Low-dose CT image denoising method based on deep learning

Abstract

The invention relates to the technical field of medical image processing and computer vision, and discloses a low-dose CT image denoising method based on deep learning, and the method comprises the following steps: obtaining a low-dose CT image and a normal-dose CT image corresponding to the low-dose CT image, carrying out the preprocessing of the low-dose CT image and the normal-dose CT image, and constructing a pairing training data set; according to the method, a denoising model based on an information multi-distillation network is constructed, and the denoising model comprises a shallow feature extraction module, a plurality of information multi-distillation modules connected in series, a contrast perception channel attention module and a reconstruction output module. And meanwhile, key structures and detail information in the image are kept. Through step-by-step extraction and screening of multi-layer features, noise and artifacts can be removed, and meanwhile, the common problem of detail loss in a traditional method can be avoided.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer vision technology, and in particular to a method for denoising low-dose CT images based on deep learning. Background Technology

[0002] With the development of CT technology, CT imaging is increasingly widely used in medical diagnosis. Compared with conventional X-ray examinations, CT scans can provide richer anatomical information, but they require a higher X-ray radiation dose during the imaging process, which may pose potential risks to human health. The radiation dose received during a single CT scan is typically tens of times that of a conventional X-ray examination. Therefore, how to effectively reduce the radiation dose during CT scans while ensuring the image quality required for medical diagnosis has become an important research direction in the field of medical imaging.

[0003] To reduce radiation dose, low-dose CT imaging technology has emerged. Low-dose CT reduces radiation dose by decreasing the X-ray tube current or other scanning parameters without altering the scanning geometry. However, due to the reduced tube current, the number of photons received by the detector is significantly reduced, leading to a marked increase in noise levels in the projected data. CT images reconstructed from this noise-contaminated projected data typically contain strong random noise and streak artifacts, affecting image contrast, sharpness, and the discernibility of detailed structures, thus adversely impacting clinical diagnosis.

[0004] Existing low-dose CT image reconstruction or denoising techniques still have shortcomings in addressing the aforementioned problems. On the one hand, traditional filtered backprojection methods or early iterative reconstruction algorithms have limited ability to suppress statistical noise, easily introducing non-uniform noise and artifacts into the image, especially in areas of high density difference, often manifesting as blurred edges and decreased contrast. On the other hand, while some post-processing denoising methods can reduce noise levels to some extent, they often come at the cost of sacrificing spatial resolution and image detail, resulting in over-smoothing of the image and loss of fine structural information. Furthermore, as algorithm complexity increases, computational costs increase significantly, making it difficult to balance image quality and processing efficiency in practical clinical applications. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a low-dose CT image denoising method based on deep learning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A deep learning-based method for denoising low-dose CT images includes the following steps:

[0008] S1. Obtain low-dose CT images and their corresponding normal-dose CT images, and preprocess the low-dose CT images and normal-dose CT images to construct a paired training dataset.

[0009] S2. Construct a denoising model based on an information multiple distillation network. The denoising model includes a shallow feature extraction module, multiple cascaded information multiple distillation modules, a contrast perception channel attention module, and a reconstruction output module.

[0010] S3. Input the preprocessed low-dose CT image into the denoising model to perform feature extraction and feature enhancement processing on the low-dose CT image. Specifically, the shallow feature extraction module extracts features from the low-dose CT image, and the extracted features are processed stepwise by the information multiple distillation module to obtain a feature representation for image reconstruction. The feature representation is then input into the contrast-aware channel attention module, where features corresponding to different channels in the feature representation are adaptively weighted. Based on the weighted feature representation, the denoised CT image is output through the reconstruction output module.

[0011] Preferably, the preprocessing in S1 includes: rigid registration of low-dose CT images with normal-dose CT images;

[0012] The pixel values ​​of low-dose CT images and normal-dose CT images are truncated, and the truncated pixel values ​​are mapped to a preset range.

