Generalized low-dose CT image denoising method based on double-space covering Transform

Abstract

The invention discloses a generalizable low-dose CT image denoising method based on a double-space covering Transform for a cross-center / cross-protocol distribution offset scene, and belongs to the field of medical image processing and intelligent reconstruction. According to the method, DSM-Transform is used as a backbone, two kinds of complementary regularization of feature space covering and token interaction covering are introduced in training, and local embedding dependency and attention propagation paths are constrained respectively, so that noise over-fitting is inhibited, and structural fidelity and generalization robustness are improved. The network firstly carries out FFT on an LDCT image, after convolution embedding, masking processing and input and output end OWAL enhancement, the LDCT image is sent to a four-layer symmetric encoder-decoder to complete feature compression and reconstruction, and finally a denoising result is output through IFFT. Training adopts a patch strategy and randomly samples a masking proportion, and preferably, only token interactive masking is started in a test stage, so that noise suppression, detail preservation and cross-center generalization capabilities are enhanced on the premise of not remarkably increasing deployment complexity.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and intelligent reconstruction technology, specifically relating to a generalizable low-dose CT (LDCT) image denoising method for cross-center / cross-protocol distributed offset scenarios. This method particularly targets the Transformer denoising backbone network, introducing a dual-space masking regularization mechanism to enhance the model's adaptability to domain offset and unseen degradation patterns, thereby improving the generalization robustness of low-dose CT denoising. Background Technology

[0002] X-ray computed tomography (CT) scans, with their advantages of high imaging speed and spatial resolution, have been widely used in clinical settings such as disease screening, tumor localization, and treatment planning. However, CT examinations rely on ionizing radiation, and radiation exposure may pose potential health risks. To reduce radiation dose, clinicians often perform low-dose CT scans by lowering tube voltage and current. However, dose reduction often significantly enhances quantum noise and artifacts, thereby reducing image quality and weakening diagnostic reliability.

[0003] With the development of deep learning, low-dose CT (LDCT) denoising based on neural networks has gradually become a mainstream research direction in this field. Early representative methods mostly used convolutional neural networks (CNNs) as the core to build the denoising framework, such as RED-CNN with a residual encoder-decoder structure, which enhances detail recovery capabilities through symmetrical encoding / decoding structures and residual connections. Subsequently, to further improve perceptual quality and visual consistency, some studies introduced generative adversarial networks (GANs) frameworks, combining adversarial loss with perceptual loss to improve the realism and texture representation of the reconstruction results. Meanwhile, some works have improved denoising reliability from the perspectives of data quality control and adaptive modeling: for example, training data purification strategies that eliminate misaligned samples by evaluating the structural similarity of paired LDCT / NDCT patches, and adaptive filtering frameworks based on pixel-level dynamic kernel prediction and adaptively determining the number of iterations. In recent years, diffusion models have also been introduced into LDCT-related tasks, achieving generative modeling across doses and anatomical regions by encoding dose and anatomical priors and introducing conditional diffusion mechanisms. Overall, these methods have continuously improved the denoising performance of LDCT, but in real clinical deployments, performance may still degrade due to the reliance on the interaction of specific noise statistics and random features in the training data, limiting its generalization ability in distribution offset scenarios.

[0004] On the other hand, the Transformer architecture, with its ability to model long-range dependencies and global context, has been gradually introduced into LDCT denoising tasks and has shown potential in low-contrast structure recovery. For example, existing studies have used a pure Transformer encoder-decoder framework to achieve good denoising results on datasets such as Mayo; other works have introduced cross-angle self-attention in the chord map domain to explicitly model the correlation between different projections; other methods enhance long-range information modeling capabilities through token expansion and translation mechanisms, and combine them with sliding window strategies to reduce edge artifacts; in addition, for 3D CT denoising, some studies have proposed combining efficient attention with convolutional feedforward structures to reduce computational overhead and improve volumetric data processing efficiency. Despite the continuous progress made by Transformer-type methods, two prominent problems still exist in practical applications: First, in order to control computational complexity, many methods adopt local windows or restricted attention mechanisms, which objectively weaken the advantage of global interaction in capturing long-range dependencies and limit the ability to model the overall consistency of low-contrast structures; Second, and more importantly, the noise statistics and degradation mechanisms of cross-center LDCT data are highly diverse. If the model relies too much on the noise patterns specific to the training domain and the accidental interactions they induce during the training phase, it will often show significant performance degradation on new center data and insufficient generalization robustness.

