A low-dose CT image denoising method based on anatomical prior guidance

Abstract

The application discloses a low-dose CT image denoising method based on anatomical prior guidance. The method obtains an initial feature map through a shallow feature extraction module, and performs hierarchical feature extraction through a residual dense encoder; a global anatomical prior modeling is performed on a bottleneck layer by using a global context module GC, a large kernel attention module LKA and a convolution feedforward network module ConvFFN; an attention gate decoder fuses the decoder up-sampling features and the corresponding scale of the skip connection features through an attention gate module AG to generate a spatial attention map, element-by-element weighting is performed on the skip connection feature map to filter noise, and the skip connection feature map is refined through an RDB-LKA basic block and then output; a residual prediction head maps the decoding features to a residual output, and adds the residual output to the low-dose CT image to obtain a denoised image. The application realizes a large receptive field long-range structure preservation through the LKA, blocks the noise leakage of the skip connection through the AG, and provides the global anatomical prior through the GC, so that the image quality and the inference efficiency are considered.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and deep learning technology, specifically relating to a low-dose CT image denoising method based on anatomical prior guidance. Background Technology

[0002] Computed tomography (CT) has become one of the most valuable imaging tools in modern medical diagnosis, playing an irreplaceable role in clinical scenarios such as early cancer screening, vascular imaging, orthopedic assessment, and emergency trauma examination. However, ionizing radiation exposure associated with CT scans remains a focal point of clinical concern. Epidemiological studies have shown a statistical correlation between cumulative radiation dose and the incidence of malignant tumors, especially for cancer patients requiring long-term follow-up and children with high radiation sensitivity, making the reduction of scanning radiation dose an urgent clinical need. While reducing tube current or tube voltage can effectively reduce radiation exposure, it leads to statistical degradation of photon counts in the projection data, resulting in severe quantum noise and streak artifacts in the reconstructed CT images, directly affecting the discernibility of tissue boundaries and the detection of low-contrast lesions. Therefore, how to maintain image diagnostic quality while significantly reducing radiation dose has become a core technical challenge that has been continuously tackled in the field of CT imaging.

[0003] To address these challenges, researchers have developed various post-processing denoising methods for low-dose CT images. Iterative reconstruction-based methods optimize the solution by constructing an objective function that includes data fidelity and regularization priors. Total variational regularization and nonlocal means are widely used priors; however, these methods rely on manually designed prior assumptions, often excessively blurring fine anatomical structures while smoothing noise. With the rise of deep learning technology, convolutional neural network-based denoising methods have gradually become mainstream. A typical approach is to construct an encoder-decoder architecture network model, using pairs of low-dose and normal-dose CT images as training samples, and learning the mapping relationship between them through the network. Some works introduce channel attention or spatial attention mechanisms to improve the network's responsiveness to key features, while others employ generative adversarial networks or diffusion models to pursue higher perceptual quality. A common feature of these deep learning methods is that they automate the denoising task using a data-driven model, avoiding reliance on manually designed priors.

[0004] However, existing methods still face several common technical bottlenecks. First, the receptive field of conventional convolutional operations is usually limited to a 3×3 or 5×5 window, making it difficult to effectively capture the continuity of anatomical structures spanning a large spatial range, resulting in breaks or blurring of organ boundaries and blood vessel trajectories in denoised images. Second, skip connections in the encoder-decoder structure directly transmit shallow features from the encoder to the decoder without any filtering mechanism for noise components mixed in the shallow features. Quantum noise therefore leaks along the skip connections into the reconstructed image, forming residual noise in uniform tissue areas. Third, the network lacks the ability to explicitly model the statistical information of the anatomical structure in the whole image, making it difficult to distinguish between normal anatomical tissue and noise artifacts, resulting in over-smoothing or uneven denoising in some areas. Fourth, while the introduction of high-performance modeling mechanisms such as self-attention can expand the receptive field, it brings a quadratic increase in spatiotemporal complexity, making it difficult to deploy efficiently with conventional clinical computing resources. How to achieve efficient and stable low-dose CT image denoising while maintaining the integrity of anatomical structures remains a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0005] The technical problem to be solved by this invention is that in existing low-dose CT image denoising methods, the limited receptive field of conventional convolution leads to the breakage or blurring of long-range anatomical structures; the skip connections of the encoder-decoder architecture lack a noise filtering mechanism, causing quantum noise to leak into the reconstructed image; the network lacks explicit modeling of global anatomical priors, resulting in uneven denoising or excessive smoothing; and the introduction of high-performance mechanisms such as self-attention brings excessive computational overhead, making it difficult to meet the real-time requirements of clinical practice.

