A Clean Gradient Map-Guided LDCT Image Denoising System and Method

Abstract

This invention discloses a clean gradient map-guided LDCT image denoising system and method, comprising: an encoder and a decoder; the encoder is used to calculate a clean gradient map of the LDCT image and fuse the features of the clean gradient map with the features of the LDCT image to obtain shallow features of the fused image; the decoder is used to extract deep features of the fused image and fuse the shallow features with the deep features to achieve LDCT image denoising. This invention uses cross-attention to fuse the gradient map with the LDCT image, which effectively supplements the missing information in the LDCT image by utilizing the structural and texture information in the gradient map.

Description

Technical Field

[0001] This invention belongs to the field of LDCT image denoising technology, specifically relating to a clean gradient map-guided LDCT image denoising system and method. Background Technology

[0002] Traditional LDCT denoising methods can be broadly categorized into three types: projection domain filtering, iterative reconstruction (IR), and image post-processing. Projection domain filtering processes the original scanned image before reconstruction. Typical algorithms include filtered backprojection reconstruction (FBP), bilateral filtering, adaptive filtering, and penalty-based weighted algorithms. While projection domain filtering effectively captures noise characteristics in the projection domain, it struggles to fully preserve edge information, reducing the resolution of the reconstructed image. Furthermore, obtaining projection domain images is difficult, limiting the applicability of this method. Iterative reconstruction combines the statistical properties of the projection domain image with prior information from the image domain to construct a physical model that suppresses noise and artifacts. Typical methods include total variation (TV), nonlocal mean (NLM), and dictionary learning. Although IR methods typically yield good reconstructed image quality, they require projection domain images, involve significant computational costs, and have slow iteration speeds, making them unsuitable for clinical applications. Image post-processing methods directly process the reconstructed LDCT image without relying on projection domain data. Examples include nonlocal mean filtering, dictionary-based learning methods, block matching algorithms, and diffusion filters, which are computationally more efficient than IR methods. However, the noise in the original LDCT image is often non-uniformly distributed, leading to more complex post-processing. Most existing techniques simply introduce gradients and edges as prior information, lacking in-depth research into the underlying mechanisms of both image processing methods. Transformer models only consider spatial self-attention, neglecting channel self-attention. Summary of the Invention

[0003] This invention aims to address the shortcomings of existing technologies by proposing a clean gradient map-guided LDCT image denoising system and method. In order to more clearly represent edge and structural texture information, a gradient denoising module is designed to obtain a clean gradient map, thereby guiding LDCT image denoising.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A clean gradient map-guided LDCT image denoising system includes: an encoder and a decoder;

[0006] The encoder is used to calculate the clean gradient map of the LDCT image and fuse the features of the clean gradient map with the features of the LDCT image to obtain the shallow features of the fused image.

[0007] The decoder is used to extract deep features from the fused image and fuse the shallow features with the deep features to achieve denoising of the LDCT image.

[0008] Preferably, the encoder includes: a gradient denoising module, a cross-dual attention module, and a cross-channel self-attention module;

[0009] The gradient denoising module is used to remove noise from the gradient domain of the LDCT image to obtain the clean gradient map;

[0010] The cross-attention module is used to integrate the texture and structural information of the clean gradient map into the features of the LDCT image to obtain preliminary fusion features of the fused image containing global and local information.

[0011] The cross-channel self-attention module is used to calculate the correlation between the edge information and structural information of the clean gradient map and the shallow features of the LDCT image based on the preliminary fusion features, so as to obtain further fusion features.

[0012] Preferably, the decoder includes: a channel self-attention module and a dual self-attention module;

[0013] The channel self-attention module is used to extract shallow features of the fused image and fuse the shallow features with the deep features to obtain deep-shallow fused features;

[0014] The dual self-attention module is used to extract global and local information of the deep and shallow layer fusion features to obtain the final features, and add the LDCT image to the final features to obtain a denoised image.

[0015] Preferably, the process of obtaining the clean gradient map is as follows:

[0016] The gradient map of the LDCT image is calculated using the Sobel edge detection operator;

[0017] Based on padding operations and batch normalization, an improved convolutional neural network is obtained to achieve gradient denoising.

