An image denoising method, device, equipment, storage medium and program product

Abstract

The application belongs to the technical field of medical image processing, and discloses an image denoising method, device, equipment, storage medium and program product, the method comprises the following steps: inputting a noisy input image into a main denoising neural network to obtain a first denoised image; calculating the difference between the noisy input image and the first denoised image to obtain a denoising residual image; calculating a spatial weight map according to a locally measurable physical quantity of the noisy input image, then weighting the denoising residual image and the spatial weight map pixel by pixel to obtain a weighted residual, wherein the physical quantity is a combination of one or more of CT values, gradient amplitudes and noise estimates; and obtaining an output image through a preset residual compensation reconstruction operator according to the first denoised image and the weighted residual. Through the residual selective compensation mechanism, the application significantly improves the preservation of high-frequency anatomical details while maintaining the noise suppression effect, effectively solving the contradiction between denoising and details.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing technology, specifically relating to an image noise reduction method, apparatus, device, storage medium, and program product. Background Technology

[0002] Medical imaging plays a central role in clinical diagnosis. Taking CT imaging as an example, low-dose CT scans are becoming increasingly common to reduce patient radiation dose. However, under low-dose conditions, the noise in reconstructed images increases significantly, affecting lesion detection and quantitative analysis. Traditional noise reduction methods (such as nonlocal mean filtering and wavelet thresholding) can suppress noise to some extent, but they are often accompanied by side effects such as edge blurring and texture loss.

[0003] In recent years, deep learning-based image denoising methods have made significant progress in the field of medical image processing. To address the issue of improving the image quality of low-dose CT images, major medical imaging equipment manufacturers and research institutions have proposed their own solutions. For example, Philips has launched iDose4, IMR, and Precise Image technologies; Neusoft Medical employs the ClearView iterative reconstruction algorithm in its CT products. These solutions have reduced CT image noise to some extent and improved image quality under low-dose scanning conditions. However, each has its own inherent technical limitations.

[0004] The main drawback of hybrid iterative reconstruction methods such as iDose4 and ClearView is that their noise reduction capabilities have a significant bottleneck. Under high-intensity application, they can easily lead to image texture distortion, producing a "wax figure" effect, and may blur or cover up small, low-contrast lesions. Although full-model iterative reconstruction methods such as IMR have stronger noise reduction capabilities, they have extremely high computational loads, slow reconstruction speeds, and their image texture styles differ significantly from traditional filtered backprojection, which is not conducive to widespread clinical adoption. Summary of the Invention

[0005] Technical objective: To address the aforementioned technical problems, this invention proposes an image noise reduction method, apparatus, device, storage medium, and program product.

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

[0007] An image denoising method includes the following steps:

[0008] Noisy input image Input main denoising neural network The first denoised image is obtained. ;

[0009] Calculate the noisy input image Compared with the first denoised image The difference is used to obtain the denoising residual map. ;

[0010] Extract the noisy input image Locally measurable quantization parameters, calculating spatial weighted graphs Then, the denoised residual image is weighted pixel by pixel. Spatial weight graph The weighted residuals are obtained. The locally measurable quantization parameter is a numerical measure obtained by mathematical transformation or mathematical statistics based on the pixel grayscale values ​​in the CT image.

[0011] According to the first noise reduction image and weighted residuals Reconstruction operator through preset residual compensation The output image is obtained. .

[0012] Furthermore, the locally measurable parameters include at least one or more combinations of the following parameters: CT value, gradient magnitude, noise estimate, local contrast, and texture entropy.

[0013] Furthermore, based on the noisy input image The spatial weight map is calculated using local CT values. :

[0014] According to the noisy input image In pixels CT value at the location Noisy input image The lungs are divided into: cortical region, nodule candidate region, soft tissue region, lung parenchyma and air region.

[0015] The spatial weights of the cortical region, nodule candidate region, soft tissue region, lung parenchyma, and air region are respectively assigned as follows: , , ,satisfy CT value spatial weight map is obtained. .

