Low-dose CT (Computed Tomography) denoising method and device based on multi-level feature perception

Abstract

The invention relates to a low-dose computed tomography (CT) denoising method and device based on multi-level feature perception, and the method comprises the steps: S1, collecting a low-dose CT picture and a high-dose CT picture, constructing a CT data set, and converting the CT data set into NPY matrix data; s2, performing multi-branch local-global feature extraction on the NPY matrix data to obtain shallow features; s3, attention information enhancement is carried out on the shallow layer features to obtain middle layer features; s4, performing feature extraction on the middle-layer features to obtain deep-layer features; s5, performing feature fusion on the shallow, middle and deep information to obtain fusion features; and S6, merging the fusion features and the NPY matrix data to obtain a CT de-noised image. By using the CT image denoising method and the CT image denoising device, a CT image denoising service which is more in line with a human eye observation effect can be provided for a low-dose CT obtained by a user through a CT scanner. The multi-level feature perception low-dose CT denoising method and device can be widely applied to the field of CT denoising.

Description

Technical Field

[0001] The present invention relates to the technical field of image denoising, and in particular to a low-dose CT image denoising method and device. Background Art

[0002] In the field of medical imaging, computed tomography (CT) plays a very important role in modern clinical medicine. However, X-rays may cause irreversible damage to the human body, including increasing the risk of cancer. When performing X-ray scans, medical staff will formulate strategies to reduce the X-ray dose based on actual conditions. For patients, lower X-ray intensity means lower cancer risk. But for medical staff, the decline in CT images caused by reduced X-ray intensity means the need for more accurate diagnosis. Therefore, it is crucial to improve LDCT images as much as possible while reducing X-ray doses.

[0003] Existing LDCT image reconstruction methods are mainly divided into three categories: denoising based on sinusoidal filtering, iterative reconstruction denoising, and image post-processing denoising. The first two methods require access to the difficult-to-obtain CT raw sinusoidal data, which limits their application. Denoising methods based on image post-processing are mainly divided into two categories: traditional machine learning and deep learning (DL). In recent years, many advances have been made in the field of image processing through the use of deep learning (DL), and the results achieved are far superior to traditional machine learning algorithms in many aspects.

[0004] In recent years, many studies on low-dose CT reconstruction have shown that neural networks trained with attention mechanisms can enhance tissue edge texture, but usually only a single attention mechanism is used, and there is no obvious improvement in obtaining attention information.

[0005] The enhancement network of low-dose CT images usually improves the reconstruction effect by increasing the network depth. Although it can continue to improve the PSNR value, it does not fully utilize the shallow, medium and deep layer information.

[0006] In summary, although the quantitative indicators PSNR and SSIM of CT images reconstructed by neural networks in the existing technology are high, there are problems such as single CT image features obtained by the network, insufficient attention information, and insufficient utilization of shallow, medium and deep layer information. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a low-dose CT denoising method and device based on multi-level feature perception, which can provide users with a CT image denoising service that is more in line with the observation effect of the human eye.

[0008] Specifically, the method comprises the following steps:

[0009] S1: Collect low-dose CT images and high-dose CT images to construct CT data sets and convert them into NPY matrix data;

[0010] S2: Perform multi-branch local-global feature extraction on NPY matrix data to obtain shallow features;

[0011] S3: Strengthen the attention information of shallow features to obtain middle-level features;

[0012] S4: Extract the middle-level features to obtain deep-level features;

[0013] S5: perform feature fusion on shallow, medium and deep information according to the feature fusion module to obtain fused features;

[0014] S6: Combine the fused features with the NPY matrix data to obtain the CT denoised image.

[0015] Preferably, S1 comprises the following steps:

[0016] S1.1: Acquire CT raw data in dicom format containing normal dose and low dose;

[0017] S1.2: Label the normal-dose CT in the CT raw data as “target” and the low-dose CT as “input”;

[0018] S1.3: Convert the labeled normal-dose CT and low-dose CT in dicom format into NPY matrix format data.

