A proton density weighted and T1 weighted MRI image conversion method and system

Abstract

The present invention belongs to the technical field of medical image processing and analysis, and discloses a proton density weighted and T1-weighted MRI image conversion method and system. After the method performs standardization processing on the PD-weighted image and the corresponding T1-weighted image, a conversion relationship from the PD-weighted image to the T1-weighted image is established based on the deep learning network prediction cSDE model; during the conversion process, an interface is provided to allow the operator to adjust the model parameters in real time according to specific clinical needs to adapt to specific clinical needs or to cope with the performance of the model under specific circumstances; the converted T1-weighted image is sharpened and contrast adjusted through a post-processing module. The present invention is based on a controlled stochastic differential equation model, and can generate a corresponding T1-weighted image from a PD-weighted image, solving the problems of insufficient generalization ability of image conversion, high dependence on large-scale training data sets, large consumption of computing resources, and lack of real-time control mechanism in the prior art.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing and analysis, and in particular relates to a proton density weighted and T1 weighted MRI image conversion method and system. Background Art

[0002] The study of conversion between proton density-weighted (PD) and T1-weighted (T1) MRI images is one of the important topics in the field of medical image processing and analysis, each with unique advantages and applications. T1-weighted images provide high-contrast anatomical structures, especially the clear boundary between gray matter and white matter, which is suitable for anatomical research and structural analysis, while PD-weighted images provide a contrast between T1 and T2, which can show more tissue details and density changes. It has important applications in fine brain tissue research and the detection of specific lesions. The information of these two techniques usually needs to be used in combination to make the most comprehensive assessment of certain diseases.

[0003] The techniques for converting PD-weighted images to T1-weighted images are mainly based on deep learning methods, such as convolutional neural networks (CNNs) and generative adversarial networks (GANs). These techniques use a large number of paired PD-weighted images and T1-weighted images for training to learn a mapping relationship from one imaging modality to another. Although these techniques have made some progress, they also have the following disadvantages. 1) A large amount of labeled data is usually required, which is very expensive to obtain in practical applications. 2) When faced with new data that is significantly different from the training set in statistical properties, the performance of existing models usually drops sharply. This limits the application of models between different machines and different patient groups. 3) The demand for computing resources during the training phase is extremely high, and expensive GPU hardware needs to be used. This is not realistic in resource-limited medical environments. 4) The conversion process lacks interpretability and control, and is usually a black box operation during the image conversion process, and the model cannot be adjusted in real time to adapt to specific clinical needs.

[0004] Through the above analysis, the problems and defects of the existing technology are: the existing technology is overly dependent on large-scale data, has weak generalization ability, requires too much computing resources, has weak interpretability and cannot be dynamically adjusted. Summary of the invention

[0005] In order to overcome the problems existing in the related art, the disclosed embodiment of the present invention provides a proton density weighted and T1 weighted MRI image conversion method and system, and the technical solution is as follows:

[0006] The present invention is implemented in this way: a proton density weighted and T1-weighted MRI image conversion method comprises:

[0007] S1, data preprocessing: for the initial PD-weighted image and the corresponding T1-weighted image, the data preprocessing module is used to standardize the data to obtain a training image pair;

[0008] S2, cSDE model training: The preprocessed training image pairs are input into the cSDE model. Based on the prediction of the deep learning network, the conversion relationship from PD-weighted image to T1-weighted image is established according to the forward statistical feature dilution process and the dynamic adjustment process of the difference between the current predicted image data and the theoretical restored image, and the cSDE image conversion model that directly converts the PD-weighted image to the corresponding T1-weighted image is obtained;

[0009] S3, image conversion execution: receiving a PD-weighted image to be converted into a corresponding T1-weighted image, and outputting a T1-weighted image generated under the guidance of the PD-weighted image through a cSDE image conversion model;

[0010] S4, post-processing: The converted T1-weighted image is optimized through a post-processing module, which includes a sharpening processing module and a contrast adjustment module, to obtain a new T1-weighted image.