[0016] Furthermore, based on the noisy input image The gradient magnitude is calculated from the spatial weight map. :

[0017] The noisy input image is computed using the Sobel operator. gradient magnitude :

[0018] ;

[0019] According to the gradient magnitude At the 70th and 30th percentiles of the entire image , The noisy input image Divided into strong edge region Weak texture areas, and A uniform region;

[0020] The spatial weights of the strong edge region, weak edge region, and uniform region are respectively assigned as follows: , , ,satisfy The gradient space weight map is obtained. .

[0021] Furthermore, based on the noisy input image The spatial weight map is calculated based on the local noise estimation level. :

[0022] In the noisy input image In pixels The standard deviation of local noise is estimated within the neighborhood using the median absolute difference (MAD). ;

[0023] According to the local noise standard deviation At the 70th and 30th percentiles of the entire image , The noisy input image Divided into High-noise areas medium noise area and The low-noise region;

[0024] The spatial weights assigned to the high-noise region, medium-noise region, and low-noise region are as follows: , , ,satisfy The noise spatial weight map is obtained. .

[0025] Furthermore, the spatial weight map is calculated using the following method. :

[0026] Based on the noisy input image Calculate the corresponding CT value spatial weight map based on the local CT value, local gradient magnitude, and local noise estimation level. Gradient space weight graph Noise Spatial Weighting Map ;

[0027] Pixel determination region type labels are set according to the principle of prioritizing skeletal regions, nodular regions, edge regions, soft tissue regions under uniform conditions or noise regions under uniform conditions, and default regions from high to low priority. Determine the noisy input image The region to which each pixel belongs;

[0028] After the determination is completed, a region adaptive fusion strategy is used to calculate the fused spatial weight map based on the region to which the pixel belongs. :

[0029]

[0030] in, This is a truncation function. , , For pixels The corresponding CT value spatial weights, gradient spatial weights, and noise spatial weights; , , Pixels The fusion weights of CT value spatial weights, gradient spatial weights, and noise spatial weights; These are the minimum and maximum values ​​of the spatial weights after fusion, respectively; , , , The values ​​are all based on the pixels. The region has been adjusted.

[0031] Furthermore, the residual compensation reconstruction operator The residual compensation function is used to output the image. ;

[0032] in, This is the residual compensation function, which is a nonlinear mapping function; This indicates a pixel-by-pixel weighted operation.

[0033] Furthermore, the residual compensation function includes a local adaptive soft thresholding term, and the output image... ;

[0034] in, It is a symbolic function;

[0035] Location-dependent adaptive threshold: , This is the global threshold coefficient. For noisy input images In pixels The noise standard deviation is estimated within the neighborhood.

[0036] Furthermore, the residual compensation reconstruction operator Using residual compensation neural network Output image or ; in, This indicates concatenation of channel dimensions.

[0037] Furthermore, the residual compensation neural network Use lightweight convolutional networks, U-Net submodules, or attention modules.

[0038] Furthermore, the residual compensation neural network With the main noise reduction neural network To construct a joint network, an aggressive training strategy is employed:

[0039] Warm-up and pre-training phase: Training the main denoising neural network separately ;

[0040] End-to-end joint training: Based on pre-warm-up and pre-training, activate reconstruction operators including residual compensation. Based on the inference structure, the joint network is optimized, and a joint loss function is designed:

[0041] ;

[0042] in, The configuration is a loss term to ensure basic noise reduction capabilities; The loss term configured to supervise the final image after residual compensation reconstruction; The loss term configured to impose explicit residual domain supervision on the residual compensation process; , , These are the weighting coefficients for each item.

[0043] Furthermore, Spatial loss weights are introduced to apply corresponding loss weights to bone edge regions, vascular regions, and low-contrast nodule regions:

[0044]

[0045] in, A noise-free, dry reference image; This is the output operator for the residual compensation path.

[0046] An image noise reduction device, comprising:

[0047] The initial noise reduction module is configured to reduce the noise in the input image. Input main denoising neural network The first denoised image is obtained. ;

[0048] The residual extraction module is configured to calculate the noisy input image. Compared with the first denoised image The difference is used to obtain the denoising residual map. According to the noisy input image Locally measurable quantization parameters, calculating spatial weighted graphs Then, the denoised residual image is weighted pixel by pixel. Spatial weight graph The weighted residuals are obtained. The locally measurable quantization parameter is a numerical measure obtained by mathematical transformation or mathematical statistics based on the pixel grayscale values ​​in the CT image.