[0018] Based on the gradient denoising network, noise in the gradient map is removed to obtain the clean gradient map.

[0019] Preferably, the process of obtaining local information in the preliminary fusion features is as follows:

[0020] The features of the clean gradient map and the features of the LDCT image are subjected to layer normalization and window partitioning to obtain the feature window of the clean gradient map and the feature window of the LDCT image.

[0021] Calculate the cross-attention between the feature windows of the clean gradient map and the feature windows of the LDCT image;

[0022] Based on the cross-attention, the attention between the feature window of the clean gradient map and the local window of the feature window of the LDCT image is calculated through a sliding window to obtain local information in the preliminary fused features.

[0023] Preferably, the process of obtaining the further fusion features is as follows:

[0024] By layer normalization, using 1x1 convolutions and 3x3 depthwise convolutions, the features I of the downsampled clean gradient map are... cg2 The Q2 tensor is generated by converting the downsampled preliminary fusion features I. l3 This converts and generates K2 tensors and V2 tensors;

[0025] The Q2 tensor, the K2 tensor, and the V2 tensor are reshaped to sizes HW×C, C×HW, and C×HW, respectively.

[0026] Based on the reshaped Q2 tensor and K2 tensor, an attention map of size C×C is obtained;

[0027] The attention map is multiplied by the reshaped V2 tensor, and then processed by convolution and residuals to obtain intermediate features;

[0028] The intermediate features are processed by layer normalization and divided into two branches. One branch is activated by the GELU function through 1×1 convolution and 3×3 depth convolution, and the other branch is activated by 1×1 convolution and 3×3 depth convolution.

[0029] By multiplying the two branches element by element and connecting them through the residuals, further fusion features are obtained.

[0030] This invention also provides a method for denoising LDCT images guided by a clean gradient map, comprising the following steps:

[0031] Acquire LDCT images;

[0032] Calculate the clean gradient map of the LDCT image, and fuse the features of the clean gradient map with the features of the LDCT image to obtain the shallow features of the fused image;

[0033] The deep features of the fused image are extracted, and the shallow features are fused with the deep features to achieve denoising of the LDCT image.

[0034] Preferably, the method for obtaining the clean gradient map is as follows:

[0035] The gradient map of the LDCT image is calculated using the Sobel edge detection operator;

[0036] Based on padding operations and batch normalization, an improved convolutional neural network is obtained to achieve gradient denoising.

[0037] Based on the gradient denoising network, noise in the gradient map is removed to obtain the clean gradient map.

[0038] Compared with existing technologies, the advantages of this invention are as follows: Since the original LDCT image contains a significant amount of noise, the correspondingly calculated gradient map also contains noise. To more clearly represent edge and structural texture information, a gradient denoising module is designed to obtain a clean gradient map, thereby guiding LDCT image denoising. This invention uses cross-attention to fuse the gradient map with the LDCT image, effectively supplementing the missing information in the LDCT image by utilizing the structural and texture information in the gradient map. Considering that using a single attention mechanism cannot effectively denoise LDCT images, and that using only a spatial attention mechanism would ignore the global nature of the channel dimension, this invention uses two dual-attention modules: spatial window attention to extract local features, and channel attention to extract global features. Attached Figure Description

[0039] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the structure of the clean gradient map-guided LDCT image denoising system according to Embodiment 1 of the present invention;

[0041] Figure 2 This is a schematic diagram of the gradient denoising module in Embodiment 1 of the present invention.

[0042] Figure 3 This is a schematic diagram of the Cross Dual Attention Module (C-DAT) module according to Embodiment 1 of the present invention. In Figure (a), it is a schematic diagram of the cross dual attention structure; in Figure (b), it is a schematic diagram of the cross attention structure of the moving window; and in Figure (c), it is a schematic diagram of the cross attention structure.

[0043] Figure 4 This is a schematic diagram of the Cross-Channel Self-Attention Module (C-CAT) according to Embodiment 1 of the present invention;

[0044] Figure 5 This is a schematic diagram of the Channel Self-Attention Transformer (CSAT) module according to Embodiment 1 of the present invention;

[0045] Figure 6 This is a schematic diagram of the spatial self-attention Transformer module for a moving window according to Embodiment 1 of the present invention, wherein Figure (a) is a schematic diagram of the spatial self-attention structure and Figure (b) is a schematic diagram of the spatial self-attention structure for a moving window.

