CT image denoising method and system based on similar block learning

Abstract

The present application relates to the technical field of CT image processing, and particularly relates to a CT image denoising method and system based on similar block learning, wherein the collected X-ray tomography data set is preprocessed based on similar block matching to obtain CT reconstructed image data with similar noise distribution; the CT reconstructed image data with similar noise distribution is input into a pre-trained CT image denoising network, and the CT image denoising network is used to perform denoising processing on the CT reconstructed image data.The present application fully learns noise distribution characteristics, protects image details while reducing noise, can quickly and effectively remove noise from CT images under the condition of a small number of parameters, is convenient for network training and will not fall into overfitting due to the use of small-scale training data; the non-local algorithm is combined with the convolutional neural network to directly denoise the reconstructed noisy CT reconstructed image, which can solve the problems of small noise and image feature difference, difficulty in stripping the reconstructed noise and image information, and texture disappearance after denoising.

Description

Technical Field

[0001] This invention relates to the field of CT image processing technology, and in particular to a CT image noise reduction method and system based on similar block learning. Background Technology

[0002] X-ray computed tomography (CT) is an imaging technique that uses X-ray projections from different angles to infer attenuation information of an object. Due to the strong penetrating power of X-rays and the intuitive three-dimensional image presentation, it is widely used in medical diagnosis, non-destructive testing, reverse engineering, materials analysis, security inspection, and many other fields. Nano-CT, to achieve nanoscale focal spot sizes, uses thin beams with high energy density, thus providing higher spatial resolution. However, due to the low-energy characteristics of the ultra-micro focal spot in nano-CT, the X-ray source power is limited, resulting in a weak photon condition during scanning, making it prone to photon starvation. Nano-CT inevitably generates noise, leading to decreased image contrast, reduced image quality, and increased noise. To improve the signal-to-noise ratio, the exposure time is generally increased, resulting in a very long scanning process. Therefore, researchers need to strike a trade-off between signal-to-noise ratio and efficiency.

[0003] A high-quality CT image involves not only the precision of machining and the performance of detector materials, but also the image post-processing. Seeking reconstruction algorithms to improve the signal-to-noise ratio of nano-CT images under weak photon conditions and addressing the efficiency issues of nano-CT in practical applications will help promote the development and application of laboratory nano-CT. Currently, methods for improving the quality of low-dose CT images can be approached from three aspects: (1) establishing a projection data restoration model based on the noise statistical characteristics of the projection data, and then using a filtered back projection (FBP) algorithm to reconstruct the restored projection data; (2) directly denoising the reconstructed low-dose CT image. Even after reconstructing the denoised projection data, reconstruction noise is still introduced. Summary of the Invention

[0004] To address this, the present invention provides a CT image denoising method and system based on similar block learning. It employs deep learning methods from artificial intelligence to directly denoise the reconstructed CT image with noise, thus resolving the issue that reconstruction noise still exists even after denoising and reconstruction of projection data.

[0005] According to the design scheme provided by this invention, a CT image noise reduction method based on similar block learning is provided, comprising:

[0006] Preprocessing of the collected X-ray tomography dataset based on similar block matching yields CT reconstructed image data with similar noise distributions.

[0007] CT reconstructed image data with similar noise distribution are input into a pre-trained CT image denoising network, which is then used to denoise the CT reconstructed image data.

[0008] As a CT image denoising method based on similar block learning in this invention, the preprocessing of the acquired X-ray tomography dataset based on similar block matching includes:

[0009] First, several reference image blocks are defined;

[0010] Then, a local neighborhood search is used to select the candidate image patch with the smallest difference from the reference image patch, and the selected candidate image patch is used to generate image data with the same properties.

[0011] As a CT image denoising method based on similarity block learning in this invention, further, the candidate image block with the smallest difference from the reference image block is selected through local neighborhood search, including:

[0012] First, set the reference image patch size, search step size, local search window size, and similarity threshold;

[0013] Then, based on the similarity between image patches, N candidate image patches that are most similar to each reference image patch are selected, and image data with the same properties and number of channels of N are constructed by returning the selected candidate image patches to their original positions, where N is an integer greater than 1.

