Low-dose CT image self-supervision denoising method based on axial structure consistency

Abstract

The invention provides a low-dose CT image self-supervision denoising method based on axial structure consistency, and belongs to the technical field of medical image processing and deep learning. The method comprises the following steps of: inputting a DICOM (Digital Imaging and Communications in Medicine) image sequence scanned by original LDCT (Level Discrete Cosine Transform), extracting three continuous axial slices to generate a triple sample data set, and calculating an axial difference image of two adjacent slices to construct a two-tuple data set; the method comprises the following steps: after constructing a double-branch neural network framework of a denoising network and a difference enhancement network, respectively outputting a denoised triple and axial structure offset estimation, and calculating a pseudo-supervision signal; and on the basis of the pseudo-supervised signal and the denoised triple, calculating by constructing a self-supervised joint loss optimization function to obtain a model for denoising the low-dose CT image. According to the method, the problems of axial structure dislocation, high-frequency detail loss, excessive smoothness and the like in an existing self-supervised denoising method are effectively solved, and the structural consistency and detail fidelity of the image are remarkably improved while noise suppression is ensured.

Description

A self-supervised denoising method for low-dose CT images based on axial structural consistency Technical Field

[0001] This invention belongs to the field of medical image processing and deep learning technology, specifically relating to a self-supervised denoising method for low-dose CT images based on axial structural consistency. Background Technology

[0002] Computed tomography (CT) is an X-ray-based three-dimensional tomographic imaging technique that has been widely used in clinical diagnosis due to its fast scanning speed, high spatial resolution, and clear tissue contrast. However, the high radiation dose during CT examinations may increase the risk of patients developing diseases such as cancer, posing potential health risks. To reduce radiation exposure, low-dose CT (LDCT) reduces the radiation dose by decreasing the X-ray tube current or exposure time. However, this results in a significant increase in quantum and electronic noise in the images, manifesting as obvious speckled artifacts, blurred tissue boundaries, and loss of texture details. This decline in image quality can easily mask key lesion features such as small nodules and early-stage tumors, thereby increasing the likelihood of missed or misdiagnosed lesions in clinical practice.

[0003] To address the denoising problem of LDCT images, existing deep learning-based methods are mainly divided into supervised and unsupervised approaches. Supervised methods typically require a large number of paired LDCT and normal-dose CT (NDCT) images as training data, achieving denoising by modeling a mapping relationship from noisy images to high-quality images. However, in actual clinical practice, obtaining paired LDCT and NDCT images of the same patient at the same anatomical location requires two independent scans, which not only violates the original intention of low-dose imaging to reduce radiation exposure but also presents ethical and operational limitations, making it difficult to obtain high-quality paired data on a large scale. Therefore, unsupervised paradigms, represented by self-supervised learning, avoid the dependence on paired data. These methods utilize only the LDCT images themselves to construct the training task, without the supervision of NDCT images, thus enhancing the feasibility of LDCT technology in clinical practice. However, existing self-supervised LDCT denoising methods still face challenges in maintaining structural consistency and fine-grained structural recovery. Meanwhile, the lack of structural modeling capabilities for CT three-dimensional volume data means that most methods only use single slices as processing units, failing to fully explore the spatial correlation and structural continuity between adjacent slices, resulting in problems such as axial structural misalignment and inter-slice information breakage in the denoised CT images.

[0004] In summary, there is an urgent need for an LDCT denoising technique that combines unpaired data dependency with 3D structural fidelity, achieving synergistic optimization of noise suppression, detail preservation, and axial structural consistency while eliminating NDCT data dependency. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a self-supervised denoising method for low-dose CT images based on axial structural consistency.