[0046] Figure 7 Flowchart of the clean gradient map-guided LDCT image denoising system according to Embodiment 1 of the present invention;

[0047] Figure 8 The following are the results of abdominal CT scans of the test set using different methods in Embodiment 2 of the present invention. Figure (a) is a low-dose CT image (LDCT), Figure (b) is a normal-dose CT image (NDCT), Figure (c) is the denoising result of RED-CNN, Figure (d) is the denoising result of CTformer, Figure (e) is the denoising result of Restormer, and Figure (f) is the denoising result obtained by the present invention (Ours).

[0048] Figure 9 Figure 1 is a schematic diagram of the region of interest (ROI) of the CT scan results in Embodiment 2 of the present invention. Figure 1(a) is a low-dose CT image (LDCT), Figure 2(b) is a normal-dose CT image (NDCT), Figure 3(c) is the denoising result of RED-CNN, Figure 4(d) is the denoising result of CTformer, Figure 5(e) is the denoising result of Restormer, and Figure 6(f) is the denoising result obtained by the present invention.

[0049] Figure 10 Figure 2 shows the results of another abdominal CT scan of the test set using different methods in Embodiment 2 of the present invention; Figure (a) is a low-dose CT image (LDCT), Figure (b) is a normal-dose CT image (NDCT), Figure (c) is the denoising result of RED-CNN, Figure (d) is the denoising result of CTformer, Figure (e) is the denoising result of Restormer, and Figure (f) is the denoising result obtained by the present invention (Ours).

[0050] Figure 11 This is Embodiment 2 of the present invention. Figure 10 The enlarged green box area is the region of interest (ROI). Figure (a) is a low-dose CT image (LDCT), Figure (b) is a normal-dose CT image (NDCT), Figure (c) is the denoising result of RED-CNN, Figure (d) is the denoising result of CTformer, Figure (e) is the denoising result of Restormer, and Figure (f) is the denoising result obtained by this invention (Ours). Detailed Implementation

[0051] 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.

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] Example 1

[0054] like Figure 1 As shown, a clean gradient map-guided LDCT image denoising system includes an encoder and a decoder.

[0055] Encoder used to compute LDCT images (low-dose CT images) l Clean gradient plot I cg The features of the clean gradient map are fused with the features of the LDCT image to obtain the shallow features of the fused image.

[0056] The decoder is used to extract deep features from the fused image and fuse shallow features with deep features to achieve denoising of LDCT images.

[0057] The specific workflow of the clean gradient map-guided LDCT image denoising system is as follows: The encoder part processes the LDCT image I... l As input, use Sobel convolution to compute I l gradient map I lg . Will I lg The input gradient denoising module obtains a clean gradient map I. cg Then I cg with I l I is generated through 3×3 convolution operation. l1 and I cg1 Then, the C-DAT module effectively fuses the two and extracts global and local features to obtain the preliminary fused feature I. l2 . Will I l2 and I cg1 After downsampling to reduce the spatial size and expanding the channel capacity, the input to C-CAT is used to extract features. This process is repeated three times to obtain the fully fused feature I. d The features generated by the encoder can all be called shallow features, but in order to distinguish the features generated by each module, they are divided into preliminary fusion features, further fusion features, and fully fusion features.

[0058] The decoder will fully fuse feature Id As input, image size and channel capacity are progressively restored through upsampling. For feature upsampling and downsampling, we use pixel-unshuffle and pixel-shuffle operations, respectively. To assist the denoising process, encoder features are concatenated with decoder features via skip connections. Following the concatenation is a 1×1 convolution operation to reduce the channel capacity of the encoder features (halving it). In the first skip connection, we use a CSAT block to aggregate shallow features from the encoder with deep features from the decoder to obtain I. d1 This is beneficial for preserving fine structural and texture details in denoised images. Next, the deep feature I... d1 Input the DSAT module to further enrich global and local information to obtain I e Finally, the convolutional layer is applied to I. e To generate the final feature I R Add LDCT images to obtain denoised images: I D =I l +I R Next, we will introduce the gradient denoising module, C-DAT, C-CAT, CSAT, and DSAT modules.