[0014] As a CT image denoising method based on similarity block learning in this invention, further, based on the similarity between image blocks, N candidate image blocks that are most similar to each reference image block are selected, and Euclidean distance is used to measure the similarity between the current window image block and the reference image block. Window image blocks with similarity greater than the similarity threshold are added to the candidate image blocks until N candidate image blocks are selected.

[0015] As a CT image denoising method based on similar block learning in this invention, the CT image denoising network further includes: a feature extraction unit for extracting feature information of the input image and a feature learning unit for selecting effective image features by utilizing the residual channel attention mechanism to enhance the weights of key features.

[0016] As a CT image denoising method based on similar block learning according to the present invention, the pre-training process of the CT image denoising network further includes:

[0017] First, X-ray image data were collected, and a sample dataset containing pairs of noisy and noiseless CT image data was created by varying the scan exposure time.

[0018] Then, the network objective function optimizer parameters are set, and the CT image denoising network is trained using the sample dataset. The final pre-trained CT image denoising network is obtained by using the preset training termination condition.

[0019] As a CT image denoising method based on similar block learning in this invention, further, a sample dataset containing noisy and noise-free data pairs is created, and the method further includes: performing data augmentation operations on the scanned CT images, wherein the data augmentation operations include, but are not limited to: rotation and scaling of the images during the scanning process.

[0020] Furthermore, the present invention also provides a CT image denoising system based on similar block learning, comprising: a preprocessing module and a denoising processing module, wherein,

[0021] The preprocessing module is used to preprocess the collected X-ray tomography dataset based on similar block matching to obtain CT reconstructed image data with similar noise distribution;

[0022] The noise reduction module is used to input CT reconstructed image data with similar noise distribution into a pre-trained CT image denoising network, and to use the CT image denoising network to perform noise reduction processing on the CT reconstructed image data.

[0023] The beneficial effects of this invention are:

[0024] This invention utilizes image features within CT reconstructed image slices to fully learn noise distribution characteristics. It employs the output of a nonlocal method as a data preprocessing algorithm, reducing noise while preserving image details. Then, a network is used to learn a nonlocal method to recover the image residuals. This enables fast and effective noise removal from CT images with a relatively small number of parameters. Because the network has few parameters and no hyperparameters that need adjustment, it is easy to train and will not fall into overfitting due to small-scale training data. By combining the nonlocal algorithm with a convolutional neural network to directly denoise the reconstructed noisy CT image, it solves problems such as the small difference between noise and image features, the difficulty in separating noise from image information after reconstruction, and the loss of texture after denoising. Attached image description:

[0025] Figure 1 This is a schematic diagram of the CT image noise reduction process based on similar block learning in the embodiment;

[0026] Figure 2 This is a schematic diagram of CT image noise reduction network data preprocessing in the embodiment;

[0027] Figure 3 This is a schematic diagram of the CT image noise reduction network architecture in the embodiment;

[0028] Figure 4 This example illustrates a comparison of the denoising results of the CT image denoising network under different methods. Detailed implementation method:

[0029] To make the objectives, technical solutions, and advantages of this invention clearer and more understandable, the invention will be further described in detail below with reference to the accompanying drawings and technical solutions.

[0030] Besides the photon starvation effect caused by the low-energy characteristics of nano-CT, the quality of CT images is also affected by factors such as the non-ideality of the hardware system, the external environment, and the reconstruction algorithm. The noise contained in the final reconstructed CT image includes quantum noise, noise introduced by the reconstruction algorithm, and hardware noise. Because the noise distribution in the reconstructed CT image is uneven, the noise model is difficult to determine. Therefore, the simulated noisy image still differs from the actual scanned noisy image in terms of noise distribution characteristics. Regarding the characteristics of CT reconstructed images, see the embodiments of this invention. Figure 1 As shown, a CT image denoising method based on similarity block learning is provided, comprising:

[0031] S101. Based on similar block matching, the collected X-ray tomography dataset is preprocessed to obtain CT reconstructed image data with similar noise distribution.

