A method and system for training a lung nodule detection model

Abstract

The application discloses a kind of lung nodule detection model training method and system, method includes: obtaining standard dose CT image, in sinusoidal domain is simulated and is reconstructed according to multidimensional physical degradation parameter, generates low dose degradation CT image sample set and corresponding physical degradation parameter set;Sample set is extracted by shared feature and three-way variation coding, obtain three kinds of hidden variables of nodule content, degradation and anatomical structure;Two two mutual information constraints are established to three hidden variables, auxiliary supervision regression constraint based on physical degradation parameter set is established to degradation hidden variable, and regular constraint is applied to the posterior distribution of three hidden variables;Cross attention is used to fuse nodule content hidden variable and anatomical structure hidden variable, input detection head calculates detection loss;Combined with the above each loss, construct joint optimization target, and the model is trained to end-to-end until convergence.This application improves the adaptability of detection model to different low dose CT conditions.

Description

Technical Field

[0001] This invention relates to the field of medical image analysis technology, and in particular to a training method and system for a lung nodule detection model. Background Technology

[0002] Low-dose computed tomography (CT) scans have been widely used for early lung cancer screening. Currently, deep learning-based computer-aided detection (CADe) of lung nodules has achieved good performance on standard-dose CT data. However, in practical clinical applications, due to differences in medical equipment, scan dose, scanning protocol, detector status, and image reconstruction algorithms, low-dose CT images exhibit complex degradation characteristics, making it difficult for detection models trained under specific data conditions to maintain stable cross-device and cross-protocol generalization performance. Existing low-dose CT methods for lung nodule detection mainly face the following problems: 1. Physical degradation simulations exhibit distributional shifts compared to actual imaging processes. Existing methods are largely confined to the image domain, employing simple perturbations such as Gaussian or Poisson noise, or data augmentation using limited high-dose / low-dose paired data. This strategy fails to deeply characterize the imaging physics within the sinusoidal domain, including X-ray photon propagation, detector statistics, and reconstruction algorithms, thus making it difficult to represent complex degradation factors such as reconstruction kernel differences and scattering effects. Consequently, significant distributional differences exist between simulated data and real clinical scan data, severely limiting the model's generalization ability across different devices and protocols.

[0003] 2. Feature representation coupling weakens the semantic stability of lesions. Traditional end-to-end detection networks encode heterogeneous information such as nodule semantics, degradation, and anatomical background in a shared feature space. Under low-dose conditions, feature fluctuations caused by noise and artifacts easily interfere with the core semantic expression of lesions, and there is a lack of explicit decoupling constraints for different information factors. This makes the model unstable when facing changes in dose or reconstruction conditions, especially unfavorable for the reliable identification of small nodules and low-contrast nodules.

[0004] 3. Neglecting anatomical spatial context priors leads to increased false positives. Existing models largely rely on local texture features, failing to fully utilize the overall anatomical structures such as pulmonary vessels, airways, and pleura as spatial constraints. Low-dose CT, due to increased noise and decreased contrast, further weakens the boundary information between nodules and normal tissues. This makes it easy for models to misclassify normal structures as candidate nodules in areas of vascular intersection or tissue overlap, resulting in a high false positive rate and affecting the accuracy of clinical auxiliary diagnosis.

[0005] In response to the limitations of the existing methods at the three levels mentioned above, the applicant proposes an improvement plan. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention proposes the following technical solutions, aiming to solve the problems of insufficient generalization ability of lung nodule detection models due to deviations between traditional image domain degradation simulation and the actual imaging process during low-dose CT lung nodule detection, as well as the increase in missed and false detections due to feature space hybrid encoding and lack of anatomical structure constraints.

[0007] In a first aspect, the present invention proposes a robust training method for low-dose CT lung nodule detection, comprising the following steps: Step 1: Acquire standard dose CT images and convert them to the sinusoidal domain through orthogonal projection. In the sinusoidal domain, perform physical degradation simulation according to multidimensional physical degradation parameters to generate degradation sinusoidal data. Then, generate a low-dose degradation CT image sample set and the corresponding physical degradation parameter set through image reconstruction. Step 2: Input the low-dose degraded CT image sample set into a shared feature extraction network to obtain basic features; input the basic features into a three-way variational coding network for processing to obtain nodule content latent variables, degraded latent variables, and anatomical structure latent variables; Step 3: Establish mutual information constraints for the pairwise combinations of the three latent variables to obtain mutual information constraint loss; establish auxiliary supervised regression constraints for the degenerate latent variables based on the physical degradation parameter set to obtain physical parameter regression loss; apply regularization constraints to the posterior distribution of the three latent variables to obtain latent variable distribution regularization loss; and use a cross-attention mechanism to fuse nodule content latent variables and anatomical structure latent variables to generate fused detection features, input the fused detection features to the 3D target detection prediction head, and calculate the 3D nodule detection loss. Step 4: Combining the mutual information constraint loss, physical parameter regression loss, latent variable distribution regularization loss, and three-dimensional nodule detection loss, construct an overall joint optimization objective function, and perform end-to-end joint parameter updates on the lung nodule detection model until the model converges, thus obtaining a trained lung nodule detection model.

[0008] Furthermore, the multidimensional physical degradation parameters include: The dose scaling factor is used to describe the proportional relationship between the simulated scan dose and the standard dose conditions. The style parameters of the reconstructed convolution kernel are used to continuously characterize the spatial frequency response features corresponding to different reconstructed convolution kernels; The detector electronic noise intensity parameter is used to quantitatively characterize the impact of electronic noise generated by the CT detector data acquisition system during signal conversion, transmission, and amplification. The scattered radiation ratio parameter is used to describe the background shift effect caused by scattered photons on the signal received by the detector; The reconstruction strategy fusion parameter describes the impact of different reconstruction strategies on the texture and noise distribution of the final image.

[0009] Furthermore, the physical degradation simulation in step 1 includes: adjusting the number of incident photons in the sinusoidal domain according to the dose scaling factor, introducing a background term for scattered photons in combination with the scattered radiation scaling parameter, simulating the statistical fluctuations of photon counts under low-dose conditions using a Poisson distribution, superimposing Gaussian electronic noise based on the detector electronic noise intensity parameter, obtaining the actual photon count measurement value received by the CT detector, and obtaining the degradation sinusoidal data by logarithmic transformation of the photon count measurement value.

[0010] Furthermore, the image reconstruction in step 1 includes: using the reconstruction strategy fusion parameters to perform weighted fusion of different reconstruction methods to generate a low-dose degraded CT image sample set corresponding to different combinations of physical degradation parameters.