[0019] Preferably, S2 comprises the following steps:

[0020] S2.1: Construct a multi-branch local-global feature extraction module Local and Global Feature Extractor (LGFE);

[0021] S2.2: According to the LGFE module, local feature extraction and calculation are performed on the NPY data to obtain the local feature f e ;

[0022] S2.3: Perform global feature extraction and calculation on NPY data according to the LGFE module to obtain the global feature f g ;

[0023] S2.4: Merge local features and global features to obtain shallow features f sp .

[0024] Preferably, S2.1 comprises the following steps:

[0025] Using the multi-branch feature extraction results, convolution calculation is performed to obtain the LGFE structure:

[0026]

[0027] f e =Relu(Conv 3×3 (Lso(x))),

[0028]

[0029] In the above formula, Relu(.) represents the Relu activation function, Conv n×n (.) represents a convolution kernel of size n, represents the splicing operation, Lso(.) represents the edge detection algorithm, and maxpooling(.) represents the maximum pooling.

[0030] Preferably, S3 comprises the following steps:

[0031] S3.1: Build the Repeated Channel Spatial Reinforcing Attention (RCSRA) module;

[0032] S3.2: According to the RCSRA module, the shallow feature f m Strengthen the attention feature and obtain the middle-level feature f m ;

[0033] Preferably, S3.1 comprises the following steps:

[0034] The convolution calculation is performed using the method of attention information enhancement-feature enhancement to obtain the RCSRA structure:

[0035] MLP(.)=Conv(Relu(Conv(.))),

[0036]

[0037]

[0038]

[0039] In the above formula, C n represents the channel attention mechanism, S n represents the spatial attention mechanism, f n represents the convolution calculation result of the nth stage, w n Represents the input of the nth stage convolution calculation.

[0040] Preferably, S4 comprises the following steps:

[0041] f d=Relu(Conv(Relu(Conv(f m )) ×3 )),

[0042] In the above formula, the subscript ×3 represents that the operation is performed 3 times.

[0043] Preferably, S5 comprises the following steps:

[0044] S5.1: Construct shallow-middle-deep Feature Fusion Module (SMDF);

[0045] S5.2: The shallow feature f s , middle-level features f m and deep features f d Input SMDF and get the fusion features.

[0046] Preferably, S5.2 comprises the following steps:

[0047] Different convolution calculation methods are used for shallow features, middle features and deep features to obtain the SMDF structure:

[0048] f s =Conv 1×1 (f sp ),

[0049] f A =Conv 3×3 (f d ),

[0050]

[0051] Among them, f s represents the shallow perceptual features obtained by convolution calculation of shallow features, f A Represents the deep enhanced features obtained after convolution calculation of deep features, Represents fusion features.

[0052] Preferably, S6 comprises the following steps:

[0053] S6.1: merge the fusion features and NPY matrix;

[0054] S6.2: Obtain a denoising model for a low-dose CT denoising method based on multi-level features;

[0055] S6.3: Evaluate the denoising effect of the model based on the calculated quantitative indicators PSNR, SSIM and qualitative human eye visual effect.

[0056] Preferably, the step of merging the fusion feature and the NPY matrix specifically includes:

[0057]

[0058] In the above formula, Represents the denoised image.

[0059] The second technical solution adopted by the present invention is: a low-dose CT denoising device based on multi-level features, comprising:

[0060] The preprocessing module is used to collect data, label the data and convert it into NPY matrix;

[0061] LGFE module, used to convert NPY matrix data into shallow features;

[0062] RCSRA module is used to transform shallow features into mid-level features after strengthening attention features, and further transform them into deep features;

[0063] The SMD module is used to fuse shallow features, middle features, and deep features to obtain fused features;

[0064] The image reconstruction module is used to merge the fusion features and the NPY matrix to generate a reconstructed CT image.

