CT sparse angle reconstruction method based on deep learning

Abstract

The invention discloses a CT sparse angle reconstruction method based on deep learning, and relates to the technical field of industrial image processing and computed tomography, and the method comprises the following steps: S1, data classification and preprocessing: obtaining original CT sinogram data, segmenting the original CT sinogram data through a 128 * 128 pixel sliding window with the overlapping rate of 15%, and obtaining a CT sparse angle; carrying out classification through a three-level classification index system to obtain 12 types of small sinogram blocks with labels; s2, sine domain deep learning data enhancement training: performing interval sampling deletion operation on the classified 12 types of sinogram small blocks, constructing a sample triple, establishing a sine domain generation network of a double-discrimination network structure, and taking the sample triple as training data to obtain a sine domain completion model after training is completed; s3, training and repairing the projection domain repair graph based on the U-Net network; s4, carrying out artifact removal processing on the reconstructed image; the method effectively optimizes the image quality.

Description

Technical Field

[0001] This invention relates to the fields of industrial image processing and computed tomography (CT) technology, specifically a deep learning-based CT sparse angle reconstruction method. Background Technology

[0002] In CT (computed tomography) imaging, sparse angle scanning can reduce radiation dose, but it leads to incomplete projection data, and the reconstructed image is prone to artifacts and loss of detail. While existing deep learning methods can complete sparse data, local biases still exist in the completed image. Traditional denoising algorithms struggle to specifically eliminate the stripe artifacts unique to sparse angle reconstruction, failing to balance image detail preservation and artifact removal, thus failing to meet the high image quality requirements of clinical or industrial testing.

[0003] To address the inherent limitations of traditional methods, this invention innovatively proposes a deep learning-based CT sparse angle reconstruction method to solve the aforementioned problems. Summary of the Invention

[0004] The purpose of this invention is to provide a deep learning-based CT sparse angle reconstruction method. Through a four-level processing flow of "data preprocessing - sinusoidal domain completion - projection domain repair - artifact removal of reconstructed image", it achieves efficient completion of sparse projection data and quality optimization of reconstructed image, thereby reducing radiation dose while improving the clarity and accuracy of reconstructed image.

[0005] This invention provides the following technical solution: a deep learning-based method for sparse angle reconstruction in CT, comprising the following steps: S1. Data Classification and Preprocessing: Obtain the original CT sinusoidal image data, use a 128×128 pixel sliding window with an overlap rate of 15% to segment the original CT sinusoidal image data, and classify it into 12 categories of labeled sinusoidal image patches through a three-level classification index system. The three-level classification index system includes spatial location stratification, information entropy classification, and local contrast grading. S2. Sine Domain Deep Learning Data Augmentation Training: Perform interval sampling deletion on the 12 classified sine graph patches to construct sample triples of "full projection-sparse projection-mask matrix" and build a sine domain generation network with a dual-discriminant network structure. The sine domain generation network includes the generation network and two sub-discriminant networks with the same structure. Using the sample triples as training data, the sine domain generation network is optimized by alternately using adversarial loss and reconstruction loss to obtain the trained sine domain completion model. S3. Training and Restoration of Projection Domain Inpainting Based on U-Net Network: Acquire new CT images, perform classification in S1 and interval sampling deletion operations in S2 sequentially to obtain sparse image patches. Input the sparse image patches into the sinusoidal domain inpainting model trained in S2 to generate complete image prediction values. Use "complete image prediction value - new image patch" as training sample pairs to train a U-Net network with a symmetric encoder-decoder structure to obtain a projection domain inpainting model. Use the projection domain inpainting model to repair the complete image prediction values ​​and output the repaired projection domain image. S4. Artifact Removal Processing of Reconstructed Image: Perform CT reconstruction on the repaired projection domain image to obtain the initial reconstructed image; sequentially perform wavelet multi-scale decomposition and adaptive threshold denoising, wavelet reconstruction, artifact direction detection, anisotropic TV filtering, image fusion and enhancement operations on the initial reconstructed image to output the final artifact-removed CT reconstructed image.