[0011] In step S1, the input of the data preprocessing module is the system input Image data pairs, where is the original T1-weighted image with noise, For The corresponding original PD-weighted image, For the sample pairs;

[0012] The output of the data preprocessing module is the paired data pairs used to train the cSDE model. ,in, is the T1-weighted image after preprocessing. is the preprocessed PD-weighted image, which serves as the controlled condition of the cSDE model input.

[0013] Further, data preprocessing includes:

[0014] Data denoising: Use the trained small-core CNN denoising module to denoise the original data pair. The variable with the symbol represents that the denoising operation has been performed; for T1-weighted images, the denoising module uses the original image with noise and the corresponding noise-free image to Conduct training, including is the original T1-weighted image with noise, is the corresponding noise-free T1-weighted image;

[0015] For PD weighted images, the denoising module uses the original noisy image and the corresponding noise-free image pair. Conduct training, including is the original PD-weighted image with noise, is the corresponding noise-free PD weighted image; through training, the model converts the original noisy Convert to noise-free ;

[0016] Data contrast enhancement: The contrast enhancement module receives the denoised data. , applying nonlinear transformation to enhance dark or bright areas of the image; the contrast enhancement module For each pixel value in ,in, is the output pixel value, is the scaling factor, is the gamma value, is the input pixel value;

[0017] After the data contrast enhancement module, Convert to ,in, After being processed by the contrast enhancement module , After being processed by the contrast enhancement module ;

[0018] Data normalization: When processing PD-weighted images and T1-weighted image data, the data normalization module uses the minimum and maximum normalization method to process the data. This module uses the normalization formula Process each pixel value, where is the original pixel value, is the minimum pixel value, is the maximum pixel value, is the normalized pixel value;

[0019] After the data preprocessing normalization module, Convert to ,in, After being processed by the normalization module , After normalization ;

[0020] The output of the data preprocessing normalization module Defined as ,in, After being processed by the preprocessing module , After being processed by the preprocessing module .

[0021] For the convenience of representation, the output of the data preprocessing normalization module is Defined as , is the input of the preprocessing module, that is, the original T1-weighted image with noise, The preprocessing module is for The final output is the preprocessed T1-weighted image; Yes and The corresponding noisy original PD weighted image is also another input data of the preprocessing module. The preprocessing module is for The final output, that is, the preprocessed PD-weighted image, are paired input data, It is the paired output data obtained after the preprocessing module, and is also the input data of the next module.

[0022] In step S2, the cSDE model training is divided into a forward diffusion stage and a reverse diffusion stage. In the forward diffusion stage, the cSDE model simulates the time state from 0 to The original data distribution is gradually transformed into a Gaussian distribution in the process of inverse diffusion. The cSDE model changes the simulation time state from The reverse process to 0.

[0023] Further, the forward diffusion stage and the reverse diffusion stage include:

[0024] Variable settings: For in time Image in state , For in time Image in state The probability density function of For images From time Time The state transition probability of , when it obeys Gaussian distribution, is , For the deadline, is the original distribution of the data, is the initial probability distribution of the system state, is the probability distribution of the system state after a series of processes; in the cSDE framework, as time goes from 0 to The system gradually evolved from Towards change, for a certain time The state of the system after adding noise and other disturbances;

[0025] Noise dilution in the diffusion stage: When calculating the diffusion process of the cSDE model, solve ,in, Representation variables Compliance interval The uniform distribution on Representation interval Uniform distribution on is the drift term, which controls how the image data evolves over time; is the diffusion term, which controls the randomness in the process; For small changes in the state of image data, It is the differential of Brownian motion, which is used to simulate the change of random fluctuations in a small time interval. In this stage, the model simulates the noise in the image acquisition process in the real world by adding randomness, which is used to restore the original image data distribution in the subsequent inverse process.

