A physical information constraint self-supervised learning method and system for CT material decomposition

Abstract

The application discloses a kind of physical information constraint self-supervised learning method and system for CT material decomposition, belong to medical image processing field.It includes obtaining dual-energy CT projection and image data;It is input into the material decomposition network trained, and outputs multi-substance density map.Network training uses unlabelled dataset, is realized by constructing the system including coding-decomposition network, differentiable physical forward projection module and anatomical perception constraint module.When training, with projection data fidelity loss as core, joint image reconstruction loss, physical consistency loss and anatomical prior adversarial loss constitute multi-task self-supervised objective and optimize.The application does not need to be paired with label data, by imaging physical model as differentiable strong constraint and with anatomical prior deep fusion, ensure the physical credibility of decomposition result, clinical rationality and high precision, with excellent generalization ability and real-time inference efficiency, overcome the key obstacle of clinical deployment of deep learning model in this field.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a physical information constraint self-supervised learning method and system for CT material decomposition. Background Technology

[0002] Computed tomography (CT) material decomposition techniques can resolve the voxel values ​​of CT images into the density distribution of various basic substances (such as soft tissue, iodine, and calcium), which is of core value for accurate diagnosis and quantitative analysis. Traditional methods mainly rely on analytical algorithms based on physical models, such as material decomposition methods in the projection domain or image domain. Although these methods have clear physical meanings, their performance is heavily dependent on accurate system modeling and calibration, and they are extremely sensitive to non-ideal factors commonly found in clinical settings, such as noise and beam hardening, resulting in insufficient stability and robustness of the decomposition results.

[0003] In recent years, supervised learning methods based on deep learning have demonstrated superior performance, implicitly compensating for complex physical effects through data-driven approaches. However, these methods rely on training with a large amount of paired data of "multi-energy CT images - real material density maps." Obtaining the true values ​​of real material density is not feasible in clinical practice; typically, limited paired data can only be generated through physical simulations or specialized phantoms. This leads to the model's tendency to overfit to ideal or specifically distributed simulation data, resulting in severely insufficient generalization ability when faced with real and diverse clinical data, greatly limiting its practical application.

[0004] To circumvent data bottlenecks, researchers have turned to unsupervised or self-supervised learning, such as using recurrent consistency generative adversarial networks to learn image domain mappings, or using the basic material decomposition formula as a network layer supplemented with image reconstruction loss. However, existing methods have fundamental flaws: the former lacks explicit physical guidance in the learned mappings, and the generated results may be visually plausible but physically unreliable and quantitatively inaccurate; the latter, although introducing physical models, has weak constraints, and the network may converge to an erroneous solution space that fits the input data well but violates fundamental physical laws such as the nonnegativity of matter and decay characteristics, resulting in decomposition results that lack clinical interpretability.

[0005] Therefore, the core contradiction facing current technology lies in the fact that clinical practice urgently needs high-precision, highly generalizable material decomposition methods, but existing technological approaches cannot simultaneously guarantee data-driven performance and adherence to physical laws in the absence of paired labeled data. How to construct a new paradigm that can effectively learn from unlabeled data while incorporating imaging physics as a strong constraint into the learning framework is a key challenge that urgently needs to be overcome in this field. Summary of the Invention

[0006] The purpose of this invention is to overcome the aforementioned shortcomings of existing technologies and provide a self-supervised learning method and system for physical information constraint-based decomposition of CT materials. This addresses the core problems of existing technologies, such as poor generalization due to reliance on paired labeled data and unreliable decomposition results due to a lack of physical constraints. By constructing a self-supervised learning framework that integrates a differentiable physical projection model with anatomical priors, the decomposition results are ensured to simultaneously satisfy the consistency of projection data, basic physical laws, and prior knowledge of human anatomical structures, without requiring any manual annotation. This achieves high-precision, strong generalization, and high-reliability quantitative decomposition of multiple materials.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a physical information-constrained self-supervised learning method for CT material decomposition, comprising the following steps:

[0009] Acquire dual-energy CT projection data of the target object and its reconstructed CT image data;

[0010] The dual-energy CT image data is input into a trained material decomposition network, which outputs density distribution maps of various basic materials of the target object; wherein, the training method of the material decomposition network includes:

[0011] Construct a training dataset without real material annotations, the dataset containing multiple sets of dual-energy CT projection data and CT image data reconstructed from the projection data;

[0012] Construct a network training system that includes an encoder-decomposer network, a differentiable physical forward projection module, and an anatomical perception constraint module;

[0013] Using the training dataset, with forward projection consistency as the core, and combining anatomical priors and physical laws, the encoder-decomposer network is trained by optimizing a multi-task self-supervised loss function consisting of projection domain data fidelity loss, image domain reconstruction loss, physical consistency loss, and anatomical prior adversarial loss.