[0005] In recent years, masking-based representation learning has received widespread attention in the field of computer vision. These methods construct missing information by proportionally and randomly masking input or intermediate features, forcing the network to rely more on global context and semantic structure rather than local texture details or noise interference, thereby improving the robustness and generalization ability of the representation. For example, high-proportion masking reconstruction paradigms such as MAE have been shown to learn transferable representations with high efficiency and have been further extended to low-level visual tasks such as denoising to enhance the model's robustness to various types of degenerate noise.

[0006] While existing research has made some progress in areas such as network architecture design, generative modeling, unsupervised / self-supervised learning, and domain adaptation, current solutions for cross-center LDCT denoising still generally suffer from the following shortcomings: First, during the training phase, models tend to learn "noise-specific patterns" that are statistically related to the noise in the training data and rely on accidental feature interactions, leading to sensitivity to out-of-distribution degradation. Second, imposing constraints only on the input layer or pixel layer often fails to simultaneously suppress "representation layer overfitting" and "attention interaction overfitting." Third, in cross-center deployments, the superposition of factors such as noise levels, scanning protocols, scanner types, and patient size differences can easily lead to problems such as insufficient detail recovery, blunted edges, and even structural distortion. Based on these limitations, there is an urgent need for a technical solution that can simultaneously suppress noise-specific overfitting from both the representation space and interaction space, and promote the network to learn structure-driven and transferable representations, without significantly increasing deployment complexity. This would improve the robustness and generalization ability of supervised LDCT denoising models in cross-center distribution migration scenarios. Summary of the Invention

[0007] Existing supervised denoising methods for low-dose CT (LDCT) typically achieve good results in co-center testing. However, when deployed across centers, scanning protocols, or devices, they often exhibit significant performance degradation due to changes in noise statistics and degradation mechanisms. A fundamental reason for this is that networks tend to over-rely on specific noise patterns and their induced random feature associations during training, forming a "memory" of domain-specific noise. This leads to unstable structure recovery, local grayscale shifts, or loss of detail under distribution shift conditions. Simultaneously, existing regularization strategies often only apply to input pixels or single feature layers, failing to simultaneously constrain the dependence on noise cues in the representation space and the unreliable token-to-token information transfer and aggregation paths in the interaction space. Based on these limitations, there is an urgent need for an LDCT denoising technique that can synergistically suppress noise-specific overfitting at both the representation and interaction levels, guiding the model to learn structure-driven and transferable representations without significantly increasing deployment complexity.

[0008] To address the aforementioned problems, this invention proposes a generalizable low-dose CT image denoising method based on dual-spatial masking Transformer. This method is based on a multi-scale contextual Transformer denoising backbone network and introduces complementary masking mechanisms—feature-space masking and token-interaction masking—at two key stages during training. These mechanisms synergistically suppress noise-specific overfitting, enhance the model's robustness in scenarios with cross-center distribution offsets, and improve the structural fidelity of the reconstruction results. This invention mainly includes the following steps:

[0009] Step 1: Acquire conventional dose CT (NDCT) projection data from different patients, and reconstruct the projection data using the filtered back projection (FBP) algorithm to obtain the corresponding NDCT images, thereby establishing an NDCT image library. Subsequently, noise is superimposed on the same set of NDCT projection data, and FBP is used again for reconstruction to simulate the corresponding quarter-dose CT images, establishing an LDCT image library. Finally, NDCT and LDCT images are paired according to a one-to-one correspondence to form training samples, constructing a paired sample set for network training; each sample is denoted as (x, y), where x represents the LDCT image and y represents the corresponding NDCT image.