[0013] Low-dose CT images and normal-dose CT images are cropped into fixed-size two-dimensional image blocks;

[0014] The cropped 2D image blocks are normalized, and normal dose CT images are used as the true labels.

[0015] Preferably, the shallow feature extraction module consists of at least one convolutional layer, used to map the input single-channel low-dose CT image into a high-dimensional feature representation.

[0016] Preferably, the information multiple distillation module employs a feature distillation mechanism of layer-by-layer channel segmentation and step-by-step refinement, specifically including:

[0017] The input features are segmented multiple times along the channel dimension. Some of the segmented features are directly retained as distilled features, while the remaining features are further subjected to convolution operations to achieve gradual filtering and compression of features used for image reconstruction.

[0018] Preferably, the information multiple distillation module includes residual connections to fuse the input and output features of the information multiple distillation module, thereby enhancing feature representation capabilities and improving network stability.

[0019] Preferably, the contrast perception channel attention module includes:

[0020] Perform global statistical analysis on the input features to obtain multi-channel descriptive information;

[0021] Generate corresponding channel weights based on the channel description information;

[0022] The input features are weighted according to the channel weights to achieve adaptive suppression of noise in structural regions with different contrasts.

[0023] Preferably, the global statistical analysis includes global average pooling and global standard deviation pooling of the input features; the channel weights are generated through at least two layers of nonlinear mapping.

[0024] Preferably, the channel weights are generated through dimensionality reduction mapping and dimensionality increase mapping.

[0025] Preferably, a global residual connection is set after the information multiple distillation module to directly add the features output by the shallow feature extraction module to the input of the reconstruction output module, thereby enhancing the detail preservation capability of the denoised CT image.

[0026] Preferably, the method further includes: evaluating the denoised CT image using at least one of peak signal-to-noise ratio, structural similarity, and root mean square error.

[0027] The beneficial effects of this invention are as follows:

[0028] This invention proposes a deep learning-based method for denoising low-dose CT images. This method effectively suppresses noise in low-dose CT images by constructing a denoising model that incorporates shallow feature extraction, information multiple distillation, and a contrast-aware channel attention mechanism, while preserving key structural and detail information. Through the stepwise extraction and filtering of multi-layer features, it can remove noise and artifacts while avoiding the detail loss problem common in traditional methods.

[0029] In the method of this invention, by adaptively weighting the importance of features from different channels, the model can perform differentiated processing for structural regions with different contrast ratios, thereby achieving a better balance between high-contrast and low-contrast regions. Furthermore, the introduction of multiple distillation and residual connection mechanisms helps to enhance feature representation capabilities and improve the stability of the network during training and inference. The denoised CT images obtained using this method can maintain the continuity and clarity of anatomical structures while reducing noise levels, which is beneficial for enhancing the practical value of low-dose CT images in medical imaging applications. Attached Figure Description

[0030] Figure 1This is a schematic diagram of the overall network structure and key modules of the denoising model used in the low-dose CT image denoising method based on deep learning of the present invention.

[0031] in, Figure 1 (a) is the overall structural flow of the information multiple distillation network;

[0032] Figure 1 (b) is a schematic diagram of the specific network structure of the information multiple distillation module;

[0033] Figure 1 (c) is an upsampling module using transposed convolution to recover image detail information;

[0034] Figure 1 (d) is a schematic diagram of the contrast perception channel attention module. Detailed Implementation

[0035] To provide a clearer understanding of the technical features, objectives, and beneficial effects of this invention, the technical solution of this invention is described in detail below, but this should not be construed as limiting the scope of implementation of this invention. Unless otherwise specified, the methods used in this invention are conventional methods in this technical field. In this invention, materials, reagents, or instruments whose manufacturers are not specified are all conventional products that can be purchased commercially.

[0036] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0037] In one embodiment, refer to Figure 1 A deep learning-based method for denoising low-dose CT images includes the following steps:

[0038] S1. Obtain low-dose CT images and their corresponding normal-dose CT images, and preprocess the low-dose CT images and normal-dose CT images to construct a paired training dataset.