[0014] Furthermore, the differentiable physical forward projection module is specifically used for:

[0015] Voxel mass density vector predicted based on encoder-decomposer network ; and the known mass decay coefficient of the base material Calculate the line attenuation coefficient including nonlinear compensation:

[0016] ;

[0017] in The energy-related residual term, predicted from the density vector using a small neural network, is used to model the beam hardening effect.

[0018] Through differentiable forward projection operator Integrating the linear attenuation coefficient yields the single-energy projection, which is then combined with the simulated X-ray tube energy spectrum S(E) and detector response D(E) to synthesize the final dual-energy projection data. and .

[0019] Furthermore, the encoder-decomposer network is a two-stream feature-aligned structure, specifically including:

[0020] Two independent shallow coding branches are used to extract the input low-energy CT image. and high-energy CT images Energy-specific characteristics;

[0021] A deformable feature alignment module, which uses the deep feature map of the low-energy branch as a reference, performs deformation alignment on the deep feature map of the high-energy branch through the learned spatial offset field to ensure accurate matching of anatomical structures during subsequent fusion;

[0022] A shared decoder is used to decode aligned and fused features into a multi-channel underlying mass density map.

[0023] Furthermore, the anatomical perception constraint module is a fixed, pre-trained multi-organ segmentation network S; the anatomical prior adversarial loss Constructed and calculated in the following manner:

[0024] Decompose the specific mass density map output by the network Compared with the original low-energy CT images The images are stitched together along the channel dimension to form a composite image;

[0025] Combine the combined image with individual Input the segmentation network S into each of the two groups of organ probability distribution maps;

[0026] Calculate the KL divergence between the two probability distributions as the loss:

[0027] This constrains the distribution of matter to conform to known anatomical priors.

[0028] Furthermore, the data fidelity loss in the projection domain Defined as:

[0029] ;

[0030] in, , For the input of real low- and high-energy projection data, , The corresponding projection data synthesized by the differentiable physical forward projection module.

[0031] Furthermore, the physical consistency loss Includes: Non-negativity constraint on matter density: This forces all predicted densities of basic substances to be non-negative.

[0032] in, The total number of basic material types is used to represent the number of basic materials to be distinguished in the decomposition task; The first prediction for the encoder-decomposer network The density value of a basic substance, its physical unit is mass / volume; To correct the linear unit function acting on The result is used to penalize only negative density values; The loss term is a non-negative constraint term for material density, and negative predicted values ​​are penalized to ensure physical plausibility.

[0033] Material distribution smoothing constraint: Apply total variational (TV) regularization to each material density map to suppress unreasonable noise and promote spatial smoothness.

[0034] Furthermore, the multi-task self-supervised loss function It is the weighted sum of the losses of each component, and its expression is:

[0035] ;

[0036] in, , , , These are the weighting coefficients for the projection domain data fidelity loss, image domain reconstruction loss, physical consistency loss, and anatomical prior adversarial loss, respectively, used to balance the contribution of each loss term to the overall objective function; This refers to the loss of fidelity in the projection domain, used to measure the difference between real and synthetic projection data; This is the image domain reconstruction loss, used to ensure that the predicted material density can accurately reconstruct the input dual-energy CT image; The physical consistency loss includes nonnegativity and smoothness constraints to ensure that the decomposition results conform to physical laws. To mitigate the loss of anatomical priors, it is used to constrain the compatibility of material distribution with anatomical structures; The image domain reconstruction loss is calculated as follows:

[0037] ;

[0038] in, The low-energy and high-energy CT image data are the inputs to the material decomposition network; , For the kth basic substance in energy and The mass attenuation coefficient is used to calculate the theoretical linear attenuation coefficient.

[0039] Furthermore, the training process of the material decomposition network is as follows:

[0040] The parameters of the differentiable physical forward projection module and the anatomical perception constraint module are fixed;

[0041] The parameters of the encoder-decomposer network are optimized solely through backpropagation to minimize the multi-task self-supervised loss function. ;

[0042] After training is completed, the differentiable physical forward projection module and the anatomical perception constraint module do not participate in the calculation during the inference phase.

[0043] Secondly, the present invention provides a physical information-constrained self-supervised learning system for CT material decomposition, comprising:

[0044] The data interface module is used to receive dual-energy projection data and reconstructed image data from the CT scanning equipment;

[0045] The network training module is used to execute the above methods to generate a trained material decomposition network;

[0046] The material decomposition and inference module integrates the trained material decomposition network, which is used to output the corresponding multi-basic material density distribution map in real time when new dual-energy CT image data is received.