[0049] The residual compensation reconstruction module is configured to reconstruct the image based on the first denoised image. and weighted residuals Reconstruction operator through preset residual compensation The output image is obtained. .

[0050] An electronic device includes at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps of the method.

[0051] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method.

[0052] A computer program product includes a computer program that, when executed by a processor, implements the steps of the method.

[0053] This invention proposes an image denoising method with a selective residual compensation mechanism. The residual from the first denoising step is treated as a mixture of noise and mistakenly deleted structural details, rather than pure noise. A spatial weighted map and a residual compensation operator are used to differentiate the residual, selectively reinjecting structural details into the denoising result while suppressing random noise backflow. This mechanism has clear estimation theoretical implications: the first denoising step primarily suppresses noise variance, while the residual compensation reconstruction operator focuses on compensating for estimation bias caused by over-smoothing. Furthermore, this invention designs a spatial weight map associated with local image physical quantities, directly linking it to these quantities (such as CT values, local gradient magnitudes, and local noise estimation levels in CT images) to achieve differentiated residual compensation for different anatomical regions. Without significantly increasing system complexity, this mechanism resolves the contradiction between denoising intensity and effective detail preservation.

[0054] Beneficial effects: Compared with the prior art, the present invention has the following technical effects:

[0055] Effectively resolves the noise reduction-detail contradiction: Through a residual selective compensation mechanism, while maintaining the noise suppression effect, it significantly improves the preservation of high-frequency anatomical details such as bone edges, blood vessel contours, and micro-nodule boundaries. This mechanism resolves the contradiction between increasing noise reduction intensity and effectively preserving details. Therefore, during high-intensity noise reduction, it can accurately remove noise while preserving or even enhancing the real anatomical texture boundaries, fundamentally avoiding the "wax figure" effect and ensuring that small, low-contrast lesions are not masked by excessive smoothing.

[0056] Explicit supervision in the residual domain: Compared to most ordinary cascaded noise reduction methods that only apply supervision to the final output, this invention establishes an additional supervision signal in the residual domain during training. This can guide the network to more accurately distinguish between noise components and structural detail components in the residual, enabling the network or function to learn to recover the accidentally deleted details from the weighted residual in the residual domain. This achieves synergistic optimization of noise suppression and image detail preservation without significantly increasing system complexity.

[0057] High versatility: This invention does not rely on a specific network architecture. The main denoising neural network can be any known convolutional neural network denoiser. The method is also applicable to other medical image denoising scenarios besides CT, such as MRI image denoising. Attached Figure Description

[0058] Figure 1 This is a flowchart of the reasoning stage of the image denoising method proposed in Embodiment 1 of the present invention.

[0059] Figure 2 This is a flowchart of a training method for a joint network proposed in Embodiment 3 of the present invention.

[0060] Figure 3This is a comparison chart showing the effects of a conventional single-pass noise reduction method and the method in Scheme 2 of Embodiment 3 of the present invention on processing noisy input images. Detailed Implementation

[0061] Example 1

[0062] Figure 1 This is a flowchart of the reasoning stage of the image denoising method proposed in Embodiment 1 of the present invention. Figure 1 As shown, this embodiment proposes an image denoising method, which includes the following steps in the inference stage:

[0063] Step 1: Initial noise reduction.

[0064] Noisy input image Feed into the noise reduction neural network The first denoised image is obtained. :

[0065] ;

[0066] in, For convolutional neural network denoisers, encoder-decoder structures such as U-Net, denoising convolutional neural networks, or fully convolutional structures can be used, but are not limited to these structures, with the aim of achieving initial denoising processing.

[0067] Step 2: Residual extraction.

[0068] Calculate noisy input image Compared with the first denoised image The difference is used to obtain the denoising residual map. :

[0069] ;

[0070] Noise reduction residual map It includes all the content removed during the first noise reduction process, including both random noise that should have been suppressed and effective anatomical details such as bone edges, blood vessel contours, and micronodule boundaries that were lost due to the network misinterpreting them as noise. Therefore, It is not a pure noise signal, but a mixed signal of noise and mistakenly deleted details, which provides important information for subsequent residual compensation reconstruction.

[0071] Step 3: Residual compensation reconstruction / selective compensation.