[0039] S2. Construct a denoising model based on an information multiple distillation network. The denoising model includes a shallow feature extraction module, multiple cascaded information multiple distillation modules, a contrast perception channel attention module, and a reconstruction output module.

[0040] S3. Input the preprocessed low-dose CT image into the denoising model to perform feature extraction and feature enhancement processing on the low-dose CT image. Specifically, the shallow feature extraction module extracts features from the low-dose CT image, and the extracted features are processed stepwise by the information multiple distillation module to obtain a feature representation for image reconstruction. The feature representation is then input into the contrast-aware channel attention module, where features corresponding to different channels in the feature representation are adaptively weighted. Based on the weighted feature representation, the denoised CT image is output through the reconstruction output module.

[0041] As a preferred embodiment of the present invention, the preprocessing in S1 includes: rigid registration of low-dose CT images with normal-dose CT images;

[0042] The pixel values ​​of low-dose CT images and normal-dose CT images are truncated, and the truncated pixel values ​​are mapped to a preset range.

[0043] Low-dose CT images and normal-dose CT images are cropped into fixed-size two-dimensional image blocks;

[0044] The cropped 2D image blocks are normalized, and normal dose CT images are used as the true labels.

[0045] In a preferred embodiment of the present invention, the preprocessing in S1 includes: first, rigidly registering the acquired low-dose CT image with its corresponding normal-dose CT image; then, truncating the pixel values ​​of the registered low-dose CT image and normal-dose CT image to limit them to a preset range, wherein the pixel value range is the HU value [-1000, 2000]; subsequently, randomly cropping the processed low-dose CT image and normal-dose CT image into two-dimensional image blocks of fixed size, wherein the size of the two-dimensional image blocks is 128×128 or 256×256; finally, performing Z-score normalization on the cropped two-dimensional image blocks, and constructing paired training data using the normal-dose CT image as the real label.

[0046] In a preferred embodiment of the present invention, normal-dose CT images obtained from the TCIA public dataset are used as the original data, and corresponding low-dose CT images are generated based on the Poisson noise model. The obtained normal-dose CT images and low-dose CT images are resized, converting the original 512×512 images into 256×256 images, and data augmentation processing is performed on the images by means of rotation, flipping, etc., to expand the training samples.

[0047] In a preferred embodiment of the present invention, the shallow feature extraction module consists of at least one convolutional layer, used to map the input single-channel low-dose CT image into a high-dimensional feature representation.

[0048] In a preferred embodiment of the present invention, the shallow feature extraction module is composed of a single 3×3 convolution, which is used to map the input single-channel low-dose CT image to a 64-dimensional feature space.

[0049] As a preferred embodiment of the present invention, the information multiple distillation module adopts a feature distillation mechanism of layer-by-layer channel segmentation and stepwise refinement, specifically including: segmenting the input features multiple times along the channel dimension, directly retaining some of the segmented features as distillation features, and continuing to perform convolution operations on the remaining features, so as to achieve stepwise screening and compression of features used for image reconstruction.

[0050] In a preferred embodiment of the present invention, the information multiple distillation module is provided with residual connections to fuse the input features and output features of the information multiple distillation module, thereby enhancing feature representation capabilities and improving network stability.

[0051] In a preferred embodiment of the present invention, the Information Multiple Distillation Module (IMDB) employs a layer-by-layer channel segmentation and progressive refinement mechanism. Specifically, the input features are first subjected to a 3×3 convolution operation and then nonlinearly transformed using the LeakyReLU activation function. Subsequently, the processed features are sequentially segmented four times along the channel dimension. In the first three channel segments, half of the input channel number is separated and directly retained as distilled features. After the channel segmentation is completed, the features corresponding to the remaining channels are expanded again by a 3×3 convolution operation and concatenated with the distilled features. The concatenated features are processed through a contrast-aware channel attention layer to compress them to the original number of channels. The processed features are then residually concatenated with the input features of the Information Multiple Distillation Module to obtain the output features of the Information Multiple Distillation Module.