[0006] To address the aforementioned technical problems, one aspect of the present invention provides a low-dose CT image denoising method based on anatomical prior guidance, comprising:

[0007] S1: Acquire low-dose CT images, and obtain an initial shallow feature map by performing shallow feature extraction on the low-dose CT images through the shallow feature extraction module;

[0008] S2: The initial shallow feature map is input into the residual dense encoder for hierarchical feature extraction. The residual dense encoder includes N cascaded encoding stages. Each encoding stage includes a cascaded downsampling module and at least one RDB-LKA base block. The RDB-LKA base block is composed of a residual dense block RDB and a large kernel attention module LKA connected in series. Each encoding stage outputs a feature map with different scales from shallow to deep. N is an integer greater than or equal to 1.

[0009] S3: Input the skip connection feature map output from the deepest encoder stage into the bottleneck layer for global dissection prior modeling to obtain the bottleneck feature map; wherein, the bottleneck layer includes multiple cascaded connection blocks, each connection block including a cascaded global context module GC, a large kernel attention module LKA and a convolutional feedforward network module ConvFFN.

[0010] S4: The bottleneck feature map is input into the attention-gated decoder for layer-by-layer reconstruction. The attention-gated decoder includes N decoding stages cascaded in sequence. Each decoding stage includes an upsampling module, an attention-gated module AG, and at least one RDB-LKA base block cascaded in sequence. The attention-gated module is used to fuse the output feature map of the upsampling module and the skip connection feature map of the corresponding scale to generate a spatial attention map. The spatial attention map is then used to perform element-wise weighting on the skip connection feature map to obtain a gated skip feature map. The gated skip feature map and the output feature map of the upsampling module are concatenated and then refined by the RDB-LKA base block to obtain the output feature map of the decoding stage.

[0011] S5: Input the output feature map of the shallowest decoding stage into the residual prediction head and map it into a residual output feature map; use the low-dose CT image and the residual output feature map to fuse them to obtain the denoised CT image.

[0012] Another aspect of the present invention provides a low-dose CT image denoising system based on anatomical prior guidance, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned low-dose CT image denoising method based on anatomical prior guidance.

[0013] Another aspect of the present invention provides a computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned low-dose CT image denoising method based on anatomical prior guidance.

[0014] The present invention has at least the following beneficial effects

[0015] This invention achieves a large receptive field with extremely low computational overhead through a large-kernel attention module, maintaining the continuity of long-range structures such as organ contours and blood vessel trajectories. An attention gating module utilizes decoder features to guide adaptive filtering of encoder skip connections, blocking noise leakage at its source. A global context module aggregates full-image statistical information at the bottleneck layer with linear complexity, eliminating uneven denoising. Combining local feature reuse of residual dense blocks with multi-dimensional constraints of multi-level joint loss functions, an optimal balance between noise suppression and structure preservation is achieved. The overall design employs a lightweight feedforward architecture, meeting real-time clinical needs in a single inference iteration, and each module exhibits good scalability and versatility. Attached Figure Description

[0016] Figure 1 This is the overall network model architecture diagram of the present invention;

[0017] Figure 2 This is a detailed structural schematic diagram of the RDB-LKA basic block of the present invention;

[0018] Figure 3 This is a detailed structural diagram of the global context module of the present invention;

[0019] Figure 4 This is a detailed structural schematic diagram of the ConvFFN convolutional feedforward network module of the present invention;

[0020] Figure 5 This is a detailed structural diagram of the attention gating module AG of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 One aspect of the present invention provides a low-dose CT image denoising method based on anatomical prior guidance, comprising:

[0023] S1: Acquire low-dose CT images, and obtain an initial shallow feature map by performing shallow feature extraction on the low-dose CT images through the shallow feature extraction module;

[0024] Preferably, after acquiring low-dose CT images, they are preprocessed, the preprocessing including: processing the original low-dose CT images in... The window width and window level are truncated within the Heinz unit range, and then linearly normalized to... The shallow feature extraction module uses a 3×3 convolutional layer.

[0025] In this embodiment, raw low-dose CT images are acquired. ,in, and These represent the height and width of the image, respectively. CT images are derived from clinical CT scanning equipment (such as GE Revolution CT, Siemens SOMATOMForce, etc.), with a scanning tube voltage range of 80~120 kVp, a tube current range of 10~50 mAs, a reconstructed image matrix size of 512×512 or larger, and a pixel pitch of 0.5~1.0 mm.

[0026] Preprocessing of raw low-dose CT images: First, in The Heinz unit is used to truncate the window width and level, limiting the CT value range to this interval to remove the influence of extreme values ​​outside the bone and air regions on network training; then the truncated image is linearly normalized to... The normalization formula for the interval is:

[0027]

[0028] in, Indicates will The value is first compared with -1024 and the larger value is taken, and then compared with 3072 and the smaller value is taken, thus achieving truncation.