[0072] The noise reduction residual map is weighted pixel by pixel. Spatial weight graph The weighted residuals are obtained. .

[0073] Based on the first noise reduction image with weighted residuals Reconstruction operator through residual compensation Obtain the final output image :

[0074] ;

[0075] in, Reconstruct the operator for residual compensation; A spatial weight map related to local image features; This indicates a pixel-by-pixel weighted operation.

[0076] Furthermore, in this embodiment, the spatial weight graph The design is key to achieving selective residual compensation in this invention. Its core idea is to use locally measurable quantization parameters of the image as surrogate indicators to optimize the residual image. The degree of information participation in different regions is adjusted to allow for the reinjection of effective anatomical details while suppressing noise backflow. The locally measurable parameters include at least one or more combinations of the following parameters: CT value, gradient magnitude, noise estimate, local contrast, and texture entropy.

[0077] Specifically, this embodiment lists the following determination methods. The method (actually, it is not limited to these methods; it can also be determined based on other types of quantization parameters) ):

[0078] Based on local CT values ​​(Henness units, HU value), the following segmentation settings are applied: higher weights are assigned to high-density cortical bone or calcified areas (high HU value areas) to enhance the restoration of edge details; lower weights are assigned to homogeneous soft tissue areas (low HU value areas) to reduce noise backflow; and weights are appropriately increased in low-contrast nodule areas to protect the boundaries of small structures.

[0079] Based on the local gradient magnitude setting: a larger weight is assigned to the image edge region (high gradient region), and a smaller weight is assigned to the uniform region.

[0080] Based on the local noise estimation level setting: according to the noisy input image The local noise standard deviation estimate σ is appropriately reduced in areas with strong noise to suppress noise backflow.

[0081] Multi-factor comprehensive determination: The spatial weight map is obtained by comprehensively weighting multiple factors such as CT value, gradient magnitude, and noise estimation.

[0082] Furthermore, in this embodiment, the residual compensation reconstruction operator... Two implementation schemes are proposed, but for the residual compensation reconstruction operator... The specific implementation is not limited to these two schemes.

[0083] Option 1: Function-based residual compensation reconstruction:

[0084] In this scheme, It consists of defined mathematical functions or combinations of functions, requires no additional trainable network parameters, has extremely low deployment costs, and can be directly inserted into existing CT reconstruction post-processing chains. Its implementation is as follows:

[0085] ;

[0086] in, The residual compensation function acting on the weighted residual can be a well-known nonlinear mapping function in the field, such as a piecewise linear function, a threshold function, or a wavelet domain filtering function.

[0087] Taking the local adaptive soft thresholding function as an example, for each pixel Compensation amount The calculation is as follows:

[0088] ;

[0089] in, For weighted residuals; It is a symbolic function; Location-dependent adaptive threshold:

[0090] ;

[0091] For noisy input images exist The noise standard deviation estimated within the neighborhood (e.g., a 7×7 pixel window); This is the global threshold coefficient, typically ranging from 1.5 to 2.5.

[0092] threshold Larger residuals (in areas of higher noise) are allowed to pass through only; Smaller areas (regions with less noise) can recover more subtle structural details. The complete closed-form of the residual compensation reconstruction operator is:

[0093] ;

[0094] This form only involves addition, subtraction, absolute value taking, sign judgment, and threshold truncation operations, requiring no additional neural network inference, making it suitable for deployment with low computational overhead in CT reconstruction post-processing chains.

[0095] Option 2: Network-based residual compensation reconstruction:

[0096] In this scheme, By residual compensation neural network The goal is to learn the optimal compensation relationship between the denoised image and the weighted residual through a data-driven approach.

[0097] ;

[0098] Residual Compensation Neural Network The input can be a denoised image. with weighted residuals Channel splicing format, weighted graph It can also be used as an auxiliary input channel to send data in together:

[0099] ;

[0100] in, This indicates concatenation of channel dimensions.

[0101] Lightweight convolutional networks, U-Net sub-modules, or attention modules can be used, and the number of parameters can be much smaller than that of the main denoising network. , can also be used with Joint end-to-end training. The method learns during training: selectively recovering effective details from weighted residuals in anatomical structures such as bone edges, blood vessels, and nodules; and suppressing noise components in residuals in homogeneous soft tissue regions. The compensation rules in this scheme are data-driven, making them more adaptable, and the main denoising network... Only minor adjustments are needed during the upgrade. This eliminates the need to redesign the compensation function.