[0047] Furthermore, the system further includes a physical prior knowledge base, which stores:

[0048] Mass decay coefficients of various basic substances at different X-ray energies ;

[0049] Typical X-ray tube energy spectrum data S(E) and detector response model D(E) used for the differentiable physics forward projection module;

[0050] The parameter file of the pre-trained multi-organ segmentation network model S required by the anatomical perception constraint module.

[0051] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0052] 1. Completely eliminates reliance on high-cost paired annotation data, making it highly practical.

[0053] The core bottleneck of existing deep learning methods lies in their heavy reliance on paired labeled data, such as "multi-energy CT images - real material density maps," which are extremely difficult to obtain. This invention creatively constructs a self-supervised closed loop centered on projection domain data consistency, requiring only a large amount of unlabeled dual-energy CT projection and image data readily available from routine clinical scans as training material. This fundamentally solves the data scarcity problem, significantly reducing the application threshold and implementation cost of the method, enabling high-performance deep learning models to be truly applicable to a wide range of clinical scenarios.

[0054] 2. By embedding a full-link physical model, the strong physical reliability and high accuracy of the decomposition results are ensured.

[0055] Traditional self-supervised methods impose weak constraints only in the image domain, easily leading to physically unreliable solutions. This invention introduces for the first time a fully differentiable physical forward projection module, which accurately simulates the complete imaging chain from matter density to the original projection data (including base matter attenuation, energy spectrum effects, detector response, etc.). By forcing the network to accurately reconstruct the original projection data from the predicted matter density map through this physical model, we transform rigorous physical laws into strong constraints in training. This ensures that the decomposition results not only mathematically fit the data but also possess uniqueness and correctness in physical essence, thus achieving quantitative decomposition accuracy close to or reaching that of supervised methods.

[0056] 3. Introducing anatomical semantic priors greatly enhances the clinical rationality and interpretability of the results.

[0057] To address the "illusion" problem caused by anatomical inconsistencies (such as iodine appearing in bones) that may arise with existing methods, this invention innovatively utilizes a pre-trained segmentation network as a fixed "anatomical referee." Through anatomical prior adversarial loss, the network's predicted distribution of specific substances (such as iodine maps) is forced to be compatible with the inherent anatomical structures of human organs. This constraint seamlessly integrates high-level semantic information into the low-level pixel-level decomposition task, ensuring that the generated substance maps are not only numerically accurate but also spatially consistent with physiological and pathological common sense, significantly improving the clinical acceptability and physician trust of the results.

[0058] 4. Excellent generalization ability and device robustness

[0059] Because the method of this invention learns the essential physical mapping relationships from the original projection data, rather than simply memorizing the distribution of a specific dataset, the learned model has strong cross-domain generalization ability. Regardless of the manufacturer or model of CT equipment, or different scanning protocols (tube voltage, filtering, etc.) and patient groups, this method maintains stable decomposition performance. The design of the dual-stream feature alignment network further eliminates the subtle misalignment effects caused by the scanning time difference between dual-energy images, enhancing the model's practicality and robustness.

[0060] 5. Highly efficient end-to-end inference, easy to integrate clinically.

[0061] While the physical forward projection and anatomical constraint modules are crucial during the training phase, only a lightweight encoder-decomposer network is needed for a single forward propagation during the final inference phase. This means that in clinical deployment, the system can perform material decomposition on newly created dual-energy CT images in real time (typically within milliseconds), much like using a conventional image processing filter, with minimal computational overhead. This "complex training, simple inference" characteristic is highly advantageous for integration into existing hospital PACS systems or image post-processing workstations, meeting the workflow requirements of real-time clinical diagnosis.

[0062] 6. It has pioneered a new paradigm that integrates "physical model-driven" and "data-driven" approaches.

[0063] This invention transcends the traditional approach of simply treating physical models as a layer or post-processing step in a network, proposing a deeply coupled fusion paradigm: a differentiable physics engine provides unshakeable underlying constraints, self-supervised learning mines complex patterns from big data, and anatomical priors are used for high-level semantic correction. This paradigm offers a universal technical approach and feasible solution for addressing other inverse problems in medical imaging lacking labeled data (such as MRI parameter quantification and ultrasound elastography), and has significant methodological implications. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the process of the present invention;

[0065] Figure 2 This is a performance comparison chart between the present invention and the prior art;

[0066] Figure 3 This is a comparison image of the material breakdown effect in clinical CT scans. Detailed Implementation

[0067] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0068] Example 1

[0069] This embodiment provides a physical information-constrained self-supervised learning method for CT material decomposition, including the following steps:

[0070] Acquire dual-energy CT projection data of the target object and its reconstructed CT image data;

[0071] The dual-energy CT image data is input into a trained material decomposition network, which outputs density distribution maps of various basic materials of the target object; wherein, the training method of the material decomposition network includes:

[0072] Construct a training dataset without real material annotations, the dataset containing multiple sets of dual-energy CT projection data and CT image data reconstructed from the projection data;

[0073] Construct a network training system that includes an encoder-decomposer network, a differentiable physical forward projection module, and an anatomical perception constraint module;

[0074] Using the training dataset, with forward projection consistency as the core, and combining anatomical priors and physical laws, the encoder-decomposer network is trained by optimizing a multi-task self-supervised loss function consisting of projection domain data fidelity loss, image domain reconstruction loss, physical consistency loss, and anatomical prior adversarial loss.