[0032] In image patch processing, regions with similar properties within the same image are classified and weighted averaged to obtain a denoised image, which can improve its noise reduction effect.

[0033] As a preferred embodiment, further, the preprocessing of the collected X-ray tomography image dataset based on similar block matching may include: first, setting several reference image blocks; second, selecting candidate image blocks with the smallest difference from the reference image blocks through local neighborhood search, and using the selected candidate image blocks to generate image data with the same properties.

[0034] Similarity block matching is a method for finding signal blocks that are similar to a given reference image block. It is achieved by comparing the similarity between a pair of reference image blocks and candidate image blocks located at different spatial locations.

[0035] Furthermore, in this embodiment, selecting the candidate image block with the smallest difference from the reference image block through local neighborhood search includes: first, setting the reference image block size, search step size, local search window size, and similarity threshold; second, selecting N candidate image blocks most similar to each reference image block based on the similarity between image blocks, and constructing image data with the same properties and N channels by returning the selected candidate image blocks to their original positions, where N is an integer greater than 1.

[0036] The reference image patch is processed using a sliding window approach. By setting several reference patches, a search is conducted within the local neighborhood of each reference patch to find N patches with the smallest differences, such as... Figure 2 As shown, Euclidean distance can be used to measure the similarity between the current window image patch and the reference image patch. Window image patches with a similarity greater than the similarity threshold are added to the candidate image patches until N candidate image patches are selected.

[0037] Only blocks with similarity greater than a threshold are matched as similar blocks and then selected into a group; blocks with low similarity are ignored. The Euclidean distance formula can be expressed as follows:

[0038]

[0039] Where ρ is the Euclidean distance between points (x1, y1) and (x2, y2), a smaller distance implies a higher similarity.

[0040] In general, global search is too costly; therefore, in this embodiment, a non-local search is used. Similar images are aggregated into a complete image by setting the number of similar blocks to be searched, N. Blocks with an ideal distance less than a threshold are matched as similar blocks, while blocks with a larger ideal distance are ignored. Each reference block searches for N similar blocks, returning these similar blocks to their original positions to construct input data with N channels.

[0041] S102. Input CT reconstructed image data with similar noise distribution into a pre-trained CT image denoising network, and use the CT image denoising network to perform noise reduction processing on the image data.

[0042] In machine vision tasks, image patch-based processing methods, which classify regions with similar properties within the same image and then use a weighted average to obtain a denoised image, often yield better noise reduction results. Convolutional Neural Networks (CNNs) can directly process two-dimensional images and have been widely used in image processing. These networks are designed to extract more abstract features from the original image using simple nonlinear models, requiring minimal human intervention throughout the process.

[0043] Furthermore, in this embodiment, the CT image denoising network may include: a feature extraction unit for extracting feature information of the input image and a feature learning unit for selecting effective image features by utilizing the residual channel attention mechanism to enhance the weights of key features.

[0044] See Figure 3As shown, a nonlocal framework for CT image denoising is constructed by combining similarity block search with CNN, and BMNet (block match network) is used to denoise CT images. Each CAB (channel attention block) consists of multiple layers of residual convolutions and channel attention mechanisms. This module uses a residual structure with local skip connections and short skip connections. The emergence of residual networks solves the problem that gradient vanishing as network depth increases will prevent the network training parameters from being updated, thus achieving a denoising effect. The first part of the CAB includes the complete receptive field of the input features, and the features are learned through a merge operation; then, in order to select effective features, the weights of important features from the mapping are enhanced by the channel attention module. Finally, through network learning, a noise-free output result is obtained.

[0045] As a preferred embodiment, the CT image denoising network pre-training process further includes:

[0046] First, X-ray image data were collected, and a sample dataset containing pairs of noisy and noiseless CT image data was created by varying the scan exposure time.

[0047] Then, the network objective function optimizer parameters are set, and the CT image denoising network is trained using the sample dataset. The final pre-trained CT image denoising network is obtained by using the preset training termination condition.