[0059] Specifically, the encoder includes: a gradient denoising module, a cross-dual attention module (C-DAT), and a cross-channel self-attention module (C-CAT);

[0060] The gradient denoising module is used to remove noise from the gradient domain of LDCT images to obtain a clean gradient map.

[0061] A cross-attention module is used to integrate texture and structural information from clean gradient maps into the features of LDCT images, obtaining preliminary fusion features (I) of the fused image containing both global and local information. l2 ;

[0062] The cross-channel self-attention module is used to calculate the correlation between the edge information and structural information of the clean gradient map and the preliminary fusion features of the LDCT image based on shallow features, so as to obtain further fusion features of the fused image.

[0063] Specifically, the decoder includes: a channel self-attention module (CAST) and a dual self-attention module (DAST);

[0064] The channel self-attention module is used to extract deep features from the fused image and fuse shallow features with deep features to obtain deep-shallow fused features; the features of the decoder can all be called deep features, but here they are used as deep-shallow fused features.

[0065] The dual self-attention module is used to extract global and local information from the deep and shallow layer fusion features to obtain the final features. The LDCT image is then added to the final features to obtain a denoised image.

[0066] The process of obtaining a clean gradient map is as follows:

[0067] The Sobel edge detection operator is used to calculate the gradient map I of the LDCT image. lg ;

[0068] Based on padding operations and batch normalization, an improved convolutional neural network is obtained to achieve gradient denoising.

[0069] A gradient denoising network is used to remove noise from the gradient map, resulting in a clean gradient map I. cg .

[0070] Specifically, the goal of the gradient denoising module is to remove noise from the gradient domain of LDCT images. Since gradient maps reveal the sharpness of each local region and can highlight areas with higher clarity and more defined structures, this invention uses gradient maps as an aid to guide denoising. However, LDCT images themselves contain a significant amount of noise, which does not disappear when converted from the image domain to the gradient domain. The presence of noise can obscure the texture and structural information contained in the gradient map to some extent, so this invention performs denoising processing on it.

[0071] The gradient denoising module will I l As input, I is computed using the Sobel operator. l gradient map I lg Then, the I-values ​​are removed by the gradient denoising module. lg Clean gradient map I obtained from noise cg Because U-Net has a simple structure, low computational cost, and can capture spatial and contextual information of images, it has a good effect on gradient domain denoising. Therefore, this invention uses U-Net as the gradient denoising module. Figure 2 As shown, the gradient denoising network consists of multiple convolutional layers. Compared with the originally proposed U-Net, this invention modifies U-Net in two aspects: using padding operations to ensure that the feature map size is the same before and after convolution, and adding batch normalization to provide regularization.

[0072] The purpose of the C-DAT module is to enable I cg1 The texture and structural information in the I l1 In this process, their global and local features are extracted. Specifically, in order to more effectively fuse I... cg1 and I l1 Using clean gradient map features I cg1 As a query, attention is applied to LDCT image features Il1 Interacting is performed. Note that the query feature has been aligned dimensionally with other features. In other words, the two features are of different sizes, which allows the cross-attention mechanism of this invention to learn clear structural information and subtle pixel-level features. Furthermore, existing Transformer-based LDCT image denoising models all compute spatial attention alone. This allows for the detection of heavily noise-affected regions in the LDCT image and assigns greater weight to these regions for denoising. However, computing spatial attention alone ignores the global interaction between channels. Specifically, when computing attention between channels, all spatial locations are considered to obtain a global representation. The dual-attention module used in this invention combines spatial attention with channel attention. Specifically, the SW-CAT module is used to compute the spatial window attention of the image to extract local features (local information), and the CSAT module computes channel attention to obtain global features (global information). The specific description is as follows:

[0073] like Figure 3 As shown in (a), C-DAT consists of the SW-CAT module and the CSAT module connected in series. Unlike ordinary modules that compute spatial attention, the SW-CAT module divides the image into windows and computes the attention of local windows, thus reducing computational cost. Specific details of the SW-CAT module are as follows... Figure 3 As shown in (b), the SW-CAT module first... cg1 and I l1 After layer normalization, and then window partitioning similar to that of the Swing Transformer, the resulting I... cg1 and I l1 Window I cgw and I lw The CA module is input to calculate cross-attention, and then the attention between local windows is calculated using a sliding window approach. The entire SW-CAT module can be represented as:

[0074] x=CA(WP(LN(x))) (1)

[0075] x1=CA(SW(X′))+I cg1 (2)

[0076] X = FFN(LN(x1)) + x1 (3)

[0077] Where CA represents cross-attention module, WP represents window partitioning, LN represents layer normalization, SW represents sliding window, and FFN represents feedforward neural network. The CA module is as follows: Figure 3 As shown in (c), I is first processed by 1×1 convolution and 3×3 depth convolution. cgw The transformation is performed to generate the Q1 tensor, and the I is then converted.lw The transformation generates tensors K1 and V1, where Q1, K1, and V1 are all of size C×H×W. Then, to perform attention calculations, Q1, K1, and V1 are reshaped to sizes HW×C, C×HW, and C×HW, respectively. Finally, cross-attention is calculated, a process that can be represented as:

[0078] x=Dwconv 3×3 (covn 1×1 (Attention(Q1,K1,V1))) (4)

[0079]

[0080] Where S represents the dimension of K1, Dwconv 3×3 (·) represents a 3×3 depthwise convolution. 1×1 (·) represents a single 1×1 convolution. Attention(·) primarily mines Ig. cg1 and I l1 The dependencies between them are used to enhance their integration.

[0081] The process of obtaining local information in shallow features is as follows:

[0082] The feature I of the clean gradient map cg1 Features of LDCT Images l1 Layer normalization is performed, and the window is partitioned to obtain the feature window I of the clean gradient map. cgw and the feature window I of LDCT image lw ;

[0083] Calculate the cross-attention between the feature windows of the clean gradient map and the feature windows of the LDCT image;

[0084] Based on cross-attention, the attention between the feature window of the clean gradient map and the local window of the feature window of the LDCT image is calculated through a sliding window to obtain local information in the shallow features.

[0085] The main function of the C-CAT module is to further fuse gradient map features with LDCT image features. The clean gradient map and LDCT image, after being processed by the C-CAT module, yield shallow features containing both global and local information. However, scaling and enlarging the channels of the clean gradient map and LDCT image results in the loss of some image information. To compensate for this loss during downsampling, the C-CAT module uses cross-attention to calculate the correlation between the edge and structural information contained in the clean gradient map and the LDCT image features. This further enhances the structural information in the LDCT image and extracts deeper features.

[0086] The structure of the C-CAT module is as follows: Figure 4 As shown. Given an input, the features I of the clean gradient map after downsampling are... cg2 Preliminary fusion features after downsampling I l3 First, by normalizing the layers, 1×1 convolutions and 3×3 depthwise convolutions are used to apply I... cg2 Transform to generate a Q2 tensor, for I l3 The transformation generates K2 and V2 tensors. These are then reshaped to sizes HW×C, C×HW, and C×HW, respectively. Next, the attention map obtained from cross-attention is C×C. This attention map is multiplied by V2 and then processed through convolution and residuals to obtain the intermediate feature F. F is normalized and split into two paths. One path is activated by the GELU function after 1×1 and 3×3 depthwise convolutions, while the other path is activated only by 1×1 and 3×3 depthwise convolutions. The two paths are multiplied element-wise, and finally, a residual connection is used to obtain the further fused feature F1. This process can be represented as:

[0087] F = conv(Attention(Q2, K) 2， V2))+I cg2 (6)

[0088]

[0089] F1 = Dwconv 3×3 (conv 1×1 (LN(F)))⊙GeLu(conv(LN(F)))+F (8)

[0090] Where D represents the dimension of V2, ⊙ represents element-wise multiplication, and GeLu(·) represents the GeLu activation function.

[0091] The process of obtaining further fusion features is as follows:

[0092] By layer normalization, using 1x1 convolutions and 3x3 depthwise convolutions, the features I of the downsampled clean gradient map are... cg2 The initial fusion features after downsampling are converted into Q2 tensors and V2 tensors.