[0075] Specifically: such as Figure 1 As shown, firstly, the raw dual-energy CT scan data of the target object are acquired, including two sets of projection data: low-energy (e.g., 80kVp) and high-energy (e.g., 140kVp). , And the low-energy and high-energy CT images reconstructed from these two sets of projection data, respectively. , .

[0076] The core step involves inputting the dual-energy CT image data into a pre-trained material decomposition network, which directly outputs density distribution maps of various basic substances (such as soft tissue, iodine, and calcium) of the target object. .

[0077] Training the material decomposition network is crucial. Before training, a large-scale training dataset needs to be constructed, which contains only a large number of clinical or simulated dual-energy CT data pairs without real material annotations. The network training system consists of three core components: a two-stream encoder-decomposer network for feature extraction and decomposition, a differentiable physical forward projection module for simulating the physical processes of imaging, and an anatomical-aware constraint module (such as a pre-trained segmentation network) for introducing high-level semantic priors. During training, the consistency between the synthesized forward projection data and the real projection data serves as the fundamental supervision signal, while also incorporating anatomical structure constraints and basic physical law constraints. Specifically, this is achieved by constructing and optimizing a multi-task self-supervised loss function. To achieve this, the various losses correspond to projection fidelity, image reconstruction quality, adherence to physical laws, and anatomical rationality, respectively.

[0078] Example 2

[0079] The differentiable physical forward projection module is specifically used for:

[0080] Voxel mass density vector predicted based on encoder-decomposer network ; and the known mass decay coefficient of the base material Calculate the line attenuation coefficient including nonlinear compensation:

[0081] ;

[0082] in The energy-related residual term, predicted from the density vector using a small neural network, is used to model the beam hardening effect.

[0083] Through differentiable forward projection operator Integrating the linear attenuation coefficient yields the single-energy projection, which is then combined with the simulated X-ray tube energy spectrum S(E) and detector response D(E) to synthesize the final dual-energy projection data. and .

[0084] Specifically: This module receives the voxel-level matter density vector predicted by the encoder-decomposer network. And perform the following differentiable calculations:

[0085] First, calculate the linear attenuation coefficient. This is done in addition to using the known mass attenuation coefficient of the base material. In addition to linear combination, this embodiment creatively introduces an energy-related residual term predicted by a small fully connected neural network. The network uses density vectors. The input is denoted as , and the output represents the residual value of the nonlinear effect. Therefore, the linear attenuation coefficient is calculated as follows: This allows the model to dynamically learn and compensate for beam hardening effects that traditional linear models cannot handle.

[0086] Secondly, a differentiable forward projection is performed. This is done using a differentiable distance-driven projection operator. Along all X-ray paths Integrate to obtain a monoenergetic projected sine curve.

[0087] Finally, the dual-energy projection is synthesized. Considering the actual X-ray tube energy spectrum S(E) and the detector energy response D(E), the signal received by the final detector unit is calculated by integrating the energy spectrum:

[0088] ;

[0089] Thus, a complete low-energy projection is synthesized. and high-energy projection The entire computational graph supports automatic differentiation, allowing gradients to backpropagate from the projected loss to the decomposed network.

[0090] Example 3

[0091] The encoder-decomposer network is a two-stream feature-aligned structure, specifically including:

[0092] Two independent shallow coding branches are used to extract the input low-energy CT image. and high-energy CT images Energy-specific characteristics;

[0093] A deformable feature alignment module, which uses the deep feature map of the low-energy branch as a reference, performs deformation alignment on the deep feature map of the high-energy branch through the learned spatial offset field to ensure accurate matching of anatomical structures during subsequent fusion;

[0094] A shared decoder is used to decode aligned and fused features into a multi-channel underlying mass density map.

[0095] Specifically, this network has a dual-input, single-output architecture. The two input branches process low-energy images independently. and high-energy images In shallow layers (e.g., the first 3 convolutional blocks), the two branches do not share weights in order to fully learn and preserve energy-specific features.