[0029] Normalized image As input to the shallow feature extraction module, the module employs a single 3×3 convolutional layer with 1 input channel and 64 output channels. The kernel size is 3×3, the stride is 1, and padding is 1 to maintain the spatial resolution of the feature map. This convolutional layer is followed by a LeakyReLU activation function with a negative slope coefficient of 0.1 to introduce non-linearity while preserving positive and negative value information. The kernel size of the shallow feature extraction module can also be 5×5 to obtain a larger initial receptive field in the shallow layer, but this increases computational parameters. In scenarios with limited computational resources, 3×3 convolution is a better choice. The negative slope coefficient of LeakyReLU can be adjusted within the range of 0.01 to 0.2. When the noise intensity in the training data is high, using a larger slope coefficient (e.g., 0.1~0.2) helps to preserve more negative value information; when the noise intensity is low, a smaller slope coefficient (e.g., 0.01~0.05) can be used. The number of output channels for the shallow feature extraction module can also be set to 32 or 128. Fewer channels result in a lighter model, while more channels provide stronger feature representation capabilities. In this embodiment, 64 channels are preferred, achieving a good balance between the number of model parameters and feature representation capabilities.

[0030] S2: The initial shallow feature map is input into the residual dense encoder for hierarchical feature extraction. The residual dense encoder includes N cascaded encoding stages. Each encoding stage includes a cascaded downsampling module and at least one RDB-LKA base block. The RDB-LKA base block is composed of a residual dense block RDB and a large kernel attention module LKA connected in series. Each encoding stage outputs a feature map with different scales from shallow to deep. N is an integer greater than or equal to 1.

[0031] Preferably, the downsampling module uses a 3×3 convolution with a stride of 2 to perform spatial downsampling on the input feature map;

[0032] Please see Figure 2 Preferably, the operation of the residual dense block RDB includes:

[0033]

[0034]

[0035]

[0036]

[0037] in, The input feature map represents the residual dense block RDB; This represents the output feature map of the residual dense block RDB; express convolution; Indicates splicing along the channel; , and These represent the output feature maps of convolutional layers 1 through 3, respectively. express convolution;

[0038] The working process of the Large Kernel Attention Module (LKA) includes:

[0039]

[0040]

[0041] in, This represents the output feature map of the Large Kernel Attention Module (LKA). This represents the input feature map of the Large Kernel Attention (LKA) module; This indicates element-wise multiplication; express Activation function; Represents a 1×1 convolution; This represents a 7×7 depthwise separable convolution; This represents a 5×5 depthwise separable convolution.

[0042] In this embodiment, the initial shallow feature map Input residual dense encoder. The residual dense encoder contains N sequentially cascaded encoding stages, as described in this embodiment. These are denoted as E1, E2, E3, and E4 respectively; each encoding stage Composed of downsampling module and It consists of RDB-LKA basic blocks connected in series.

[0043] For the Each encoding stage, the input feature map is denoted as... After this stage, the feature map is output. ,in, , , (Except for the first encoding stage, its input) Output The reason for first halving and then doubling is to control the model complexity.

[0044] Each encoding stage first uses a downsampling module to halve the spatial resolution and adjust the number of channels of the input feature map. The downsampling module employs a 3×3 convolution with a stride of 2, a kernel size of 3×3, a stride of 2, and padding of 1. The downsampled feature map is represented as follows: A 3×3 convolution with a stride of 2 can simultaneously achieve spatial downsampling and channel dimension transformation, offering higher computational efficiency compared to the two-step operation of pooling followed by convolution. The spatial resolution of the feature map is halved layer by layer. The number of channels changes layer by layer. .

[0045] Feature map after downsampling Passing in sequence Feature extraction and refinement are performed on each RDB-LKA basic block. Each RDB-LKA basic block is constructed by concatenating RDB and LKA in sequence; the number of RDB-LKA blocks in each encoding stage is 4, 4, 2 and 2 respectively; it can also be set to 1, 2, 2 and 2, and the specific setting can be selected according to the experimental results of those skilled in the art.

[0046] Furthermore, after each convolution within the RDB module, InstanceNorm+LeakyReLU(0.1) is used. InstanceNorm normalizes the single-sample statistics, which is more suitable for noise pattern variations caused by differences in scanning parameters in LDCT; that is, the working process of the residual dense block RDB also includes:

[0047]

[0048]

[0049]

[0050]

[0051] in, The input feature map represents the residual dense block RDB; This represents the output feature map of the residual dense block RDB; express convolution; Indicates splicing along the channel; , and These represent the output feature maps of convolutional layers 1 through 3, respectively. express convolution; ;growth rate .