[0026] Image restoration in the inverse diffusion stage: When calculating the inverse diffusion process of the cSDE model, solve , used to simulate from any time to 0, where represents the inverse Brownian motion, is the gradient term, which represents the trend of the system state evolution over time and is used to find the current probability density function of given data A more appropriate evolution direction in the system, guiding the system to follow the current probability density function The logarithmic gradient of the image data is adjusted; Represents the current probability density function Take the logarithm value;

[0027] cSDE model training optimization: The cSDE model design objective function is , where For expectations, Indicates at time The expected value of Representing images The expected value of the initial time, Representing images From time Return to the expected value of the initial time, is a weighting function greater than 0, used to adjust the importance; , using a neural network approximate ,in, are the parameters of the neural network, For images From time status The gradient of the probability density function when restored to the initial time state 0 indicates the optimal direction of recovery. The task of the model is to perform a high-precision image restoration process by approximating the gradient. Represents the probability density distribution in the original state of the data.

[0028] In step S3, the cSDE model uses The method uses supervised learning training on image data; introduces a prediction-correction sampling method, and combines basic prediction steps and multiple correction steps to gradually reversely reconstruct the process from the terminal noise state to the initial true state; the process includes continuous-time simulation, solving the cSDE model through numerical methods, and generating T1-weighted images; during the conversion process, an interface is provided to allow the operator to adjust the model parameters according to specific clinical needs. The model parameters include diffusion intensity. The model parameters can be adjusted in real time to adapt to specific clinical needs or to cope with the performance of the model in specific situations.

[0029] In step S3, image conversion is performed, including:

[0030] Guidance information setting: Add guidance information during the forward and inverse training of diffusion, using PD weighted images As guidance information, and train the network in a supervised manner;

[0031] Computing resource dependency reduction settings: Settings , using Gaussian noise at different scales Dilute the original data distribution and model the data, then express yes About time status Function of For images In Gaussian noise The probability density function of image The probability density function of For images In Gaussian noise The probability density function of

[0032] Given , for The square root of the derivative of the rate of change of , then , by calculating The time derivative of the diffusion rate ; Due to the addition of guidance information during training , neural network Updated to ; During guided training, the solution in the positive diffusion process is Simplified to , the solution in the inverse diffusion process is Simplified to , given It means the mean , the covariance is The multivariate normal distribution of is the identity matrix, then , so the objective function is ;

[0033] Prediction Correction Sampling: From Prior Distribution Take a sample from , and then numerically solve ;when When it is large, the mean of the prior distribution is close to 0, and the prior distribution is approximated as ; The predictor of the prediction-correction sampling method adopts a fixed discretization strategy and uses small increments replace And use Gaussian noise Alternative ,therefore ,in ; The corrector uses annealed Langevin dynamics and sets the Langevin step size ,in, is the signal-to-noise ratio, According to the annealing Langevin kinetic formula Take samples, For time The state variables at for The state variables at is a random variable with standard normal distribution The relevant disturbance term, Adjust the amplitude of random perturbations; represents the multivariate normal distribution from 0 to the identity matrix, represents Gaussian noise, Indicates time The Gaussian noise function is For a simplified inverse Brownian motion, Represents the probability density distribution of the original state of the data; For neural networks, The Gaussian noise function is represents the state variable at time i, Gaussian noise at time i;

[0034] Image conversion training: The cSDE model is trained repeatedly until the loss function converges; the diffusion control term The operator adjusts the model parameters according to specific clinical needs to adjust the model optimization direction in real time, based on the input PD-weighted image Generate the corresponding T1-weighted image .

[0035] In step S4, the converted T1-weighted image is enhanced via a post-processing module.

[0036] In step S4, post-processing includes:

[0037] Sharpening: Design a sharpening module, use a high-pass filter, enhance the high-frequency components to highlight the details and boundaries in the image, and the input of the sharpening module is the T1-weighted image output in step S3 , the output of the sharpening module is , which is also the input of the contrast adjustment module; represents a weighted image after sharpening;

[0038] Contrast adjustment: Design a contrast adjustment module and use adaptive histogram equalization technology to adjust the contrast. Partition block processing is divided into , and then for each independent block Perform contrast enhancement; set the window size parameters of adaptive histogram equalization to adapt to different image areas and diagnostic requirements; finally, the enhanced independent blocks Reorganize and output the final desired image.

