A method for low-dose CT image reconstruction
By using unsupervised learning algorithms and self-supervised image processing networks, the problems of noise adaptation and image quality degradation in low-dose CT image reconstruction are solved, achieving adaptive denoising and image quality improvement, which is suitable for low-dose CT image reconstruction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN YANGQI MEDICAL CORE INTELLIGENT TECH CO LTD
- Filing Date
- 2022-09-23
- Publication Date
- 2026-05-26
AI Technical Summary
Existing low-dose CT image reconstruction techniques cannot adapt to the noise characteristics of different CT equipment and scanning parameters, resulting in decreased image quality and affecting the accuracy of clinical diagnosis. Furthermore, existing algorithms struggle to accommodate different noise ratios in unsupervised learning.
An unsupervised learning algorithm is adopted. By randomly segmenting and erasing image blocks, combined with controllable modules and multi-head attention feature extraction, a self-supervised image processing network is designed. The VGG-16 network is used to extract texture features and adjust the noise reduction ratio to achieve adaptive noise reduction.
It enables adaptive adjustment of the denoising ratio to improve image signal-to-noise ratio and detail texture without the need for high-dose CT data, thus avoiding medical ethics issues and improving the accuracy of clinical diagnosis.
Smart Images
Figure CN115953487B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing technology, and specifically relates to a method for low-dose CT image reconstruction. Background Technology
[0002] Computed tomography (CT) technology is widely used in clinical screening, diagnostic imaging, and image-guided radiotherapy due to its ease of operation and low cost. However, excessive CT doses can induce leukemia, cancer, and other hereditary diseases. Therefore, reducing the CT dose in practical applications mitigates radiation risks. However, reducing the radiation dose can lead to severe artifacts and noise contamination in the images, resulting in reduced edge and corner contrast and a decline in CT image quality, seriously affecting the accuracy of downstream medical image processing tasks and clinical diagnosis. To address these issues, existing low-dose CT image denoising techniques mainly include traditional filtering methods, deep supervised learning denoising methods, and deep unsupervised learning denoising methods, each with its own advantages and disadvantages:
[0003] 1. Traditional Filtering Method: Based on manually designed convolutional kernels, this method performs a single convolution operation on the input image to achieve a filtering effect. The size, parameters, and stride of the convolutional kernel are fixed. Its problems include: because the parameters of the convolutional kernel, such as size and stride, are fixed due to the quantitative data, it can only adapt to the noise characteristics of the acquired data. It cannot adapt to the noise characteristics of images obtained from other CT devices or under different scanning parameters, thus lacking widespread applicability and scalability.
[0004] 2. Deep Supervised Learning Denoising Method: This method uses convolutional neural networks or Tranformer models to construct low-dose CT denoising algorithms. Supervised model training is performed using simultaneously acquired low- and high-dose paired images (or only high-dose images are acquired, and the algorithm simulates low-dose noise to modify the high-dose image, thus obtaining a low-dose image). The network model employs an encoder-decoder structure. The encoder network extracts important feature vectors (e.g., edges, texture, content), while the decoder network reconstructs the image from the encoded feature vectors. Problems exist: In practical clinical applications, it is impossible to directly acquire low- and high-dose paired images simultaneously (medical ethics issues), and simulating low-dose noise characteristics cannot completely fit the real low-dose noise characteristics. Therefore, supervised deep learning algorithms are difficult to promote and apply in clinical practice.
[0005] 3. Deep Unsupervised Learning Denoising Method: The network structure design in this method is not significantly different from that of supervised learning. The main difference lies in the acquisition of supervisory information. Unsupervised learning algorithms typically only require the low-dose image to be denoised. They construct an algorithm, such as a similarity test, and calculate the similarity with one or more reconstructed low-dose images. The similarity test algorithm is related to the image evaluation method. After obtaining the similarity result, it is used as a loss for backpropagation to update the network parameters. The problem is that current low-dose CT image denoising methods cannot simultaneously address both unsupervised learning algorithms and denoising different noise levels. In clinical use, different locations, tube currents, and radiation times result in varying noise levels.
[0006] In summary, the existing low-dose CT denoising algorithms mentioned above usually filter out noise in high-frequency components, which leads to excessive filtering of edge and detail texture information, resulting in a decrease in the sharpness of the reconstructed image (i.e., a decrease in the signal-to-noise ratio), causing additional impact on clinical diagnosis. Summary of the Invention
[0007] In view of this, the purpose of the present invention is to provide a low-dose CT image reconstruction method.