[0006] To achieve the above objectives, the specific steps include: S1, Data Preparation and Sample Construction: Input the original LDCT scan DICOM image sequence, perform preprocessing operations, extract three consecutive axial slices to generate a triplet sample dataset, and then calculate the axial difference map of two adjacent slices to construct a binary dataset; the preprocessing operations include: HU value conversion and amplitude limiting, linear normalization, image size unification, and manual screening; the specific preprocessing operation of the original DICOM file in this invention is as follows: after stripping non-image metadata, the HU value window is limited to [ Within the range of [1000, 1000], and linearly normalized to the interval [0, 1]; the unified image size is 512×512, and bilinear interpolation is used to process the original resolution differences; then, slices containing motion or metal artifacts are manually screened out, the CT axial slice thickness is 1mm, and adjacent slices have the medical characteristic of high axial structural continuity; the original DICOM file includes the LDCT dataset and the NDCT dataset; this invention uses a sliding window traversal sequence with a size of 3 and a step size of 1 to sample the CT slices of each patient's continuous scans, and extracts three consecutive axial slices to form a triplet. And calculate the axial difference between two adjacent slices to form a binary pair. ,in , .

[0007] S2. Constructing a two-branch neural network framework for denoising and interpolation enhancement: Based on a triplet sample dataset, a denoising network is constructed to generate denoised triplets; based on a binary dataset, an interpolation enhancement network is constructed to generate axial structural offset estimates; the denoised triplets include: a denoised front slice, a denoised center slice, and a denoised back slice; this invention constructs a two-branch deep neural network joint learning framework, with one branch being a denoising network, and the other branch being a triplet... As input, the data is processed through a denoising network, and the output is a denoised triplet. Another branch is the difference enhancement network, which uses binary pairs. As input, the interpolation enhancement network outputs an estimate of the axial structural offset. The two network parameters are independent of each other and do not share weights.

[0008] The denoising network in this invention employs the RED-CNN network, which consists of a symmetrical encoder-decoder structure, including four levels of downsampling and upsampling modules. It fuses shallow details and deep semantic information through skip connections, and introduces a residual learning mechanism to stabilize training and improve denoising performance. This branch uses triples... As input, the corresponding denoising result is output. .

[0009] The difference enhancement network in this invention is based on the Attention U-Net design. Its backbone is a classic U-Net structure, containing four levels of downsampling and upsampling modules. Squeeze-and-Excitation (SE) channel attention modules are embedded at the skip connections in the encoder-decoder path, which can adaptively enhance the responses of salient features related to structural changes. The input to this branch is the structural difference binary calculated from the original triples. The output is the axial offset estimated by the network. The denoising network and the interpolation enhancement network operate completely independently during training and inference, without sharing weights. The denoising network employs a convolutional neural network structure based on residual connections, while the interpolation enhancement network uses a variant of U-Net with an encoder-decoder structure. Independent parameter optimization for both networks allows for specialized learning targeting global noise suppression and high-frequency detail enhancement, respectively.

[0010] S3. Constructing a pseudo-supervision signal: Based on the denoised triplet and axial structural offset estimation, a pseudo-supervision signal is constructed and calculated to output the pseudo-supervision signal; this invention is based on the denoised triplet output by the denoising network. Axial structural offset estimation from the output of the sum-difference enhancement network Construct a pseudo-monitor signal S, and calculate its expression as follows: in, The front slice after noise reduction. The slices are after denoising. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

[0011] This invention observes that the difference in axial displacement between adjacent slices cannot be ignored. Combined with the setting of "noise being independently statistically analyzed between different slices", the average of the "front-to-back" values ​​in the triplet output by the denoising network is calculated. A high-quality pseudo-supervisory signal S is generated through mathematical reasoning to replace NDCT in supervised learning. A self-supervised loss function is set based on the pseudo-supervisory signal.

