spatial ratio of the three orthogonal MRI images is consistent, and obtain images I1, I2 and I3.

[0047] (3) After linear interpolation, select the image with the best quality among the three images I1, I2 and I3 as the reference image R1, and the other two images I2 and I3 as floating images. Select a small number of feature points representing the image direction on the three images R1, I2 and I3 respectively, and use the 2.5D Harris corner detection algorithm to extract the feature points on the reference image R1.

[0048] (4) The CLM coherent marker mapping matching algorithm is used to search for feature points on two floating images I2 and I3 that match the feature points on the reference image R1. The affine transformation matrix of the reference image and the floating image is calculated using the correspondence between the two sets of feature point pairs corresponding to the reference image and the floating image. The CLM coherent marker mapping matching algorithm comes from the paper: Cross-Modality Coherent Registration of Whole Mouse Brains.

[0049] (5) Use the affine transformation matrix to perform an affine transformation on the two floating images, so that the two floating images I2 and I3 are completely aligned with the reference image R1, and two generated images R2 and R3 are obtained.

[0050] (6) Using the brightness distribution of the reference image R1 as a reference, adjust the brightness of the two generated images R2 and R3 to make the brightness distribution of the three consistent, and obtain the three images R1', R2' and R3' after brightness adjustment;

[0051] (7) For the three images R1', R2' and R3', the three images R1', R2' and R3' are fused and reconstructed according to the 3D gradient enhancement fusion method to obtain a 3D isotropic image, which is the final output image I. o .

[0052] Step (2) specifically includes the following steps:

[0053] (2a) Convert an anisotropic 3D nii format image to v3draw format;

[0054] (2b) Calculate the upsampling factor between image slices. Based on the ratio of intra-slice resolution to inter-slice resolution, obtain the intra-slice resolution of the image as n*n micrometers and the inter-slice resolution as m micrometers, where m>n. Calculate the inter-slice upsampling factor as α=m / n.

[0055] (2c) Calculate the image size after upsampling based on the upsampling factor: If the initial anisotropic image size is X*Y*Z, then the output image size after interpolation is X*Y*(Z*α);

[0056] (2d) Perform interlayer upsampling on the three-dimensional anisotropic image according to the output image size. The interpolation method is bilinear interpolation. Perform the same interpolation method on the three images to obtain three images I1, I2 and I3 with the same spatial ratio.

[0057] Step (3) specifically includes the following steps:

[0058] (3a) Compare the three images after linear interpolation, select the image with the best quality I1 as the reference image and name it R1, and the two floating images are I2 and I3; the best quality means the largest field of view and uniform brightness distribution;

[0059] (3b) Select 8 feature points on each of the three images R1, I2 and I3 that can represent the approximate directional distribution of the three images;

[0060] (3c) Using the 2.5D Harris corner detection algorithm, non-maximum suppression windows are set respectively. The radii of the three-dimensional non-maximum suppression window and the two-dimensional non-maximum suppression window are set to 25 and 20 respectively. 700 feature points are extracted on the reference image R1.

[0061] (3d) Filter the 700 generated feature points and finally retain 500 feature points.

[0062] Step (4) specifically includes the following steps:

[0063] (4a) Using the selected feature points, calculate the coarse-precision three-dimensional affine transformation matrix from floating image I2 to reference image R1, and from floating image I3 to reference image R1. Use the coarse-precision three-dimensional affine transformation matrix to perform a preliminary affine transformation on the two floating images I2 and I3 to reference image R1 to obtain floating image I2' and floating image I3'. The basic directions of the three images R1, I2' and I3' are consistent.

[0064] (4b) Input the reference image R1 and extract the feature points on the reference image R1 using the 2.5D Harris corner detection algorithm, as well as the floating images I2' and I3' after orientation adjustment. Use the CLM coherent marker mapping matching algorithm, set the search radius to 10, and find the feature points on the floating images I2' and floating I3' that match the feature points on the reference image R1 through iterative search.