[0008] The objective of this invention can be achieved through the following technical solution: a low-dose CT image reconstruction method, characterized by comprising the following steps:
[0009] Step A: Obtain two different low-dose, high-noise CT image data, one as input image I and the other as self-supervised image T;
[0010] Step B: Randomly select one of the four ratios: 3x3, 7x7, 11x11, and 15x15, and randomly divide the input image I into equally proportioned image blocks;
[0011] Step C: Randomly erase 60% of the image blocks in Step B, randomly sort the remaining 40% of the image blocks, and input the position information of the sorted image block sequence into the algorithm model.
[0012] Step D: After reconstruction using the algorithm model, the reconstructed image is obtained. ;right and Calculate the L1 loss.
[0013] In the aforementioned low-dose CT image reconstruction method, 60% of the image blocks are erased using salt-and-pepper noise or Gaussian white noise.
[0014] The low-dose CT image reconstruction method described above also includes the step of setting a controllable module to adjust different noise reduction ratios, including: increasing the input noise reduction ratio factor β, where β ranges from 0 to 1, 0 indicates no noise reduction at all, and 1 indicates the maximum noise reduction level.
[0015] In the aforementioned low-dose CT image reconstruction method, the process of setting a controllable module to adjust different noise reduction ratios further includes: after inputting the β parameter into the algorithm model, it is transformed through a two-layer fully connected neural network. The first fully connected layer contains 20 neurons, which are then activated by Leaky-ReLU nonlinearity, followed by a second fully connected layer containing 20 neurons, which are then activated by Leaky-ReLU nonlinearity, and finally output through a fully connected layer containing 9 neurons. The output is then normalized by Sigmoid. The normalized 9 fully connected layers are reset into a 3x3 matrix as the convolution kernel for adjusting the noise reduction ratio. This convolution kernel is then convolved with four layers of feature maps of different resolutions extracted by multi-head attention features to finally obtain a feature vector that has been modulated by the noise reduction ratio.
[0016] In the above-mentioned low-dose CT image reconstruction method, the controllable module setting also includes its learning process: Step 1: Set β to 0 so that the feature extraction network can fully extract CT image features, and the self-supervised image in this step is the original image; Step 2: Set β to 1 so that the feature extraction network learns the maximum proportion of noise reduction features, and the self-supervised image in this step is other images.
[0017] In the aforementioned low-dose CT image reconstruction method, the steps further include setting a fine-grained texture extraction step:
[0018] Step E1: Extract four downsampled feature maps of different resolutions from the input low-dose image;
[0019] Step E2: Extract four downsampled feature maps from the self-supervised data, and calculate four L1 losses for each of the four pairs of feature maps.
[0020] Step E3: Sum the four losses and calculate the average value;
[0021] Step E4: Take a weighted average of the average loss and the unsupervised L1 loss:
[0022] In the low-dose CT image reconstruction method described above, the downsampled feature map in step a is obtained by extracting it using a VGG-16 ImageNet pre-trained network.
[0023] In the aforementioned low-dose CT image reconstruction method, the algorithm model is trained through the following steps:
[0024] Step F1: Obtain low-dose CT DICOM data 1; parse the internal pixel data and convert it into an image matrix; the size specification is 512 * 512.
[0025] Step F2: Set the noise reduction scaling factor β to 0, and use data 1 as the supervision data;
[0026] Step F3: Divide the input data into image blocks using a random ratio, and erase 60% of the image blocks using another random variable; encode the remaining 40% of the image blocks into a matrix, and obtain the image block position codes corresponding to the matrix.
[0027] Step F4: Encode the original image patches that were not erased after segmentation in step F3, and use them as input to the first attention module of the model decoder;
[0028] Step F5: Simultaneously train the feature parameters of the input encoding through an attention encoder-decoder network and a VGG-16 texture extraction network; update the network parameters through backpropagation by iteratively calculating the loss.
[0029] Step F6: During the second iteration training process, set the noise reduction scaling factor β to 1, and randomly sample the supervision data from the dataset again; complete the second iteration training in the manner of steps F3-F5.
[0030] Step F7: The final network is used for prediction.
[0031] Compared with existing technologies, this low-dose CT image reconstruction method has the following advantages:
[0032] 1. It adopts an unsupervised algorithm, which does not require high-dose CT supervision data, completely avoiding medical ethics issues; and it can continuously combine existing data to iterate the model offline, continuously optimizing the denoising effect of the algorithm on real CT data.