[0012] S4. Constructing a self-supervised joint loss optimization function and training the model: Based on the pseudo-supervised signal and the denoised triples, a self-supervised joint loss optimization function is constructed and calculated to minimize the joint loss optimization function and simultaneously optimize the network parameters of the denoising network and the interpolation enhancement network. Training parameters are then set for training to obtain a model for low-dose CT image denoising, thus completing the self-supervised denoising method for low-dose CT images. The self-supervised joint loss optimization function includes: a self-supervised guided loss function, a pixel-level reconstruction loss function, and a structural similarity loss function. The training parameters include: an initial learning rate set to 1×10⁻⁶. 4 The batch size is 8, the maximum number of training epochs is 100,000, and an early stopping mechanism is designed; the expression for the self-supervised guided loss function is: in, The front slice after noise reduction. This is the center slice after denoising. The slices are after denoising. This is a false monitoring signal. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

[0013] The expression for the pixel-level reconstruction loss function is: The expression for the structural similarity loss function is: in, Let represent the structural similarity index; ultimately, the self-supervised joint loss optimization function is a weighted sum of the above three terms: in, To preset the self-supervised guided loss weight coefficient, For pixel-level reconstruction loss weight coefficients, The preset structural similarity loss weight coefficients are used.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention proposes a self-supervised denoising method for low-dose CT images based on axial structural consistency constraints. By constructing a dual-branch neural network (denoising network and interpolation enhancement network) and utilizing the prior characteristics of approximately consistent axial structure and statistically independent noise between adjacent slices, it explicitly models the structural continuity between slices and suppresses random noise. Simultaneously, it introduces a pseudo-supervisory signal generated by mathematical derivation as a training target, eliminating the need for paired conventional-dose CT images. This invention effectively solves the problems of axial structural misalignment, loss of high-frequency details, and over-smoothing in existing self-supervised denoising methods. While ensuring noise suppression, it significantly improves the structural consistency and detail fidelity of the image. On the AAPM-Mayo dataset, it achieves denoising performance superior to mainstream self-supervised methods. Visual evaluation shows that the nodule boundaries are clear and the inter-slice continuity is good after denoising, demonstrating good clinical applicability. Attached Figure Description

[0015] Figure 1 is a flowchart of a self-supervised denoising method for low-dose CT images based on axial structural consistency according to the present invention; Figure 2 shows the results of denoising the 170th to 172nd consecutive slices of L333 patients according to the present invention, wherein, Figure (a) is the 170th LDCT image sampled by the sliding window, Figure (b) is the 171st LDCT image sampled by the sliding window, Figure (c) is the 172nd LDCT image sampled by the sliding window, Figure (d) is the denoised 170th result image according to the present invention, Figure (e) is the denoised 171st result image according to the present invention, and Figure (f) is the denoised 172nd result image according to the present invention; Figure 3 shows the results of denoising the 170th LDCT image sampled by the present invention at AAPM-Mayo The visual contrast results on the 2016 dataset are shown in Figure (a), which is the input LDCT image, Figure (b) is the reference NDCT image, Figure (c) is the output image of Noise2Noise, Figure (d) is the output image of Noise2Sim, Figure (e) is the output image of the present invention, Figure (f) is a magnified view of the input LDCT image, Figure (g) is a magnified view of the reference NDCT image, Figure (h) is a magnified view of the output image of Noise2Noise, Figure (i) is a magnified view of the output image of Noise2Sim, and Figure (j) is a magnified view of the output image of the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] To enable those skilled in the art to clearly understand and implement the technical solution of the present invention, the specific implementation process of the present invention will be described in detail below with reference to the publicly available dataset of AAPMLow-Dose CT Grand Challenge 2016 (AAPM-Mayo 2016); the entire implementation process strictly follows the four core steps described in the invention content.