[0011] Furthermore, step 1 also includes dynamically sampling and scheduling the continuous degradation parameter space constituted by the multidimensional physical degradation parameters, specifically configured as follows: The continuous degradation parameter space is discretized into multiple degradation sub-regions; In the initial stage of training, uniform sampling is performed on each degenerate sub-region; A portion of samples are reserved from each degenerate sub-region as validation samples. During the training process, after each round of training, the three-dimensional nodule detection loss of the lung nodule detection model on the corresponding validation samples in each degenerate sub-region is calculated, and the average detection loss of multiple validation samples in each degenerate sub-region is calculated respectively. The sampling weights of each degenerate sub-region in the next training phase are adjusted based on the average detection loss of each degenerate sub-region. Degenerate sub-regions with relatively larger average detection losses receive relatively higher sampling weights.

[0012] Furthermore, the three-way variational coding network includes a nodule content variational coding head, a degeneracy variational coding head, and an anatomical structure variational coding head. The three coding heads respectively map the basic features to variational approximate posterior distributions of nodule content latent variables, degeneracy latent variables, and anatomical structure latent variables, and obtain the nodule content latent variables, degeneracy latent variables, and anatomical structure latent variables by sampling through a reparameterization method.

[0013] Furthermore, the establishment of mutual information constraints in step 3 specifically includes: using the CLUB estimator to construct conditional probability estimation networks for three pairs of combinations: nodule content latent variable and degeneration latent variable, nodule content latent variable and anatomical structure latent variable, and degeneration latent variable and anatomical structure latent variable, respectively, estimating the upper limit of mutual information of the combinations, and reducing the information correlation between the three latent variables by minimizing the upper limit of mutual information of each combination.

[0014] Furthermore, the establishment of auxiliary supervised regression constraints specifically includes: obtaining physical degradation parameter estimates based on the predicted degradation latent variables, and constructing a physical parameter regression loss based on the difference between the physical degradation parameter estimates and the actual physical degradation parameters corresponding to the physical degradation parameter set.

[0015] Furthermore, the cross-attention mechanism in step 3 includes: expanding the nodule content latent variable and the anatomical structure latent variable into token sequences respectively, using the nodule content latent variable as query information and the anatomical structure latent variable as key information, calculating attention weights, and retaining the original nodule content information through residual connections.

[0016] Secondly, the present invention provides a training system for a lung nodule detection model to implement the above-mentioned training method, comprising: The sinusoidal domain multidimensional degradation manifold generation module is used to acquire standard dose CT images and perform physical degradation simulation in the sinusoidal domain according to multidimensional physical degradation parameters. The low dose degradation CT image sample set and the corresponding physical degradation parameter set are generated through image reconstruction. The three-way feature variational orthogonal decoupled coding module is used to extract nodule content latent variables, degradation latent variables, and anatomical structure latent variables from the low-dose degraded CT image sample set through shared feature extraction and a three-way variational coding network; establish mutual information constraints on the pairwise combination relationships between the three latent variables to obtain mutual information constraint loss; establish auxiliary supervised regression constraints on the degradation latent variables based on the physical degradation parameter set to obtain physical parameter regression loss; and apply regularization constraints to the posterior distribution of the three latent variables to obtain latent variable distribution regularization loss. The cross-path feature fusion module is used to employ a cross-attention mechanism, using nodule content latent variables as query information and anatomical structure latent variables as key-value information, to fuse nodule content latent variables and anatomical structure latent variables to generate fused detection features. The 3D nodule detection and prediction module is used to decode the fused detection features using a 3D target detection prediction head, and output the 3D bounding box of the lung nodule and the nodule presence confidence score; calculate the 3D bounding box regression loss based on the difference between the 3D bounding box and the corresponding ground truth bounding box, calculate the classification loss based on the difference between the nodule presence confidence score and the corresponding ground truth class label, and obtain the 3D nodule detection loss from the 3D bounding box regression loss and the classification loss. The joint optimization module is used to perform a weighted summation of the mutual information constraint loss, physical parameter regression loss, latent variable distribution regularization loss, and three-dimensional nodule detection loss to construct an overall joint optimization objective function and perform end-to-end joint parameter updates for the lung nodule detection model.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention establishes a five-dimensional continuous physical degradation parameter space by moving the degradation simulation forward to the sinusoidal domain, thereby realizing the parametric simulation of the main degradation factors in CT imaging. By continuously sampling degradation parameters to generate training samples covering different imaging conditions, the adaptability of the detection model to different low-dose CT imaging conditions is improved.

[0018] 2. This invention explicitly decouples nodule content, degradation, and anatomical structure information to reduce the interference of degradation factors on detection features; by fusing cross-path attention of anatomical structure and content latent variables, it introduces original lung nodule content information to assist in detection, thereby improving the discrimination stability of the detection model in complex areas such as low-dose noise and vascular crossings.

[0019] 3. This invention uses physical degradation parameters as self-supervised information to constrain the learning of degradation latent variables, eliminating the need for additional manual annotation; during the inference phase, degradation latent variables are not invoked, and only nodule content latent variables and anatomical structure latent variables are used for detection, which improves the efficiency and applicability of actual deployment. Attached Figure Description

[0020] Figure 1 This is a block diagram showing the overall composition of the training system for a lung nodule detection model proposed in this invention. Figure 2 This is a flowchart of the processing flow of the sinusoidal graph domain multidimensional degenerate manifold generation module of the present invention; Figure 3 This is a flowchart of the processing flow of the three-way feature variational orthogonal decoupling coding module of the present invention; Figure 4 This is a flowchart of the cross-path feature fusion module of the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0022] like Figure 1 As shown, the training system for a lung nodule detection model proposed in this invention includes a sinusoidal domain multidimensional degenerate manifold generation module, a three-way feature variational orthogonal decoupling coding module, a cross-path feature fusion module, a three-dimensional nodule detection and prediction module, and a joint optimization module.

[0023] The sinusoidal domain multidimensional degradation manifold generation module, serving as the physical data generation module, establishes a K-dimensional continuous physical degradation parameter space in the sinusoidal domain, including dose scaling factors, reconstruction convolution kernel style parameters, detector electronic noise intensity parameters, scattered radiation scaling parameters, and reconstruction strategy fusion parameters. This module generates a low-dose CT image sample set covering continuous degradation states through forward projection of standard-dose CT images, physical degradation simulation, and image reconstruction. This provides physically consistent training data for subsequent detection models, addressing the technical problem of distortion in traditional image domain degradation simulation.

[0024] The three-way feature variational orthogonal decoupled coding module, acting as a feature representation constraint module, utilizes a shared feature extraction network and a three-way variational coding network to map input image features to three latent variable representation spaces: nodule content, degradation, and anatomical structure, obtaining latent variables for nodule content, degradation, and anatomical structure. It then establishes mutual information constraints on the pairwise combinations of the three latent variables using a CLUB (Contrastive Log-ratio Upper Bound) estimator, establishes auxiliary supervised regression constraints on the degradation latent variables based on the physical degradation parameter set to obtain physical parameter regression loss, and applies regularization constraints to the posterior distributions of the three latent variables to obtain latent variable distribution regularization loss. This reduces information redundancy between different paths, effectively decoupling degradation factors from lesion semantic features, and addresses the problem of decreased detection performance caused by feature mixing coding.