[0033] 2. By designing a controllable adjustable parameter network and convolution module, the algorithm can control the noise reduction ratio by adjusting parameter values in practical scenarios, which meets the needs of doctors.
[0034] 3. When examining lesions using CT images, the signal-to-noise ratio is crucial. Improving texture reproduction and signal-to-noise ratio can significantly reduce the risk of misdiagnosis or uncertainty for doctors. Attached Figure Description
[0035] Figure 1 This is a diagram of the neural network structure corresponding to the reconstruction method in the embodiment. Detailed Implementation
[0036] The following are specific embodiments of the present invention, which are described in conjunction with the accompanying drawings. However, the present invention is not limited to these embodiments. Example
[0037] Combination Figure 1 This low-dose CT image reconstruction method is implemented using a neural network. The neural network structure specifically includes three parts: a controllable module, an image processing part, and a self-supervised image processing part.
[0038] The specific steps of low-dose CT image reconstruction are as follows:
[0039] Step A: Obtain two different low-dose, high-noise CT image data, one as input image I and the other as self-supervised image T;
[0040] Step B: Randomly select one of the four ratios: 3x3, 7x7, 11x11, and 15x15, and randomly divide the input image into equally sized image blocks; the four ratios can be set according to actual needs and are not specifically limited.
[0041] Step C: Randomly erase 60% of the image blocks in Step B using salt and pepper noise or Gaussian white noise, randomly sort the remaining 40% of the image blocks, and input the position information of the sorted image block sequence into the algorithm model.
[0042] The algorithm model in this step is obtained through the following steps:
[0043] F1. Obtain low-dose CT DICOM data 1; parse the internal pixel data and convert it into an image matrix; the image matrix size is specified as 512*512.
[0044] Step F2: Set the noise reduction scaling factor β to 0, and use data 1 as the supervision data. The noise reduction scaling factor β ranges from 0 to 1, where 0 represents no noise reduction and 1 represents the maximum level of noise reduction. Figure 1 As shown, after inputting the β parameter, the image is transformed through a two-layer fully connected neural network. The first fully connected layer contains 20 neurons, which are then activated by Leaky-ReLU nonlinearity. This is followed by a second fully connected layer containing 20 neurons, again activated by Leaky-ReLU, and finally output through a fully connected layer containing 9 neurons. The output is then normalized using Sigmoid. The normalized 9 fully connected layers are then reconfigured as a 3x3 matrix, serving as the convolution kernel to adjust the noise reduction ratio. This convolution kernel is then convolved with four layers of feature maps at different resolutions extracted by multi-head attention, resulting in a feature vector that has been modulated for noise reduction. Finally, the image is decoded using a stacked multi-head attention module to obtain the denoised output CT image.
[0045] Step F3: Divide the input data into image blocks using a random ratio, and erase 60% of the image blocks using another random variable; encode the remaining 40% of the image blocks into a matrix, and obtain the image block position codes corresponding to the matrix.
[0046] Step F4: Encode the original image patches that were not erased after segmentation in step F3, and use them as input to the first attention module of the model decoder;
[0047] Step F5: Simultaneously train the feature parameters of the input encoding through an attention encoder-decoder network and a VGG-16 texture extraction network; update the network parameters through backpropagation by iteratively calculating the loss.
[0048] Step F6: During the second iteration training process, set the noise reduction scaling factor β to 1, and randomly sample the supervision data from the dataset again; complete the second iteration training in the manner of steps F3-F5.
[0049] Step F7: The final network is used for prediction.
[0050] When using the model, load the network parameters and adjust the noise reduction scaling factor β as needed. Set β to 0 so that the feature extraction network can fully extract CT image features; the self-supervised image in this first step is the original image. Set β to 1 so that the feature extraction network learns the maximum scale noise reduction features; the self-supervised image in this second step is another image.
[0051] Step D: After reconstruction using the algorithm model, the reconstructed image is obtained. ;right and The L1 loss is calculated using the following formula:
[0052] Among them, for Batch size within the current training iteration cycle, The first in the current batch Zhang image.
[0053] Combination Figure 1 To improve the detail and texture information of reconstructed CT images and increase the image signal-to-noise ratio, this paper proposes to calculate texture loss as additional loss information by using the output of the downsampling feature extraction layer with four resolution dimensions. Specific method:
[0054] E1. The VGG-16 ImageNet pre-trained network is used to extract four downsampled feature maps of different resolutions from the input low-dose image;
[0055] E2. Similarly, extract four downsampled feature maps from the self-supervised data, and calculate four L1 losses for each of the four pairs of feature maps.
