Sparse view SPECT reconstruction method based on mixed feature hash table and wavelet implicit neural representation
Through the method of mixing feature hash table and wavelet implicit neural representation, combined with ray sampling and multi-layer perceptron network, the poor effect of traditional SPECT reconstruction in sparse views and high noise is solved, and high-quality three-dimensional image reconstruction is achieved , and reduces the computing resource requirements.
Patent Information
- Application Number
- CN202410964708.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-12
AI Technical Summary
Traditional SPECT reconstruction methods have poor reconstruction effects under sparse views and high noise conditions, and deep learning-based methods require a large amount of data and high computing resources.
Using a sparse view SPECT reconstruction method based on a mixed feature hash table and wavelet implicit neural representation, three-dimensional images under low dose and high noise are realized through ray sampling, multi-resolution hash encoding, Gabor wavelet activation function and Preconditioned Alternating Projection Algorithms reconstruction.
The image quality of sparse view SPECT reconstruction is improved, and the effect of fault reconstruction is further improved compared with traditional methods and reduced computing resource requirements.
Smart Images

Figure CN118941662A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of biomedical image reconstruction, and in particular to a sparse view SPECT reconstruction method based on a hybrid feature hash table and wavelet implicit neural representation. Background Art
[0002] In the field of medical imaging, Single-Photon Emission Computed Tomography (SPECT) is a commonly used nuclear medicine imaging technique used to detect and diagnose a variety of diseases. SPECT obtains information about tissue and organ function by measuring gamma rays emitted from the patient's body. However, due to the noise in the SPECT imaging process (the image may have Poisson noise), scattering, and limited angle sampling, reconstructing accurate three-dimensional images has always been a challenging task.
[0003] Traditional SPECT reconstruction methods can be divided into three categories: analytical methods, iterative methods, and learning-based methods. Analytical methods estimate the distribution of radionuclides by solving Radon transform and its inverse problem. Among them, the classic FDK algorithm performs well in ideal conditions, but the reconstruction effect is poor in the case of incomplete data such as sparse views. Iterative methods are another commonly used reconstruction method, which models the reconstruction problem as a minimization process. These methods use an optimization framework combined with a regularization module for reconstruction. Although iterative methods perform well when dealing with incomplete data problems, they require long computing time and a large amount of memory. With the rise of artificial intelligence, learning-based methods have become increasingly popular in SPECT reconstruction. These methods use deep neural networks to predict and extrapolate projection data, and estimate the distribution of radionuclides through regression of similar data to make the reconstruction process differentiable. Among them, most methods require large-scale data sets for network training and rely on neural networks to memorize the features of SPECT images. Therefore, a sparse view SPECT reconstruction method based on a hybrid feature hash table and wavelet implicit neural representation is invented. Summary of the invention
[0004] In view of the above problems and / or the existing problems in the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation, the present invention is proposed.
[0005] Therefore, the purpose of the present invention is to provide a sparse view SPECT reconstruction method based on a hybrid feature hash table and wavelet implicit neural representation, which can solve the above-mentioned existing problems.
[0006] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:
[0007] The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation includes the following specific steps:
[0008] S1: Obtain sampling points in three-dimensional space through ray sampling;
[0009] S2: Position encoding of the coordinates of the sampling points using a multi-resolution hash encoder with a hybrid feature hash table;
[0010] S3: The hash-encoded sparse view is fed into a fully connected multilayer perceptron, using Gabor wavelet as the activation function to learn high-frequency functions in a low-dimensional problem domain.
[0011] S4: The network is trained by minimizing the error between the real and synthesized projections, and dense views are generated when predicting that do not contain the training views;
[0012] S5: The sparse views during training and the generated dense views are combined to achieve SPECT reconstruction of low-dose and high-noise sparse views through the Preconditioned Alternating Projection Algorithms reconstruction algorithm.
[0013] As a preferred solution of the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation described in the present invention, wherein: for each input projection image in S1, 1024 rays will be generated, each ray will be sampled at uniform intervals, and N points will be selected, where the value of N must exceed the maximum dimension of the image.
[0014] As a preferred solution of the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation described in the present invention, wherein: the two-dimensional coordinate x in each of the projection images , =(i,j) through a rotation matrix R∈R 4×4 , thus converting to the three-dimensional coordinates x=(x,y,z) in the world coordinate system, its three-dimensional space coordinates x=(x,y,z)∈R 3 Can be normalized.
[0015] As a preferred solution of the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation described in the present invention, wherein: the multi-resolution hash coding with hybrid feature hash table is used as a position encoder in S2, and its implementation process can be described by the following three core formulas:
[0016]
[0017] Where p represents the spatial point to be queried, and Θ refers to the set of feature lookup tables of size T learned for each voxel grid. By using a multi-resolution hash encoder with a shared hash table to query the feature lookup table corresponding to the query point p, the corresponding feature vectors can be retrieved for the adjacent vertices c of point p at L multi-resolution levels.