[0025] The cross-path feature fusion module, as an anti-degeneration feature fusion module, is used to establish the association between lesion semantics and lung spatial structure by utilizing the cross-attention mechanism, and to fuse latent variables of nodule content and latent variables of anatomical structure to improve the detection stability in complex regions, thereby addressing the problem of increased false positives caused by the lack of structural priors.

[0026] The 3D nodule detection and prediction module, as a task output module, is used to perform 3D target detection on the fused robust features. It outputs the 3D bounding box of the lung nodule and the nodule presence confidence score through the 3D target detection prediction head. It calculates the 3D bounding box regression loss based on the difference between the 3D bounding box and the corresponding ground truth bounding box, and calculates the classification loss based on the difference between the nodule presence confidence score and the corresponding ground truth class label. The 3D nodule detection loss is composed of the 3D bounding box regression loss and the classification loss.

[0027] The multi-objective joint optimization module, as the overall training control module, is used to perform weighted summation of the mutual information constraint loss, physical parameter regression loss, latent variable distribution regularization loss, and three-dimensional nodule detection loss. By constructing an overall joint optimization objective function, the parameters of each network module are updated collaboratively, so that the detection model can achieve a balance between detection accuracy and degradation robustness.

[0028] The specific implementation methods of the above modules are described in detail below. It should be noted that this implementation method, while describing the specific workflow of each functional module, also corresponds to each step of the training method of this invention. Specifically, the workflow of the sinusoidal graph domain multidimensional degenerate manifold generation module corresponds to step 1; the workflow of the three-way feature variational orthogonal decoupling coding module corresponds to the constraint establishment in steps 2 and 3; the workflow of the cross-path feature fusion module and the three-dimensional nodule detection and prediction module corresponds to the feature fusion and three-dimensional nodule detection loss calculation in step 3; and the workflow of the multi-objective joint optimization module corresponds to step 4.

[0029] I. Module for Generating Multidimensional Degenerate Manifolds in the Sine Graph Domain This module is primarily used to construct training samples with multidimensional physical degradation features during the training phase of a low-dose CT lung nodule detection model. Unlike existing methods that typically directly superimpose artificial noise into the reconstructed image domain or employ discrete data augmentation under finite conditions, this module moves the degradation process forward to the sinusoidal domain within the CT physical imaging chain. By parametrically modeling factors such as dose ratio changes, photon statistical fluctuations, detector electronic noise, the effects of scattered radiation, and differences in reconstruction strategies during the projection acquisition process, a continuous physical degradation representation space is established.

[0030] Specifically, this module first defines a continuous degradation parameter space composed of multiple physical degradation parameters. By sampling different combinations of physical degradation parameters, physical degradation states covering different scanning conditions are formed. Subsequently, standard-dose CT images are converted to the sinusoidal domain, and physical degradation simulation is performed in the sinusoidal space according to the sampled physical degradation parameters. The corresponding low-dose degradation CT image sample set is obtained through image reconstruction. Since each degradation generation process is controlled by the corresponding physical degradation parameters, the generated low-dose degradation images can not only serve as input samples for subsequent detection networks, but also retain the corresponding physical degradation parameter information as auxiliary supervision signals, providing a data foundation for subsequent learning of degradation-related features.

[0031] The specific process of this module is as follows: Figure 2 As shown, it includes four sub-steps: continuous degradation parameter space definition, physical degradation simulation, image reconstruction, and dynamic sampling scheduling.

[0032] 1.1 Continuous Degradation Parameter Space In actual clinical applications, the quality of low-dose CT images is affected by a variety of factors, including scan dose, detector performance, scanning protocol, and image reconstruction method, and these factors exhibit continuous variation. To describe the complex degradation state resulting from the combined effects of these factors, this module constructs a five-dimensional continuous physical degradation parameter space.

[0033] Define the physical degradation parameter vector:

[0034] in: This represents the dose scaling factor, used to describe the proportional relationship between the simulated scan dose and standard dose conditions. Its value range is: . Specifically, This indicates the lowest simulated dose ratio. Through continuous adjustment... It can change the number of incident photons in the sinusoidal domain, thereby simulating the statistical changes of photons at different dose levels; This represents the style parameter of the reconstruction convolution kernel, used to continuously characterize the spatial frequency response features corresponding to different reconstruction convolution kernels. Its value range is: Specifically, lower Higher values ​​correspond to smoother reconstruction characteristics. The value corresponds to a sharper high-frequency enhancement characteristic. By representing the continuous parameter, it is possible to simulate the image texture changes caused by differences in reconstruction kernels under different devices or clinical protocols.

[0035] This parameter represents the electronic noise intensity of the detector, used to quantitatively characterize the impact of electronic noise generated by the CT detector data acquisition system (DAS) during signal conversion, transmission, and amplification. Its value range is as follows: . Specifically, and These represent the lower and upper limits of the detector's electronic noise intensity parameter, respectively.

[0036] The scattered radiation ratio parameter describes the background shift caused by scattered photons on the detector's received signal. Its value range is: By adjusting It can simulate the deviation of projection data caused by changes in scattering intensity under different scanning conditions.

[0037] This represents the reconstruction strategy fusion parameter, used to describe the impact of different reconstruction strategies on the texture and noise distribution of the final image. Its value range is: Different values ​​correspond to different combinations of reconstruction strategies.

[0038] This represents the continuous physical degradation parameter space composed of the five degradation parameters mentioned above.

[0039] Furthermore, to describe the continuous state changes of the physical degradation parameters in the sinusoidal graph space, this module defines a sinusoidal graph degenerate manifold:

[0040] in: This represents a degenerate manifold of a sine graph; This represents a sine curve degradation generation mapping driven by physical degradation parameters; This represents sine wave data under standard dosage conditions.

[0041] The sinic graph degenerate manifold is used to describe the set of continuous degenerate states formed in the sinic graph space by different combinations of physical degenerate parameters. Through... Spatial sampling can obtain the degradation state corresponding to different locations, thereby achieving coverage of complex degradation changes in low-dose CT imaging.

[0042] 1.2 Simulation of Physical Degradation To convert anatomical information in the image domain into attenuation information along the ray propagation path, this module utilizes a ray-tracing-based forward projection operator to process the input standard-dose CT image. Perform a discrete Radon transform to obtain the standard dose sine curve:

[0043] in, This represents the sine wave data corresponding to a standard dose CT image; This represents the input standard dose CT image; Indicates the detector channel index; Indicates the projection angle index; This represents the forward projection operator based on ray geometry.

[0044] Through the above process, the image space is transformed. To the sine graph space The mapping allows the subsequent degradation process to be completed in a projection space that is closer to the CT physical acquisition process.

[0045] After obtaining the standard dose sine curve, this module further simulates the X-ray propagation and detection process.