[0065] The beneficial effects of the method and device of the present invention are as follows: the present invention first extracts shallow features by performing multi-branch local-global feature extraction on the NPY matrix, then performs attention information enhancement on the shallow features to obtain middle-level features, and obtains deep features by convolution calculation, and finally establishes feature connections between shallow features, middle-level features and deep features through feature fusion, so as to achieve the purpose of improving the CT human eye observation effect. At the same time, the present invention can effectively cope with the common challenges of existing neural network models in CT denoising. On the one hand, the multi-branch feature extraction is used to effectively promote the feature extraction richness of the model for CT images; on the other hand, the attention enhancement method can effectively improve the model's attention to pixel-level features and alleviate data redundancy pressure; at the same time, the feature fusion method is used to improve the model's utilization efficiency of features at each stage, and finally improve the visualization effect of the denoised CT image. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the prior art and the drawings required for use in the embodiments. The following drawings are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0067] Figure 1It is a flow chart of a low-dose CT denoising method based on multi-level feature perception of the present invention;

[0068] Figure 2 It is a schematic diagram of a specific embodiment of a low-dose CT denoising method based on multi-level feature perception of the present invention;

[0069] Figure 3 It is a schematic diagram of the LGFE module structure of a low-dose CT denoising method based on multi-level feature perception of the present invention;

[0070] Figure 4 It is a schematic diagram of the RCSRA module structure of a low-dose CT denoising method based on multi-level feature perception of the present invention;

[0071] Figure 5 It is a schematic diagram of the SMDF module structure of a low-dose CT denoising method based on multi-level feature perception of the present invention;

[0072] Figure 6 It is a comparison chart of the CT denoising results of the animal data set of the specific embodiment of the present invention and the traditional machine learning algorithm and deep learning network. Specific implementation plan

[0073] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all other embodiments obtained by ordinary technicians in this field without creative work based on the embodiments of the present invention are within the scope of protection of the present invention.

[0074] The embodiments of the present invention provide a low-dose CT denoising method and device based on multi-level feature perception, which are used to reduce the noise level caused by photon starvation in CT at the algorithm level, thereby improving the visualization effect of low-dose CT.

[0075] A typical embodiment of the present invention, taking an animal data set as an example, refers to Figure 1 , Figure 2 , the method comprises the following steps:

[0076] S1: Collect low-dose CT images and high-dose CT images to construct CT data sets and convert them into NPY matrix data;

[0077] S2, perform multi-branch local-global feature extraction on NPY matrix data to obtain shallow features;

[0078] S3, strengthen the attention information of shallow features to obtain middle-level features;

[0079] S4, extracting the middle-level features to obtain deep-level features;

[0080] S5, performing feature fusion on shallow, medium and deep information according to the feature fusion module to obtain fusion features;

[0081] S6. Merge the fused features with the NPY matrix data to obtain a CT denoised image.

[0082] Further, as a preferred embodiment of the method, S1 comprises the following steps:

[0083] S1.1: Acquire CT raw data in dicom format containing normal dose and low dose;

[0084] Further, the step of acquiring the DICOM format CT raw data including normal dose and low dose specifically includes:

[0085] Animal data sets were acquired at a source voltage of 100 kVp and a slice thickness of 0.625 mm, with tube current adjusted from 300 mAs down to 15 mAs.

[0086] S1.2. Label the normal-dose CT in the CT raw data with the label “target” and the low-dose CT with the label “input”;

[0087] Furthermore, the step of marking the normal-dose CT in the CT raw data with a "target" label and the low-dose CT with an "input" label specifically includes:

[0088] The CT images acquired under the imaging conditions of tube current 300mAs and source voltage 100kVp were labeled with the label “target”;

[0089] The CT images acquired under the imaging conditions of tube current 15mAs and source voltage 100kVp are labeled with the label “input”;

[0090] Convert normal-dose CT and low-dose CT in dicom format with labels into NPY matrix format data.

[0091] S1.3: Convert the labeled normal-dose CT and low-dose CT in dicom format into NPY matrix format data.

[0092] Further, refer to Figure 3 , S2 includes the following steps:

[0093] S2.1: Construct a multi-branch local-global feature extraction module Local and Global Feature Extractor (LGFE);

[0094] S2.2: According to the LGFE module, local feature extraction and calculation are performed on the NPY data to obtain the local feature f e ;

[0095] S2.3: Perform global feature extraction and calculation on NPY data according to the LGFE module to obtain the global feature f g ;

[0096] S2.4: Merge local features and global features to obtain shallow features f sp .

