Image transformation method for three-dimensional image

Abstract

The invention belongs to the technical field of medical image processing, and relates to a three-dimensional image transformation method based on recursive enhancement. The method comprises the following steps: firstly, carrying out feature transformation on an input three-dimensional image by adopting a recursive algorithm, and dividing the image into a plurality of local cubes for processing; then, the side length and voxel pitch of the cube are gradually adjusted through an iteration process, image features are continuously updated, and the transformation efficiency is improved; finally, a group of transformed three-dimensional images are generated, and new information is provided for analysis and prediction of brain images. The method can be applied to the fields of brain classification, brain disease diagnosis and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, in particular to a recursive enhancement algorithm for three-dimensional images, belonging to the field of computer vision. Background Art

[0002] With the development of medical image analysis, industrial inspection and computer vision technology, efficient extraction of 3D image features has become a research hotspot in 3D image recognition and classification.

[0003] Traditional methods such as Fourier transform and wavelet transform are mainly used for two-dimensional images and have insufficient feature extraction capabilities for three-dimensional images.

[0004] Performing feature transformation recursively can effectively improve the computational efficiency of the algorithm and reduce the time complexity of the algorithm.

[0005] The present invention combines recursive technology with linear transformation, and provides a method for extracting three-dimensional image features by enhancing global features at multiple levels. Summary of the Invention

[0006] The method of the present invention is roughly divided into the following steps:

[0007] Step (1): Initialize the image. Assume that the input image is represented as A x，y，z , (1≤x≤m, 1≤y≤n, 1≤z≤t), where m, n, t are the length, width, and height of the three-dimensional image respectively. The image after feature transformation is represented as B x,y,z , (1≤x≤m, 1≤y≤n, 1≤z≤t).

[0008] Step (2): Initialize the parameters. Side is set to 2, indicating the side length of the cube for generating features. Gap is set to 1, indicating the distance between the voxels that make up the Side*Side*Side cube. Iteration is set to 1, indicating the number of iterations. Count is set to in Indicates rounding up, and Count indicates the number of cubes contained in each axis of the X axis, Y axis, and Z axis.

[0009] Step (3.1): Start the first iteration. First, segment the image into cubes of size side×side×side. The 3D image contains Count×Count×Count cubes.

[0010] Step (3.2): Perform feature transformation on each cube. For each side×side×side cube block, make the following changes:

[0011]

[0012]

[0013]

[0014]

[0015]

[0016]

[0017]

[0018]