[0056] E3. Sum the four losses and calculate the average value.
[0057] E4. Take a weighted average of the average loss and the unsupervised L1 loss: .
[0058] The low-dose CT image reconstruction method of the present invention can still improve the image signal-to-noise ratio and detail texture during the denoising process, and can actively adjust the noise reduction ratio, which is practical.
[0059] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.
Claims
1. A method for low-dose CT image reconstruction, characterized in that, Includes the following steps: Step A: Obtain two different low-dose, high-noise CT image data, one as input image I and the other as self-supervised image T; Step B: Input image The image is randomly divided into proportionally sized blocks. Step C: Randomly erase 60% of the image blocks in Step B, randomly sort the remaining 40% of the image blocks, and input the position information of the sorted image block sequence into the algorithm model. Step D: After reconstruction using the algorithm model, the reconstructed image is obtained. ;right and Calculate L1 loss; It also includes the step of setting a controllable module to adjust different noise reduction ratios, including: increasing the input noise reduction ratio factor β, where β ranges from 0 to 1, 0 indicates no noise reduction at all, and 1 indicates the maximum noise reduction level; The noise reduction scaling factor β learning process consists of two steps: Step 1: Set β to 0 so that the feature extraction network can fully extract CT image features. The self-supervised image in this step is the original image. Step 2: Set β to 1 so that the feature extraction network learns the noise reduction features with the maximum proportion. The self-supervised image in this second step is another image. The algorithm model is obtained through the following steps: Step F1: Obtain low-dose CT DICOM data 1; parse the internal pixel data and convert it into an image matrix; Step F2: Set the noise reduction scaling factor β to 0, and use data 1 as the supervision data; Step F3: Divide the input data into image blocks using a random ratio, and erase 60% of the image blocks using another random variable; encode the remaining 40% of the image blocks into a matrix, and obtain the image block position codes corresponding to the matrix. Step F4: Encode the original image patches that were not erased after segmentation in step F3, and use them as input to the first attention module of the model decoder; Step F5: Simultaneously train the feature parameters of the input encoding through an attention encoder-decoder network and a VGG-16 texture extraction network; update the network parameters through backpropagation by iteratively calculating the loss. Step F6: During the second iteration training process, set the noise reduction scaling factor β to 1, and randomly sample the supervision data from the dataset again; complete the second iteration training in the manner of steps F3-F5. Step F7: Use the final network for prediction.
2. The low-dose CT image reconstruction method according to claim 1, characterized in that, The 60% of image blocks were removed using salt-and-pepper noise or Gaussian white noise.
3. The low-dose CT image reconstruction method according to claim 1 or 2, characterized in that, The process of setting the controllable module to adjust different noise reduction ratios further includes: after inputting the β parameter into the algorithm model, it is transformed through a 2-layer fully connected neural network with hidden layers. The first fully connected layer contains 20 neurons, which are then activated by Leaky-ReLU nonlinearity. After that, it is activated by a second fully connected layer containing 20 neurons, which are then activated by Leaky-ReLU nonlinearity. Finally, it is output through a fully connected layer containing 9 neurons, and the output is normalized by Sigmoid. The normalized 9 fully connected layers are reset into a 3x3 matrix as the convolution kernel for adjusting the noise reduction ratio. The convolution kernel is then convolved with the 4 layers of feature maps with different resolutions extracted by multi-head attention, and the final feature vector is the vector after the noise reduction ratio is modulated.
4. The low-dose CT image reconstruction method according to claim 1 or 2, characterized in that, The steps also include setting a fine-grained texture extraction step: Step E1: Extract four downsampled feature maps of different resolutions from the input low-dose image; Step E2: Extract four downsampled feature maps from the self-supervised data, and calculate four L1 losses for each of the four pairs of feature maps. Step E3: Sum the four losses and calculate the average value; Step E4: Perform a weighted average of the average loss and the unsupervised L1 loss.
5. The low-dose CT image reconstruction method according to claim 4, characterized in that, The downsampled feature map in step E1 is obtained by extracting it using a VGG-16 ImageNet pre-trained network.
6. The low-dose CT image reconstruction method according to claim 1 or 2, characterized in that, The image matrix is specified to be 512 * 512 pixels.