[0018] As shown in Figure 1, a self-supervised denoising method for low-dose CT images based on axial structural consistency includes the following steps: S1, Data preparation and sample construction: Input the original LDCT scan DICOM image sequence, perform preprocessing operations, extract three consecutive axial slices to generate a triplet sample dataset, and calculate the axial difference map of adjacent two slices to construct a binary dataset; the preprocessing operations include: HU value conversion and amplitude limiting, linear normalization, image size unification, and manual screening; the preprocessing operation of the original DICOM file in this invention specifically includes: after stripping non-image metadata, limiting the HU value window to [ Within the range of [1000, 1000], and linearly normalized to the interval [0, 1]; the unified image size is 512×512, and bilinear interpolation is used to process the original resolution differences; then, slices containing motion or metal artifacts are manually screened out, the CT axial slice thickness is 1mm, and adjacent slices have the medical characteristic of high axial structural continuity; the original DICOM file includes the LDCT dataset and the NDCT dataset; this invention uses a sliding window traversal sequence with a size of 3 and a step size of 1 to sample the CT slices of each patient's continuous scans, and extracts three consecutive axial slices to form a triplet. And calculate the axial difference between two adjacent slices to form a binary pair. ,in , This embodiment uses the AAPM-Mayo 2016 clinical LDCT dataset, which contains paired low-dose and conventional-dose CT scans of 10 patients, with a slice thickness of 1 mm and an image size of 512×512 pixels. In the experiment, only low-dose CT (LDCT) images were used for training, while conventional-dose CT (NDCT) images were used only for performance evaluation. Specifically, continuous CT scans from 8 patients were used as the training set, continuous CT scans from 1 patient were used as the validation set, and continuous CT scans from 1 patient were used as the test set.

[0019] After preprocessing, taking advantage of the continuity of axial structural height in CT volumetric data, three consecutive slices were extracted from the standardized sequence of each patient using a sliding window strategy, and then arranged into triplets according to spatial order. The window step size was set to 1 to ensure that there was overlap between adjacent samples but structural continuity was preserved. In the end, about 4,000 training samples were constructed, covering typical anatomical regions of the lung.

[0020] S2. Constructing a two-branch neural network framework for denoising and interpolation enhancement: Based on a triplet sample dataset, a denoising network is constructed to generate denoised triplets; based on a binary dataset, an interpolation enhancement network is constructed to generate axial structural offset estimates; the denoised triplets include: a denoised front slice, a denoised center slice, and a denoised back slice; this invention constructs a two-branch deep neural network joint learning framework, with one branch being a denoising network, and the other branch being a triplet... As input, the data is processed through a denoising network, and the output is a denoised triplet. Another branch is the difference enhancement network, which uses binary pairs. As input, the interpolation enhancement network outputs an estimate of the axial structural offset. The two network parameters are independent of each other and do not share weights.

[0021] In this embodiment, the denoising network adopts the RED-CNN network, which consists of a symmetrical encoder-decoder structure, including four levels of downsampling and upsampling modules. It fuses shallow details and deep semantic information through skip connections, and introduces a residual learning mechanism to stabilize training and improve denoising performance. This branch uses triples... As input, the corresponding denoising result is output. .

[0022] In this embodiment, the sum-difference enhancement network is based on the Attention U-Net design. Its backbone is a classic U-Net structure, containing four levels of downsampling and upsampling modules. Squeeze-and-Excitation (SE) channel attention modules are embedded at the skip connections in the encoder-decoder path, which can adaptively enhance the responses of salient features related to structural changes. The input to this branch is the structural difference tuple calculated from the original triples. The output is the axial offset estimated by the network. The denoising network and the interpolation enhancement network operate completely independently during training and inference, without sharing weights. The denoising network employs a convolutional neural network structure based on residual connections, while the interpolation enhancement network uses a variant of U-Net with an encoder-decoder structure. Independent parameter optimization for both networks allows for specialized learning targeting global noise suppression and high-frequency detail enhancement, respectively.