[0010] Step 2: Construct the DSM-Transformer denoising network, which mainly consists of three parts: input embedding, latent feature encoding-decoding, and output projection. The latent feature encoding-decoding part is composed of multiple Transformer modules stacked together, forming a four-layer symmetrical encoder-decoder structure.

[0011] Step 3: During the training phase, enable the following two masking mechanisms simultaneously:

[0012] Applying feature-space masking to the early convolutional feature maps obtained during the input embedding stage randomly obscures some local embedding cues, forcing the network to rely on the visible context and globally consistent anatomical structure to complete reconstruction during training. This suppresses the "memory" of noisy patterns and promotes representation learning dominated by structural information.

[0013] In the self-attention computation process of Transformer, token-interaction masking is introduced to randomly mask some attention-map entries in the attention map in order to regularize token-to-token interactions and suppress unreliable feature association propagation paths, thereby guiding the network to pay more attention to transferable contextual structure information.

[0014] Step 4: In a preferred implementation, to improve training efficiency, a patch training strategy is adopted: the 512×512 image is cropped into 64×64 patches, and the patches are randomly sampled to form training batches; the optimizer is Adam, and the initial learning rate is 1×10⁻⁶. -4Furthermore, the learning rate is halved every 6000 training iterations, with a total training duration of approximately 200 epochs. Simultaneously, the feature space masking ratio and the token-interaction masking ratio are randomly sampled within preset ranges to enhance training stability and avoid model adaptation bias to single masking intensities. Experimental results show that randomly sampled masking ratios outperform fixed-ratio training. More preferably, randomly sampling the feature space masking ratio within the 75%-85% range yields better overall performance; the optimal range for token-interaction masking is 60%-70%.

[0015] Step 5: Use the trained model to perform denoising inference on the input LDCT image and output the denoising result. In a preferred embodiment, only token-interaction masking is enabled during the testing phase to stabilize attention interaction behavior and improve output consistency across center-of-care scenarios.

[0016] The advantages of this invention compared to the prior art are as follows:

[0017] (1) The generalizable low-dose CT image denoising method based on dual-space occlusion Transformer proposed in this invention introduces complementary masking constraints in the representation space and interaction space to suppress the model's "memory" of noise patterns in the training domain and the propagation of noise-related pseudo-associations, thereby prompting the model to learn a representation dominated by structural information, thus significantly improving its robustness and generalization ability in cross-center / outside-distribution scenarios.

[0018] (2) Feature space masking provides a structural consistency learning objective of “missing completion”, while token-interaction masking further standardizes the attention interaction path and token aggregation behavior. The two work together to enable the model to achieve a better trade-off between noise / artifact suppression and structural fidelity, and improve the ability to preserve diagnostic-related low-contrast information.

[0019] (3) The dual-space masking denoising method proposed in this invention is a pluggable solution that can be flexibly adapted to denoising backbone networks such as multi-scale context Transformer; without changing the overall training supervision signal (with paired NDCT as a reference), it can effectively improve the stability of cross-center deployment. Attached Figure Description

[0020] Figure 1 : A generalizable low-dose CT image denoising network based on dual spatial occlusion Transformer (DSM-Transformer).

[0021] Figure 2: A Transformer representation module (masking block) that integrates token-interaction masking;

[0022] Figure 3 Multi-scale Context Feedforward Network (MCFN) and its Enhanced Context Module (eCoT); Detailed Implementation

[0023] To make the above-mentioned objectives, features, and advantages of the present invention more readily understood, the specific embodiments of the present invention will be described in detail below with reference to examples. The following content is only for illustrative purposes and explanation of the inventive concept. Those skilled in the art can make various modifications, additions, or similar substitutions to the specific embodiments described, as long as they do not deviate from the core concept of the present invention or exceed the scope defined by the claims, they should all be included in the protection scope of the present invention.