[0046] According to Lambert-Beer's law, a standard dose sine curve can represent the attenuation integral information of radiation after it passes through the human body. Firstly, based on the dose scaling factor... The number of incident photons is adjusted, and then combined with the scattered radiation ratio parameter. A background term for scattered photons is introduced, and the statistical fluctuations in photon counting under low-dose conditions are simulated using a Poisson distribution. Finally, Gaussian electronic noise is superimposed on the detector's electronic noise intensity parameter to calculate the actual photon count measurement value received by the CT detector.

[0047] in, This indicates that after physical degradation simulation, the CT detector array in The actual photon count measurement value obtained at the location; This represents the Poisson random sampling process, used to simulate statistical fluctuations in X-ray photon counting; Indicates the dose proportionality factor; This represents the initial photon count without human attenuation. Represented by natural constant An exponential function with base 0; This represents the standard dose sine curve data; Indicates the proportion of scattered radiation; This represents the scattered photon background term, used to describe the low-frequency background effect formed by scattered radiation in the detector space; Indicates that it follows the mean. variance is The additive Gaussian electronic noise model. This refers to the electronic noise intensity parameter of the detector.

[0048] Through the above cascaded process, multiple degradation factors such as dose variation, photon statistical fluctuations, scattering effects, and electronic noise can be considered simultaneously in the sinusoidal domain, enabling the generated samples to contain degradation characteristics under different physical conditions.

[0049] After obtaining the detector photon count measurement values, which include the effects of photon statistical fluctuations, scattered radiation, and electronic noise, this module further performs logarithmic transformation on the photon count measurement values ​​to obtain degraded sine curve data containing physical degradation information.

[0050] Specifically, the logarithmic transformation is expressed as:

[0051] in, This represents the degraded sine curve data after physical degradation simulation; This represents the actual photon count measurement value received by the CT detector; This represents the initial photon count before attenuation by the human body. This represents the dose-proportioning factor.

[0052] Through the aforementioned logarithmic transformation process, factors such as photon statistical fluctuations, scattering shifts, and electronic noise in the sine diagram are propagated nonlinearly to the projection data space, enabling the generated degraded sine diagram to retain the degradation features caused by the physical acquisition process during low-dose CT imaging. Compared to existing techniques that typically add Gaussian or Poisson noise directly to the reconstructed image domain, this module first simulates the X-ray detection process in the sine diagram space before performing image reconstruction, thus more fully characterizing the coupling effect between the projection acquisition process and the reconstruction process.

[0053] 1.3 Image Reconstruction After obtaining the degradation sinusoidal data, this module further converts the projection space data to image space and generates corresponding low-dose degradation CT images through the reconstruction process. .

[0054] To simulate the impact of different reconstruction strategies on image noise, texture, spatial resolution, and edge features, this module introduces a reconstruction strategy fusion parameter. Different reconstruction strategies are fused together. As an example, the reconstruction strategies include the Filtered Back Projection (FBP) algorithm and the Model-Based Iterative Reconstruction (MBIR) algorithm based on statistical models.

[0055] Low-dose degraded CT images obtained through fusion Represented as:

[0056] in, The representation is derived from the degenerate parameter vector. Control the generation of low-dose degraded CT image samples; This parameter represents the fusion ratio between different reconstruction strategies; This represents the analytical reconstruction operator based on the filtered back projection algorithm; This represents an iterative reconstruction operator based on statistical model optimization. Represents degraded sine curve data; Indicates the style parameters of the reconstructed convolution kernel. A controlled frequency domain filtering response model is used to adjust the smoothness and edge preservation capabilities of the reconstructed image.

[0057] Through the above reconstruction process, different combinations of physical degradation parameters generate low-dose CT image samples with different noise characteristics, texture features, and spatial resolution properties:

[0058] in, This represents the degradation generation map determined by the standard dose image and physical degradation parameters.

[0059] Since each generated low-dose CT image sample is composed of a corresponding degradation parameter vector Control, therefore A one-to-one data relationship is formed, and the resulting set of physical degradation parameters can participate in subsequent auxiliary supervision regression constraints.

[0060] 1.5 Dynamic Sampling Scheduling Due to the continuous physical degradation parameter space It contains multiple intercoupled degenerate dimensions, and its corresponding degenerate state space. It has high dimensionality. If only a fixed uniform random sampling strategy is used to generate training samples, some easily learnable regions may be repeatedly sampled, while some complex and degraded regions may not be adequately covered, thus affecting the detection model's ability to adapt to complex low-dose conditions.

[0061] To improve the effectiveness of degenerate manifold coverage during training, this module further performs dynamic sampling and scheduling on the continuous degenerate parameter space constituted by the multidimensional physical degenerate parameters, specifically including: The continuous degradation parameter space is discretized into multiple degradation sub-regions; In the initial training phase, uniform sampling is performed on each degenerate sub-region, enabling the model to initially acquire the basic representational capability of the entire degenerate space. Specifically, uniform sampling is represented as: ,in, Indicates the number of degenerate subregions; Indicates the first The initial sampling probability of each degenerate sub-region; A portion of samples is reserved as validation samples from each degraded sub-region. Specifically, for each degraded sub-region, when generating low-dose degraded CT image samples, a portion of the samples are allocated to the validation set, and the remaining samples are allocated to the training set. Preferably, the validation samples account for 10% to 30% of the total samples in each degraded sub-region. The validation samples do not participate in the gradient update of model parameters; they are only used to evaluate the model's detection performance on that degraded sub-region.

[0062] During training, after each training round, the validation samples corresponding to each degenerate sub-region are input into the current detection model. The 3D nodule detection loss of the detection model on each validation sample is calculated, and the average detection loss of multiple validation samples within each degenerate sub-region is calculated separately. The sampling weights of each degenerate sub-region in the next training phase are adjusted based on the average detection loss of each degenerate sub-region. Specifically, this module updates the sampling weights for the next phase using an exponential weighting method.

[0063] in, Indicates the first training phase The sampling weights corresponding to each degenerate sub-region; This represents the average detection loss of the m-th degenerate sub-region; This represents the temperature adjustment parameter, used to control the degree of change in the sampling weight distribution; This represents the index of the degenerate subregion.

[0064] When the average detection loss of a certain degraded sub-region is greater than that of other degraded sub-regions, its corresponding sampling weight will be increased accordingly, so that the training process can increase the frequency of occurrence of such difficult degraded samples; when the lung nodule detection model of a certain sub-region has good adaptability, its sampling weight will be reduced accordingly.

[0065] Through the aforementioned dynamic sampling and scheduling mechanism, training samples can be adaptively distributed and adjusted within the entire sinusoidal degenerate manifold space, enabling the model to learn more fully the feature change patterns under different degenerate states within a limited number of training rounds, thereby improving its adaptability to complex low-dose imaging conditions.

[0066] Ultimately, the resulting degraded CT images All of them are their corresponding physical degradation parameters Establish data binding relationships and use them as supervisory labels for subsequent constraint losses.