[0023] S3. Constructing a pseudo-supervision signal: Based on the denoised triplet and axial structural offset estimation, a pseudo-supervision signal is constructed and calculated to output the pseudo-supervision signal; this invention is based on the denoised triplet output by the denoising network. Axial structural offset estimation from the output of the sum-difference enhancement network Construct a pseudo-monitor signal S, and calculate its expression as follows: in, The front slice after noise reduction. The slices are after denoising. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

[0024] This invention observes that the difference in axial displacement between adjacent slices cannot be ignored. Combined with the setting of "noise being independently statistically analyzed between different slices", the average of the "front-to-back" values ​​in the triplet output by the denoising network is calculated. A high-quality pseudo-supervisory signal S is generated through mathematical reasoning to replace NDCT in supervised learning. A self-supervised loss function is set based on the pseudo-supervisory signal.

[0025] In this embodiment, during training, the system dynamically generates a high-quality pseudo-supervisory signal to replace the real, clean NDCT. Specifically, firstly, utilizing the medical priors of "statistical independence of noise in slices of different axes" and "approximately equal structure in adjacent slices," it is deduced that the ideal noise-free state of the central slice can be approximated by the mean of the preceding and following slices. Furthermore, combining the denoised triplet from the output of the dual-branch network and the axial structural offset estimation, a corrected pseudo-supervisory signal S is constructed. S not only integrates the similarity information of preceding and following slices to suppress noise, but also compensates for the geometric deviation caused by small axial displacements through structural offset estimation, significantly outperforming simple mean estimation and providing a high-fidelity supervisory target for self-supervised training.

[0026] S4. Constructing a self-supervised joint loss optimization function and training the model: Based on the pseudo-supervised signal and the denoised triples, a self-supervised joint loss optimization function is constructed and calculated to minimize the joint loss optimization function and simultaneously optimize the network parameters of the denoising network and the interpolation enhancement network. Training parameters are then set for training to obtain a model for low-dose CT image denoising, thus completing the self-supervised denoising method for low-dose CT images. The self-supervised joint loss optimization function includes: a self-supervised guided loss function, a pixel-level reconstruction loss function, and a structural similarity loss function. The training parameters include: an initial learning rate set to 1×10⁻⁶. 4The invention employs a batch size of 8, a maximum training epoch of 100,000, and an early stopping mechanism. It aims to optimize the difference between the denoised triples output by the dual-branch network and the pseudo-supervisory signal S by defining a joint loss optimization function. This function is a weighted sum of a self-supervised guided loss function, a pixel-level reconstruction loss function, and a structural similarity loss function. The pixel-level reconstruction loss improves pixel-level denoising accuracy, while the structural similarity loss constrains the consistency between the prediction result and the pseudo-supervisory signal in local structure. By minimizing this loss function, the network parameters of the two independent branches are simultaneously optimized, ultimately achieving a denoising model that simultaneously suppresses noise and restores high-fidelity details without paired NDCT. It is important to note that no NDCT images are used as supervision signals during any training process; the self-supervised signal is constructed entirely from the LDCT images themselves.

[0027] The construction of the self-supervised joint loss optimization function includes the following steps: (1) The self-supervised guided loss function is used to constrain the consistency between the central slice output by the denoising network and the pseudo-supervision signal. It effectively alleviates the structural ambiguity caused by small axial displacement by fusing the information of the slices before and after and the structural offset correction. The loss term is measured by the L2 norm, and the expression is as follows: in, The front slice after noise reduction. This is the center slice after denoising. The slices are after denoising. This is a false monitoring signal. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

[0028] (2) The pixel-level reconstruction loss function is used to constrain the consistency of the structural difference representation between the denoising network and the interpolation enhancement network. Specifically, the denoised triples output by the denoising network... The structural difference between the front and back slices was calculated. and the structure shift estimate directly output by the difference enhancement network Pixel-by-pixel comparison is performed; the loss term uses the L1 norm as a metric, and its mathematical expression is: By minimizing this loss term, the two independent branches are encouraged to work together to optimize the modeling of axial structural changes, effectively enhancing the recovery accuracy of lesion edges and minute textures.