[0102] Example 2

[0103] This embodiment proposes a spatial weighting graph. The specific implementation plan.

[0104] Before feeding the residual compensation reconstruction operator, first analyze the noisy input image. Computational Spatial Weighted Graph and the residual plot The weighted residual is obtained by multiplying pixel by pixel. This invention explicitly correlates spatial weights with three observable physical quantities in CT images: anatomical density partitioning, edge structure intensity, and local signal-to-noise ratio. It employs piecewise assignment or continuous mapping for each sub-region, enabling residual compensation reconstruction to more effectively restore details and suppress invalid noise.

[0105] The four specific calculation schemes for the spatial weight map M are as follows, each method is applied to pixels. Calculate weights The typical range of values ​​is =[0.1, 1.0], the neighborhood window defaults to 7×7.

[0106] Scheme (1) Regional spatial weighting based on local CT values ​​(HU)

[0107] remember For noisy input images The CT values ​​at the location are divided into four regions based on the threshold: cortical bone / calcification region ( , Take 300HU), low-contrast nodule candidate region ( And local contrast , Take -100HU), soft tissue area ( and , Take -500HU), lung parenchyma / air region ( Local contrast Defined as The standard deviation of CT values ​​within the central neighborhood. This is the local contrast threshold.

[0108] Segmented assignment: Cortical region Nodule candidate region soft tissue area Lung parenchyma / air zone =0.20, which satisfies CT value spatial weight map is obtained. Optionally, for Apply Gaussian smoothing to eliminate abrupt transitions at the piecewise boundaries.

[0109] Scheme (2) Regional spatial weights based on local gradient magnitude

[0110] Calculating gradient magnitude using the Sobel operator :

[0111] .

[0112] according to At the 70th and 30th percentiles of the entire image , Delineate strong edge regions ( ), weak texture areas ( ), uniform region ( ), respectively assigned , , The gradient space weight map is obtained. .

[0113] Scheme (3) Regional Spatial Weights Based on Local Noise Estimation

[0114] exist The standard deviation of local noise is estimated within the neighborhood using the median absolute difference (MAD). .according to 70th and 30th percentiles , Divide the area into high-noise, medium-noise, and low-noise zones, and assign them the following characteristics: , , The noise spatial weight map is obtained. .

[0115] Unlike existing diffusion denoising schemes that increase degradation weights based on local variance, this method reduces residual compensation weights in regions with large noise estimates to prevent random noise from being reinjected with the residuals.

[0116] Scheme (4) Multi-factor regional integration of comprehensive spatial weights

[0117] Obtained respectively by methods (1) to (3) above. , , Set pixel determination region type label :

[0118]

[0119] in, Indicates the cortical bone or calcified area. Indicates the nodule region. Indicates the edge region. Indicates soft tissue area, Indicates a uniform region Indicates the noise region.

[0120] After the determination is completed, a comprehensive spatial weight map is calculated using a region adaptive fusion strategy based on the anatomical / physical region to which the pixel belongs:

[0121]

[0122] in This is a truncation function used to restrict the merged weight values ​​between a preset lower bound and an upper bound.

[0123] , , For pixels The corresponding CT value spatial weights, gradient spatial weights, and noise spatial weights. , , Pixels The fusion weights of CT value spatial weights, gradient spatial weights, and noise spatial weights; These are the minimum / lower bound and the maximum / upper bound of the fused spatial weights, respectively.

[0124] , , , The values ​​are all based on the pixels. The region has been adjusted.

[0125] Specifically:

[0126] ① Cortical bone / calcified area ( =Bone), :

[0127]

[0128] The bone region is dominated by HU and gradient, while the noise factor is only weakly constrained.

[0129] ② Low-contrast nodule candidate region ( =Nodule), :

[0130]

[0131] The nodular region emphasizes the geometric consistency between the HU contrast and the edge gradient, and sets a lower bound of 0.6 for the weight to protect weak boundaries.