[0097] Further, S2.1 includes the following steps:

[0098] Using the multi-branch feature extraction results, convolution calculation is performed to obtain the LGFE structure:

[0099]

[0100] f e =Relu(Conv 3×3 (Lso(x))),

[0101]

[0102] Among them, Relu(.) represents the Relu activation function, Conv n×n (.) represents a convolution kernel of size n, represents the splicing operation, Lso(.) represents the edge detection algorithm, and maxpooling(.) represents the maximum pooling.

[0103] Further, refer to Figure 4 , S3 includes the following steps:

[0104] S3.1: Build the Repeated Channel Spatial Reinforcing Attention (RCSRA) module;

[0105] S3.2: According to the RCSRA module, the shallow feature f m Strengthen the attention feature and obtain the middle-level feature f m ;

[0106] Further, S3.1 includes the following steps:

[0107] The convolution calculation is performed using the method of attention information enhancement-feature enhancement to obtain the RCSRA structure:

[0108] MLP(.)=Conv(Relu(Conv(.))),

[0109]

[0110]

[0111]

[0112] Among them, C n represents the channel attention mechanism, S n represents the spatial attention mechanism, f n represents the convolution calculation result of the nth stage, w n Represents the input of the nth stage convolution calculation.

[0113] Further, S4 comprises the following steps:

[0114] f d =Relu(Conv(Relu(Conv(f m )) ×3 )),

[0115] In the above formula, the subscript ×3 represents that the operation is performed 3 times.

[0116] Further, refer to Figure 5 , S5 comprises the following steps:

[0117] S5.1: Construct shallow-middle-deep Feature Fusion Module (SMDF);

[0118] S5.2: The shallow feature f s , middle-level features f m and deep features f d Input SMDF and get the fusion features.

[0119] Further, S5.1 includes the following steps:

[0120] Different convolution calculation methods are used for shallow features, middle features and deep features to obtain the SMDF structure:

[0121] f s =Conv 1×1 (f sp ),

[0122] f A =Conv 3×3 (f d ),

[0123]

[0124] Among them, f srepresents the shallow perceptual features obtained by convolution calculation of shallow features, f A Represents the deep enhanced features obtained after convolution calculation of deep features, Represents fusion features.

[0125] Further, S6 comprises the following steps:

[0126] S6.1: merge the fusion features and NPY matrix;

[0127] S6.2: Obtain a denoising model for a low-dose CT denoising method based on multi-level features;

[0128] S6.3: Evaluate the denoising effect of the model based on the calculated quantitative indicators PSNR, SSIM and qualitative human eye visual effect.

[0129] Furthermore, the step of merging the fusion features and the NPY matrix specifically includes:

[0130]

[0131] In the above formula, Represents the denoised image.

[0132] Reference Figure 6 , which is the result of the denoising CT comparison test between this example and the traditional method on the animal data set. The experiment selected the same CT slices with rich tissues for comparison.

[0133] The comparative tests were all carried out using an Ubuntu 20.04 64-bit computer equipped with an Intel(R) Core(TM) i9-10900X CPU@3.70GHz, 32G RAM, and an NVIDIA GeForce GTX1080Ti with 11GB memory.

[0134] The comparative experiments were all implemented in Pytorch-1.9.0, and the network was optimized using the Adam optimizer.

[0135] The denoised image obtained in this example has a lower degree of noise than the denoised images obtained by all experimental methods, and the blood vessels and tissue areas are clearer than those of other methods, indicating that this method provides a CT image denoising service that is more consistent with human eye observation.

[0136] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make other equivalent modifications or substitutions without violating the spirit of the invention, and these equivalent modifications or substitutions are included in the scope defined by the application claims.

Claims

1. A low-dose CT denoising method based on multi-level features, characterized in that: The method comprises the following steps: S1: Collect low-dose CT images and high-dose CT images to construct CT data sets and convert them into NPY matrix data; S2: Perform multi-branch local-global feature extraction on NPY matrix data to obtain shallow features; S3: Strengthen the attention information of shallow features to obtain middle-level features; S4: Extract the middle-level features to obtain deep-level features; S5: perform feature fusion on shallow, medium and deep information according to the feature fusion module to obtain fused features; S6: Combine the fused features with the NPY matrix data to obtain the CT denoised image.