[0029] (3) The structural similarity loss function is used to maintain the consistency of local texture, edge and axial structure and prevent over-smoothing. Its expression is: in, The structural similarity index measures the local consistency of two images in terms of brightness, contrast, and structural information, thereby effectively preserving lesion edges and texture details and suppressing excessive smoothing.

[0030] Finally, the self-supervised joint loss optimization function is a weighted sum of the above three terms: in, To preset the self-supervised guided loss weight coefficient, For pixel-level reconstruction loss weight coefficients, The pre-defined structural similarity loss weight coefficients are used; by minimizing this joint loss end-to-end, the two branches achieve synergistic optimization of noise suppression, detail preservation, and structural spatial continuity under the condition of unpaired conventional dose CT.

[0031] In this embodiment, the weight coefficients are set to 0.6, 0.3, and 0.1, respectively. This embodiment prioritizes ensuring the dominance of the self-supervised guided loss to drive the model to learn the core denoising mapping. Simultaneously, pixel consistency constraints and structure-aware regularization are used to effectively balance detail preservation and noise suppression while ensuring the correctness of the main denoising direction, preventing the model from falling into trivial solutions or excessive smoothing. This weight ratio is determined after obtaining the optimal PSNR and SSIM tradeoff on the validation set through grid search. The model is trained end-to-end using the Adam optimizer, with an initial learning rate set to 1×10⁻⁶. 4 The batch size is 8, the maximum training epochs are 100,000, and other parameters are default. To avoid overfitting, an early stopping mechanism is introduced: training is terminated when the PSNR improvement on the validation set is less than 0.05 dB for 5 consecutive epochs. Training begins by loading the deep learning network model and loading images in batches according to the batch size. Based on the set parameters, weights are output once after the set number of iterations, and the optimal weights are selected. The performance of the weights is quantitatively evaluated using the validation set, with evaluation metrics including Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). Finally, LDCT image denoising is performed, and the obtained optimal weights are loaded into the deep learning network framework. The experimental results after denoising are shown in Figure 2. The content circled by the ellipse is the region of interest. Subjective visual evaluation shows that the proposed method can significantly suppress noise, and the slice structure is continuous with no misalignment between layers, fully verifying the comprehensive advantages of this invention in noise suppression, detail preservation, and 3D structural consistency.

[0032] To verify and evaluate the noise reduction performance of this invention, a comparative experiment was conducted on the same test set with the Noise2Noise and Noise2Sim self-supervised methods. Indicators for the lung and abdominal regions were tested, and the quantitative results are shown in Table 1. This invention significantly outperformed the comparative methods in both PSNR and SSIM: in the lung region of the test set, this invention achieved PSNR = 37.68 dB and SSIM = 0.9107, while Noise2Noise's PSNR / SSIM was 32.63 dB / 0.7551, and Noise2Sim's PSNR / SSIM was 36.39 dB / 0.8606.

[0033] Table 1 Comparative Experimental Results Experimental results show that, on the AAPM-Mayo dataset, the PSNR of the lung region reached 37.68 dB and the SSIM was 0.9107, which is higher than the comparison method. Visual assessment also shows that the lung nodules have clear boundaries, continuous blood vessels, and consistent interlayer structure, which helps doctors to judge the morphology of lesions.