[0067] Two-way feature variational orthogonal decoupling coding module Under low-dose CT imaging conditions, factors such as reduced photon count, increased detector noise, and differences in reconstruction algorithms cause noise textures, artifact features, and reconstruction style variations in the image to co-integrate with the semantic features of the nodule target into the latent space representation of the deep network. Traditional end-to-end detection networks typically encode both target content and imaging degradation information simultaneously through a single shared feature space. This makes the feature representation learned by the detection model susceptible to changes in scan dose, equipment differences, and reconstruction protocols, thereby reducing cross-condition detection performance.

[0068] To address the aforementioned issues, this module first utilizes a shared feature extraction network to obtain the basic visual representation of the input image. Then, through three parallel variational coding paths, the basic features are mapped to three latent variable representation spaces with different semantic functions. Furthermore, variational probability modeling is used to describe the distribution characteristics of the three latent spaces, and a mutual information constraint mechanism based on a CLUB estimator is introduced to reduce the information correlation between different feature paths, making each latent variable space more inclined to express its corresponding information factors. Simultaneously, variational regularization constraints maintain the continuity of the latent variable space distribution, improving the detection model's adaptability to different low-dose degradation conditions.

[0069] See Figure 3 Specifically, this module includes five sub-steps: shared feature extraction, three-way variational coding, mutual information constraints, auxiliary supervised regression constraints, and distribution regularization constraints.

[0070] 2.1 Shared Feature Extraction This module uses a shared feature extraction network to encode the input low-dose CT images to obtain basic features:

[0071] in, Indicates a shared feature extraction network; This represents the basic characteristics of the network output; This represents the learnable parameters of a shared feature extraction network.

[0072] Specifically, the shared feature extraction network can adopt a convolutional neural network structure, a visual Transformer structure, or other deep network structures with hierarchical feature extraction capabilities.

[0073] 2.2 Three-way variational coding The three-way variational coding network includes a nodule content variational coding head, a degeneracy variational coding head, and an anatomical structure variational coding head. The three coding heads respectively map the basic features into variational approximate posterior distributions of nodule content latent variables, degeneracy latent variables, and anatomical structure latent variables, and obtain the nodule content latent variables, degeneracy latent variables, and anatomical structure latent variables by sampling through a reparameterization method.

[0074] 2.2.1 Latent Variables in Node Content Specifically, nodule content latent variables are used to represent semantic information related to the lung nodule target, including but not limited to nodule boundary morphology, internal density variations, spatial location, and local texture structure. Nodule content variational coding head. Receive basic features Output the corresponding latent variable distribution parameters. Sum of variance vectors Therefore, the variational approximation posterior distribution corresponding to the latent variables of nodule content is defined as follows:

[0075] in, This represents the variational approximation posterior distribution corresponding to the latent variables of nodule content; Represents implicit variables indicating nodule content; Represents the posterior mean vector of latent variables related to nodule content; This represents the posterior standard deviation vector of the latent variables in the nodule content.

[0076] To achieve gradient backpropagation during random variable sampling, a reparameterization method is used to obtain latent variables within the node content:

[0077] in, ; This represents the Hadamard element-wise product; This represents a standard multivariate normally distributed random variable.

[0078] The basic features are mapped to a continuous probability latent space through the above process, enabling the nodule content features to be represented in the form of a probability distribution.

[0079] 2.2.2 Degenerate Latent Variables Degradation latent variables are used to characterize degradation-related information in low-dose CT images caused by the imaging process, including photon statistical noise, detector electronic noise, scattering effects, and texture changes due to differences in reconstruction algorithms. Degradation coding head Receive shared basic features and output the corresponding latent variable distribution parameters. and Its variational approximate posterior distribution is defined as:

[0080] in, This represents the variational posterior distribution corresponding to the degenerate latent variable; Indicates degenerate hidden variables; This represents the posterior mean vector of the degenerate latent variables; This represents the posterior standard deviation vector of the degenerate latent variable.

[0081] Obtaining degenerate latent variables through reparameterization:

[0082] in, This represents the random sampling variable corresponding to the degradation path.

[0083] Since the path is simultaneously supervised by the physical degradation parameters provided by the sinusoidal graph domain degradation manifold generation module, the degradation latent variable representation can learn the feature change patterns related to the imaging degradation state.

[0084] 2.2.3 Latent variables in anatomical structure Anatomical latent variables are used to characterize the overall spatial structure of the lungs, providing anatomical context constraints for subsequent detection. This approach primarily focuses on structural layout features in non-lesion regions, such as the spatial distribution of the vascular tree, airway structure, and pleural boundaries. Anatomical structure encoding head. Receive shared basic features Output and Its variational approximate posterior distribution is expressed as:

[0085] in, This represents the variational posterior distribution corresponding to the latent variables of anatomical structure. Represents latent variables related to anatomical structure; Represents the mean vector of latent variables in anatomical structures; This represents the standard deviation vector of the latent variables related to anatomical structure.

[0086] Obtained through reparameterization:

[0087] in: This represents the random sampling variable corresponding to the anatomical structure path.

[0088] Through the above three-way encoding process, the shared basic features of the input image are mapped to three latent variable spaces with different semantic functions: These are used to represent nodule content, degeneration, and anatomical structure information, respectively.

[0089] 2.3 Mutual Information Constraints All three latent variable spaces originate from shared basic features. To avoid feature information overlap between different encoding paths (e.g., allowing degenerate information to enter the nodule content path or allowing target semantic information to be mixed into the degenerate path), this module adopts a mutual information constraint method to reduce the information correlation between different latent variable spaces, estimates the upper bound of mutual information between different paths, and uses it as a constraint term in the training process.

[0090] Specifically, due to the precise mutual information between high-dimensional continuous random variables This module utilizes parameterized conditional distributions, which are typically difficult to calculate directly. An approximate model is performed on the distribution of the true conditions.

[0091] Hidden variables in nodule content With degenerate latent variables Taking the constraints between them as an example, this module is based on constructing a conditional probability estimation network using the CLUB estimator:

[0092] in, Indicates by parameters The control conditional probability estimation network can be implemented using a multilayer perceptron or other learnable mapping structure.

[0093] Correspondingly, the nodule content contains hidden variables. With degenerate latent variables The mutual information estimation loss between them is expressed as:

[0094] in, This represents the mutual information constraint term between the latent variables of nodal content and the latent variables of degradation. This represents the total number of samples in the current training batch; This represents the traversal index of the target sample in the current batch; Indicates the index of unpaired samples within a batch; Indicates that the parameter is The log-conditional probability density function value output by the conditional probability estimation network; and They represent the first Latent variable features and degraded latent variable features extracted from each paired sample; This represents the degenerate latent variables corresponding to different samples.

[0095] By comparing the conditional probabilities between the same sample paths with the conditional probabilities between randomly combined samples, constraints can be imposed on the statistical correlation between different characteristic paths.