[0096] Prior to deep fusion, the key innovation of this embodiment lies in the introduction of a deformable feature alignment module (DFAM). This module uses deep feature maps with low-energy branches. It serves as a spatial reference standard. Internally, it contains a light quantum network, characterized by its assembled features. As input, learn a dense two-dimensional spatial migration field. Then, using deformable convolution operations, this offset field is applied. High-energy feature map Perform non-rigid spatial deformation to obtain aligned feature maps. This operation can automatically correct for subtle anatomical misalignments between dual-energy images caused by differences in patient breathing, heart rate, or scanning timing.

[0097] Finally, the aligned features With reference features Element-wise addition is performed, and then the data is fed into a shared decoder for upsampling and convolution, ultimately outputting a multi-channel (K-channel) basic mass density map.

[0098] Example 4

[0099] The anatomical perception constraint module is a fixed, pre-trained multi-organ segmentation network S; the anatomical prior adversarial loss Constructed and calculated in the following manner:

[0100] Decompose the specific mass density map output by the network Compared with the original low-energy CT images The images are stitched together along the channel dimension to form a composite image;

[0101] Combine the combined image with individual Input the segmentation network S into each of the two groups of organ probability distribution maps;

[0102] Calculate the KL divergence between the two probability distributions as the loss:

[0103] This constrains the distribution of matter to conform to known anatomical priors.

[0104] Specifically: The anatomical perception constraint module is a semantic segmentation network S pre-trained on a large public dataset (such as MSD or CT-ORG) that is capable of segmenting multiple organs (such as liver, kidney, bone, aorta, etc.). Its parameters are frozen during training and do not participate in updates.

[0105] Taking the constraint of iodine distribution as an example, the specific process is as follows:

[0106] Path A (Baseline): Transform the raw low-energy CT images By inputting a segmentation network S alone, a set of baseline organ probability distribution maps are obtained. .

[0107] Path B (Combination): Decomposes the iodine density map predicted by the network. With the same original image The input channels are concatenated along the channel dimension to form a new two-channel input. Then input the same segmentation network S again to obtain another set of probability distribution diagrams. .

[0108] Loss calculation: For each organ category o (e.g., the vascularized liver), calculate the KL divergence between the probability distributions of the outputs of path B and path A:

[0109] ;

[0110] Summing over all organ categories yields... .

[0111] The significance of this loss function lies in the fact that it mandates that when the segmentation network "sees" an image overlaid with iodine map information, the probability distribution of organs it determines should not change. This is equivalent to requiring that the predicted iodine distribution pattern be highly consistent with the actual anatomical structure (e.g., iodine should be concentrated in blood vessels and blood-rich organs, and should not appear in bones or air), thereby using high-level semantic priors to guide low-level pixel decomposition.

[0112] Example 5

[0113] The data fidelity loss in the projection domain Defined as:

[0114] ;

[0115] in, , For the input of real low- and high-energy projection data, , The corresponding projection data synthesized by the differentiable physical forward projection module.

[0116] Specifically, this loss is the most crucial and binding element in the self-supervised signal. In each forward propagation during the training phase, the differentiable physics forward projection module outputs a synthesized low-energy projection. and high-energy projection .

[0117] Projection domain data fidelity loss Directly compare these synthetic projections with the raw projection data obtained from real CT scan equipment. , The difference between them. This embodiment uses the L1 norm (sum of absolute errors) for calculation:

[0118] ;

[0119] in, This represents the summation of the absolute values ​​of all elements in the matrix. Compared to the L2 norm (mean squared error), the L1 norm is less sensitive to outliers (such as metal artifacts), provides a more robust gradient, and contributes to training stability and clarity of decomposition results.

[0120] Example 6

[0121] The physical consistency loss Includes: Non-negativity constraint on matter density:

[0122] Force all predicted densities of basic materials to be non-negative; Material distribution smoothing constraint: Apply total variational (TV) regularization to each material density map to suppress unreasonable noise and promote spatial smoothness.

[0123] Specifically: The first item is the non-negativity constraint loss of material density. In physics, the density (or concentration) of any fundamental substance cannot be negative. Therefore, for each voxel and each fundamental substance k predicted by the network, the density value is... We apply a non-negative penalty to it. Specifically, the implementation is as follows:

[0124] ;

[0125] here, The function will handle all negative values. The output values ​​are left as positive numbers, while positive values ​​are changed to zero. Then, the squared L2 norm of these positive values ​​is calculated. This loss term drives the network to correct all predictions that produce negative density during training.

[0126] The second item is the loss due to the smoothing constraint of material distribution. We assume that within a small, localized region, the density of the same substance changes continuously, without drastic or meaningless oscillations. Therefore, for each substance's density map... Apply anisotropic total variation (TV) regularization:

[0127] ;

[0128] This loss effectively suppresses noise in the resulting image while maintaining sharp edges and promoting spatial smoothness. The final physical consistency loss is a weighted sum of the two:

[0129] .