[0093] Reshape the Q2 tensor, K2 tensor, and V2 tensor to sizes HW×C, C×HW, and C×HW, respectively;

[0094] Based on the reshaped Q2 tensor and K2 tensor, an attention map of size C×C is obtained;

[0095] The attention map is multiplied by the reshaped V2 tensor, and then processed by convolution and residuals to obtain the intermediate feature F.

[0096] By normalizing the layers, F is divided into two branches. One branch is activated by the GELU function through 1×1 convolution and 3×3 depth convolution, and the other branch is activated by 1×1 convolution and 3×3 depth convolution.

[0097] The two branches are multiplied element by element, and then connected by residuals to obtain the further fusion feature F1.

[0098] The CSAT module, with its design limiting the aggregation of global contextual information for images, allows the model to learn long-range dependencies while maintaining computational efficiency. The backbone network of this model consists of Transformer Blocks, with the following structure: Figure 5 As shown. Given an input X, query(Q), key(K), and value(V) tensors are first generated through layer normalization, 1×1 convolutions, and 3×3 depthwise convolutions. The 1×1 convolutions aggregate cross-channel context information in the pixel direction, and the 3×3 depthwise convolutions encode spatial context information in the channel direction. Then, the query, key, and value are reshaped into 2D tensors, and multi-head self-attention is computed. Note that the attention map generated by the dot product of query and key is C×C in size. Next, for the attention output, it is first split into two branches after layer normalization. One branch is activated by the GeLu function after passing through 1×1 and 3×3 depthwise convolutions, while the other branch is activated only through 1×1 and 3×3 depthwise convolutions, where the 1×1 convolutions control the number of feature channels. Then, the two paths are multiplied element-wise, and finally, the output is obtained through a residual concatenation.

[0099] The DSAT module is structurally similar to the C-DAT module, both being dual-attention modules. The difference is that the C-DAT module is a cross-attention module, while the DSAT module is a self-attention module. Its function is to further extract deep features. d1 Its global and local features. Its structure is as follows: Figure 1 As shown, it consists of one SW-SSAT module and one CSAT module. The SW-SSAT module calculates the spatial window self-attention, and the CSAT module calculates the channel self-attention. The SW-SSAT module is shown below. Figure 6 As shown in (b), firstly, the deep feature I d1 After layer normalization, the deep feature window I is obtained by partitioning the layer into windows. dw The SSA module is input to compute cross-attention, and then the attention between local windows is computed using a sliding window approach. The entire SW-SSAT can be represented as:

[0100] y = SSA(WP(LN(I) dw ))) (9)

[0101] y1=SSA(SW(y))+I d1 (10)

[0102] I S =FFN(LN(y1))+y1 (11)

[0103] In this module, SSA stands for Spatial Self-Attention Module, WP represents Window Partitioning, LN represents Layer Normalization, SW represents Sliding Window, and FFN represents Feedforward Network Layer. y and y1 are intermediate features. The specific structure is as follows: Figure 6 As shown in (a), I is first processed by 1×1 convolution and 3×3 depth convolution. dw The transformation generates Q3, K3, and V3 tensors, all of which have the shape C×H×W. Then, to perform attention calculations, Q3, K3, and V3 are reshaped into HW×C, C×HW, and C×HW, respectively. Finally, cross-attention is calculated, which can be represented as:

[0104] y = Dwconv 3×3 (conv 1×1 (Attention(Q3,K3,V3))) (12)

[0105]

[0106] Where G is the dimension of K3.

[0107] Specifically, the loss function and training process used in the clean gradient map-guided LDCT image denoising system (model) constructed in this invention are as follows:

[0108] Loss Function: The proposed model uses L1 loss as the total loss function of this invention. L1 loss includes L... L1C and L L1G L L1C The difference between the input LDCT and the ground truth NDCT is measured to reduce noise between the input LDCT images. L1G The difference between the LDCT and NDCT gradients is measured to reduce noise in the gradient domain and obtain a clean gradient map. Formally, L1 is defined as follows:

[0109] L L1C =||I n -I l ||1 (14)

[0110] L L1G =||I ng -I lg ||1 (15)

[0111] L1 = LL1C +L L1G (16)

[0112] Where I n For normal dose CT images, I l For low-dose CT images, I ng For the gradient of a normal dose CT image, I lg The gradient of a low-dose CT image.