[0048] A ZEISS Xradia 510Versa 3D X-ray scanner was used to scan bronze coins, and the reconstructed images were used as the training dataset for the network. During dataset creation, noisy and noise-free data pairs were obtained by varying the scan exposure time. During the scan, different exposure times were set to obtain normal and low-dose data. The long exposure time was set to 5 seconds, and the short exposure time to 0.1 seconds. The tube voltage was set to 140kV, the X-ray source power to 10W, the distance from the source to the rotation axis (SOD) to 80.01mm, and the distance from the source to the detector (SDD) to 200.01mm. Furthermore, the scanned CT images were augmented using data enhancement operations, including but not limited to image rotation and scaling during the scan. Flipping and rotating the scanned data expanded the training data, improving the network training and optimization effect.

[0049] It should be noted that the settings of various parameters during the scanning process in this embodiment are not within the scope of protection of this case. In actual use, the relevant parameters can be adjusted according to data requirements.

[0050] Furthermore, based on the above method, this embodiment of the invention also provides a CT image denoising system based on similar block learning, comprising: a preprocessing module and a denoising processing module, wherein,

[0051] The preprocessing module is used to preprocess the collected X-ray tomography dataset based on similar block matching to obtain CT reconstructed image data with similar noise distribution;

[0052] The noise reduction module is used to input CT reconstructed image data with similar noise distribution into a pre-trained CT image denoising network, and to use the CT image denoising network to perform noise reduction processing on the CT reconstructed image data.

[0053] To verify the effectiveness of this solution, the following explanation is based on experimental data:

[0054] To address the issue of noise affecting image quality during low-dose CT scanning, the experiment set the block size to 16×16, the search step size to 5, and the local search window size to 100×100. Euclidean distance was used to measure similarity, with a threshold of 2500. The number of similar blocks searched was set to 5, and similar images were aggregated into a complete image, thus obtaining training data with 5 input channels. The denoising model network architecture is as follows... Figure 3 As shown, feature extraction consists of a 3×3 convolution and a ReLU activation function, which extracts feature information from the noisy image. The extracted feature information is then learned by a four-residual-channel attention module. In subsequent noise feature processing, the acquired high-frequency features are enhanced to improve noise removal capabilities.

[0055] First, noisy CT images are preprocessed by associating global image information through similar block search. Second, the preprocessed data is input into the denoising model network for training. The training dataset contains 4728 training pairs and 2640 validation pairs.

[0056] The network was trained and tested using the PyTorch framework on an AMAX workstation with an Intel Xeon Gold 5118 CPU and 64GB of available memory. A GeForce RTX 2080 Ti GPU was used for both training and testing. The Adam optimizer was chosen as the objective function, with a learning rate starting from 0.0001. The momentum parameters β1 and β2 of the Adam optimizer were set to 0.9 and 0.999, respectively. The batch size was set to 24 for all network parameter training optimization experiments. The training process consisted of 80 iterations and took approximately 6 hours to train.

[0057] Two metrics widely used in CT denoising are compared and analyzed. Peak Signal-to-Noise Ratio (PSNR) evaluates the model's noise suppression capability by calculating the pixel difference between the generated image and the label image. Structural Similarity (SSIM) assesses the structural similarity between the generated image and the reference image based on the image's brightness, contrast, and structure. Better models aim for higher PSNR and SSIM to achieve more significant denoising performance. See also Figure 4 The image shows the denoising results for two different slices. The first column shows the corresponding noisy image; the second column shows the corresponding reference image; the third column shows the output of the comparison method CPCE; the fourth column shows the output of the comparison method CBDNet; and the fifth column shows the output of the proposed method. While both CPCE and CBDNet can suppress noise to some extent, the denoised outputs exhibit blurring and loss of some image information. Visual inspection shows that the reconstruction performance of the proposed denoising network is close to that of the reference image, effectively removing noise from CT images.

[0058] Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0059] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0060] The units and method steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations are not considered to be beyond the scope of this invention.