[0052] Furthermore, a residual connection is established between the outputs of RDB and LKA, i.e., the output characteristic representation of the RDB-LKA base block. for:

[0053]

[0054] The input to LKA is the output feature map of RDB. LKA employs a decomposed large-kernel convolution design; and For depthwise separable convolution, its This indicates that the convolution operation is performed independently for each input channel; This indicates an expansion rate of 3, further expanding the effective receptive field; express The activation function maps the attention weights to... Interval; two-level depthwise separable dilated convolution cascaded to achieve an effective receptive field of approximately 25×25, capable of capturing anatomical contextual information spanning a large spatial range:

[0055]

[0056] Attention Map The medium-to-high response regions correspond to the locations of anatomical structures, while the low response regions correspond to the locations of noise, thus achieving selective enhancement of effective information in the feature map.

[0057] After each encoding stage is completed, the output feature map of that stage is... Save as a skip connection feature map This is used for attention-gated fusion in the corresponding stages of the subsequent decoder. The dimensions of the skip connection feature maps for each stage are as follows: , , and Finally, the output feature map of the deepest encoding stage, E4. A global dissection prior model is performed on the input bottleneck layer. Skip connection feature maps are saved at each stage. Attention gating fusion will be used for the corresponding decoding stage.

[0058] Preferably, the number of coding stages It can be adjusted according to the input image resolution and computing resources. For a 512×512 input image, This results in a deepest feature map resolution of 32×32, balancing receptive field coverage and computational efficiency. For higher resolution images (e.g., 1024×1024), the resolution can be appropriately increased. The number of RDB-LKA blocks in each encoding stage can preferably be adjusted according to the task complexity.

[0059] S3: Input the skip connection feature map output from the deepest encoder stage into the bottleneck layer for global dissection prior modeling to obtain the bottleneck feature map; wherein, the bottleneck layer includes multiple cascaded connection blocks, each connection block including a cascaded global context module GC, a large kernel attention module LKA and a convolutional feedforward network module ConvFFN.

[0060] Please see Figure 3 Preferably, the working process of the Global Context Module (GC) includes:

[0061]

[0062]

[0063]

[0064] in, Represents the transformation characteristics of the global context module; The context vector representing the global context module. express Transpose of; Indicates tensor reshaping; This represents the input feature map of the global context module; This represents the output feature map of the global context module; Indicates the activation function; Indicates the number of feature channels; Indicates feature height; Indicates the feature width; Represents a 1×1 convolution; Indicates the normalization layer; This represents tensor broadcasting, used to broadcast vectors to and from... Same spatial dimensions.

[0065] Please see Figure 4 Preferably, the operation of the ConvFFN convolutional feedforward network module includes:

[0066]

[0067] in, This represents the output feature map of the ConvFFN convolutional feedforward network module; This represents the input feature map of the ConvFFN convolutional feedforward network module; express convolution; express Depth-separable convolution; This represents the activation function.

[0068] In this embodiment, the skip connection feature map output by the deepest encoding stage E4 is used. Input bottleneck layer. The bottleneck layer is located between the encoder and decoder. It performs global dissection prior modeling at the lowest spatial resolution. At this point, the spatial size of the feature map is the smallest, the number of channels is the largest, the computational cost of global modeling is the smallest, and the channel dimension contains the richest semantic information.

[0069] The bottleneck layer contains M cascaded connection blocks, where M is 1 in this embodiment. Each connection block is internally connected in the following order: Global Context Module (GC) → Large Kernel Attention Module (LKA) → Convolutional Feedforward Network (ConvFFN).

[0070] The GC module aggregates statistical information about anatomical structures from the entire map, generating a compact global descriptor. The input feature map is... , here , and As only The dimension mentioned here is not directly related to the aforementioned features; this description is only for the purpose of facilitating understanding and explaining the working process of the GC module. The specific values ​​of H and W are different for different GC modules, which should be understood by those skilled in the art; their specific values ​​vary depending on the location of the GC module.

[0071] Step 1: Spatial Feature Compression. The input feature map is compressed using a 1×1 convolution. The channel is compressed into a 1-channel space and then reorganized into a two-dimensional matrix. , ; Compress C channels into a single channel. Will Reorganization ; Normalize along the spatial dimension so that the sum of the attention weights for all spatial locations is 1;

[0072] Step 2: Global context aggregation. This involves combining the original features... Perform matrix multiplication with the transpose of the context vector to obtain the global context descriptor:

[0073]

[0074] This aggregated vector encodes the weighted statistical information of all spatial locations in the entire image, and can be regarded as a compact representation of the anatomical distribution pattern of the entire image.