[0132] ③ Non-bone marginal area ( =Edge, such as the outline of blood vessels or organs, :

[0133]

[0134] The edge region is dominated by gradient and multiplied with noise weight to achieve the dual constraint of "compensation at the edge and suppression at high noise".

[0135] ④ Uniform soft tissue area ( =Soft), :

[0136]

[0137] The soft tissue homogeneous region is dominated by noise weight, and an upper limit of 0.45 is set to prevent noise backflow.

[0138] ⑤ High-noise uniform region ( =Noisy):

[0139]

[0140] Pixels with high noise and low gradient are directly clamped to the lowest weight to suppress noise residual backflow;

[0141] ⑥ Default area ( =Default), :

[0142] Default coefficient .

[0143] Example 3

[0144] Figure 2 This is a flowchart of a training method for a joint network proposed in Embodiment 3 of the present invention. In this embodiment, the joint network is the residual compensation neural network described in Embodiment 1. With the main denoising neural network constitute.

[0145] This embodiment employs a progressive end-to-end training framework. If end-to-end training is performed directly on the complete inference structure from random initialization, early... Poor quality, residual The image is approximately equal to the original noisy image, but noise dominates the weighted residual, resulting in chaotic gradient signals and difficulty in convergence. Therefore, this invention designs a progressive training strategy, such as... Figure 2 As shown.

[0146] Phase 1: Warm-up and Pre-training

[0147] This phase involves training the main noise reduction network separately. The residual compensation reconstruction operator will not be activated for the time being. This enables the network to have basic noise reduction capabilities:

[0148] in, For noisy input images, This is the corresponding clean reference image. (To be continued) After converging to a stable stage, it enters stage two.

[0149] The function-based residual compensation reconstruction scheme in this embodiment can achieve zero additional inference parameters, without introducing any additional trainable parameters, and relies entirely on mathematical operations. It can be directly inserted into the existing CT reconstruction post-processing chain, with extremely low deployment costs and no need to retrain the existing noise reduction network.

[0150] Phase Two: End-to-End Joint Training

[0151] Based on preheating and pre-training, activate the reconstruction operator including residual compensation. The complete reasoning structure, for and the residual compensation neural network in Scheme 2 (If adopted) joint optimization.

[0152] The joint loss function consists of three terms:

[0153] ;

[0154] ;

[0155] ;

[0156] in, The output operator for the residual compensation path, and the inference stage or The corresponding residual processing branch.

[0157] Ensure that basic noise reduction capabilities do not degrade; Supervision is applied to the final image after residual compensation reconstruction to optimize the overall noise reduction and detail preservation effect; Applying explicit residual domain supervision to the residual compensation process enables the network or function to learn from the weighted residuals to selectively recover the effective details that were mistakenly deleted.

[0158] Furthermore, the joint loss function is:

[0159] ;in, , , These are the weighting coefficients for each item, which can be adjusted according to the training phase. For example, in the initial stage after warm-up, a value of [value] can be used. = 0.3、 = 0.5、 = 0.2; to be determined Gradually increase after convergence Strengthen detailed compensation learning.

[0160] Furthermore, regarding CT images, Spatial loss weights can be introduced to apply higher weights to the bone edge regions (high CT value gradient regions). Weights are applied to constrain the similarity of superimposed structures in the vascular region, and specific loss weights are applied to low-contrast nodule regions:

[0161] ;

[0162] During end-to-end training, the loss gradient is backpropagated through two main paths to Weight: once the path has been , pass Direct return; Path 2 → → Return. Path 2 makes The update takes into account both noise reduction accuracy and residual compensation correction effect, thereby achieving joint optimization in the original image domain and residual domain.

[0163] Furthermore, in this embodiment, similar to standard convolutional neural network noise reduction training, a noisy image is used. Corresponding clean image This constitutes paired training data. For CT images, training data can be obtained in the following ways:

[0164] 1. Acquire paired data from low-dose CT and standard-dose CT scans;

[0165] 2. Add synthetic noise conforming to a CT physical noise model (such as Poisson-Gaussian mixed noise) to clean CT images to generate paired data;

[0166] 3. Use the average image from multiple scans as a reference clean image and pair it with the single scan image.