[0024] The preparation method of the present invention will be described in detail below through specific embodiments.

[0025] Example

[0026] This embodiment provides a generalizable low-dose CT image denoising network (DSM-Transformer) employing a dual-spatial-masking Transformer. The network is based on a frequency-domain multi-scale contextual Transformer denoising backbone and consists of three parts: input embedding, latent feature encoding-decoding, and output projection. The encoder-decoder employs a four-layer symmetrical structure, formed by stacking multiple Transformer modules. Compared to existing common structures, this network differs in that it introduces Operationally Aware Attention Layers (OWALs) at both the input and output ends to enhance its adaptability to heterogeneous noise patterns. Furthermore, it introduces complementary dual-spatial-masking regularization mechanisms at two key locations: feature-space masking and token-interaction masking. These mechanisms suppress overfitting to specific noise patterns and accidental interactions at the "representation space" and "interaction space" levels, respectively.

[0027] like Figure 1As shown, the input LDCT image is first subjected to a Fast Fourier Transform (FFT) to obtain a frequency domain representation. Then, it is embedded through a 3×3 convolutional layer, and a feature space mask is applied to the early feature map after embedding. The masked shallow features are then enhanced by several OWAL modules before being introduced into a token-interactive masking Transformer encoder-decoder to complete the compression and reconstruction of latent features. The reconstructed deep features are further refined by several OWAL modules. Finally, a 3×3 convolutional mapping is used to obtain the residual spectrum, which is added to the original spectrum to obtain the denoised spectrum. This denoised spectrum is then restored to the spatial domain image using an Inverse Fast Fourier Transform (IFFT), thus outputting the final denoised result.

[0028] In this embodiment, feature space masking is used to suppress the network's over-reliance on noise statistics in the training data, prompting the model to learn more anatomically driven and transferably applicable reconstructed representations in cross-center scenes. Specifically, during the training phase, early convolutional feature maps after input embedding are randomly masked: let the input embedding features be f∈R H×W×C In the spatial dimension, with an occlusion ratio p IM By randomly sampling certain locations and setting the feature vectors at those locations to zero, a "feature missing completion" learning task is constructed. Through this mechanism, the network can still reconstruct the entire image while retaining only some visible embedding cues, thereby reducing the tendency to memorize local noise patterns and enhancing the utilization of globally consistent structural cues.

[0029] like Figure 2 As shown, a Transformer representation module integrating token-interaction masking is used as the basic unit in the latent feature encoding-decoding backbone. Each module first performs layer normalization on the input features; then, token-interaction masking is performed before self-attention computation to randomly mask some entries in the attention map, thereby constraining unreliable token-to-token information transmission paths and adjusting the attention distribution; then, multi-head deep transposed attention (MDTA) is introduced to model long-range dependencies and global context; finally, a multi-scale context feedforward network (MCFN) is used to further characterize multi-scale local dynamic context information, thereby enhancing the ability to express fine-grained structures.

[0030] Token-interaction masking is performed before computing the MDTA attention graph: with a given probability p TMBy dynamically sampling some attention-map entries and setting them to zero in the attention map, the corresponding token-to-token information transmission path is blocked, limiting the impact of unreliable interactions on feature aggregation. Through this mechanism, attention computation is forced to rely more on unmasked interaction relationships to complete information integration, thereby forming a more robust and transferable attention distribution that focuses more on universal contextual cues.