[0130] Example 7

[0131] The multi-task self-supervised loss function It is the weighted sum of the losses of each component, and its expression is:

[0132] ;

[0133] in, , , , These are preset positive weighting coefficients; The image domain reconstruction loss is calculated as follows:

[0134] .

[0135] Specifically: Total Loss Function It is a weighted sum of four sub-loss terms, each responsible for supervising a different aspect of the model's learning:

[0136] ;

[0137] Image domain reconstruction loss, used as auxiliary supervision, ensures that the decomposition result can linearly reconstruct the input image. Its calculation method is as follows: Weight 0.5 is acceptable.

[0138] Physical consistency loss, enforces adherence to fundamental physical laws, weights 0.2 is acceptable.

[0139] : Dissect prior adversarial loss, introduce high-level semantic constraints, and weights 0.1 is acceptable.

[0140] In the early stages of training, the intensity can be appropriately increased. and The weights are adjusted to allow the network to quickly converge to an initial solution that conforms to both physical and image reconstruction. In the later stages of training, the weights are gradually increased. and The weights are adjusted for fine-tuning and to dissect reasonable constraints. This dynamic or phased weighting strategy helps optimize overall performance.

[0141] Example 8

[0142] The training process of the material decomposition network is as follows:

[0143] The parameters of the differentiable physical forward projection module and the anatomical perception constraint module (segmentation network S) are fixed.

[0144] The parameters of the encoder-decomposer network are optimized solely through backpropagation to minimize the multi-task self-supervised loss function. ;

[0145] After training is completed, the differentiable physical forward projection module and the anatomical perception constraint module do not participate in the calculation during the inference phase.

[0146] Specifically: Training phase:

[0147] A phased, modular training strategy is adopted. First, the physical parameters in the differentiable physics forward projection module (such as...) are fixed. All parameters of the encoder-decomposer network (S(E) model) and the anatomical-aware constraint module (segmentation network S) are calculated. Then, training data is batch-fed into the network training system. In forward propagation, the encoder-decomposer network outputs a density map, and each loss is calculated via the forward projection module and the anatomical constraint module, respectively. In backpropagation, only the gradients of the encoder-decomposer network parameters are calculated, and their parameters are updated using the Adam optimizer. The small residual network in the differentiable physics forward projection module, as part of the encoder-decomposer network, also has its parameters updated.

[0148] Reasoning stage:

[0149] After training, only the optimized encoder-decomposer network is retained and packaged into the final material decomposition inference model. At this point, the differentiable physical forward projection module and the segmentation network S are no longer needed. For new dual-energy CT data, only the image pairs need to be processed. By inputting this inference model and performing a simple forward propagation, a mass density map can be output within milliseconds. This ensures the efficiency and practicality of clinical deployment.

[0150] Example 9

[0151] This embodiment provides a physical information-constrained self-supervised learning system for CT material decomposition, including:

[0152] The data interface module is used to receive dual-energy projection data and reconstructed image data from the CT scanning equipment;

[0153] The network training module is used to execute the above methods to generate a trained material decomposition network;

[0154] The material decomposition and inference module integrates the trained material decomposition network, which is used to output the corresponding multi-basic material density distribution map in real time when new dual-energy CT image data is received.

[0155] Specifically, the system consists of three main modules:

[0156] Data interface module: Responsible for interfacing with CT scanning equipment or PACS (Picture Archiving and Communication System) to receive raw dual-energy projection data streams. , It then calls built-in or integrated reconstruction algorithms (such as FBP or iterative reconstruction) to generate corresponding dual-energy CT images. , This involves standardizing and preprocessing the data.

[0157] Network Training Module: This is an offline working module. It integrates all the aforementioned training components: encoder-decomposition network, differentiable physical forward projection library, anatomical prior model library, and multi-task loss calculator. Users or administrators can import new unlabeled datasets, start the training process, and output trained material decomposition network model files (such as .pt or .onnx format).

[0158] Material decomposition inference module: This is the core of the online service. It loads the optimal model file generated by the training module. When the data interface module sends the dual-energy CT images of the patient to be analyzed, the inference module calls the model to perform real-time calculations, quickly generating and outputting multi-material decomposition results such as iodine maps, calcium maps, and water maps, which can be directly displayed on the workstation or transferred to post-processing software.

[0159] Example 10

[0160] The system further includes a physical prior knowledge base, which stores:

[0161] Mass decay coefficients of various basic substances at different X-ray energies ;

[0162] Typical X-ray tube energy spectrum data S(E) and detector response model D(E) used for the differentiable physics forward projection module;

[0163] The parameter file of the pre-trained multi-organ segmentation network model S required by the anatomical perception constraint module.