[0061] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. Optionally, all or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module / unit in the above embodiments can be implemented in hardware or as a software functional module. This invention is not limited to any particular combination of hardware and software.

[0062] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, 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, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A CT image denoising method based on similarity block learning, characterized in that, Include: The acquired X-ray tomography dataset is preprocessed based on similar block matching to obtain CT reconstructed image data with similar noise distribution. The preprocessing process includes: first, setting several reference image blocks; then, setting the reference image block size, search step size, local search window size, and similarity threshold; using Euclidean distance to measure the similarity between the current window image block and the reference image block; adding window image blocks with similarity greater than the similarity threshold to the candidate image blocks; until N candidate image blocks are selected; and constructing image data with N channels and similar noise distribution by returning the selected candidate image blocks to their original positions, where N is an integer greater than 1. CT reconstructed image data with similar noise distributions are input into a pre-trained CT image denoising network. The CT image denoising network is used to denoise the CT reconstructed image data. The CT image denoising network includes: a feature extraction unit for extracting feature information of the input image and a feature learning unit for selecting effective image features by using the residual channel attention mechanism to enhance the weight of key features. The feature extraction unit comprises a 3×3 convolution and a ReLU activation function, and the feature learning unit comprises four residual channel attention modules; wherein, the first part of the residual channel attention module comprises the complete receptive field of the input features, and the features are learned by merging and running, the weights of the key features from the mapping are enhanced by the channel attention modules, and finally the network learns to obtain a noise-free output result; The pre-training process of the CT image denoising network includes: First, X-ray tomography data were collected, and a sample dataset containing pairs of noisy and noiseless CT image data was created by varying the scan exposure time. Then, the network objective function optimizer parameters are set, and the CT image denoising network is trained using the sample dataset. The final pre-trained CT image denoising network is obtained by using the preset training termination condition.

2. The CT image denoising method based on similar block learning according to claim 1, characterized in that, Creating a sample dataset containing noisy and noise-free data pairs, and further including: performing augmentation processing on scanned CT images through data augmentation operations, wherein the data augmentation operations include, but are not limited to: rotation and scaling of images during the scanning process.

3. A CT image denoising system based on similar block learning, characterized in that, The method described in claim 1 includes: a preprocessing module and a noise reduction module, wherein... The preprocessing module is used to preprocess the acquired X-ray tomography dataset based on similar block matching to obtain CT reconstructed image data with similar noise distribution. The preprocessing process includes: first, setting several reference image blocks; then, setting the reference image block size, search step size, local search window size, and similarity threshold; using Euclidean distance to measure the similarity between the current window image block and the reference image block; adding window image blocks with similarity greater than the similarity threshold to the candidate image blocks; until N candidate image blocks are selected; and constructing image data with N channels and similar noise distribution by returning the selected candidate image blocks to their original positions, where N is an integer greater than 1. The noise reduction module is used to input CT reconstructed image data with similar noise distribution into a pre-trained CT image denoising network, and to use the CT image denoising network to perform noise reduction on the CT reconstructed image data. The CT image denoising network includes: a feature extraction unit for extracting feature information of the input image and a feature learning unit for using the residual channel attention mechanism to enhance the weight of key features to select effective features of the image. The feature extraction unit comprises a 3×3 convolution and a ReLU activation function, and the feature learning unit comprises four residual channel attention modules; wherein, the first part of the residual channel attention module comprises the complete receptive field of the input features, and the features are learned by merging and running, the weights of the key features from the mapping are enhanced by the channel attention modules, and finally the network learns to obtain a noise-free output result; The pre-training process of the CT image denoising network includes: First, X-ray tomography data were collected, and a sample dataset containing pairs of noisy and noiseless CT image data was created by varying the scan exposure time. Then, the network objective function optimizer parameters are set, and the CT image denoising network is trained using the sample dataset. The final pre-trained CT image denoising network is obtained by using the preset training termination condition.

4. An electronic device, characterized in that, The system includes a memory and a processor, which communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can execute the steps of the method as described in any one of claims 1 to 2 by calling the program instructions.

5. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the method described in any one of claims 1 to 2.

Images