[0167] The training set data can be global images or image patches extracted from regions of interest, such as the skull margins or the blood vessel regions of the lung parenchyma. Training uses the Adam optimizer with an initial learning rate of 0.001; data augmentation strategies such as affine transformation and elastic deformation can be employed during training.

[0168] This embodiment offers a modular and upgradeable solution, when the main noise reduction network... During the upgrade, only minor adjustments to the lightweight residual compensation network are required. No need to redesign the entire solution, and maintenance costs are low.

[0169] The end-to-end training framework based on explicit supervision in the residual domain, as constructed in this invention, introduces a residual domain loss function during the training phase. This invention applies explicit supervision to the residual compensation process, enabling the network to learn in the residual domain to selectively recover useful anatomical details from the discarded denoising residuals. While conventional methods only apply overall error supervision to the final output, this invention additionally constrains the network in the residual domain to learn how to recover useful parts from mistakenly deleted content, thus establishing a supervisory path for residual information separation at the mechanism level.

[0170] Figure 3 This is a comparison chart showing the effects of a conventional single-pass noise reduction method and the method in Scheme 2 of Embodiment 2 of this invention on processing noisy input images. Figure 3As shown, this invention, through a residual selective compensation mechanism, significantly improves the preservation of high-frequency anatomical details such as bone edges, blood vessel contours, and micronodule boundaries while maintaining noise suppression effects, compared to ordinary methods. It solves the over-smoothing problem of existing deep learning noise reduction methods from a mechanistic perspective, resulting in better noise reduction performance.

[0171] Example 4

[0172] This embodiment proposes an image noise reduction device to implement the methods in Embodiments 1 to 3. The device includes:

[0173] The initial noise reduction module is configured to reduce the noise in the input image. Input main denoising neural network The first denoised image is obtained. ;

[0174] The residual extraction module is configured to calculate the noisy input image. Compared with the first denoised image The difference is used to obtain the denoising residual map. According to the noisy input image Locally measurable quantization parameters, calculating spatial weighted graphs Then, the denoised residual image is weighted pixel by pixel. Spatial weight graph The weighted residuals are obtained. Among them, the locally measurable quantization parameters are numerical measures obtained based on the pixel grayscale values ​​in the CT image through mathematical transformation or mathematical statistics.

[0175] The residual compensation reconstruction module is configured to reconstruct the image based on the first denoised image. and weighted residuals Reconstruction operator through preset residual compensation The output image is obtained. .

[0176] Another embodiment of the present invention provides an electronic device including at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps of the methods described in Embodiments 1 to 3.

[0177] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the methods described in Embodiments 1 to 3.

[0178] Another embodiment of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the methods described in Embodiments 1 to 3.

[0179] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. An image noise reduction method, characterized in that, Including the following steps: Noisy input image Input main denoising neural network The first denoised image is obtained. ; Calculate the noisy input image Compared with the first denoised image The difference is used to obtain the denoising residual map. ; Extract the noisy input image Locally measurable quantization parameters, calculating spatial weighted graphs Then, the denoised residual image is weighted pixel by pixel. Spatial weight graph The weighted residuals are obtained. The locally measurable quantization parameter is a numerical measure obtained by mathematical transformation or mathematical statistics based on the pixel grayscale values ​​in the CT image. According to the first noise reduction image and weighted residuals Reconstruction operator through preset residual compensation The output image is obtained. .

2. The image denoising method as described in claim 1, characterized in that: The locally measurable parameters include at least one or more combinations of the following parameters: CT value, gradient magnitude, noise estimate, local contrast, and texture entropy.

3. The image denoising method as described in claim 1, characterized in that, Based on the noisy input image The spatial weight map is calculated using local CT values. : According to the noisy input image In pixels CT value at the location Input image with noise The lungs are divided into: cortical region, nodule candidate region, soft tissue region, lung parenchyma and air region. The spatial weights of the cortical region, nodule candidate region, soft tissue region, lung parenchyma, and air region are respectively assigned as follows: , , , ,satisfy CT value spatial weight map is obtained. .

4. The image noise reduction method as described in claim 1, characterized in that, Based on the noisy input image The gradient magnitude is calculated from the spatial weight map. : The noisy input image is computed using the Sobel operator. gradient magnitude : ； According to the gradient magnitude At the 70th and 30th percentiles of the entire image , The noisy input image Divided into strong edge region Weak texture areas, and A uniform region; The spatial weights of the strong edge region, weak edge region, and uniform region are respectively assigned as follows: , , ,satisfy The gradient space weight map is obtained. .