[0096] Furthermore, this module establishes mutual information constraints for the pairwise combinations of the three latent variable spaces:

[0097] in, This represents the overall mutual information constraint loss among the three latent variables; This represents the constraint term between the latent variables of nodule content and the latent variables of anatomical structure; This represents the constraint term between the degenerate latent variable and the anatomical structure latent variable.

[0098] The aforementioned mutual information constraints enable a clearer division of information among the three latent variable spaces during feature representation, reducing feature confounding among different factors. By minimizing the upper limit of mutual information for each combination, the information correlation among the three latent variables can be reduced.

[0099] 2.4 Auxiliary Supervision Regression Constraints Furthermore, although mutual information constraints between different latent variables can reduce the information overlap between content features, degradation features, and anatomical features, the degradation path may still learn redundant information irrelevant to the task due to the strong representational freedom of deep neural networks, or reduce the mutual information constraint term by reducing its own feature expressive power.

[0100] Therefore, this module utilizes the physical degradation parameter vector synchronously recorded during sample generation in the sinusoidal graph domain multidimensional degenerate manifold generation module. As auxiliary monitoring information for degenerate latent variables, it guides the degenerate latent variables. Learn feature representations related to imaging degradation factors to further clarify the semantic function of degradation pathways.

[0101] Specifically, this module sets a degradation parameter regression prediction head at the output of the degradation coding path. The prediction head consists of a cascaded linear layer and activation functions. The degenerate parametric regression prediction head receives degenerate latent variables. It also outputs the predicted estimates of the physical degradation parameters:

[0102] in, The learnable parameters of the regression prediction head representing the degradation parameters; This represents the physical degradation parameters predicted by the latent degradation variables.

[0103] A physical parameter regression loss is constructed by minimizing the difference between the estimated physical degradation parameters and the corresponding true physical degradation parameters in the physical degradation parameter set:

[0104] in, This represents the regression loss of physical parameters; The total dimension of the physical degradation parameter manifold is represented in this invention. ; The dimension traversal index represents the physical degradation parameter; The first value in the regression prediction head output of the degradation parameter represents the first value. Estimates of physical degradation parameters; This indicates the corresponding dynamically sampled number from the physical degenerate manifold module. Reference values ​​for true physical degradation parameters.

[0105] Through the aforementioned supervised constraints, the degradation path is guided to learn feature information related to physical imaging degradation, enabling the latent variable representation of degradation to reflect changes in the degradation state in the input image. Simultaneously, this is combined with the aforementioned mutual information constraints. This can reduce the propagation of degraded information into the content path and anatomical structure path, and enable a clearer division of information in different latent variable spaces.

[0106] 2.5 Distribution Regularization Constraints To ensure that the latent variable space generated by the three variational coding paths has continuous and smooth distribution characteristics, this module further applies variational regularization constraints to the posterior distributions of the three latent variables. Since the degradation process of low-dose CT images usually exhibits continuous changes rather than abrupt changes between several discrete categories, it is necessary to ensure that the latent variable space can continuously represent different degradation states.

[0107] Specifically, this module uses the Kullback-Leibler (KL) divergence as a distance metric between the latent variable posterior distribution and the standard multivariate normal prior distribution, expressed as:

[0108] in, This represents the distributional regularization loss of the three-way latent variable space; The Kullback-Leibler divergence operator represents the distance between two probability distributions. This represents the standard multivariate normal prior distribution.

[0109] This regularization constraint maintains a continuous probability structure in the three latent variable spaces, preventing overly discrete feature distributions from arising in the latent space for different input samples.

[0110] Furthermore, the combined effect of KL constraints and mutual information constraints makes the latent variables of nodule content pay more attention to the semantic information of the target, the latent variables of degradation pay more attention to the imaging degradation information, and the latent variables of anatomical structure pay more attention to the spatial context information of the lungs, thereby improving the stability of the feature representation of the detection model under different scanning conditions and different degrees of degradation.

[0111] After three-way variational coding and constraint optimization, this module outputs... Three latent variables represent nodule content features, degenerative features, and anatomical structure features, respectively. During the training phase, these three pathways jointly participate in feature learning: utilizing... Constrain the information correlation between different paths; utilize The correspondence between constrained degradation paths and actual physical degradation parameters; utilizing Maintain the continuous distribution of the latent variable space. In subsequent detection stages, selectively fuse features from different paths according to the requirements of the detection task. Among these, the latent variables of nodule content... Anatomical structure latent variables serve as the primary semantic representation for object detection. Used to provide contextual information about the spatial structure of the lungs, degenerative latent variables Primarily used for degradation constraints and model robustness learning during the training phase. Through the aforementioned three-way decoupling mechanism, the detection network can reduce the impact of different low-dose imaging conditions on feature representation while preserving the target semantic information, providing potential feature representations with clear semantic division of labor for subsequent cross-path feature fusion modules and 3D nodule detection and prediction modules.

[0112] III. Cross-path feature fusion module To fully utilize the important auxiliary role of lung anatomy in target localization, this module designs a cross-path feature fusion method based on the cross-attention mechanism. This method enables the latent variables of nodule content to actively query the spatial context information in the anatomical features, thereby enhancing the expressive ability of the fused detection features for complex tissue regions.

[0113] Specifically, such as Figure 4 As shown, this module constructs a fusion detection feature based on a cross-attention mechanism. Because... It mainly describes the local semantic information of the nodules, while It contains information about the overall spatial structure of the lungs, and the two are complementary. Therefore, this module uses the latent variables of nodule content as query information to actively retrieve relevant spatial context information from the latent variables of anatomical structure, thereby achieving cross-path feature fusion.

[0114] Specifically, Represents implicit variables indicating nodule content; This represents latent variables related to anatomical structure. Among them, Indicates the number of feature channels; These represent the dimensions in three-dimensional space.

[0115] Before performing attention calculations, the latent variables of nodule content and anatomical structure are expanded into token sequences respectively: .in, This represents the number of tokens after the three-dimensional spatial location is unfolded. This represents the feature dimension of a single token. Then, a learnable mapping matrix is ​​used: Generate the query matrix, key matrix, and value matrix respectively:

[0116] in, This represents the query matrix generated from the implicit variables of the node content; This represents the key matrix generated by latent variables of anatomical structure; This represents a value matrix generated by latent variables of anatomical structure. This represents the learnable mapping parameters.

[0117] Calculate attention weights using the query matrix and key matrix:

[0118] in, This indicates that the detection features are fused and the original nodule content information is preserved through residual connections, thereby improving the stability of network training.

[0119] Through the aforementioned fusion process, the fused detection features simultaneously contain semantic information about the nodule itself and contextual information about the overall spatial structure of the lungs. For example, in locations prone to false detection, such as vascular crossing areas or near the airways, anatomical latent variables can provide additional spatial constraints, enabling the network to make judgments by combining local target features with the overall structural relationship; while in independent nodule regions, nodule content latent variables can maintain a strong response, thereby improving the discriminative power of the fused detection features.