[0039] The object of the present invention is to provide a proton density weighted and T1-weighted MRI image conversion system, which is used to control the proton density weighted and T1-weighted MRI image conversion method, and the system comprises:

[0040] A data preprocessing module, used for standardizing the data of the PD-weighted image and the corresponding T1-weighted image using the data preprocessing module, including denoising, contrast enhancement and normalization;

[0041] cSDE model training module, cSDE sets drift term and diffusion term to control the smooth transition path of image data and the introduced random disturbance respectively; cSDE model is predicted based on deep learning network and dynamically adjusted according to the difference between the current predicted image data and the theoretical restored image;

[0042] The image conversion execution module is used to establish the conversion relationship from PD-weighted images to T1-weighted images by training the cSDE model. During the conversion process, an interface is provided to allow the operator to adjust the model parameters in real time according to specific clinical needs to adapt to specific clinical needs or to cope with the performance of the model in specific situations.

[0043] The post-processing module is used to optimize the converted T1-weighted image through the post-processing module. The post-processing module includes a post-processing sharpening processing module and a post-processing contrast adjustment module.

[0044] In combination with all the above technical solutions, the beneficial effects of the present invention are as follows:

[0045] In view of the defects and shortcomings of the existing technology, such as excessive reliance on large-scale data, weak generalization ability, excessive demand for computing resources, weak interpretability and inability to be dynamically adjusted, the present invention carries out the following four optimization designs: 1) By introducing controlled stochastic differential equations and utilizing the dynamic modeling capabilities of cSDE, the system can achieve better learning effects on limited training samples and reduce dependence on large-scale data sets. 2) The dynamic modeling framework provided by cSDE enhances the model's adaptability to new data, thereby improving its generalization performance in different clinical environments. 3) Through the designed numerical solution strategy, the cSDE model can run on standard computing resources, reducing dependence on high-end hardware. 4) cSDE allows real-time adjustment of model parameters and customization of the image conversion process according to specific clinical needs, providing higher controllability and flexibility.

[0046] Existing technologies, especially deep learning-based image conversion methods (such as CNN and GAN), require a large amount of paired labeled data for training, which is often difficult to obtain in practical applications and is costly. The present invention reduces the dependence on large-scale labeled data sets by introducing a controlled stochastic differential equation (cSDE) model. cSDE can effectively use limited samples for more accurate image generation and conversion, optimize the information utilization of data through noise perturbations in the simulation and inverse simulation process, and improve the performance of the model on a small amount of data.

[0047] The performance of traditional image conversion models usually drops sharply when faced with new data that is statistically significantly different from the training set. This limits the wide application of the model between different machines and different patient groups. The cSDE model in the present invention provides a dynamic modeling framework that enhances the model's adaptability to new data. By adjusting the drift term and diffusion term in real time during model training and execution, the cSDE model can adapt to changes in various clinical environments, thereby improving its generalization performance.

[0048] Deep learning models, especially during the training phase, have extremely high requirements for computing resources and often rely on expensive GPU devices, which is not practical in resource-constrained medical environments. The present invention uses a designed numerical solution strategy to enable the cSDE model to run on standard computing resources, reducing the reliance on high-end hardware. In addition, the controlled perturbation and reverse recovery steps introduced during the training process of the cSDE model of the present invention make the model more efficient in utilizing computing resources.

[0049] The image conversion process is often a black box operation in the prior art, and it is difficult to adjust in real time to adapt to specific clinical needs in practical applications. The cSDE model of the present invention provides higher controllability and flexibility by allowing the operator to adjust model parameters (such as diffusion intensity) in real time during the conversion process. The present invention not only increases the interpretability of the model, but also enables clinicians to customize the image conversion process according to specific medical needs, thereby improving the effectiveness and safety of clinical applications.

[0050] The above innovations not only improve the performance and applicability of image conversion technology, but also ensure the feasibility and effectiveness of the technology in actual medical environments, significantly improving the quality and efficiency of medical image processing.

[0051] The PD-weighted image and T1-weighted image conversion method based on the controlled stochastic differential equations (cSDE) model of the present invention can effectively generate corresponding T1-weighted images from PD-weighted images, solving the problems of insufficient generalization ability of image conversion, high dependence on large-scale training data sets, large consumption of computing resources and lack of real-time control mechanism in the prior art.