[0006] Compared with the prior art, the beneficial effects achieved by the present invention are: The data classification and preprocessing stage of this invention adopts a three-level classification index system, which utilizes spatial location stratification, information entropy classification, and local contrast grading to provide targeted input for subsequent network training, thereby improving model training efficiency and completion accuracy. This invention ensures the authenticity and accuracy of sparse projection completion data by constructing a sinusoidal domain generation network with dual discriminant network and by alternately optimizing the sinusoidal domain generation network with adversarial loss and reconstruction loss, thus laying the foundation for subsequent repair. This invention sets up a training and repair stage for the projection domain inpainting map of the U-Net network. By fusing multi-scale features through skip connections, it further corrects the completion bias of the generated network and improves the quality of the projection domain image. The artifact removal processing, which combines traditional algorithms with orientation-sensitive filtering, can specifically eliminate the strip artifacts in sparse angle reconstruction while preserving image edges and details, and finally outputs high-quality CT reconstructed images to meet the needs of clinical diagnosis or industrial testing. Attached Figure Description

[0007] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the network architecture of the sinusoidal domain generation network of the present invention; Figure 2 This is a schematic diagram of the network architecture of the U-Net of this invention; Figure 3 This is a schematic diagram of the image artifact removal process of the present invention; Figure 4 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0008] The following detailed, non-limiting description of the technical solution of the present invention, in conjunction with preferred embodiments and accompanying drawings, is provided. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0009] Please see Figure 1-4 This invention provides a technical solution: a deep learning-based CT sparse angle reconstruction method, comprising the following: S1. Data Classification and Preprocessing: Obtain the original CT sine wave data, use a 128×128 pixel sliding window with an overlap rate of 15% to segment the original CT sine wave data, and classify it into 12 categories of labeled sine wave blocks through a three-level classification index system to provide standardized and highly regular input data for subsequent deep learning training; The three-level classification index system includes spatial location stratification, information entropy classification, and local contrast grading. S2. Sine Domain Deep Learning Data Augmentation Training: Perform interval sampling and deletion operations on the 12 classified sine graph patches to construct sample triples of "full projection-sparse projection-mask matrix". Build a sine domain generation network with a dual-discriminator network structure, which includes the generation network and two sub-discriminator networks with the same structure. Using the sample triples as training data, optimize the sine domain generation network alternately through adversarial loss and reconstruction loss to obtain the trained sine domain completion model, thus achieving the completion of sparse projection data. S3. Training and Restoration of Projection Domain Inpainting Based on U-Net Network: Acquire new CT images, and sequentially perform the classification in S1 and the interval sampling deletion operation in S2 to obtain sparse image patches. Input the sparse image patches into the sinusoidal domain inpainting model trained in S2 to generate complete image prediction values. Use "complete image prediction value - new image patch" as training sample pairs to train a U-Net network with a symmetric encoder-decoder structure to obtain a projection domain inpainting model. Use the projection domain inpainting model to repair the complete image prediction values ​​and output the repaired projection domain image to improve the consistency between the inpainted image and the real image. S4. Artifact Removal Processing of Reconstructed Image: Perform CT reconstruction on the repaired projection domain image to obtain the initial reconstructed image; sequentially perform wavelet multi-scale decomposition and adaptive threshold denoising, wavelet reconstruction, artifact direction detection, anisotropic TV filtering, image fusion and enhancement operations on the initial reconstructed image to output the final artifact-removed CT reconstructed image, eliminating the strip artifacts unique to sparse angle reconstruction and optimizing image brightness and details.

[0010] Regarding the data classification and preprocessing implementation in S1, the original complete CT sinusoidal image data is divided into small blocks according to preset sizes and rules, providing standardized input units for subsequent classification and network training. The specific implementation process is as follows: S101. The original CT sine wave data is segmented, using 128×128 pixels as the fixed size of the sine wave blocks; First, feature extraction is balanced: a size of 128×128 pixels can fully preserve the local projection features of the sine wave (such as edge contours and grayscale gradients), while avoiding the dilution of local features due to excessive size, ensuring that the network can accurately capture the detailed information of the projection data; Secondly, the computational efficiency is well-suited: the size matches the memory capacity of mainstream GPUs, allowing more samples to be loaded in a single training session, thus improving the training iteration speed and avoiding memory overflow issues caused by excessive size. Finally, spatial information preservation: Compared to smaller sizes such as 64×64, 128×128 pixels can preserve richer spatial correlations, which is especially suitable for correlation analysis of adjacent sampling points in CT projection data, reducing spatial information breaks caused by segmentation.

[0011] S102. Segmentation method design: A sliding window segmentation method with a 15% overlap rate is adopted. Specific operation: Using a window size of 128×128 pixels, the window is moved in the horizontal (width) and vertical (height) directions of the original CT sine wave data according to a sliding step size. The formula for calculating the sliding step size is: Sliding step size = window size × (1 - overlap rate) = 128 × (1 - 15%) = 108 pixels; Each time the window is moved, the pixel area within the current window is captured as a small sine wave patch until the entire area of ​​the original CT sine wave data is covered. For the edge areas of the original CT sine wave data (the parts of the window that extend beyond the image boundary), pixels are added using a mirror-fill method to ensure that the size of the edge patches is still 128×128 pixels, thus avoiding the loss of edge information. In the 12 categories of sine wave patches after classification, the number of samples in each category is no less than 500, and the total number of samples is no less than 10,000. This segmentation method achieves complete coverage of the original data and enhances the feature correlation between adjacent small blocks through overlapping regions, providing high-quality input data for subsequent classification and training.