[0120] It should be noted that in low-dose CT detection tasks, although degradation information can reflect the imaging state of the input image, its changes do not directly correspond to the target detection task. Therefore, this module adopts differentiated feature path utilization strategies based on the different task requirements of the training and inference phases.

[0121] Specifically, during the training phase, three latent variables are used to participate in feature constraints, enabling the network to learn the feature division of labor among nodule content, degradation factors, and anatomical structures. During the inference phase, based on the requirements of the target detection task, the influence of degradation paths on the prediction results is selectively reduced, and only nodule content latent variables and anatomical structure latent variables are used to complete the construction of fused detection features. This reduces the direct dependence of the detection prediction process on degradation changes and improves the stability of the detection model when facing different low-dose scanning conditions.

[0122] IV. Three-dimensional nodule detection and prediction module Specifically, for fusion detection features This module utilizes a 3D nodule detection and prediction head. Decoding is performed, and its parameters are represented as follows: Output the 3D bounding box of the lung nodule. The confidence level of the nodules is high. .

[0123] Among them, the three-dimensional bounding box of the lung nodule To predict the coordinates of the nodule center, These represent the dimensional parameters of the 3D bounding box of the lung nodule in three spatial directions. The confidence level of nodule presence is also indicated. This indicates the predicted probability that the current candidate region belongs to a lung nodule. These two types of prediction results enable the detection model to utilize both classification and spatial location information simultaneously, improving the completeness of nodule detection results in 3D CT scenes.

[0124] Specifically, since lung nodule detection tasks often suffer from an imbalance between the number of target samples and background samples, this module uses the Focal Loss function to calculate the classification loss based on the difference between the nodule's confidence level and its corresponding true class label. This reduces the impact of a large number of easily classifiable background samples on the training process, allowing the detection model to focus more on difficult-to-classify candidate regions. The classification loss is expressed as:

[0125] in, Represents classification loss. The true category label for the presence or absence of nodules; parameters This represents the balanced weights for positive and negative samples, and the parameters are... This indicates the adjustment factor for samples of varying difficulty.

[0126] Specifically, to improve the spatial overlap between the predicted bounding box and the actual nodule region, this module calculates the 3D bounding box regression loss based on the difference between the 3D bounding box and the corresponding actual bounding box, expressed as:

[0127] in, The predicted 3D bounding box; This is the corresponding truth value bounding box; This represents the volume of the 3D spatial overlap region between the predicted 3D bounding box and the actual bounding box; This represents the volume of the union of the two in three-dimensional space.

[0128] Construct a joint detection loss function based on 3D bounding box regression loss and classification loss:

[0129] In the formula, This indicates the loss in three-dimensional nodule detection; This represents the positioning loss balance coefficient.

[0130] The joint detection loss function optimizes the 3D spatial overlap between the 3D bounding box and the ground truth bounding box, enabling the model to learn more accurate nodule location and scale information.

[0131] V. Joint Optimization Module This module optimizes the aforementioned mutual information constraint loss, physical parameter regression loss, latent variable distribution regularization loss, and 3D nodule detection loss in a coordinated manner as optimization objectives.

[0132] Specifically, the overall joint optimization objective function is expressed as:

[0133] in, This represents the overall joint optimization loss; This represents the 3D nodule detection loss, used to constrain the difference between the detection prediction results and the actual nodule annotations, enabling the detection model to learn accurate 3D localization and classification capabilities; This represents the mutual information constraint loss among the three latent variables, which is used to reduce the information correlation among latent variables of nodule content, degradation, and anatomical structure, and to promote the separate representation of different semantic factors in the latent space. The regression loss of physical parameters corresponding to the degradation latent variables is used to apply auxiliary constraints to the degradation latent variables using the set of physical degradation parameters, so that they learn feature representations related to imaging degradation factors. This represents the latent variable distribution regularization loss, used to constrain the distance between the posterior distribution of the three-way variational latent variables and the preset prior distribution, thereby improving the continuity and stability of the latent space representation. This represents the weight coefficient corresponding to each constraint term, used to adjust the degree of influence of different optimization objectives in the overall training process.

[0134] By jointly optimizing the overall objective function, the detection task objective and feature representation constraints jointly participate in the optimization of network parameters, enabling the model to improve detection performance while obtaining a more stable feature representation ability for low-dose CT degradation changes.

[0135] During the training phase, the system uses an end-to-end approach to jointly update the parameters of each module, and the parameters of each module are optimized through the overall joint objective function. The generated gradient information is updated synchronously, enabling the system to form a closed-loop optimization process consisting of physical degradation modeling, feature decoupling, anatomical structure constraints, and target detection. Through the above joint optimization mechanism, this invention can simultaneously utilize physical degradation prior information and detection task supervision information during the training phase, enabling the network to learn a more stable and degradation-adaptive lung nodule detection feature representation until the model converges, thus obtaining a well-trained lung nodule detection model.

[0136] This invention addresses the problems of insufficient physical degradation simulation, mixed feature representation, and false detection in complex tissue regions in existing low-dose CT lung nodule detection methods, and has the following technical effects: 1. In terms of physical degradation simulation, this invention achieves physical degradation simulation of the main degradation factors in CT imaging by moving the degradation injection process forward to the sinusoidal domain and constructing a continuous physical degradation parameter space composed of multi-dimensional physical degradation factors. Generating low-dose CT training samples with different degradation characteristics through continuous degradation parameter sampling helps reduce the distribution difference between simulated degradation data and real clinical scan data, improving the adaptability of the detection model to different low-dose CT imaging conditions.

[0137] 2. Regarding feature representation and robust detection, this invention utilizes variational information bottleneck constraints and the CLUB mutual information estimation mechanism to explicitly decouple nodule content information, degradation information, and anatomical structure information, forming latent variable representations with different semantic functions, thus reducing the interference of degradation factors on the expression of detection features. Simultaneously, through cross-path attention fusion between anatomical structure latent variables and nodule content latent variables, original lung nodule content information is introduced to assist detection, enabling the detection model to maintain relatively stable target discrimination ability in complex regions such as low-dose noise and vascular crossings.

[0138] 3. In engineering applications, this invention utilizes the set of physical degradation parameters obtained during the generation of low-dose degraded CT image sample sets as self-supervised information. This eliminates the need for additional manual degradation labels to constrain the learning of degradation latent variables, improving the flexibility of the training process. During the inference phase, the training system does not invoke degradation paths and their auxiliary prediction branches; it only utilizes nodule content latent variables and anatomical structure latent variables for feature fusion and detection, reducing the computational processes related to degradation modeling and improving the applicability of this invention in practical deployment environments.