[0075] Step 3: Feature Transformation and Broadcasting. After transforming the aggregated vector through 1×1 convolution and layer normalization, it is expanded to each spatial location through tensor broadcasting; Will Global descriptor copy expansion to The core function of the GC module is to aggregate anatomical statistics from all locations across the entire image into a global descriptor, which is then broadcast back to each spatial location, providing the network with a global criterion for distinguishing between normal anatomical structures and random noise. The computational complexity of the GC module is far lower than that of traditional self-attention, with minimal computational overhead at the bottleneck layer resolution. The LKA module in the bottleneck layer has the same structure as the LKA in step S2.3.2, but its target differs—here, LKA acts on the global enhanced feature map output by the GC module, further strengthening long-range structural responses in the spatial dimension. Using LKA again in the bottleneck layer, guided by global anatomical priors, further enhances the spatial responses of large-scale structures such as organ contours and bone boundaries, preventing the dilution of global contextual information during channel transformation.

[0076] ConvFFN employs an inverted bottleneck structure for information fusion and enhancement in the channel dimension. ConvFFN complements LKA: LKA focuses on long-range structural modeling in the spatial dimension, while ConvFFN focuses on information fusion and enhancement in the channel dimension. Together, they complete the global anatomy prior modeling of the bottleneck layer. After processing through M connection blocks, the bottleneck layer outputs a feature map. This feature map contains both the global anatomical prior information aggregated by the GC module and the channel dimension enhancement from LKA (Large Spatial Receptive Field) and ConvFFN (Conv FFN), serving as the input to the attention-gated decoder in step S4. The number M of connection blocks in the bottleneck layer can be adjusted according to the model depth requirements. When M=1, a good balance is achieved between computational efficiency and global modeling capability; for more complex denoising tasks, M=2 or M=3 can be set to further deepen the extraction of global priors.

[0077] S4: The bottleneck feature map is input into the attention-gated decoder for layer-by-layer reconstruction. The attention-gated decoder includes N decoding stages cascaded in sequence. Each decoding stage includes an upsampling module, an attention-gated module AG, and at least one RDB-LKA base block cascaded in sequence. The attention-gated module is used to fuse the output feature map of the upsampling module and the skip connection feature map of the corresponding scale to generate a spatial attention map. The spatial attention map is then used to perform element-wise weighting on the skip connection feature map to obtain a gated skip feature map. The gated skip feature map and the output feature map of the upsampling module are concatenated and then refined by the RDB-LKA base block to obtain the output feature map of the decoding stage.

[0078] Preferably, the upsampling module uses bilinear interpolation upsampling to double the spatial resolution of the input feature map, and then performs channel projection alignment through 1×1 convolution;

[0079] Please see Figure 4 Preferably, the working process of the attention gating module AG includes:

[0080]

[0081] in, This indicates bilinear interpolation upsampling; This represents the skip connection feature map of the input attention gating module AG; This represents the output feature map of the upsampling module during the decoding stage; express convolution; Indicates the activation function; express Activation function; This indicates element-wise multiplication; This represents the output feature map of the attention gating module AG.

[0082] The bottleneck feature map obtained in this embodiment Input attention-gated decoder. The attention-gated decoder contains N sequentially cascaded decoding stages, as described in this embodiment. , respectively denoted as D1, D2, and D3 correspond one-to-one with encoder stages E4, E3, E2, and E1. Each decoding stage... It consists of an upsampling module, an attention-gating module (AG), and several RDB-LKA base blocks connected in series. The decoder employs a step-by-step reconstruction strategy from deep to shallow layers, gradually restoring the spatial resolution of the feature maps. At each stage, the attention-gating module fuses the skip connection features of the corresponding levels of the encoder, achieving accurate transmission of structural information and effective noise filtering. The channel configuration for each decoding stage is as follows: D4 outputs 256 channels, D3 outputs 128 channels, D2 outputs 64 channels, and D1 outputs 64 channels, with the spatial resolution doubling step by step, ultimately restoring the image to the same level as the input image. .

[0083] Each decoding stage begins by restoring the spatial resolution of the feature map output from the previous decoding stage through an upsampling module. The upsampling module uses bilinear interpolation to double both the height and width of the feature map, followed by channel projection alignment using 1×1 convolutions. This allows the decoder's upsampled features to be fused with the encoder's skip connection features along the same channel dimension. Bilinear interpolation upsampling does not introduce learnable parameters, thus avoiding upsampling artifacts such as checkerboard artifacts; the 1×1 convolutions are used to flexibly adjust the number of channels to adapt to the fusion dimension requirements of different decoding stages.

[0084] The AG (Augmented Link) is the core innovative module of this invention, used to solve the long-standing problem of noise leakage from skip connections in encoder-decoder architectures. The inputs to the AG include: decoder upsampled features, carrying coarse-grained contextual information about "what is currently being reconstructed"; and encoder skip connection features, carrying detailed information from the input image.