2. A low-dose CT denoising method based on multi-level features according to claim 1, characterized in that: S1 includes the following steps: S1.1: Acquire CT raw data in dicom format containing normal dose and low dose; S1.2: Label the normal-dose CT in the CT raw data as "target" and the low-dose CT as "input"; S1.3: Convert the labeled normal-dose CT and low-dose CT in dicom format into NPY matrix format data.

3. A low-dose CT denoising method based on multi-level features according to claim 2, characterized in that: S2 includes the following steps: S2.1: Construct a multi-branch local-global feature extraction module Local and Global Feature Extractor (LGFE); S2.2: According to the LGFE module, local feature extraction and calculation are performed on the NPY data to obtain the local feature f e ; S2.3: Perform global feature extraction and calculation on NPY data according to the LGFE module to obtain the global feature f g ; S2.4: Merge local features and global features to obtain shallow features f sp ; Among them, S2.1 includes the following steps: Using the multi-branch feature extraction results, convolution calculation is performed to obtain the LGFE structure: f e =Relu(Conv 3×3 (Lso(x))), In the above formula, Relu(.) represents the Relu activation function, Conv n×n (.) represents a convolution kernel of size n, represents the splicing operation, Lso(.) represents the edge detection algorithm, and maxpooling(.) represents the maximum pooling.

4. A low-dose CT denoising method based on multi-level features according to claim 4, characterized in that: S3 includes the following steps: S3.1: Constructing the Repeated Channel Spatial ReinforcingAttention (RCSRA) module; S3.2: According to the RCSRA module, the shallow feature f s Strengthen the attention feature and obtain the middle-level feature f m ; Among them, S3.1 includes the following steps: The convolution calculation is performed using the method of attention information enhancement-feature enhancement to obtain the RCSRA structure: MLP(.)=Conv(Relu(Conv(.))), In the above formula, C n represents the channel attention mechanism, S n represents the spatial attention mechanism, f n Represents the convolution calculation result of the nth stage.

5. A low-dose CT denoising method based on multi-level features according to claim 6, characterized in that: S4 includes the following steps: f d =Relu(Conv(Relu(Conv(f m )) ×3 )), In the above formula, the subscript ×3 represents that the operation is performed 3 times.

6. A low-dose CT denoising method based on multi-level features according to claim 6, characterized in that: S5 includes the following steps: S5.1: Construct Shallow-middle-deep Feature Fusion Module (SMDF); S5.2: The shallow feature f s , middle-level features f m and deep features f d Input SMDF to obtain fusion features; Among them, S5.1 includes the following steps: Different convolution calculation methods are used for shallow features, middle features and deep features to obtain the SMDF structure: f s =Conv 1×1 (f sp ), f A =Conv 3×3 (f d ), Among them, f s represents the shallow perceptual features obtained after convolution calculation of shallow features, f A Represents the deep enhanced features obtained after convolution calculation of deep features, Represents fusion features.

7. A low-dose CT denoising method based on multi-level features according to claim 7, characterized in that: S6 includes the following steps: S6.1: merge the fusion features and NPY matrix; S6.2: Obtain a denoising model for a low-dose CT denoising method based on multi-level features; S6.3: Evaluate the denoising effect of the model based on the calculated quantitative indicators PSNR, SSIM and qualitative human visual effect; Wherein, S6.1 comprises the following steps: In the above formula, Represents the denoised image.

8. A low-dose CT denoising device based on multi-level features, characterized in that: include: Preprocessing module: used to collect data, label the data and convert it into NPY matrix; LGFE module: used to convert NPY matrix data into shallow features; RCSRA module: used to transform shallow features into mid-level features after strengthening attention features, and further transform them into deep features; SMD module: used to fuse shallow features, middle features and deep features to obtain fused features; Image reconstruction module: used to merge the fusion features and NPY matrix to generate the reconstructed CT image.

Images