[0164] Specifically: This knowledge base is a structured database or file storage system that provides verifiable physical and anatomical priors for the entire methodology, including:

[0165] Mass decay coefficient library: Stores mass decay coefficients of basic substances such as soft tissue, fat, iodine, calcium, water, and bone in a continuous energy range from 10 keV to 200 keV, provided by international standards organizations (such as NIST), in the form of lookup tables (LUTs) or parametric models. .

[0166] Imaging System Characteristic Library: Stores typical X-ray tube energy spectrum data for different brands and models of CT equipment (such as Siemens Somatom Force, GE Revolution). , The system provides the energy response function D(E) model for the corresponding detector. Based on the input DICOM header file information, the system can automatically match or select the closest model.

[0167] Anatomical Prior Model Library: Stores weight files of one or more robust multi-organ segmentation network models S pre-trained on different datasets. Depending on the clinical application scenario (e.g., chest and abdomen, head), different specialized models can be called to provide the most relevant anatomical distribution prior knowledge.

[0168] Example 11

[0169] This embodiment is illustrated by... Figure 2 The performance comparison chart shown provides a quantitative evaluation and comparative analysis of the overall performance of the present invention. Figure 2 The performance differences between the method of this invention and existing mainstream technologies were compared from different dimensions.

[0170] like Figure 2 (a) shows a comparison of the accuracy of substance quantification. The method of the present invention is significantly better than the traditional physical method (15.2 / 25.8 mg / mL) and the existing self-supervised method (12.4 / 21.5 mg / mL) in iodine quantification (RMSE 6.3 mg / mL) and calcium quantification (RMSE 9.7 mg / mL), with an improvement in quantification accuracy of 58%-62%.

[0171] like Figure 2 (b) shows the comparison of generalization ability. The performance retention rate of the method of the present invention in cross-device generalization (88.6%) and cross-protocol generalization (85.3%) is significantly higher than that of supervised learning methods (62.8% / 54.6%) and traditional methods (45.2% / 38.7%), indicating that it has stronger clinical adaptability and robustness.

[0172] like Figure 2 (c) in the figure presents a comparison of training data requirements. The method of this invention only requires unlabeled data (100%), which completely avoids the heavy dependence of supervised learning methods on paired labeled data (95%), and fundamentally solves the data acquisition bottleneck.

[0173] like Figure 2 (d) In the comprehensive performance radar chart, the method of the present invention is comprehensively evaluated from five dimensions: quantitative accuracy (9 points), generalization ability (9 points), data efficiency (10 points), physical reliability (9 points), and computational efficiency (8 points). The method of the present invention has balanced and excellent performance in all indicators, with a comprehensive score of 9.0 points, which is superior to other methods in all aspects.

[0174] The comparison chart visually verifies that while ensuring high precision and strong generalization, the present invention completely eliminates the dependence on paired data, achieving a comprehensive improvement in technical performance.

[0175] Example 12

[0176] This embodiment is illustrated by... Figure 3 The comparison of material decomposition results shown in the clinical CT images intuitively demonstrates the significant advantages of the method of this invention in actual clinical applications. The first row of the image compares the decomposition results of different methods: the original dual-energy CT image shows the anatomical structure of the liver, the traditional method results in obvious noise and blurred boundaries, the supervised learning method has some improvement but still has diffusion artifacts, while the method of this invention presents the clearest and most accurate material decomposition results, with iodine distribution (red) and calcification (green) highly consistent with the real anatomical structure.

[0177] The second line details the breakdown of the method of the present invention: the soft tissue density map clearly outlines the liver contour; the iodine distribution map accurately displays vascular structures and tumor enhancement areas; the calcification distribution map not only detects major plaques but also clearly shows microcalcifications; the breakdown error map shows an average error of only 0.059 and a PSNR of 42.6 dB; and the clinical diagnostic markers accurately locate the tumor, blood vessels, and calcification areas.