[0034] The qualitative visual comparison is shown in Figure 3, with the second row showing a magnified view of the region of interest within the red box. In the LDCT images recovered by this invention, the boundaries of calcifications on the liver are clear and sharp, and the slices show no obvious structural misalignment or inter-slice breaks. In contrast, other methods generally suffer from problems such as excessive smoothing, nodule blurring, or structural loss. Experimental results show that, without paired NDCT supervision, this method outperforms the comparative methods in terms of noise suppression, detail preservation, and axial structural continuity. The lung nodule boundaries in the denoised images are clear, and there are no obvious structural breaks between adjacent slices. The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the invention. Various modifications and improvements to the technical solutions of the present invention made by those skilled in the art without departing from the spirit of the invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A self-supervised denoising method for low-dose CT images based on axial structural consistency, characterized in that, Includes the following steps: S1. Data Preparation and Sample Construction: Input the original LDCT scan DICOM image sequence, perform preprocessing operations, and extract three consecutive axial slices to generate triplets. After processing the sample dataset, calculate the axial difference map between two adjacent slices to construct a tuple. Dataset; the preprocessing operations include: HU value conversion and amplitude limiting, linear normalization, image size unification, and manual screening; the tuple Specifically defined as: and The LDCT images were derived from lung and abdominal scans, with a slice thickness of 1. S2. Constructing a dual-branch neural network framework for denoising and interpolation enhancement: Based on the triplet sample dataset, a denoising network is constructed to generate denoised triplets; based on the binary dataset, an interpolation enhancement network is constructed to generate axial structural offset estimates; the denoised triplets include: a denoised front slice, a denoised center slice, and a denoised back slice; S3. Constructing a pseudo-supervisory signal: Based on the denoised triplets and axial structural offset estimates, a pseudo-supervisory signal is constructed and calculated to output the pseudo-supervisory signal; S4. Constructing a self-supervised joint loss optimization function and model training: Based on the pseudo-supervisory signal and the denoised triplets, a self-supervised joint loss optimization function is constructed and calculated to minimize the joint loss optimization function and simultaneously optimize the network parameters of the denoising network and the interpolation enhancement network, and training parameters are set for training to obtain a model for low-dose CT image denoising, thus completing the self-supervised denoising method for low-dose CT images; the self-supervised joint loss optimization function includes: a self-supervised guided loss function, a pixel-level reconstruction loss function, and a structural similarity loss function; the training parameters include: an initial learning rate set to 1×10⁻⁶. 4 The batch size is 8, the maximum number of training rounds is 100,000, and an early stopping mechanism is designed.

2. The self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 1, characterized in that, In step S1, the preprocessing operation specifically involves limiting the HU value window to [ Within the range of [1000, 1000], linearly normalized to the interval [0, 1], with a uniform image size of 512×512, and slices containing motion and metal artifacts are removed.

3. The self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 1, characterized in that, In S2, the denoising network adopts the RED-CNN network. The RED-CNN network consists of a symmetrical encoder-decoder structure, including 4 levels of downsampling and upsampling modules, and introduces skip connections and residual learning mechanisms.

4. The self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 1, characterized in that, In S2, the difference enhancement network adopts the Attention U-Net network. The Attention U-Net network has U-Net as its backbone and includes 4 levels of downsampling and upsampling modules. The SE channel attention module is embedded at the jump connection of the encoder-decoder path.

5. The self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 1, characterized in that, In step S3, the calculation expression for the output pseudo-monitoring signal is as follows: in, The front slice after noise reduction. The slices are after denoising. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

6. The self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 1, characterized in that, In S4, the expression for the self-supervised joint loss optimization function is: in, To pre-determine the weighting coefficients for the self-supervised guided loss, These are the pixel-level reconstruction loss weight coefficients. To preset the structural similarity loss weight coefficients, For self-supervised guided loss function, For pixel-level reconstruction loss function, This is the structural similarity loss function.

7. The self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 6, characterized in that, In S4, the expression for the self-supervised guided loss function is: in, The front slice after noise reduction. This is the center slice after denoising. The slices are after denoising. This is a false monitoring signal. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

8. A self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 6, characterized in that, In S4, the expression for the pixel-level reconstruction loss function is: in, The front slice after noise reduction. This is the center slice after denoising. The slices are after denoising. and This is the axial structural offset estimate of the output of the interpolation enhancement network.

9. A self-supervised denoising method for low-dose CT images based on axial structural consistency according to claim 6, characterized in that, In S4, the expression for the structural similarity loss function is: in, Represents the structural similarity index. This is the center slice after denoising. This is a false monitoring signal.

Images