[0065] (4c) Use the inverse matrix of the coarse-precision three-dimensional affine transformation matrix to inversely transform the feature points on floating images I2' and I3 to floating images I2 and I3, and obtain the feature points on floating images I2 and I3 that accurately match the reference image R1. Using these two sets of accurately matched feature point pairs, calculate the affine transformation matrix of floating image I2 to reference image R1 and floating image I3 to reference image R1.

[0066] Step (5) specifically includes the following steps:

[0067] (5a) Generate two empty images with the same size as the reference image R1 and a pixel value of 0. Obtain the pixel value of each pixel in the two empty images and calculate the position of each pixel in the two empty images in the original floating images I2 and I3 according to the affine transformation matrix.

[0068] (5b) Linear interpolate the pixel values ​​of the generated image using the pixel values ​​of the eight neighboring locations around this position in the original image; if the position of the target image pixel point in the original image exceeds the boundary, the pixel value of the part exceeding the boundary is taken as the pixel value of the reference image R1 at that position, the generated image R2 is obtained according to the floating image I2, and the generated image R3 is obtained according to the floating image I3.

[0069] Step (6) specifically includes the following steps:

[0070] (6a) Select a sliding one-dimensional window with a radius of 5, and perform mean smoothing filtering on the X, Y and Z views of the three images R1, R2 and R3 respectively to obtain three smoothed images S1, S2 and S3.

[0071] (6b) Traverse the three images R1, R2, and R3 pixel by pixel, adjust the brightness values ​​of the generated images R2 and R3, and after brightness matching, the pixel values ​​at the corresponding positions are R2(x, y, z) = R2(x, y, z) * (S1(x, y, z) / S2(x, y, z)) and R3(x, y, z) = R3(x, y, z) * (S1(x, y, z) / S3(x, y, z)), thus obtaining the three images R1', R2', and R3' with consistent brightness after brightness adjustment.

[0072] Step (7) specifically includes the following steps:

[0073] (7a) Generate an empty image with the same size as the three images R1', R2' and R3' and a pixel value of 0;

[0074] (7b) For the completely empty image in step (7a), iterate through each pixel position to obtain the final output image I. oFor each pixel position (x,y,z) in the image, extract the corresponding pixel values ​​R'(x,y,z), R'(x+1,y,z), R'(x-1,y,z), R'(x,y+1,z), R'(x,y-1,z), R'(x,y,z+1), and R'(x,y,z-1) from the three images R1', R2', and R3' respectively.

[0075] (7c) The three images are fused using a three-dimensional gradient enhancement fusion method, as follows: For the final output image I o Assigning pixel values ​​at each pixel location in the data, and taking... Where a = 4b, and b ranges from 0 to 1, adjusted according to the fusion effect. In the above formula, if I o (x,y,z)<0 or I o If (x,y,z)>255, then take By iterating through the images one by one, we can obtain the final output image I. o .

[0076] like Figure 2 As shown, the images are three approximately orthogonal MRI images of the tree shrew brain acquired under a high-intensity magnetic field, with a resolution down to the micrometer level. It can be seen from the images that the resolution of each image within the slice is much higher than the resolution between image slices, and the three-dimensional view shows that the images are significantly inconsistent in spatial proportion.

[0077] like Figure 3 As shown, the three images were upsampled by calculation and inter-layer pixels were filled by bilinear interpolation. The three images obtained were of the best quality in the view with the higher original resolution, while the images obtained by inter-layer interpolation were of lower quality in the other two views. The middle image had the largest imaging coverage area and the best imaging quality, and was used as the reference image. The other two images were used as floating images.

[0078] like Figure 4 As shown, in order to achieve subsequent image fusion, the two floating images are precisely registered to the reference image. After the transformation, the feature positions of the three images are completely aligned; however, the brightness distribution of the three images is not consistent at this time.