[0052] In a preferred embodiment of the present invention, the contrast-aware channel attention module includes: performing global statistical analysis on the input features to obtain multiple channel description information; generating corresponding channel weights based on the channel description information; and weighting the input features according to the channel weights to achieve adaptive suppression of noise in different contrast structure regions.

[0053] In a preferred embodiment of the present invention, the global statistical analysis includes global average pooling and global standard deviation pooling of the input features; the channel weights are generated through at least two layers of nonlinear mapping.

[0054] In a preferred embodiment of the present invention, the channel weights are generated through dimensionality reduction mapping and dimensionality increase mapping.

[0055] In a preferred embodiment of the present invention, the specific implementation of the contrast-aware channel attention module includes: simultaneously performing global average pooling and global standard deviation pooling on the input features to obtain corresponding channel statistics; concatenating the global average pooling result and the global standard deviation pooling result, and generating channel weights through at least two layers of nonlinear mapping. In this embodiment, the nonlinear mapping consists of two 1×1 convolutional layers, wherein the first 1×1 convolutional layer is used to reduce the dimensionality of the concatenated features, and the second 1×1 convolutional layer is used to increase the dimensionality of the features. The ReLU activation function is used between the two 1×1 convolutional layers; the dimensionality-increased features generate weights for each channel through the Sigmoid function, and the input features are weighted according to the channel weights, thereby adaptively suppressing noise in different contrast structure regions.

[0056] In a preferred embodiment of the present invention, a global residual connection is set after the information multiple distillation module to directly add the features output by the shallow feature extraction module to the input of the reconstruction output module, thereby enhancing the detail preservation capability of the denoised CT image.

[0057] As a preferred embodiment of the present invention, it further includes: evaluating the denoised CT image using at least one of peak signal-to-noise ratio, structural similarity, and root mean square error.

[0058] In this embodiment, Peak Signal-to-Noise Ratio (PSNR) is a commonly used metric for measuring the quality of reconstructed images. It is evaluated by calculating the PSNR between the reconstructed image and the real image. The higher the PSNR value, the smaller the difference between the reconstructed image and the real image, and the better the image quality.

[0059] The Structural Similarity Index (SSIM) measures the structural similarity between two images, taking into account brightness, contrast, and structural information. SSIM values ​​range from 0 to 1; a value closer to 1 indicates a higher structural similarity between the two images. This index assesses the degree of similarity between the reconstructed image and the original image at the structural level.

[0060] The root mean square error (RMSE) is a commonly used metric to measure the deviation between predicted and observed values. It is calculated by dividing the sum of the squares of the differences between the predicted and actual values ​​by the number of observations and then taking the square root. The smaller the RMSE value, the smaller the deviation between the predicted and actual values, and the higher the prediction accuracy.

[0061] These three metrics comprehensively evaluate the differences between the reconstructed image and the real image from two levels: pixel-level differences and structural similarity.

[0062] In a series of specific experiments, the above-mentioned indicators were calculated for low-dose CT images and denoised CT images obtained based on the method of this invention, and the evaluation results are shown in Table 1:

[0063] Table 1

[0064]

[0065] As shown in Table 1, the evaluation results demonstrate that, compared to low-dose CT images, the denoised CT images obtained using the method of this invention show improvements in peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean square error (RMSE) under this implementation method. Specifically, the PSNR value increased from 20.35 dB to 30.05 dB, a gain of 9.7 dB, indicating an effective balance between noise suppression and detail preservation under this implementation method. Regarding the SSIM, the denoised CT images obtained using the method of this invention improved structural fidelity, with the SSIM value increasing from 0.7136 to 0.8625, an increase of 20.9%, indicating higher visual consistency between the reconstructed image and the original image. Furthermore, in terms of the root mean square error (RMSE), the denoised CT images obtained using the method of this invention reduced the error value from 11.95 to 6.97, a reduction of 41.7%, demonstrating its advantage in pixel-level accuracy. In summary, the method of this invention achieves simultaneous improvements in three key indicators—PSNR, SSIM, and RMSE—in the denoised CT images, fully demonstrating its excellent application effect in image quality enhancement tasks.