[0031] The MDTA module aims to reduce the time and storage overhead of traditional spatial self-attention on high-resolution features, which increases quadratically with the number of pixels. Let the input feature map X∈R C×H×W MDTA first normalizes X, then generates Q, K, and V through 1×1 convolution and 3×3 depthwise separable convolution. Its attention calculation can be expressed as: ,

[0032] Here, α is a learnable scaling factor that controls the numerical stability of the attention weights; the operation φ(⋅) represents reshaping the output to R. C×H×W Then, a 1×1 convolution is applied to enhance the feature fusion capability.

[0033] like Figure 3 As shown, to more fully model multi-scale local dynamic context information in the feedforward branch, we designed a Multi-Scale Context Feedforward Network (MCFN). The MCFN consists of two cascaded Enhanced Context Modules (eCoT) to simultaneously capture multi-scale static context and dynamic adaptive context. Given input features F... C ∈R H×W×C eCoT generates Q, K, and V through 1×1 convolutions and extracts multi-scale local context using parallel 3×3 convolutions and dilated convolutions. It then fuses multi-scale information to form attention-enhanced features and uses Selective Kernel Attention (SK) for adaptive weighting to improve texture and detail representation. As a complement to MDTA, this design can more stably fuse multi-scale local dynamic context information in LDCT scenes with varying noise distributions, thereby improving detail recovery and artifact suppression. Its computational process can be summarized as follows: ,

[0034] in, ξ represents a dilated convolution operation with a dilation rate of 2, Conv represents a standard convolution operation, ξ represents the ReLU activation function, and ⊛ represents a local matrix multiplication operation.

[0035] To enhance the network's adaptability to complex degradations such as granular noise and stripe artifacts in LDCT, Operation-Aware Attention Layers (OWALs) are set at the input and output ends of the model. OWALs employ a multi-branch parallel operator structure, with branches covering operations such as depthwise convolution, dilated depthwise convolution, and average pooling with different kernel sizes. Learnable attention weights are used to dynamically select and weightedly fuse the outputs of each branch. The fusion result is then channel-integrated by (1×1) convolution and residually superimposed with the input features for output, thereby achieving adaptive enhancement and robust suppression of different noise patterns.

[0036] In one embodiment, to balance pixel-level fidelity and structure-aware quality, a weighted composite loss function is used during training. The loss is jointly optimized using structural similarity (SSIM) loss, and the total loss takes the form of: ,

[0037] Where λ is the weighting coefficient used to balance the contributions of the two.

[0038] During the training phase, paired LDCT and NDCT images were used as the training set to train and evaluate the constructed model. Images were cropped into 64×64 pixel patches as input to improve training performance, and the initial learning rate was set to 1×10⁻⁶. -4 The learning rate is halved after every 6000 iterations. The batch size is set to 4, and the network continues to be trained until the loss shows no improvement after 200 iterations. Simultaneously, dual-space masking regularization is introduced during training, where the masking ratios for feature-space masking and token-interaction masking are preferably 75%-80% and 60%-70%, respectively.

[0039] In summary, this invention proposes a DSM-Transformer method for robust low-dose CT denoising across centers. By simultaneously applying complementary constraints to the representation space and interaction space, it suppresses overfitting of the model to specific noise patterns and accidental feature interactions in the training data. Specifically, this invention introduces feature-space masking on early features after input embedding during the training phase and token-interaction masking during the Transformer self-attention interaction process, prompting the network to rely more on globally consistent anatomical cues and transferable contextual information for reconstruction. Simultaneously, the network backbone employs an efficient Transformer denoising architecture: modeling global contextual dependencies through multi-head deep transposed attention (MDTA) and capturing multi-scale local contextual features through a multi-scale contextual feedforward network (MCFN). Therefore, more stable detail recovery and artifact suppression can be achieved in LDCT scenarios with varying noise distributions, and the generalization ability under cross-center testing conditions can be further improved.