[0079] like Figure 5 As shown, to ensure that the brightness distribution of the three images is consistent, the brightness distribution of the reference image is used as a benchmark, and the brightness of the other two images is matched. After brightness adjustment, the brightness distribution of the three images is consistent.

[0080] like Figure 6As shown, for three perfectly aligned and brightness-adjusted images, only one view still has superior quality; the other two views remain of lower quality due to interpolation. After fusion using a 3D gradient enhancement method, the result is... Figure 6 The results shown demonstrate that the image achieves optimal performance across all three views.

[0081] In summary, this invention precisely stitches and aligns three images from different perspectives, complementing images with different field of view acquired from multiple perspectives and integrating the effective information contained in the three images. By fusing and reconstructing multiple orthogonal low-resolution interlayer images, the clarity of the originally low-resolution images is greatly improved, providing high-definition image data support for the study of organ function in small mammals.

Claims

1. A method for fusing reconstruction of multi-view anisotropic three-dimensional images, characterized in that: The method comprises the following steps in sequence: (1) obtaining three orthogonal MRI images of a small mammal organ; (2) converting the three orthogonal MRI images from an nii format into a v3draw format, calculating the ratio of an in-slice resolution to an inter-slice resolution, using a linear interpolation method to interpolate inter-slice pixels, and making the spatial ratio of the three orthogonal MRI images consistent to obtain images I1, I2 and I3; (3) after linear interpolation, selecting one of the three images I1, I2 and I3 with the best quality as a reference image R1, and the other two images I2 and I3 as floating images, respectively selecting a small number of feature points representing the image direction on the three images R1, I2 and I3, and using a 2.5D Harris corner point detection algorithm to extract the feature points on the reference image R1; (4) using a CLM coherent marker point mapping matching algorithm to search for feature points matching the feature points on the reference image R1 on the two floating images I2 and I3, and using the correspondence between the two sets of feature points obtained from the reference image and the floating image to calculate the affine transformation matrix of the reference image and the floating image; (5) using the affine transformation matrix to perform affine transformation on the two floating images I2 and I3 to make them completely aligned with the reference image R1, and obtaining two generated images R2 and R3; (6) taking the brightness distribution of the reference image R1 as a reference, adjusting the brightness of the two generated images R2 and R3 to make the brightness distribution of the three images consistent, and obtaining three images R1', R2' and R3' after brightness adjustment; (7) For the three images R1', R2' and R3', the three images R1', R2' and R3' are fused and reconstructed according to a three-dimensional gradient enhanced fusion method to obtain a three-dimensional isotropic picture, i.e. a final output image I o .

2. The method of claim 1, wherein: The step (2) specifically comprises the following steps: (2a) converting the anisotropic three-dimensional nii format image into a v3draw format; (2b) calculating the up-sampling multiple of the image slices, obtaining the in-slice resolution of the image as n*n microns and the inter-slice resolution as m microns according to the ratio of the in-slice resolution to the inter-slice resolution, and calculating the inter-slice up-sampling multiple as α=m / n, wherein m>n; (2c) calculating the image size after up-sampling according to the up-sampling multiple: the initial anisotropic image size is X*Y*Z, and the output image size after interpolation is X*Y*(Z*α); (2d) performing inter-slice up-sampling on the three-dimensional anisotropic image according to the output image size, using a bilinear interpolation method, and interpolating the three images in the same way to obtain three images I1, I2 and I3 with consistent spatial ratios.

3. The method of claim 1, wherein: The step (3) specifically comprises the following steps: (3a) comparing the three images after linear interpolation, selecting the image I1 with the best quality as the reference image R1, and the two floating images as I2 and I3; the best quality refers to the largest field of view and the most uniform brightness distribution; (3b) selecting eight feature points representing the approximate direction distribution of the three images on the three images R1, I2 and I3, respectively; (3c) using 2.5D Harris corner detection algorithm, setting non-maximum suppression window respectively, wherein, the radius of three-dimensional non-maximum suppression window and two-dimensional non-maximum suppression window is set as 25 and 20 respectively, extracting 700 feature points on the reference image R1; (3d) screening the generated 700 feature points, finally retaining 500 feature points.