[0113] Training Details: During training, all networks were implemented using an NVIDIA GeForce RTX 3090 GPU within the PyTorch framework. Adam served as the optimizer for all networks, with a learning rate of 0.0001 and a batch size of 2. Before training, the initial weights for convolutions and batch normalization were random numbers following normal distributions of N(0, 0.02) and N(1.0, 0.02), respectively, with initial biases set to 0. All networks underwent 200 training epochs.

[0114] The overall process is as follows Figure 7 As shown.

[0115] Example 2

[0116] Based on the clean gradient map-guided LDCT image denoising system (model) constructed in this invention, experiments were conducted and the experimental results were analyzed:

[0117] Dataset: A publicly released dataset from the 2016 NIH-AAPM Mayo Clinic LDCT Grand Challenge was used for model training and testing. This dataset includes 5410 low-dose (quarter-dose) and normal-dose (full-dose) CT images with a slice thickness of 1 mm from 10 anonymized patients. This invention selected patient L506's data for evaluation, while the remaining 9 patients were used for model training. Furthermore, to verify the generalization ability of the method, this invention evaluated the trained model on low-dose images with a slice thickness of 3 mm from the aforementioned dataset.

[0118] Denoising Performance: The performance of the proposed model was compared with other state-of-the-art techniques, such as RED-CNN, WGAN-VGG, and CTformer. The selected models were all popular low-dose CT models or the latest Transformer models. All models were retrained based on the officially released code.

[0119] To evaluate the system constructed in this invention, the results of two typical low-dose CT images from the test set were compared. The denoising results using different methods are shown below. Figure 8 and Figure 10 As shown. Their enlarged partial views are as follows. Figure 9 and Figure 11 .

[0120] Due to insufficient photons in the incident X-rays, low-dose CT images suffer severe degradation, such as... Figure 8 (a) and Figure 10 As shown in (a). The two low-dose CT images make it difficult to discern structures and details, from... Figure 8 and Figure 10 As can be seen, all methods suppress noise to some extent. The denoising results of the deep learning method RED-CNN are shown below. Figure 8 (c) and Figure 10 As shown in (c), some noise and artifacts still exist in the entire denoised image. The Transformer-based method is as follows... Figure 8 (d)-(f) and Figure 10 As shown in (d)-(f), it can be seen that Transformer-based methods effectively reduce noise and significantly outperform RED-CNN, improving denoising performance and suppressing most artifacts. Figure 8 (f) and Figure 10 As can be seen in (f), the system of the present invention produces better content detail and texture information compared with other methods.

[0121] exist Figure 9 (a) and Figure 11 In (a), it is difficult to observe the metastatic lesions enclosed by the green box. Figure 9 and Figure 11 Compared to other methods shown, the network restoration results of this invention are the most obvious. Compared to other methods studied in this invention, the method of this invention is not only clearer but also has minimal interference with the surrounding environment; the white dots indicated by the blue arrows are restored by the model constructed in this invention.

[0122] Quantitative Evaluation: This invention uses three metrics—Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Root Mean Square Error (RMSE)—to quantitatively evaluate the quality of denoised images. Table 1 shows the average PSNR, SSIM, and RMSE for all slices in L506. Among state-of-the-art methods, only Restormer has a PSNR exceeding 29, and only Restormer has an RMSE score below 14. In contrast, the network proposed in this invention achieves a maximum PSNR score of 29.4010, a maximum SSIM score of 0.8678, and a minimum RMSE of 13.8756. Both Restormer and the model in this invention use a Transformer architecture. Compared to convolutional neural networks like RED-CNN, Transformer networks can better learn global contextual information, thus achieving better denoising results. The proposed network differs from CTformer in that the input image for CTformer training is segmented into blocks, rather than being input as a whole image. While this reduces computation, it also loses the connections between blocks. This may explain why CTformer, despite being a Transformer network, does not perform as well as Restormer. The proposed network, compared to Restormer, incorporates gradient-guided denoising after denoising to better learn structural texture information and includes a dual attention module to compensate for Restormer's limited ability to process local information. Compared to its competitors, the proposed network exhibits the best performance. Table 1: Quantitative evaluation results of different methods for L506 using PSRN, SSIM, and RMSE. The last row shows the best results.