Claims

1. A generalizable low-dose CT image denoising method based on dual-spatial occlusion Transformer, characterized in that, Includes the following steps: Step 1: Collect conventional dose CT (NDCT) projection data from different patients and reconstruct NDCT images using filtered back projection (FBP) to establish an NDCT image library; superimpose noise on the same set of NDCT projection data and reconstruct using FBP again to simulate and generate quarter-dose CT images to establish a low-dose CT (LDCT) image library; pair LDCT images with their corresponding NDCT images in a one-to-one correspondence to form paired samples (x, y), where x is the LDCT image and y is the corresponding NDCT image; Step 2: Construct a DSM-Transformer denoising network. The denoising network includes at least input embedding, latent feature encoding-decoding, and output projection. The latent feature encoding-decoding consists of a four-layer symmetrical encoder-decoder structure formed by stacking multiple Transformer modules. Step 3: During the training phase, both feature-space masking and token-interaction masking are enabled simultaneously: feature-space masking is applied to the early convolutional feature maps obtained during the input embedding phase, and token-interaction masking is introduced during the Transformer self-attention calculation process to randomly mask some attention entries in the attention map. Step 4: Train the DSM-Transformer denoising network using a patch training strategy, and randomly sample the feature space masking ratio and the token-interaction masking ratio within a preset range during the training process. Step 5: Use the trained model to perform denoising inference on the input LDCT image and output the denoising result.

2. The generalizable low-dose CT image denoising method based on dual-spatial occlusion Transformer according to claim 1, characterized in that, The DSM-Transformer denoising network in step 2 includes the following sub-steps: Step 2.1: Perform a Fast Fourier Transform (FFT) on the input LDCT image to obtain its frequency domain representation; Step 2.2: Input embedding of the frequency domain representation through a convolutional layer, and apply feature space masking to the early feature map after embedding; Step 2.3: After the masked shallow features are enhanced by inputting several Operational Aware Attention Layers (OWALs), they are input into the Transformer encoder-decoder with token-interaction masking to complete the latent feature compression and reconstruction. The Transformer module includes a multi-head deep transposed attention (MDTA) module and a multi-scale context feedforward network (MCFN) module. The MDTA module first fuses local context information through deep convolution, and then performs attention interaction in the channel dimension to capture global correlations with linearly scalable computational complexity. The MCFN module consists of cascaded enhanced context modules eCoT, which are used to simultaneously model multi-scale static contexts and dynamic adaptive contexts, and achieve adaptive fusion through selective kernel attention. Step 2.4: Input the reconstructed deep features into several OWALs for feature refinement; Step 2.5: Obtain the residual spectrum through convolution mapping and add it to the original spectrum to obtain the denoised spectrum. Then, restore it to the spatial domain through inverse fast Fourier transform (IFFT) and output the denoised CT image.

3. The generalizable low-dose CT image denoising method based on dual-spatial occlusion Transformer according to claim 1, characterized in that, Step 3 includes: Step 3.1: During the training phase, the early convolutional feature maps f∈R after input embedding are processed. H×W×C According to the occlusion ratio p IM Several positions are randomly sampled in the spatial dimension, and the feature vectors corresponding to the sampled positions are set to zero vectors to form a learning constraint for feature missing completion; its p IM Random sampling was conducted within the 75%-85% range. Step 3.2: During the training phase, before calculating the MDTA attention map, according to the occlusion probability p... TM Randomly sample and set some attention entries in the attention graph to zero to block the corresponding token-to-token information transmission path and regularize the attention aggregation process; where p TM Random sampling was conducted within the 60%-70% range.

4. The generalizable low-dose CT image denoising method based on dual-spatial occlusion Transformer according to claim 1, characterized in that, Step 4 includes: cropping the 512×512 image into 64×64 patches and randomly sampling them to form training batches; using the Adam optimizer with an initial learning rate of 1×10⁻⁶. −4 The learning rate is reduced to half every 6000 training iterations, with a total training time of approximately 200 epochs. A composite loss function consisting of L1 loss and structural similarity SSIM loss is used to supervise the learning of the output denoised image and the corresponding NDCT image.

Images