[0012] S103. Sine graph data classification: A three-level classification index system was used to finely classify the segmented 128×128 sine graph blocks from three dimensions: spatial location stratification, information entropy classification, and local contrast grading. A total of 12 categories were divided and labeled accordingly. The classification method is as follows: Spatial location stratification is a primary classification: The upper and lower regions of the CT detector exhibit symmetrical differences in path length and absorption coefficient when X-rays penetrate an object. The upper region corresponds to the projection data of the upper part of the object penetrated by X-rays, while the lower region corresponds to the projection data of the lower part. Their grayscale distribution characteristics are significantly different and require separate classification to improve the targeting of subsequent network feature extraction. Spatial location stratification is based on the acquisition geometry of the CT detector. Utilizing the symmetrical differences in the X-ray penetration path, each sine wave block is divided into upper and lower regions along the vertical direction (detector height direction). Specific division rules are as follows: Let the height of the small block of the sine graph be... (CT detector height), the layer number where the sine wave patch is located is defined as... The upper area is the label. The vertical coordinates satisfy The area below; the lower area is the label. The vertical coordinates satisfy The area; Based on primary classification, all sine graph blocks are initially divided into... , Two categories.

[0013] Information entropy is classified into two levels: the information entropy of each sine wave block is calculated based on the improved information entropy formula, and further classified according to the entropy value distribution. Specific methods are as follows: An improved information entropy calculation introduces a gradient energy term to optimize traditional information entropy, thereby more accurately quantifying the information content of small patches in a sine wave. The improved formula for calculating information entropy is as follows: ; The parameters are defined as follows: X: The set of pixels in the current sine wave block, where x is the pixel grayscale value; p(x): Normalized histogram probability of pixel grayscale value, i.e., p(x) = number of pixels with this grayscale value / total number of pixels in the small block; G(X): Gradient energy term, calculated as G(X)=(x,y)∈X|▽I(x,y)|, where ▽I(x,y) is the gradient value of pixel (x,y), which is calculated by taking the modulus of the gradient in the x and y directions using the Sobel operator. G(X) is used to characterize the edge and detail richness of small blocks. λ: Adjustment coefficient, preferably 0.3. This value can balance the weights of information entropy and gradient energy, avoiding classification bias caused by a single indicator.

[0014] Calculate the improved information entropy H(X) of all sine graph patches and statistically obtain the information entropy distribution of the entire sine graph patch; determine the classification thresholds Q1 and Q2: Q1 is the 33rd quantile of the entropy distribution, and Q2 is the 66th quantile of the entropy distribution.

[0015] The small blocks are divided into three categories according to the entropy value: The low-information region is labeled k1: H(X) < Q1. In this region, the pixel gray value distribution is uniform, with fewer edges and details; The medium-information region is labeled k2: Q1 ≤ H(X) ≤ Q2. This region contains certain edges and details; The high-information region is labeled k3: H(X) > Q2. In this region, the edges are dense, the details are rich, and the gray value distribution has significant differences; Through secondary classification, the two categories l1 and l2 are further subdivided into six categories: l1k1, l1k2, l1k3, l2k1, l2k2, l2k3.

[0016] The local contrast grading is a three-level classification: Based on the six categories of the secondary classification, it is further subdivided based on the adaptive contrast metric to distinguish regions with different local contrasts. The specific method is as follows: Adaptive contrast calculation, using a contrast metric formula based on the gray value quantile, to avoid the problem of being interfered by extreme values in traditional contrast calculation. The formula for adaptive contrast is: ; The definitions of each parameter are as follows: : The mean value of the top 10% high-gray-value pixels in the current sine map small block, that is, select the pixels with the top 10% gray values in the small block and calculate their gray average value; : The mean value of the bottom 10% low-gray-value pixels in the current sine map small block, that is, select the pixels with the bottom 10% gray values in the small block and calculate their gray average value; : A small constant to prevent division by zero, with a value of , to avoid The calculation exception caused when it approaches 0; Calculate the adaptive contrast C(X) of each small block after secondary classification, and statistically analyze the contrast distribution within each category; Using the 50% quantile of the contrast distribution of each category as the threshold, each category is further divided into two categories: The low-contrast region is labeled c1: C(X) < the 50% quantile of the contrast of this category. The gray value difference within the region is small, and it is visually relatively smooth; The high-contrast region is labeled c2: C(X) ≥ the 50% quantile of the contrast of this category. The gray value difference within the region is large, and the edges are visually clear.