[0018] As a preferred solution of the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation described in the present invention, wherein: in S3, the hash-encoded sparse view is input into a fully connected multilayer perceptron network, the network includes 4 fully connected layers, each layer of the fully connected multilayer perceptron network has 32 neurons, there is a residual block after the second layer, the residual block is used as the input of the fourth layer, and except for the last layer, each fully connected layer is followed by a Gabor wavelet activation function, and the function expression formula is:
[0019]
[0020] where ω i Used to control the frequency of the wavelet, s i Used to control the Gaussian falloff width.
[0021] As a preferred solution of the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation described in the present invention, wherein: the real projection I in S4 r and Synthetic Projection I s The loss function between is:
[0022]
[0023] Among them, projection I is used To express, I0 is the initial strength, δ i is the distance between adjacent points, where B is a batch of rays, and the fully connected multilayer perceptron network parameters θ are updated during the training process.
[0024] As a preferred solution of the sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation described in the present invention, wherein: in S5, the corresponding SPECT image is obtained by the Preconditioned Alternating Projection Algorithms reconstruction algorithm.
[0025] Compared with existing technologies:
[0026] The present invention uses a wavelet implicit neural representation method to enable the method to reconstruct three-dimensional images under the condition of sparse view projection. Secondly, by using a multi-resolution hash encoder with a mixed feature hash table, features at multiple resolution levels share the same hash table, and the scene can be encoded at multiple resolution levels, aiming to quickly and adaptively capture high-frequency and low-frequency features in three-dimensional space by flexibly adjusting the resolution of the data. Finally, through a multi-layer perceptron, using the Gabor wavelet activation function, combined with the Preconditioned Alternating Projection Algorithms reconstruction algorithm, low-dose sparse view SPECT reconstruction is achieved, thereby providing high-quality reconstruction results. Therefore, the beneficial effect of the present invention is to achieve high-quality results for sparse view SPECT reconstruction, which can further improve the quality of tomographic reconstructed images compared to traditional deep learning methods. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a schematic diagram of the overall model framework of the present invention;
[0028] Figure 2 The architecture diagram of the fully connected multilayer perceptron of the present invention;
[0029] Figure 3 It is a flow chart of the present invention. DETAILED DESCRIPTION
[0030] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.
[0031] The present invention provides a sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation, please refer to Figure 1-Figure 3 , including the following specific steps:
[0032] S1: Obtain sampling points in three-dimensional space through ray sampling;
[0033] Among them: For each input projection image, 1024 rays will be generated, and each ray will be sampled at uniform intervals to select N points, where the value of N must exceed the maximum dimension of the image. For example, for a 128×128×128 pixel chest CT image, 192 points will be sampled on each ray; the two-dimensional coordinates x in each training set projection image , =(i,j) through a rotation matrix R∈R 4×4 , thus converting to the three-dimensional coordinates x=(x,y,z) in the world coordinate system, its three-dimensional space coordinates x=(x,y,z)∈R 3 Ability to normalize;
[0034] S2: Position encoding of the coordinates of the sampling points using a multi-resolution hash encoder with a hybrid feature hash table;
[0035] Among them: using multi-resolution hash coding with a hybrid feature hash table as a position encoder, its implementation process can be described by the following three core formulas:
[0036]
[0037] Where p represents the spatial point to be queried, and Θ refers to the set of feature lookup tables of size T learned for each voxel grid. By using a multi-resolution hash encoder with a shared hash table to query the feature lookup table corresponding to the query point p, the corresponding feature vectors for the adjacent vertices c of the point p can be retrieved at L multi-resolution levels. Specifically, at each resolution level: (1) the index of the query point p mapped to the hash table is found through the mapping function T, i.e., formula (3); (2) after obtaining the correct index, the corresponding feature value H can be determined for each adjacent vertex c of the query point p in the three-dimensional space through the hash function h, i.e., formula (2); (3) based on the relative position of the query point p in its surrounding voxels (defined by the adjacent vertices c), the above retrieved features are linearly interpolated. (4) After the interpolated feature vectors obtained at all resolution levels are merged together, that is, formula (1), the final multi-resolution feature representation of the query point p is constructed;
[0038] S3: The hash-encoded sparse view is fed into a fully connected multilayer perceptron, using Gabor wavelet as the activation function to learn high-frequency functions in a low-dimensional problem domain.