[0066] In a preferred embodiment of the present invention, normal dose CT image data and low dose CT image data obtained after processing are used to form a paired image, wherein the low dose CT image is used as input and the normal dose CT image is used as the real label to construct a training sample set for supervised learning.

[0067] In this process, low-dose CT images are first subjected to shallow feature extraction through 3×3 convolution to obtain a 64-dimensional feature representation. Subsequently, the features are sequentially input into multiple information multiple distillation modules for processing, and deep feature information is mined through channel segmentation, multi-scale convolution, and contrast-aware channel attention mechanism. After feature extraction is completed, the features are upsampled through transposed convolution to restore them to the target resolution. Finally, a multi-channel denoised CT image is reconstructed and output through 3×3 convolution.

[0068] During model training, the L1 loss function is used to optimize the denoising model, and the Adam optimizer is used for parameter updates. The parameters β1 and β2 of the Adam optimizer are set to 0.9 and 0.999, respectively, and the initial learning rate is set to 5 × 10⁻⁶. -4 And every 2×10 5The learning rate is decayed after each small batch update.

[0069] As a preferred embodiment of the present invention, the calculation process of the deep learning-based low-dose CT image denoising method described in the present invention is as follows:

[0070] like Figure 1 As shown in (a), the overall forward computation process of the low-dose CT image denoising model used in this invention can be expressed as follows:

[0071]

[0072] in, To input low-dose CT images, To output a denoised image, This is the result after deep feature extraction. for Figure 1 (c) shows the upsampling module (reused as a fine reconstruction head in this invention), and the last item is the global residual connection.

[0073] like Figure 1 As shown in (a), the shallow feature extraction and deep feature extraction processes can be represented as follows:

[0074]

[0075] Wherein, Conv-3 is a 64-channel 3×3 convolution, Conv-1 is a 1×1 convolution, and IMDB1~IMDB4 are... Figure 1 (b) shows the four information multiple distillation modules; the shallow features and the features processed by the information multiple distillation modules are fused through residual connections.

[0076] like Figure 1 As shown in (b), the calculation process of a single information multiple distillation module is as follows:

[0077]

[0078] in and These represent the input and output features of the information multiple distillation module, respectively. The Conv-3 convolution operation maintains the current number of input channels in each layer; the ChannelSplit operation divides the features along the channel dimension. express Figure 1 (d) shows the contrast perception channel attention module, which is used to perform channel weighting on the spliced ​​features; the information multiple distillation module fuses the weighted features with the module input features through residual connection to obtain the output features.

[0079] like Figure 1As shown in (d), the calculation process of the contrast-aware channel attention module can be represented as follows. First, for the input features... Global statistical analysis is performed in the spatial dimension, calculating the global mean and global standard deviation. The global mean is... and global standard deviation The calculation method is as follows:

[0080]

[0081] The global average and global standard deviation are concatenated to obtain the channel statistics vector:

[0082]

[0083] The channel statistical vectors are used to generate channel weights through nonlinear mapping, and the calculation method is as follows:

[0084]

[0085] in This represents the Sigmoid function. Based on the generated channel weights, the input features are subjected to channel-weighted processing, calculated as follows:

[0086]

[0087] Here, ⊙ represents element-wise multiplication along the channel dimension.

[0088] like Figure 1 As shown in (c), the upsampling module is directly used as the reconstruction output module in this invention, and its calculation process can be expressed as follows:

[0089]

[0090] The transposed convolution is set with a stride of 1 to achieve the reconstruction output of denoised CT images through channel adjustment and feature refinement while maintaining the spatial resolution.