[0017] Through the three-level classification, all 128×128 sine map small blocks are finally divided into 12 categories, and further subdivided into: l1k1c1, l1k1c2, l1k2c1, l1k2c2, l1k3c1, l1k3c2, l2k1c1, l2k1c2, l2k2c1, l2k2c2, l2k3c1, l2k3c2.

[0018] This classification method fully considers the inherent differences in spatial distribution (differences between upper and lower regions of the detector), information content (differences in entropy), and local features (differences in contrast) of industrial CT sinusoidal image data. It divides the originally complex data into 12 categories with strong regularity and significant feature differences. Subsequently, deep learning networks can design targeted training strategies for the features of each data category, effectively improving feature extraction efficiency and model training accuracy, and avoiding the problem of decreased network generalization ability due to data mixing.

[0019] The specific details of data augmentation training for deep learning in the sinusoidal domain of S2 are as follows: S201. Based on S1, data classification and preprocessing are performed to obtain a small dataset of CT sinusoidal images after previous classification and processing, denoted as... Then, an interval sampling deletion operation is performed to construct sample triples. Each sample triple contains fully projected data, sparsely projected data, and a mask matrix, i.e., { ; Each original CT sinusoidal patch is a fully projected data, defined as follows: The parameters are explained below: The height dimension of the small blocks in the sine wave corresponds to the number of projection angles during the CT scan. The width dimension of the sine wave patch corresponds to the number of detector sampling points at a single projection angle; : No. line (corresponding to the first) (the projection angle), the first Column (corresponding to the first) The projected intensity value of (each detector sampling point), , ; right Each of them Perform an interval sampling deletion operation, specifically following these rules: delete samples at fixed intervals, retaining odd-indexed sampling points from the original sampling sequence and deleting even-indexed sampling points, resulting in a sequence similar to... One-to-one corresponding sparse projection sine curve patches ; Introducing a binary mask matrix The definition is as follows: ; at this time The relationship satisfies: ,Right now Only keep The data in the middle is a valid area of ​​0, and missing data is recorded as 0 or null.

[0020] S202, Construction of training and test sets, including all The sample triplets are divided into training sets according to a preset ratio. and test set ; Preferably, the division ratio is as follows: And it meets the following conditions: training set It contains no fewer than 500 samples to ensure the generalization ability of the model training; the test set Sample distribution and training set Consistency is maintained to avoid test bias caused by differences in data distribution.

[0021] S203. Construct a sinusoidal domain generator network with a dual-discriminant network structure to complete missing projection information from sparse projection data. The network architecture includes a generator network G and a dual-discriminant network D. Refer to [reference needed] for the specific structure. Figure 1 ; ① The generator network G adopts an "encoder-decoder" architecture, and the input is a small patch of sparse projected sine graph. The output is a small patch of the predicted fully projected sine curve after completion. It is divided into an encoding part and a decoding part. The encoding part consists of 7 convolutional layers ( Composed of ) and sequentially processing the inputs Multi-scale feature extraction and dimensionality compression are performed, and the parameter settings for each convolutional layer are shown in Table 1 below.

[0022] Table 1 Generates network convolutional parameter table The mathematical expression of the encoding process is as follows: This is the encoded high-dimensional feature vector. The 8x downsampling is achieved by performing three convolutions with a stride of 2.

[0023] The decoding part consists of six deconvolutional layers (D1~D6) that encode features. Upsampling and feature reconstruction are performed, and the final output is the predicted full projection. The parameter settings for each deconvolution layer are shown in Table 2 below;

[0024] Table 2. Parameters of Deconvolutional Layers in Generative Networks The mathematical expression for the decoding process is as follows: ,in And it satisfies the constraint of only filling in the missing regions: ,in This is the generated projection data for the missing region.

[0025] ② The dual-discriminator network D contains two sets of sub-discriminator networks D1 and D2 with the same structure. Both are full-convolution integral classifiers, which are used to distinguish the authenticity of generated missing regions and real missing regions, respectively. The input to D1 is the missing region data from the generator network output. The output is the probability of the data being genuine, with 1 indicating authenticity and 0 indicating forgery. It consists of 6 convolutional layers. The composition and parameter settings are shown in Table 3 below;

[0026] Table 3 Discriminant Network Parameter Table The output of D1 is: ,in ; The input for D2 is the entire region data. Its structure is exactly the same as D1, consisting of 6 convolutional layers. The composition and parameter settings are the same as in Table 3; the output is the probability of the truth of the real data. ,in .