[0123] Table 1

[0124] Moudule PSNR SSIM RMSE LDCT 24.4688 0.8246 24.6370 RED-CNN 28.6384 0.8536 15.1123 CTformer 28.6373 0.8628 15.2101 Restormer 29.3334 0.8670 13.9791 Proposed 29.4010 0.8678 13.8756

[0125] Example 3

[0126] This invention also provides a method for denoising LDCT images guided by a clean gradient map, comprising the following steps:

[0127] Acquire LDCT images, calculate the clean gradient map of the LDCT images, and fuse the features of the clean gradient map with the features of the LDCT images to obtain the shallow features of the fused image.

[0128] Deep features are extracted from the fused image, and shallow features are fused with deep features to achieve denoising of LDCT images.

[0129] The method for obtaining a clean gradient map is as follows:

[0130] The gradient map of the LDCT image is calculated using the Sobel edge detection operator;

[0131] Based on padding operations and batch normalization, an improved convolutional neural network is obtained to achieve gradient denoising.

[0132] A gradient denoising network is used to remove noise from the gradient map and obtain a clean gradient map.

[0133] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A clean gradient map guided LDCT image denoising system, characterized in that, The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The cross-bi-attention module is composed of a SW-CAT module and a CSAT module in series; the SW-CAT module first converts the clean gradient map features I cg1 and the LDCT image features I l1 After layer normalization processing, the clean gradient map features are then windowed and partitioned, and the feature windows of the clean gradient map I cgw and the LDCT image I lw The CA module is inputted to calculate cross-attention, and then the attention between local windows is calculated through a sliding window; the CA module first converts the feature windows of the clean gradient map I cgw to generate Q 1 tensors, respectively I lw to generate K 1 tensors and V 1 tensors, where the sizes of Q 1, K 1 and V 1 are all C × H × W ; then, in order to perform attention calculation, the Q 1, K 1 and V 1 are respectively reshaped to sizes of The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. × C , C × The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. and C × The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. ; finally, cross-attention is calculated and represented as: wherein S represents K 1 dimension, represents a 3x3 depthwise convolution once, represents a 1x1 convolution once, mining I cg1 and I l1 dependency between them to enhance the fusion of both.

2. The clean gradient map guided LDCT image denoising system of claim 1, wherein, The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images.

3. The clean gradient map guided LDCT image denoising system of claim 1, wherein, The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images.

4. The clean gradient map guided LDCT image denoising system of claim 1, wherein, The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. By layer normalization, using 1x1 convolution and 3x3 depth convolution, features of the down-sampled clean gradient map I cg2 , are converted to generate Q 2 a tensor, the down-sampled preliminary fused features I l3 , are converted to generate K 2 tensors and V 2 tensors; resized to Q 2 tensor, the K 2-tensor and the V 2-tensor, respectively, to The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. , The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. and The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. size; based on the reshaping Q 2 tensor and the K 2 tensor, an attention map of size CxC is obtained; The attention map is combined with the reshaped intermediate features V 2. The tensor is dot-multiplied with the intermediate features, which are convolved and residual processed to obtain the final features. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images.

5. A clean gradient map guided LDCT image denoising method, applying the system of any one of claims 1-4, characterized in that, The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising method and device for low-dose computed tomography (LDCT) images. The application relates to a denoising A clean gradient map of the LDCT image is calculated, and features of the clean gradient map are fused with features of the LDCT image to obtain shallow features of a fused image; Deep features of the fused image are extracted, and the shallow features are fused with the deep features to realize denoising of the LDCT image.

6. The clean gradient map guided LDCT image denoising method of claim 5, wherein, A method for obtaining the clean gradient map is as follows: A Sobel edge detection operator is used to calculate a gradient map of the LDCT image; A convolutional neural network is improved based on padding operation and batch normalization to obtain a gradient denoising network; Based on the gradient denoising network, noise of the gradient map is removed to obtain the clean gradient map.

Images