[0178] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A self-supervised learning method for physical information constraints in CT material decomposition, characterized in that, Includes the following steps: Acquire dual-energy CT projection data of the target object and its reconstructed dual-energy CT image data; The dual-energy CT image data is input into a trained material decomposition network, which outputs density distribution maps of various basic materials of the target object; wherein, the training method of the material decomposition network includes: Construct a training dataset without real material annotations, the dataset containing multiple sets of dual-energy CT projection data and CT image data reconstructed from the projection data; Construct a network training system that includes an encoder-decomposer network, a differentiable physical forward projection module, and an anatomical perception constraint module; The encoder-decomposer network is a two-stream feature-aligned structure, specifically including: Two independent shallow coding branches are used to extract the input low-energy CT image. and high-energy CT images Energy-specific characteristics; A deformable feature alignment module, which uses the deep feature map of the low-energy branch as a reference, performs deformation alignment on the deep feature map of the high-energy branch through the learned spatial offset field to ensure accurate matching of anatomical structures during subsequent fusion; A shared decoder is used to decode the aligned and fused features into a multi-channel underlying mass density map; The differentiable physical forward projection module is specifically used for: Voxel mass density vector predicted based on encoder-decomposer network and known mass decay coefficient of the base material Calculate the line attenuation coefficient including nonlinear compensation: ; in The energy-related residual term, predicted from the density vector using a small neural network, is used to model the beam hardening effect. Through differentiable forward projection operator Integrating the linear attenuation coefficient yields the single-energy projection, which is then combined with the simulated X-ray tube energy spectrum S(E) and detector response D(E) to synthesize the final dual-energy projection data. and ; The anatomical perception constraint module is a fixed, pre-trained multi-organ segmentation network S; Using the training dataset, with forward projection consistency as the core, and combining anatomical priors and physical laws, the encoder-decomposer network is trained by optimizing a multi-task self-supervised loss function consisting of projection domain data fidelity loss, image domain reconstruction loss, physical consistency loss, and anatomical prior adversarial loss.

2. The method according to claim 1, characterized in that, The anatomical prior adversarial loss It is constructed and calculated in the following manner: Decompose the specific mass density map output by the network Compared with the original low-energy CT images The images are stitched together along the channel dimension to form a composite image; Combine the combined image with the individual Input the segmentation network S into each of the two groups of organ probability distribution maps; Calculate the KL divergence between the two probability distributions as the loss: This constrains the distribution of matter to conform to known anatomical priors.

3. The method according to claim 1, characterized in that, The data fidelity loss in the projection domain Defined as: ； in, , For the input of real low- and high-energy projection data, , The corresponding projection data synthesized by the differentiable physical forward projection module.

4. The method according to claim 1, characterized in that, The physical consistency loss Includes: Non-negativity constraint on matter density: Force all predicted densities of basic substances to be non-negative; in, The total number of basic material types is used to represent the number of basic materials to be distinguished in the decomposition task; The first prediction for the encoder-decomposer network The density value of a basic substance, its physical unit is mass / volume; To correct the linear unit function acting on The result is used to penalize only negative density values; The loss term is a non-negative constraint term for material density, and negative predicted values ​​are penalized to ensure physical plausibility. Material distribution smoothing constraint: Apply total variational (TV) regularization to each material density map to suppress unreasonable noise and promote spatial smoothness.

5. The method according to claim 4, characterized in that, The multi-task self-supervised loss function It is the weighted sum of the losses of each component, and its expression is: ; in, , , , These are the weighting coefficients for the projection domain data fidelity loss, image domain reconstruction loss, physical consistency loss, and anatomical prior adversarial loss, respectively, used to balance the contribution of each loss term to the overall objective function; This refers to the loss of fidelity in the projection domain, used to measure the difference between real and synthetic projection data; This is the image domain reconstruction loss, used to ensure that the predicted material density can accurately reconstruct the input dual-energy CT image; The physical consistency loss includes nonnegativity and smoothness constraints to ensure that the decomposition results conform to physical laws. To mitigate the loss of anatomical priors, it is used to constrain the compatibility of material distribution with anatomical structures; The image domain reconstruction loss is calculated as follows: ; in, The low-energy and high-energy CT image data are the inputs to the material decomposition network; For the kth basic substance in energy and The mass attenuation coefficient is used to calculate the theoretical linear attenuation coefficient.

6. The method according to claim 1, characterized in that, The training process of the material decomposition network is as follows: The parameters of the differentiable physical forward projection module and the anatomical perception constraint module are fixed; The parameters of the encoder-decomposer network are optimized solely through backpropagation to minimize the multi-task self-supervised loss function. ; After training is completed, the differentiable physical forward projection module and the anatomical perception constraint module do not participate in the calculation during the inference phase.

7. A self-supervised learning system for physical information constraints in CT material decomposition, characterized in that, include: The data interface module is used to receive dual-energy projection data and reconstructed image data from the CT scanning equipment; A network training module is used to perform the method of any one of claims 1 to 6 to generate a trained material decomposition network; The material decomposition and inference module integrates the trained material decomposition network, which is used to output the corresponding multi-basic material density distribution map in real time when new dual-energy CT image data is received.

8. The system according to claim 7, characterized in that, The system further includes a physical prior knowledge base, which stores: Mass decay coefficients of various basic substances at different X-ray energies ; Typical X-ray tube energy spectrum data S(E) and detector response model D(E) used for the differentiable physics forward projection module; The parameter file of the pre-trained multi-organ segmentation network model S required by the anatomical perception constraint module.

Images