[0027] S204. The network is trained by alternately optimizing the generator network G and the discriminator network D to achieve accurate completion from sparse projection to complete projection. The specific training process and the setting of the optimal parameters are as follows: Optimizer: Employs the Adam optimizer, with a learning rate of [missing information]. Momentum parameters , Training batch size is 32; training epochs are 100; balance factor. It is set to 100 to balance the losses from adversarial efforts and reconstruction efforts.

[0028] S205. Define the loss function, which is divided into generator network adversarial loss and discriminant network adversarial loss, used to constrain the consistency of the distribution of generated data and real data. Generative adversarial loss is used to improve the generation of missing region data. To fool D1 as much as possible, the expression is: ,in Indicate the mathematical expectation of the training set; Determining network adversarial losses enables D1 to distinguish D2 can accurately identify The expression is: , The sub-discriminator network's predicted ground truth value for generating missing regions. The sub-discriminator network's predicted realism value for the truly missing regions; Reconstruction loss is used to constrain the pixel-level error between generated and real data, calculating only the mean square error of missing regions. Its expression is: ; in For mask matrix The number of 1s in the missing area is the total number of pixels in the missing region.

[0029] S206, Alternately begin training; The first step is training the discriminator network: from the training set A batch of samples were randomly selected from the middle. ;Will enter ,get ,extract Extract the actual missing regions ,calculate Updated via backpropagation Parameters; Then comes the training of the generative network: using the samples extracted in S1; calculating... The parameters of the encoded and decoded parts of G are updated through backpropagation; Iteration termination condition: When the training epochs reach a preset 100; or when the number of training epochs on the test set reaches a preset 100; No decline after 10 consecutive training rounds, with an error fluctuation of less than [value missing]. .

[0030] S207. Verify compliance: Verify compliance every 10 training rounds using the test set. Verify model performance and calculate the average reconstruction error on the test set. To ensure the model's generalization ability on non-training data, when At that time, the model is deemed to have passed training.

[0031] The specific implementation steps for training and repairing the projection domain patching map of the U-Net network in S3 are as follows: S301. Data Acquisition and Preprocessing: Acquire new CT images and crop them into 128×128 blocks according to S1. Perform the same interval sampling deletion operation as S201 to simulate a scenario with missing projection domain data and obtain sparse image patches. With mask ; S302, The sparse image patches obtained in S301 are... The input sinusoidal domain completion model of S2 generates an image and completes it to restore the full image prediction value in the projection domain. This provides foundational data for subsequent fine-grained repair of U-Net; S303, Training data construction, predicting complete image values With the actual complete image Composition of training sample pairs The sample pairs were divided into training sets in a 7:3 ratio. and verification set ; training set Used for model parameter updates; validation set Used for model performance evaluation and early stopping detection.

[0032] The S304 and U-Net network architecture design aims to achieve fine restoration of the projection domain inpainted image. The U-Net network with a symmetric encoder-decoder structure is adopted, and the multi-scale features of the encoder output are fused through skip connections to improve the detail accuracy of the restored image. The network architecture is adapted to a maximum input size of 512x512, and is specifically divided into three parts: encoder, decoder, and output layer; see the overall framework for details. Figure 2 ; The encoder consists of alternating convolutional layers and max pooling layers, which progressively downsample and extract features from the input image. The sampling size is halved and the number of feature extraction channels is doubled, realizing the transformation from pixel-level details to high-level semantic features. The parameters of each layer are shown in Table 4 below.

[0033] Table 4 U-Net Network Encoder Parameter Table The decoder consists of deconvolutional layers (upsampling) and convolutional layers. The feature map size is doubled by deconvolution, and it is concatenated with the feature map of the corresponding layer of the encoder through skip connections. The number of channels is superimposed to achieve the fusion of high-level semantic features and low-level detailed features, thereby improving the repair accuracy. The parameters of each layer are shown in Table 5 below.

[0034] Table 5 U-Net Network Decoder Parameter Table The output layer is 1. The convolutional layer compresses the 64-channel feature map output from the last layer (level 8) of the decoder into a 1-channel image, ultimately outputting the repaired image. The expression is:

[0035] in, This indicates the deconvolution operation. surface Convolution operation, express Convolution operation; S305. To ensure stable model convergence and improve training efficiency, the following optimizer and training parameters are preferred: Optimizer: Adam optimizer; learning rate To balance convergence speed and stability; momentum parameter , To accelerate gradient descent convergence; the batch size is 32, which improves training efficiency and gradient estimation accuracy within the limits of GPU memory capacity; the training epochs are 50, which is the preset maximum number of training epochs to avoid overfitting.