5. The image denoising method as described in claim 1, characterized in that, Based on the noisy input image The spatial weight map is calculated based on the local noise estimation level. : In the noisy input image In pixels The standard deviation of local noise is estimated using the absolute median difference within the neighborhood. ; According to the local noise standard deviation At the 70th and 30th percentiles of the entire image , The noisy input image Divided into High-noise areas medium noise area and The low-noise region; The spatial weights assigned to the high-noise region, medium-noise region, and low-noise region are as follows: , , ,satisfy The noise spatial weight map is obtained. .

6. The image denoising method as described in claim 1, characterized in that, The spatial weight map is calculated as follows: : Based on the noisy input image Calculate the corresponding CT value spatial weight map based on the local CT value, local gradient magnitude, and local noise estimation level. Gradient space weight graph Noise Spatial Weighting Map ; Pixel determination region type labels are set according to the principle of prioritizing skeletal regions, nodular regions, edge regions, soft tissue regions under uniform conditions or noise regions under uniform conditions, and default regions from high to low priority. Determine the noisy input image The region to which each pixel belongs; After the determination is completed, a region adaptive fusion strategy is used to calculate the fused spatial weight map based on the region to which the pixel belongs. : in, This is a truncation function. , , For pixels The corresponding CT value spatial weights, gradient spatial weights, and noise spatial weights; , , Pixels The fusion weights of CT value spatial weights, gradient spatial weights, and noise spatial weights; These are the minimum and maximum values ​​of the spatial weights after fusion, respectively; , , , The values ​​are all based on the pixels. The region has been adjusted.

7. The image denoising method as described in claim 1, characterized in that: The residual compensation reconstruction operator The residual compensation function is used to output the image. ; in, This is the residual compensation function, which is a nonlinear mapping function; This indicates a pixel-wise weighted operation; The residual compensation function includes a local adaptive soft thresholding term, and the output image is: ; in, It is a symbolic function; Location-dependent adaptive threshold: , This is the global threshold coefficient. For noisy input images In pixels The noise standard deviation is estimated within the neighborhood.

8. The image denoising method as described in claim 1, characterized in that: The residual compensation reconstruction operator Using residual compensation neural network Output image or ; in, Indicates concatenation of channel dimensions; The residual compensation neural network Use lightweight convolutional networks, U-Net submodules, or attention modules; The residual compensation neural network With the main noise reduction neural network To construct a joint network, an aggressive training strategy is employed: Warm-up and pre-training phase: Training the main denoising neural network separately ; End-to-end joint training: Based on pre-warm-up and pre-training, activate reconstruction operators including residual compensation. Based on the inference structure, the joint network is optimized, and a joint loss function is designed: ； in, The configuration is a loss term to ensure basic noise reduction capabilities; The loss term configured to supervise the final image after residual compensation reconstruction; The loss term configured to impose explicit residual domain supervision on the residual compensation process; , , These are the weighting coefficients for each item; Spatial loss weights are introduced to apply corresponding loss weights to bone edge regions, vascular regions, and low-contrast nodule regions: in, A noise-free, dry reference image; This is the output operator for the residual compensation path.

9. An image noise reduction device, characterized in that, include: The initial noise reduction module is configured to reduce the noise in the input image. Input main denoising neural network The first denoised image is obtained. ; The residual extraction module is configured to calculate the noisy input image. Compared with the first denoised image The difference is used to obtain the denoising residual map. ; According to the noisy input image Locally measurable quantization parameters, calculating spatial weighted graphs Then, the denoised residual image is weighted pixel by pixel. Spatial weight graph The weighted residuals are obtained. The locally measurable quantization parameter is a numerical measure obtained by mathematical transformation or mathematical statistics based on the pixel grayscale values ​​in the CT image. The residual compensation reconstruction module is configured to reconstruct the image based on the first denoised image. and weighted residuals Reconstruction operator through preset residual compensation The output image is obtained. .

10. An electronic device, characterized in that: The method includes at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the steps of the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computational program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 8.

12. A computer program product, characterized in that: Includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8.

Images