[0085] The AG process works as follows: the decoder upsampled features and encoder skip connection features are projected onto the same gated fusion dimension through 1×1 convolutions, element-wise summed, activated by ReLU, and then mapped to a single-channel spatial attention map through 1×1 convolutions. Finally, the attention map is normalized to the [0,1] interval by the Sigmoid activation function. This attention map is then multiplied element-wise with the encoder skip connection features to generate a gated skip feature map.

[0086] The core principle of Asynchronous Gradient Attention (AG) lies in the fact that decoder features carry semantic information about "what is being reconstructed," while encoder features carry detailed information from the input. The resulting attention map, generated by fusing the two, adaptively determines whether each spatial location belongs to an anatomical structure or noise—high-response regions (close to 1) correspond to anatomical structures and are preserved; low-response regions (close to 0) correspond to noise and are suppressed. The four decoder stages of AG form a multi-layered gating system from coarse to fine: D4 filters high-level abstract features (overall organ shape), D3 and D2 filter mid-level structural features, and D1 filters the lowest-level detailed features (texture, small blood vessels). This multi-layered gating blocks the path of noise leakage into the reconstructed image via skip connections at the source, fundamentally solving the problem of noise leakage through skip connections.

[0087] The gated skip feature map output by AG is concatenated with the decoder upsampled features along the channel dimension. After channel projection through a 1×1 convolution, it is input into the RDB-LKA base block for feature refinement. The number of RDB-LKA base blocks in each decoding stage are 2, 2, 4, and 4 (or can be set to 2, 2, 2, 1). RDB fully extracts local texture features through dense connections, LKA maintains the spatial continuity of the anatomical structure through a large receptive field, and block-level residual connections ensure smooth gradient propagation in each stage of the decoder.

[0088] For decoding stages D4, D3, and D2, a side output branch is added after refining the RDB-LKA base block. This branch maps the output feature map of the current decoding stage to a single-channel image using a 1×1 convolution. After bilinear downsampling to the same resolution as the real image, the L1 loss is calculated using this image and compared to a real normal-dose CT image, serving as a depth supervision signal. This depth supervision accelerates gradient propagation to the shallow layers of the decoder through the side output, improving training stability and convergence speed. The side output is only enabled during the training phase and not during the inference phase.

[0089] S5: Input the output feature map of the shallowest decoding stage into the residual prediction head and map it into a residual output feature map; use the low-dose CT image and the residual output feature map to fuse them to obtain the denoised CT image.

[0090] Preferably, the residual prediction head includes cascaded 3×3 convolutions and 1×1 convolutions.

[0091] The feature map output from the shallowest stage D1 of the decoder is input into the residual prediction head. The residual prediction head employs a residual learning strategy to map the decoded features into a residual map of the same size as the input image, which is used to compensate for missing structural and detail information in low-dose CT images.

[0092] The residual prediction head consists of two cascaded 3×3 convolutional layers (followed by a LeakyReLU activation function with a negative slope coefficient of 0.1) and a 1×1 convolutional layer. The two 3×3 convolutional layers are responsible for performing deep nonlinear transformations and refinements on the decoded features. The LeakyReLU activation function preserves weak gradients in the negative region, ensuring that detailed information is not lost during feature propagation. The 1×1 convolutional layer fuses and compresses the 64-channel features into a single channel, outputting a residual output feature map. This feature map has the same spatial size and number of channels as the input low-dose CT image, representing the structural compensation information predicted by the network.

[0093] Since the main differences between low-dose and normal-dose CT images lie in noise interference and the weakening of some structures, and low-dose images already contain complete anatomical contours, the network only needs to learn the residual mapping from low-dose input to normal-dose output. This means predicting the missing or weakened structural detail components in the input image, and then reconstructing a high-quality CT image through an additive operation. Compared to directly learning the complete image mapping, residual learning makes the optimization objective closer to the identity mapping, significantly reducing the learning difficulty of the network, accelerating model convergence, and effectively avoiding structural distortion caused by over-smoothing.

[0094] Preferably, the number of 3×3 convolutional layers in the residual prediction head can be set to 1 to 3 layers as needed. The more convolutional layers, the stronger the nonlinear expressive power, but the number of parameters also increases accordingly. In this embodiment, two 3×3 convolutional layers and one 1×1 convolutional layer are used to achieve a good balance between expressive power and parameter efficiency.

[0095] As an alternative implementation, the residual prediction head can also adopt a subtraction form of residual learning strategy, that is, let the residual feature map output by the residual prediction head represent the noise component in the input image, and subtract the input low-dose CT image from the residual feature map element by element to obtain the denoised image. The essence of both implementations belongs to residual learning, the only difference is that the target predicted by the residual prediction head is different (predicting structural compensation amount or predicting noise component). Mathematically, the two are equivalent and can be interchanged by the labeling method of training data.