[0036] S306. Define the loss function. To balance pixel-level restoration accuracy and structural texture similarity, a hybrid loss function combining mean squared error (MSE) and structural similarity (SSIM) is adopted. Mean squared error loss (MSE Loss) measures the pixel-level error between the restored image and the original image, reflecting the overall deviation of pixel values. Its expression is: ; in, For the sample size, For the first The restored image of each sample, For the first The actual complete image of each sample; Structural similarity loss (SSIM Loss) measures the structural and texture similarity between the restored image and the original image, avoiding image blurring caused by mean square error. The expression is: ; in This is the function for calculating the structural similarity index, with a value range of [-1, 1]. The closer the value is to 1, the more similar the structures are.

[0037] Total loss function: Introducing a balance coefficient Values A value of 0.2 is preferred, used to fuse the mean squared error loss and the structural similarity loss, expressed as: ; S307. Training Iteration Process: The training process employs a batch stochastic gradient descent strategy; ① Data Input: From the training set... A batch of samples was randomly selected from the middle { ① Input to the U-Net network; ② Forward propagation: The U-Net network processes the input... Perform feature extraction and restoration, and output the restored image. ③ Loss Calculation: Calculate the loss of the current batch based on the loss function defined in S306. , ④ Backpropagation: based on total loss The U-Net network's trainable parameters, including the weights and biases of convolutional and deconvolutional layers, are updated using gradient descent. ⑤ Validation and early stopping: After each training epoch, the validation set... Calculate and verify the loss The preferred method is to verify that the loss decrease is less than 10% over five consecutive training epochs. -5 If the early stop mechanism is triggered, training will continue; otherwise, training will continue.

[0038] The specific implementation method for artifact removal processing of reconstructed images in S4 is as follows, with the process referenced below. Figure 3 : S401. Input the reconstructed image to be denoised. Then, normalization preprocessing is performed to map the image pixel values ​​to a specific range, such as normalizing the pixel values ​​to the [0, 1] interval, as shown in the formula: ; in, This represents the pixel value of the preprocessed image at coordinates (x, y). and The input images are respectively The minimum and maximum pixel values ​​are determined, and this step provides a uniform and suitable pixel value basis for subsequent processing.

[0039] S402, Wavelet Multiscale Decomposition and Adaptive Threshold Denoising; First, the db4 wavelet basis is used to analyze the preprocessed image. Perform a three-level wavelet multi-scale decomposition; let the approximate coefficients after wavelet decomposition be... The detail factor is , =1,2,3, representing detail sub-bands in the horizontal, vertical, and diagonal directions, respectively; Then, the noise reduction threshold for each sub-band is dynamically calculated based on the noise level. Using a threshold calculation method such as the Stein Unbiased Risk Estimation (SURE) threshold calculation, the formula for the noise reduction threshold is: ; in, The standard deviation of noise. The number of coefficients in the subband, for high-frequency detail coefficients. Thresholding is performed using a soft threshold function: ; This thresholding process effectively distinguishes between real details and noise components, thus achieving noise suppression.

[0040] S403. Wavelet Reconstruction: Perform wavelet reconstruction on the wavelet coefficients after high-frequency coefficient thresholding. The wavelet coefficients are approximate coefficients. and the processed detail coefficient The reconstructed image after preliminary denoising is obtained. The reconstruction process is the inverse process of wavelet decomposition, and can be represented as: ; in, This indicates the inverse wavelet transform operation.

[0041] S404, artifact direction detection, employs a gradient field-based approach. Transformation on the reconstructed image after wavelet reconstruction Analyze the image to accurately locate the path of the strip-shaped artifacts; First, calculate the gradient field of the image. The gradient can be calculated using operators such as the Sobel operator. The gradients in the horizontal and vertical directions are respectively... and Then the magnitude of the gradient and direction for: ; Then use The transformation maps gradient information from image space to parameter space, detecting the direction of stripe artifacts. ; S405, anisotropic TV filtering, based on the detected artifact direction Implement direction-sensitive anisotropic TV filtering; the energy functional of the anisotropic TV filter is: ; in, The filtered image. The input image is, here it is , Let be the unit vector in the direction of the artifact. Let be a unit vector perpendicular to the artifact direction. These are the weighting coefficients. This is used to control the constraint strength in the vertical and parallel artifact directions. For regularization parameters; By minimizing this energy functional, in the direction perpendicular to the artifact... Apply stronger constraints in the direction to specifically eliminate stripes while preserving other directions. The edge structure (direction) is used to obtain the filtered image. .