4. The method of claim 1, wherein: The step (4) specifically comprises the following steps: (4a) using the selected feature points, calculating the coarse accuracy three-dimensional affine transformation matrix of the floating image I2 to the reference image R1 and the floating image I3 to the reference image R1, using the coarse accuracy three-dimensional affine transformation matrix to make the preliminary affine transformation of the two floating images I2 and I3 to the reference image R1, obtaining the floating image I2' and the floating image I3', the basic direction of the three images R1, I2' and I3' remains consistent; (4b) inputting the reference image R1 and the feature points on the reference image R1 extracted using the 2.5D Harris corner detection algorithm, and the floating image I2' and I3' after direction adjustment, using the CLM coherent marker point mapping matching algorithm, setting the search radius as 10, finding the feature points on the floating image I2' and the floating image I3' matched with the feature points on the reference image R1 through the iterative search mode; (4c) using the inverse matrix of the coarse accuracy three-dimensional affine transformation matrix to inversely transform the feature points on the floating image I2' and the floating image I3' to the floating image I2 and the floating image I3, obtaining the feature points on the floating image I2 and the floating image I3 accurately matched with the reference image R1, using the two sets of accurately matched feature point pairs to calculate the affine transformation matrix of the floating image I2 to the reference image R1 and the floating image I3 to the reference image R1.

5. The method of claim 1, wherein: The step (5) specifically comprises the following steps: (5a) generating two full empty images with the same size as the reference image R1 and the pixel value of 0, obtaining the pixel value of each pixel point of the two full empty images, and according to the affine transformation matrix, obtaining the position of each pixel point in the original image of the two floating images I2 and I3; (5b) using the pixel value of the eight neighborhood positions around the position in the original image to perform linear interpolation on the pixel value of the generated image; if the position of the target image pixel point in the original image exceeds the boundary, the pixel value of the part exceeding the boundary is taken as the pixel value of the reference image R1 at the position, obtaining the generated image R2 according to the floating image I2 and obtaining the generated image R3 according to the floating image I3.

6. The method of claim 1, wherein: The step (6) specifically comprises the following steps: (6a) selecting the radius of the sliding one-dimensional window as 5, and performing mean value smoothing filtering on the X, Y and Z three views of the three images R1, R2 and R3 respectively, obtaining three smoothed filtered images S1, S2 and S3; (6b) traversing the three images R1, R2, R3 pixel by pixel, adjusting the luminance values of the generated images R2 and R3, after the luminance matching, the pixel values of the corresponding positions are R2(x, y, z) = R2(x, y, z)*(S1(x, y, z) / S2(x, y, z)), R3(x, y, z) = R3(x, y, z)*(S1(x, y, z) / S3(x, y, z)), obtaining three images R1', R2' and R3' after the luminance adjustment and the consistent luminance.

7. The method of claim 1, wherein: The step (7) specifically comprises the following steps: (7a) generating a full empty image with the same size as the three images R1', R2' and R3', and the pixel value is 0; (7b) For the full empty image of step (7a), iterate over each pixel position, for each pixel position (x,y,z) in the final output image I o R'(x,y,z), R'(x+1,y,z), R'(x-1,y,z), R'(x,y+1,z), R'(x,y-1,z), R'(x,y,z+1), R'(x,y,z-1) of the corresponding position and neighborhood of the three images R1', R2' and R3' respectively; (7c) according to the three-dimensional gradient enhanced fusion method to fuse three images, as follows: for the pixel value assignment of each pixel position in the final output image I o , take , where a = 4b, b is in the range of 0 to 1, and is adjusted according to the fusion effect, and in the above formula, if I o (x,y,z)<0 or I o (x,y,z)>255, take , and then iterate to obtain the final output image I o .

Images