[0096] Preferably, the loss function used during the training phase includes:

[0097]

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] in, Represents the total loss function; Indicates pixel-level L1 loss; Represents structural similarity loss; This represents the edge-preserving loss based on the Sobel operator; Indicates frequency domain loss; The decoder outputs a depth supervision loss; Indicates the number of decoding stages involved in loss calculation; This indicates an adjustable weight parameter; Indicates the first Adjustable depth supervision weight parameters for each decoding stage; Indicates the first Deep supervision loss in each decoding stage; This represents a normal-dose CT image after noise reduction. These are true normal dose CT images; Indicates the total number of pixels; Represents the structural similarity index; Indicates taking the absolute value; This represents the gradient magnitude calculated by the Sobel operator; The amplitude spectrum of the two-dimensional discrete Fourier transform; Indicates the first The single-channel image obtained by mapping the output feature maps of each decoding stage through a 1×1 convolution; Bilinear downsampling to Same spatial resolution; express Total number of pixels; and The pixel index is represented. In this embodiment, there are four stages in the decoder stage. D4, D3, and D2 have side output branches. Therefore, the decoding stages involved in the loss calculation are D4, D3, and D2. Thus, the value of D is 3. When d equals 2 to 4, it corresponds to the decoding stages D2 to D4.

[0105] In this embodiment, the training process of the method of the present invention is as follows: A training dataset is constructed by collecting pairs of low-dose CT images and corresponding normal-dose CT images. The images are preprocessed by window width and level truncation and linear normalization, and data augmentation is performed using random flipping and rotation. The network is trained end-to-end using a multi-level joint loss function. This loss function comprehensively constrains the learning direction of the network from multiple dimensions, including pixel fidelity, structural consistency, edge sharpness, frequency domain fidelity, and supervision of the decoder's intermediate layers, enabling the network to simultaneously achieve effective noise suppression and preservation of anatomical structure. During training, the AdamW optimizer is used in conjunction with a cosine annealing learning rate decay strategy. Mixed-precision training is used to accelerate convergence and save GPU memory. Peak signal-to-noise ratio and structural similarity indices are monitored on the validation set to preserve the optimal model weights. The entire training process is an end-to-end single-stage training, requiring no staged or pre-training operations. After training, the model can perform a single forward inference on the input low-dose CT images and output high-quality denoised images.

[0106] Another aspect of the present invention provides a low-dose CT image denoising system based on anatomical prior guidance, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned low-dose CT image denoising method based on anatomical prior guidance.

[0107] Another aspect of the present invention provides a computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned low-dose CT image denoising method based on anatomical prior guidance.

[0108] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0109] In summary, this invention maps low-dose CT images into high-dimensional features through a shallow feature extraction module. A residual dense encoder extracts hierarchical features step-by-step and outputs skip connection feature maps at each stage. A global context module in the bottleneck layer aggregates full-image anatomical statistics with linear complexity to provide global priors. A large-kernel attention module achieves long-range structural modeling of a large receptive field through decompositional depthwise separable convolutions, and a convolutional feedforward network enhances channel dimensions. An attention-gated decoder reconstructs spatial resolution step-by-step. At each decoding stage, the attention-gated module fuses the decoder's upsampled features with the corresponding encoder's skip connection features to generate a spatial attention map. The skip connection features are element-wise weighted to adaptively filter noise, and then refined by RDB-LKA base blocks. The output of the deepest decoding stage is mapped to residual components by a residual prediction head and added to the original low-dose CT image to obtain a denoised image. This invention achieves efficient local texture extraction through dense connections of residual dense blocks, effectively suppressing noise and maintaining anatomical structural integrity while reducing radiation dose. The inference speed meets real-time clinical needs and possesses good scalability and versatility.