[0042] S406. Image fusion and enhancement: The results of wavelet denoising and TV filtering are fused, preferably using a 7:3 weighted ratio for merging. The formula is as follows: ; in, Let be the pixel value at coordinates (x, y) in the fused image. Then, an improved Retinex algorithm is used to correct for uneven brightness. The Retinex algorithm, based on the color constancy of the human visual system, decomposes the image into reflection and illumination components. The improved Retinex algorithm can be expressed as: ; in, To enhance the image in coordinates Pixel value at that location, The weighting function is Gaussian. The normalization coefficients are used to obtain the optimized denoised image. .

[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features, and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based CT sparse angle reconstruction method, characterized in that: Includes the following steps: S1. Data classification and preprocessing: Obtain the original CT sine wave data, and use a 128×128 pixel sliding window with an overlap rate of 15% to segment the original CT sine wave data. Then, classify the data into 12 categories of labeled sine wave blocks through a three-level classification index system. The three-level classification index system includes spatial location stratification, information entropy classification, and local contrast grading. S2. Sine Domain Deep Learning Data Augmentation Training: Perform interval sampling and deletion operations on the 12 categories of sine graph patches after classification, construct sample triples of "full projection-sparse projection-mask matrix", and build a sine domain generation network with a dual-discriminant network structure. The sine domain generation network includes a generation network and two sub-discriminant networks with the same structure. Using the sample triples as training data, the sine domain generation network is optimized alternately by adversarial loss and reconstruction loss to obtain the trained sine domain completion model. S3. Training and repair of projection domain inpainting based on U-Net network: Acquire new CT images, perform classification in S1 and interval sampling deletion in S2 in sequence to obtain sparse image patches, input the sparse image patches into the sinusoidal domain inpainting model trained in S2 to generate complete image prediction values. Using "predicted full image value - new image patch" as training sample pairs, a U-Net network with a symmetric encoder-decoder structure is trained to obtain a projection domain inpainting model; the projection domain inpainting model is used to inpaint the predicted full image value, and the inpainted projection domain image is output. S4. Artifact removal processing of reconstructed image: Perform CT reconstruction on the repaired projection domain image to obtain an initial reconstructed image; sequentially perform wavelet multi-scale decomposition and adaptive threshold denoising, wavelet reconstruction, artifact direction detection, anisotropic TV filtering, image fusion and enhancement operations on the initial reconstructed image to output the final artifact-removed CT reconstructed image.

2. The CT sparse angle reconstruction method based on deep learning according to claim 1, characterized in that: In S1, a sliding window with an overlap rate of 15% is used for segmentation. The sliding step size of the sliding window is calculated as follows: sliding step size = window size × (1 - overlap rate). The edge regions of the original CT sine wave data are supplemented with pixels using a mirror filling method to ensure that the size of the edge blocks remains 128×128 pixels.

3. The CT sparse angle reconstruction method based on deep learning according to claim 1, characterized in that: The specific implementation method of the three-level classification index system is as follows: Spatial stratification: Let the height of the sine graph block be H, and divide the sine graph block into an upper region along the vertical direction. With the lower region ; Information entropy classification: The information entropy of each sine graph block is calculated using an improved information entropy formula. The information entropy distribution of the entire sine graph block is statistically obtained. The 33rd percentile Q1 and the 66th percentile Q2 are taken as thresholds to divide the sine graph block into low information region k1, medium information region k2, and high information region k3. Local contrast grading is performed by calculating the adaptive contrast of each sine graph patch after classification by the information entropy. Using the 50th percentile of the adaptive contrast distribution of each class of sine graph patches as the threshold, each class is further divided into low contrast region c1 and high contrast region c2, resulting in 12 labeled sine graph patches.

4. The CT sparse angle reconstruction method based on deep learning according to claim 3, characterized in that: The improved formula for calculating information entropy is as follows: ； in The set of pixels for the current sine wave, where x is the pixel grayscale value; The normalized histogram probability of pixel grayscale values; This is the gradient energy term, used to characterize the richness of edges and details in small patches; This is an adjustment coefficient used to balance the weights of information entropy and gradient energy.

5. The CT sparse angle reconstruction method based on deep learning according to claim 3, characterized in that: The formula for the adaptive contrast ratio is: ； in This is the average value of the top 10% of high grayscale pixels in the current sine wave patch. This is the average value of the last 10% of low grayscale pixels in the current sine wave patch. To prevent division by zero of small constants.