[0039] Among them: the hash-encoded sparse view is input into a fully connected multilayer perceptron network, and the network includes 4 fully connected layers, wherein each layer of the fully connected multilayer perceptron network has 32 neurons, and there is a residual block after the second layer, and the residual block is used as the input of the fourth layer. Since the Gabor wavelet activation function combines the complex exponential term and the Gaussian attenuation term, it can capture the high-frequency components of the input signal and has good localization characteristics. Therefore, except for the last layer, each fully connected layer is followed by a Gabor wavelet activation function, and the function expression formula is:
[0040]
[0041] where ω i Used to control the frequency of the wavelet, s i Used to control the Gaussian attenuation width;
[0042] S4: The network is trained by minimizing the error between the real and synthesized projections, and dense views are generated when predicting that do not contain the training views;
[0043] Real Projection I r and Synthetic Projection I s The loss function between is:
[0044]
[0045] Among them, projection I is used To express, I0 is the initial strength, δ i is the distance between adjacent points, where B is a batch of rays, and the fully connected multilayer perceptron network parameters θ are updated during the training process;
[0046] S5: Combine the sparse views during training with the generated dense views, and use the Preconditioned Alternating Projection Algorithms reconstruction algorithm to achieve low-dose and high-noise sparse view SPECT reconstruction;
[0047] Among them: the corresponding SPECT image is obtained through the Preconditioned Alternating Projection Algorithms reconstruction algorithm.
[0048] The entire algorithm of this implementation is tested in Ubuntu 16.04.1 (64-bit) system, where the CPU is Intel Xeon Silver 4210R (2.4GHz) and the graphics card model is GeForce RTX 3090 (8GB video memory). In programming, Pytorch 1.11.0 platform is used to build the network model.
[0049] The experimental data used in the present invention was obtained by Monte Carlo simulation software Simind, and the size of each low-dose view was 128 × 128. There were a total of 120 projection views in the 360-degree scan, of which the training set was 1 view at every 6-degree interval, that is, 20 views were selected as the training set, and then all the remaining 100 projection views were rendered and generated.
[0050] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation, characterized in that: The specific steps are as follows: S1: Obtain sampling points in three-dimensional space through ray sampling; S2: Position encoding of the coordinates of the sampling points using a multi-resolution hash encoder with a hybrid feature hash table; S3: The hash-encoded sparse view is fed into a fully connected multilayer perceptron, using Gabor wavelet as the activation function to learn high-frequency functions in a low-dimensional problem domain. S4: The network is trained by minimizing the error between the real and synthesized projections, and dense views are generated when predicting that do not contain the training views; S5: The sparse views during training are combined with the generated dense views, and the low-dose and high-noise sparse views of SPECT reconstruction are achieved through the Preconditioned Alternating Projection Algorithms reconstruction algorithm.
2. The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation according to claim 1, characterized in that: In S1, for each input projection image, 1024 rays are generated, each ray is sampled at uniform intervals, and N points are selected, where the value of N must exceed the maximum dimension of the image.
3. The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation according to claim 2, characterized in that: The two-dimensional coordinates x'=(i,j) in each projection image are transformed into a rotation matrix R∈R 4 ×4 , thus converting to the three-dimensional coordinates x=(x,y,z) in the world coordinate system, its three-dimensional space coordinates x=(x,y,z)∈R 3 Can be normalized.
4. The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation according to claim 1, characterized in that: In S2, multi-resolution hash coding with a hybrid feature hash table is used as a position encoder, and its implementation process can be described by the following three core formulas: Where p represents the spatial point to be queried, and Θ refers to the set of feature lookup tables of size T learned for each voxel grid. By using a multi-resolution hash encoder with a shared hash table to query the feature lookup table corresponding to the query point p, the corresponding feature vectors can be retrieved for the adjacent vertices c of point p at L multi-resolution levels.
5. The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation according to claim 1, characterized in that: In S3, the hash-encoded sparse view is input into a fully connected multilayer perceptron network, which includes 4 fully connected layers, wherein each layer of the fully connected multilayer perceptron network has 32 neurons, and there is a residual block after the second layer, which is used as the input of the fourth layer. Except for the last layer, each fully connected layer is followed by a Gabor wavelet activation function, and the function expression formula is: where ω i Used to control the frequency of the wavelet, s i Used to control the Gaussian falloff width.
6. The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation according to claim 1, characterized in that: The real projection I in S4 r and Synthetic Projection I s The loss function between is: Among them, projection I is used To express, I0 is the initial strength, δ i is the distance between adjacent points, where B is a batch of rays, and the fully connected multilayer perceptron network parameters θ are updated during the training process.
7. The sparse view SPECT reconstruction method based on hybrid feature hash table and wavelet implicit neural representation according to claim 1, characterized in that: In S5, the corresponding SPECT image is obtained by using the Preconditioned Alternating Projection Algorithms reconstruction algorithm.
Citation Information
Cited By
Sparse view reconstruction system and method based on neural attenuation field
CN119180883A
Cryoelectron microscope density map data compression and decompression method and system
CN120416496A
A method and system for cryo-em density map data compression and decompression
CN120416496B