[0110] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for denoising low-dose CT images based on anatomical prior guidance, characterized in that, include: S1: Acquire low-dose CT images, and obtain an initial shallow feature map by performing shallow feature extraction on the low-dose CT images through the shallow feature extraction module; S2: The initial shallow feature map is input into the residual dense encoder for hierarchical feature extraction. The residual dense encoder includes N cascaded encoding stages. Each encoding stage includes a cascaded downsampling module and at least one RDB-LKA base block. The RDB-LKA base block is composed of a residual dense block RDB and a large kernel attention module LKA connected in series. Each encoding stage outputs a feature map with different scales from shallow to deep. N is an integer greater than or equal to 1. S3: Input the skip connection feature map output from the deepest encoder stage into the bottleneck layer for global dissection prior modeling to obtain the bottleneck feature map; wherein, the bottleneck layer includes multiple cascaded connection blocks, each connection block including a cascaded global context module GC, a large kernel attention module LKA and a convolutional feedforward network module ConvFFN. S4: The bottleneck feature map is input into the attention-gated decoder for layer-by-layer reconstruction. The attention-gated decoder includes N decoding stages cascaded in sequence. Each decoding stage includes an upsampling module, an attention-gated module AG, and at least one RDB-LKA base block cascaded in sequence. The attention-gated module is used to fuse the output feature map of the upsampling module and the skip connection feature map of the corresponding scale to generate a spatial attention map. The spatial attention map is then used to perform element-wise weighting on the skip connection feature map to obtain a gated skip feature map. The gated skip feature map and the output feature map of the upsampling module are concatenated and then refined by the RDB-LKA base block to obtain the output feature map of the decoding stage. S5: Input the output feature map of the shallowest decoding stage into the residual prediction head and map it into a residual output feature map; use the low-dose CT image and the residual output feature map to fuse them to obtain the denoised CT image.

2. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, After acquiring low-dose CT images, they are preprocessed. The preprocessing includes: processing the original low-dose CT images... The window width and window level are truncated within the Heinz unit range, and then linearly normalized to... Interval.

3. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, The shallow feature extraction module uses a 3×3 convolutional layer; the downsampling module uses a 3×3 convolution with a stride of 2 to perform spatial downsampling on the input feature map; the upsampling module uses bilinear interpolation upsampling to double the spatial resolution of the input feature map, and then performs channel projection alignment through a 1×1 convolution; the residual prediction head includes cascaded 3×3 convolutions and 1×1 convolutions.

4. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, The working process of the residual dense block RDB includes: in, The input feature map represents the residual dense block RDB; This represents the output feature map of the residual dense block RDB; express convolution; Indicates splicing along the channel; , and These represent the output feature maps of convolutional layers 1 through 3, respectively. express convolution; The working process of the Large Kernel Attention Module (LKA) includes: in, This represents the output feature map of the Large Kernel Attention Module (LKA). This represents the input feature map of the Large Kernel Attention (LKA) module; This indicates element-wise multiplication; express Activation function; Represents a 1×1 convolution; This represents a 7×7 depthwise separable convolution; This represents a 5×5 depthwise separable convolution.

5. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, The working process of the Global Context GC module includes: in, Represents the transformation characteristics of the global context module; The context vector representing the global context module. express Transpose of; Indicates tensor reshaping; This represents the input feature map of the global context module; This represents the output feature map of the global context module; Indicates the activation function; Indicates the number of feature channels; Indicates feature height; Indicates the feature width; Represents a 1×1 convolution; Presentation layer normalization layer; This represents tensor broadcasting, used to broadcast vectors to and from... Same spatial dimensions.

6. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, The operation of the ConvFFN convolutional feedforward network module includes: in, This represents the output feature map of the ConvFFN convolutional feedforward network module; This represents the input feature map of the ConvFFN convolutional feedforward network module; express convolution; express Depth-separable convolution; This represents the activation function.

7. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, The working process of the attention gating module AG includes: in, This indicates bilinear interpolation upsampling; This represents the skip connection feature map of the input attention gating module AG; This represents the output feature map of the upsampling module during the decoding stage; express convolution; Indicates the activation function; express Activation function; This indicates element-wise multiplication; This represents the output feature map of the attention gating module AG.

8. The method for denoising low-dose CT images based on anatomical prior guidance according to claim 1, characterized in that, The loss functions used during the training phase include: in, Represents the total loss function; Indicates pixel-level L1 loss; Represents structural similarity loss; This represents the edge-preserving loss based on the Sobel operator; Indicates frequency domain loss; The decoder outputs a depth supervision loss; Indicates the number of decoding stages involved in loss calculation; This indicates an adjustable weight parameter; Indicates the first Adjustable depth supervision weight parameters for each decoding stage; Indicates the first Deep supervision loss in each decoding stage; This represents a normal-dose CT image after noise reduction. These are true normal dose CT images; Indicates the total number of pixels; Represents the structural similarity index; Indicates taking the absolute value; This represents the gradient magnitude calculated by the Sobel operator; The amplitude spectrum of the two-dimensional discrete Fourier transform; Indicates the first The single-channel image obtained by mapping the output feature maps of each decoding stage through a 1×1 convolution; Bilinear downsampling to Same spatial resolution; express Total number of pixels; and Indicates the pixel index.

9. A low-dose CT image denoising system based on anatomical prior guidance, the system comprising a memory and a processor; characterized in that, The memory is used to store the application program; the processor is used to run the application program and execute the low-dose CT image denoising method based on anatomical prior guidance as described in any one of claims 1 to 8.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements a low-dose CT image denoising method based on anatomical prior guidance as described in any one of claims 1 to 8.

Images