6. The CT sparse angle reconstruction method based on deep learning according to claim 1, characterized in that: The sample triplet in S2 is represented as follows: These correspond to fully projected data, sparse projected data, and mask matrix, respectively. Defined as , , ,in The height of the small block in the sine wave corresponds to the number of projection angles during the CT scan. The width of the sine wave block corresponds to the number of detector sampling points at a single projection angle.

7. The CT sparse angle reconstruction method based on deep learning according to claim 6, characterized in that: The specific rules for the interval sampling deletion operation in S2 are as follows: at a fixed interval of deleting data once every time, retain the sampling points with odd indices in the original sampling sequence and delete the sampling points with even indices, resulting in the same result as... One-to-one correspondence A binary mask matrix M is introduced, where the sample points deleted from the mask matrix are denoted as 1, and the sample points retained are denoted as 0. in The relationship satisfies: , , representing the sequence number of the projection angles. , which represents the sequence number of the detector sampling points.

8. The CT sparse angle reconstruction method based on deep learning according to claim 7, characterized in that: The specific structure and training process of the sinusoidal domain generator network in S2 are as follows: The generator network is divided into an encoding part and a decoding part. The encoding part includes 7 convolutional layers, and the decoding part includes 6 deconvolutional layers. The two sets of sub-discriminator networks with the same structure each include 6 convolutional layers. The generator network and the discriminator network are trained using the Adam optimizer, with 100 training epochs. The adversarial loss includes generative network adversarial loss and discriminative network adversarial loss, and the expression for generative network adversarial loss is as follows: The discriminant network adversarial loss expression is as follows: ,in The sub-discriminator network's predicted ground truth value for generating missing regions. The sub-discriminator network's predicted realism value for the truly missing regions; The expression for the reconstruction loss is: ,in For mask matrix The number of 1s in the middle, Ŝ full To generate the predicted full projection of the network output, For fully projected data; alternately fix the generator network to optimize the discriminator network, and fix the discriminator network to optimize the generator network, until the training epochs reach 100 or the reconstruction loss on the test set fluctuates less than 1 / 2 for 10 consecutive training epochs. .

9. The CT sparse angle reconstruction method based on deep learning according to claim 1, characterized in that: The specific structure and training process of the U-Net network in S3 are as follows: The U-Net network adopts a symmetric encoder-decoder structure. The encoder consists of alternating convolutional layers and max pooling layers, which progressively downsample and extract features from the input image. The sampled size is halved while the number of channels for feature extraction is doubled. The decoder consists of deconvolutional layers and convolutional layers. The feature map size is doubled through deconvolution and then concatenated with the feature map of the corresponding layer of the encoder through skip connections, thus stacking the number of channels. The training process of the U-Net network involves normalizing the sparse image patches and inputting them into a sinusoidal domain completion model. Training sample pairs are formed by combining the predicted values ​​of the complete image with the actual complete image. These sample pairs are then divided into training and validation sets at a 7:3 ratio. The Adam optimizer is used with a batch size of 32 and 50 training epochs. The expression for the total loss function is as follows: ,in For mean square error loss, For structural similarity loss; an early stopping mechanism is set, whereby the validation set loss decreases by less than [a certain value] for five consecutive training epochs. Training should be stopped at this time.

10. The CT sparse angle reconstruction method based on deep learning according to claim 1, characterized in that: The specific steps of S4 are as follows: S401. Input the reconstructed image to be denoised. And perform normalization preprocessing; S402, Wavelet Multi-Scale Decomposition and Adaptive Threshold Denoising: First, a three-level wavelet multi-scale decomposition is performed using the db4 wavelet basis; approximation coefficients and detail coefficients in the horizontal, vertical, and diagonal directions are obtained; and then the denoising threshold is calculated. S403. Wavelet Reconstruction: Wavelet reconstruction is performed on the wavelet coefficients after high-frequency coefficient thresholding to obtain the reconstructed image after preliminary denoising. ; S404, Artifact Direction Detection: The Sobel operator is used to calculate the gradients and artifact directions in the horizontal and vertical directions. S405, Anisotropic TV filtering, performing direction-sensitive anisotropic TV filtering based on the detected artifact direction; S406. Image fusion and enhancement: The processing results of wavelet denoising and TV filtering are fused and merged using a 7:3 weighted ratio. Then, the brightness unevenness problem is corrected by the improved Retinex algorithm, and finally the optimized denoised image is obtained.

Images