[0011] -x in

[0018] LUF ,y LUF , z LUF Represents the X, Y, and Z coordinates of the leftmost, topmost, and frontmost voxels of the cube; x, y, and z are non-negative integers that satisfy 0≤x≤Gap-1, 0≤y≤Gap-1, and 0≤z≤Gap-1. Step (3.3): Start a new iteration. Update the following parameters: Gap is twice the original value; Side is twice the original value; Iteration is 1 plus the original value; Count is 2 divided by the original value. Step (3.4): Repeat steps (3.1) and (3.2). Repeat steps (3.3) and (3.4) until the value of Count is less than 2, and the iteration ends. BRIEF DESCRIPTION OF THE DRAWINGS Figure 1 The following exemplarily shows different viewing angles of the brain image after image filling according to an embodiment of the present invention. Figure 1 (Left) is the axial view. Figure 1 (Middle) is the coronal view. Figure 1 (Right) Sagittal view. Figure 2 The different viewing angles of the embodiment of the present invention after image transformation are exemplarily shown. Figure 2 (Left) is the axial view. Figure 2 (Middle) is the coronal view. Figure 2 (Right) Sagittal view. DETAILED DESCRIPTION In order to further explain the present invention to researchers in this technical field, the following is a complete and detailed description of the specific embodiments of the present invention in combination with the technical solution of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. (1) Image Acquisition and Preprocessing: The multimodal medical brain image data selected in this embodiment is derived from the publicly available Alzheimer's Disease Neuroimaging Initiative (ADNI) database, specifically T1-weighted MRI images. The MRI images undergo the following preprocessing steps to meet the data requirements of step (1) of the present invention. (1.1) Image acquisition: The data comes from the ADNI-1 dataset. This example uses the 1_003_S_1059.mat file, with an image size of 160 × 192 × 224 (1 ≤ x ≤ 160, 1 ≤ y ≤ 192, 1 ≤ z ≤ 224). (1.2) Brain extraction: The BET algorithm of FSL software was used to perform skull stripping on the MRI images to remove the skull from the source image space. (1.3) Image alignment: The FLIRT linear registration algorithm was used to align MRI images to the Montreal Neurological Institute (MNI) T1 standard template space (MNI152_T1_1mm). (1.4) Image cropping and downsampling: The registered MRI and PET images were cropped to 152 × 188 × 152 (1 ≤ x ≤ 152, 1 ≤ y ≤ 188, 1 ≤ z ≤ 152) and further downsampled to 76 × 94 × 76 (1 ≤ x ≤ 76, 1 ≤ y ≤ 94, 1 ≤ z ≤ 76) to reduce computational complexity. (1.5) Image padding: pad the downsampled image to 128×128×128 (1≤x≤128,1≤y≤128,1≤z≤128). Figure 1 Different viewing angles of the embodiment of the present invention after image filling are exemplarily shown. (2) Model initialization process, corresponding to step (2) of the invention content. Initialization parameter settings include gap = 1, side = 2 (initial block size is 2×2×2), and number of recursions = 5. (3) Model calculation process, corresponding to step (3) of the invention content. (3.1) First iteration, corresponding to step (3.1) of the invention. Block processing: Divide the 128×128×128 image into 2×2×2 blocks, for a total of 64×64×64=262144 blocks. (3.2) Block-level transformation, corresponding to step (3.2) of the Summary of the Invention. Flatten each small block A (2×2×2, containing a total of 8 voxels) into a one-dimensional vector. Substitute the one-dimensional vector into the formula in step (3.2) of the Summary of the Invention, and fill the transformed vector B back into the corresponding small block position in the image. (3.3) The second iteration corresponds to step (3.3) of the invention. Parameter update: gap = 2, side = 4. Block processing: Divide the image according to the new gap = 2, and each block is 4×4×4 in size, for a total of 32×32×32=32768 blocks. Block-level transformation: For each 4x4×4 block, extract 64 voxel values and flatten them into a vector. Substitute the flattened vector into the formula in step (3.2) of the invention, and fill the transformed vector B back into the corresponding small block position in the image. (3.4) The third iteration corresponds to steps (3.3) and (3.4) of the invention. Parameter update: gap = 4, side = 8. Block processing: Divide the image into blocks of 8 × 8 × 8 blocks, for a total of 16 × 16 × 16 = 4096 blocks. Block-level transformation: Each block contains 512 voxel values, flattened into a vector. Substitute the flattened vector into the formula in step (3.2) of the invention, and fill the transformed vector B back into the corresponding small block position. (3.5) The fourth iteration corresponds to steps (3.3) and (3.4) of the invention. Parameter update: gap = 8, side = 16. Block processing: Divide the image into blocks of size 16 × 16 × 16 at intervals of gap = 8, for a total of 8 × 8 × 8 = 512 blocks. Block-level transformation: Each block contains 4096 voxel values, flattened into a vector. Substitute the flattened vector into the formula in step (3.2) of the invention, and fill the transformed vector B back into the corresponding small block position. (3.6) The fifth iteration corresponds to steps (3.3) and (3.4) of the invention. Parameter update: gap = 16, side = 32. Block processing: Divide the image into 32×32×32 blocks at intervals of gap = 16, for a total of 4×4×4 = 64 blocks. Block-level transformation: Each block contains 32,768 voxel values, flattened into a vector. Substitute the flattened vector into the formula in step (3.2) of the invention, and fill the transformed vector B back into the corresponding small block position. After 5 recursive linear transformations, the image is transformed from the initial A x，y,z (1≤x≤128, 1≤y≤128, 1≤z≤128) is converted into a new image Bx,y,z (1≤x≤128,1≤y≤128,1≤z≤128). After implementing the above steps, Figure 2 Different viewing angles of the embodiment of the present invention after image transformation are exemplarily shown.

Claims

1. A method for image transformation for three-dimensional images, characterized in that: The method comprises the following unique steps: Step (1): a method for performing feature transformation on each local cube, wherein the specific transformation formula is as follows: where x LUF ,y LUF ,Z LUF Represents the X, Y, and Z coordinates of the leftmost, topmost, and frontmost voxels of the local cube; x, y, and z are non-negative integers, satisfying 0≤x≤Gap-1, 0≤y≤Gap-1, and 0≤z≤Gap-1; the input image is represented by A x,y,z (1≤x≤m, 1≤y≤n, 1≤z≤t), the image after feature transformation is represented as B x,y,z (1≤x≤m, 1≤y≤n, 1≤z≤t), where m, n, and t are the length, width, and height of the three-dimensional image, respectively; Step (2): Iterative method of image transformation; initialization parameters: Side is set to 2, indicating the side length of the cube generating the feature; Gap is set to 1, indicating the distance between the voxels that constitute the Side*Side*Side cube; Iteration is set to 1, indicating the number of iterations; Count is set to in Indicates rounding up, Count indicates the number of cubes contained in each axis of the X-axis, Y-axis, and Z-axis. The parameter update method in subsequent iterations is as follows: each iteration, the new value of Gap is twice the original value, the new value of Side is twice the original value, the new value of Iteration is the original value plus 1, and the new value of Count is the original value divided by 2. The judgment condition for the end of iteration is: the value of Count is less than 2.

Images