[0139] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A training method for a lung nodule detection model, characterized in that, Includes the following steps: Step 1: Acquire standard dose CT images and convert them to the sinusoidal domain through orthogonal projection. In the sinusoidal domain, perform physical degradation simulation according to multidimensional physical degradation parameters to generate degradation sinusoidal data. Then, generate a low-dose degradation CT image sample set and the corresponding physical degradation parameter set through image reconstruction. Step 2: Input the low-dose degraded CT image sample set into a shared feature extraction network to obtain basic features; input the basic features into a three-way variational coding network for processing to obtain nodule content latent variables, degraded latent variables, and anatomical structure latent variables; Step 3: Establish mutual information constraints for the pairwise combinations of the three latent variables to obtain mutual information constraint loss; establish auxiliary supervised regression constraints for the degenerate latent variables based on the physical degradation parameter set to obtain physical parameter regression loss; apply regularization constraints to the posterior distribution of the three latent variables to obtain latent variable distribution regularization loss; and use a cross-attention mechanism to fuse nodule content latent variables and anatomical structure latent variables to generate fused detection features, input the fused detection features to the 3D target detection prediction head, and calculate the 3D nodule detection loss. Step 4: Combining the mutual information constraint loss, physical parameter regression loss, latent variable distribution regularization loss, and three-dimensional nodule detection loss, construct an overall joint optimization objective function, and perform end-to-end joint parameter updates on the lung nodule detection model until the model converges, thus obtaining a trained lung nodule detection model.

2. The training method according to claim 1, characterized in that, The multidimensional physical degradation parameters include: The dose scaling factor is used to describe the proportional relationship between the simulated scan dose and the standard dose conditions. The style parameters of the reconstructed convolution kernel are used to continuously characterize the spatial frequency response features corresponding to different reconstructed convolution kernels; The detector electronic noise intensity parameter is used to quantitatively characterize the impact of electronic noise generated by the CT detector data acquisition system during signal conversion, transmission, and amplification. The scattered radiation ratio parameter is used to describe the background shift effect caused by scattered photons on the signal received by the detector; The reconstruction strategy fusion parameter describes the impact of different reconstruction strategies on the texture and noise distribution of the final image.

3. The training method according to claim 2, characterized in that, The physical degradation simulation in step 1 includes: adjusting the number of incident photons in the sinusoidal domain according to the dose scaling factor, introducing a background term for scattered photons in combination with the scattered radiation scaling parameter, simulating the statistical fluctuations of photon counts under low-dose conditions using the Poisson distribution, superimposing Gaussian electronic noise based on the detector electronic noise intensity parameter, obtaining the actual photon count measurement value received by the CT detector, and obtaining the degradation sinusoidal data after the photon count measurement value is logarithmically transformed.

4. The training method according to claim 3, characterized in that, The image reconstruction in step 1 includes: using the reconstruction strategy fusion parameters to perform weighted fusion of different reconstruction methods to generate a low-dose degradation CT image sample set corresponding to different combinations of physical degradation parameters.

5. The training method according to claim 2, characterized in that, Step 1 further includes dynamically sampling and scheduling the continuous degradation parameter space constituted by the multidimensional physical degradation parameters, specifically configured as follows: The continuous degradation parameter space is discretized into multiple degradation sub-regions; In the initial stage of training, uniform sampling is performed on each degenerate sub-region; A portion of samples are reserved from each degenerate sub-region as validation samples. During the training process, after each round of training, the three-dimensional nodule detection loss of the lung nodule detection model on the corresponding validation samples in each degenerate sub-region is calculated, and the average detection loss of multiple validation samples in each degenerate sub-region is calculated respectively. The sampling weights of each degenerate sub-region in the next training phase are adjusted based on the average detection loss of each degenerate sub-region. Degenerate sub-regions with relatively larger average detection losses receive relatively higher sampling weights.

6. The training method according to claim 1, characterized in that, The three-way variational coding network includes a nodule content variational coding head, a degeneracy variational coding head, and an anatomical structure variational coding head. The three coding heads respectively map the basic features to variational approximate posterior distributions of nodule content latent variables, degeneracy latent variables, and anatomical structure latent variables, and obtain the nodule content latent variables, degeneracy latent variables, and anatomical structure latent variables by sampling through a reparameterization method.

7. The training method according to claim 1, characterized in that, The establishment of mutual information constraints in step 3 specifically includes: using the CLUB estimator to construct a conditional probability estimation network for three pairs of combinations: nodule content latent variable and degeneration latent variable, nodule content latent variable and anatomical structure latent variable, and degeneration latent variable and anatomical structure latent variable, respectively, to estimate the upper limit of mutual information of the combinations, and to reduce the information correlation between the three latent variables by minimizing the upper limit of mutual information of each combination.

8. The training method according to claim 1, characterized in that, The establishment of auxiliary supervised regression constraints specifically includes: obtaining physical degradation parameter estimates based on the predicted degradation latent variables, and constructing a physical parameter regression loss based on the difference between the physical degradation parameter estimates and the corresponding real physical degradation parameters in the physical degradation parameter set.

9. The training method according to claim 1, characterized in that, The cross-attention mechanism in step 3 includes: expanding the nodule content latent variable and the anatomical structure latent variable into token sequences, using the nodule content latent variable as query information and the anatomical structure latent variable as key information, calculating attention weights, and retaining the original nodule content information through residual connections.

10. A training system for a lung nodule detection model, characterized in that, include: The sinusoidal domain multidimensional degradation manifold generation module is used to acquire standard dose CT images and perform physical degradation simulation in the sinusoidal domain according to multidimensional physical degradation parameters. The low dose degradation CT image sample set and the corresponding physical degradation parameter set are generated through image reconstruction. The three-way feature variational orthogonal decoupled coding module is used to extract nodule content latent variables, degradation latent variables, and anatomical structure latent variables from the low-dose degraded CT image sample set through shared feature extraction and a three-way variational coding network; establish mutual information constraints on the pairwise combination relationships between the three latent variables to obtain mutual information constraint loss; establish auxiliary supervised regression constraints on the degradation latent variables based on the physical degradation parameter set to obtain physical parameter regression loss; and apply regularization constraints to the posterior distribution of the three latent variables to obtain latent variable distribution regularization loss. The cross-path feature fusion module is used to employ a cross-attention mechanism, using nodule content latent variables as query information and anatomical structure latent variables as key-value information, to fuse nodule content latent variables and anatomical structure latent variables to generate fused detection features. The 3D nodule detection and prediction module is used to decode the fused detection features using a 3D target detection prediction head, and output the 3D bounding box of the lung nodule and the nodule presence confidence score; calculate the 3D bounding box regression loss based on the difference between the 3D bounding box and the corresponding ground truth bounding box, calculate the classification loss based on the difference between the nodule presence confidence score and the corresponding ground truth class label, and obtain the 3D nodule detection loss from the 3D bounding box regression loss and the classification loss. The joint optimization module is used to perform a weighted summation of the mutual information constraint loss, physical parameter regression loss, latent variable distribution regularization loss, and three-dimensional nodule detection loss to construct an overall joint optimization objective function and perform end-to-end joint parameter updates for the lung nodule detection model.

Images