[0091] During model training, a composite loss function is used to optimize the denoising model. The expression for the composite loss function is as follows:

[0092]

[0093] in, This represents the denoised CT image output by the model. This represents the corresponding normal dose CT image. This represents the perceptual loss calculated based on high-frequency features from a pre-trained VGG network. This represents the loss of uniformity in the noise power spectrum.

[0094] The above mathematical expression and Figure 1 (a) Figure 1 (b) Figure 1 (c) and Figure 1 The network structure shown in (d) corresponds one-to-one and is used to fully describe the calculation process of the deep learning-based low-dose CT image denoising method of the present invention in a specific embodiment.

[0095] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for denoising low-dose CT images based on deep learning, characterized in that, Includes the following steps: S1. Obtain low-dose CT images and their corresponding normal-dose CT images, and preprocess the low-dose CT images and normal-dose CT images to construct a paired training dataset. S2. Construct a denoising model based on an information multiple distillation network. The denoising model includes a shallow feature extraction module, multiple cascaded information multiple distillation modules, a contrast perception channel attention module, and a reconstruction output module. S3. Input the preprocessed low-dose CT image into the denoising model to perform feature extraction and feature enhancement processing on the low-dose CT image. Specifically, the shallow feature extraction module extracts features from the low-dose CT image, and the extracted features are processed stepwise by the information multiple distillation module to obtain a feature representation for image reconstruction. The feature representation is then input into the contrast-aware channel attention module, where features corresponding to different channels in the feature representation are adaptively weighted. Based on the weighted feature representation, the denoised CT image is output through the reconstruction output module.

2. The method for denoising low-dose CT images based on deep learning according to claim 1, characterized in that, The preprocessing in S1 includes: rigid registration of low-dose CT images with normal-dose CT images; The pixel values ​​of low-dose CT images and normal-dose CT images are truncated, and the truncated pixel values ​​are mapped to a preset range. Low-dose CT images and normal-dose CT images are cropped into fixed-size two-dimensional image blocks; The cropped 2D image blocks are normalized, and normal dose CT images are used as the true labels.

3. The method for low-dose CT image denoising based on deep learning according to claim 2, characterized in that, The shallow feature extraction module consists of at least one convolutional layer, used to map the input single-channel low-dose CT image into a high-dimensional feature representation.

4. The method for denoising low-dose CT images based on deep learning according to claim 1, characterized in that, The information multiple distillation module employs a feature distillation mechanism of layer-by-layer channel segmentation and step-by-step refinement, specifically including: The input features are segmented multiple times along the channel dimension. Some of the segmented features are directly retained as distilled features, while the remaining features are further subjected to convolution operations to achieve gradual filtering and compression of features used for image reconstruction.

5. A low-dose CT image denoising method based on deep learning according to claim 4, characterized in that, The information multiple distillation module is equipped with residual connections to fuse the input and output features of the information multiple distillation module.

6. The method for denoising low-dose CT images based on deep learning according to claim 1, characterized in that, The contrast perception channel attention module includes: Perform global statistical analysis on the input features to obtain multi-channel descriptive information; Generate corresponding channel weights based on the channel description information; The input features are weighted according to the channel weights.

7. A method for denoising low-dose CT images based on deep learning according to claim 6, characterized in that, The global statistical analysis includes global average pooling and global standard deviation pooling of the input features; the channel weights are generated through at least two layers of nonlinear mapping.

8. A method for denoising low-dose CT images based on deep learning according to claim 6, characterized in that, The channel weights are generated through dimensionality reduction mapping and dimensionality increase mapping.

9. A method for denoising low-dose CT images based on deep learning according to claim 1, characterized in that, A global residual connection is set after the information multiple distillation module to directly add the features output by the shallow feature extraction module to the input of the reconstruction output module.

10. A method for denoising low-dose CT images based on deep learning according to claim 1, characterized in that, Also includes: The denoised CT images are evaluated using at least one of peak signal-to-noise ratio, structural similarity, and root mean square error.

Images