[0052] As the positive effects of the present invention, it is also embodied in the following important aspects:

[0053] 1) Improve diagnostic accuracy: Using advanced algorithms to improve image quality can more clearly display details such as the edges and internal structures of tumors. This helps doctors make more accurate diagnoses, especially in early disease detection and identification of small lesions. The application of this technology is expected to increase treatment success rates and reduce medical costs caused by misdiagnosis.

[0054] 2) Improve medical efficiency: Through real-time image processing technology, high-quality medical images can be obtained in a short time, thereby speeding up the diagnosis process and improving medical efficiency. This is especially important for emergency and intensive care, and can provide timely medical intervention for patients at critical moments.

[0055] 3) Reduced operation and maintenance costs: This technical solution reduces the reliance on high-performance hardware and can be run in a standard computing environment. This not only reduces the initial hardware investment cost, but also reduces the long-term maintenance and upgrade costs. In addition, due to the reduced reliance on high-cost medical imaging equipment, small and medium-sized medical institutions can also afford advanced medical image processing technology.

[0056] 4) Expansion to new clinical applications: Improved image processing technology can be applied to a variety of clinical scenarios, such as radiology, neurology, and cardiology, to enhance diagnostic capabilities in these fields. In addition, high-quality image conversion technology can also be used to support the development of machine learning and artificial intelligence algorithms to provide support for future automated diagnosis and treatment.

[0057] 5) Increase market competitiveness and brand value: The use of advanced technology solutions can enhance the market competitiveness of medical institutions and attract more patients to choose high-standard diagnosis and treatment services. In addition, technological innovation can also enhance the brand image of enterprises and increase their visibility and influence in the industry.

[0058] In general, the commercial application of this technical solution is expected to bring significant economic benefits and social value, especially in improving medical quality, reducing costs and promoting medical technology innovation. This will not only help promote the advancement of related technologies, but also help improve the service level and efficiency of the entire medical industry.

[0059] 6) Reduced reliance on large-scale datasets: Existing image conversion technologies, such as deep learning-based convolutional neural networks (CNNs) and generative adversarial networks (GANs), generally rely on a large amount of paired labeled data for training. This is often difficult to achieve in practical applications because the cost of obtaining high-quality paired labeled data is very high. This invention reduces the reliance on large-scale labeled datasets by introducing a controlled stochastic differential equation (cSDE) model, achieving more accurate image generation and conversion.

[0060] 7) Improve the generalization and adaptability of the model: Traditional image conversion models usually experience a sharp drop in performance when faced with new data that is statistically significantly different from the training set. The cSDE model of the present invention provides a dynamic modeling framework that enhances the model's adaptability to new data, allowing the model to have better generalization performance and wide application across different machines and patient groups.

[0061] 8) Reduce the demand for computing resources: The existing technology has extremely high demand for computing resources during the training phase, and usually relies on expensive GPU hardware. The numerical solution strategy designed by the present invention enables the cSDE model to run on standard computing resources, significantly reducing the dependence on high-end hardware.

[0062] 9) Enhanced interpretability and real-time control mechanism: The existing image conversion process is often a black box operation, and it is difficult to adjust in real time to adapt to specific clinical needs in practical applications. The cSDE model of the present invention allows the operator to adjust the model parameters (such as diffusion intensity) in real time during the conversion process, providing higher controllability and flexibility. This not only increases the interpretability of the model, but also enables clinicians to customize the image conversion process according to specific medical needs, thereby improving the effectiveness and safety of clinical applications.

[0063] 10) The technical solution of the present invention successfully solves several long-standing technical problems by introducing a controlled stochastic differential equation model, including reducing the reliance on large-scale annotated data sets, enhancing the adaptability and generalization ability of the model to new data, reducing the reliance on high-end hardware, and improving the real-time control ability and interpretability of the model. These innovations not only improve the efficiency and quality of medical image processing, but also open up new possibilities for future research and application of medical image technology, and have significant scientific research and commercial application value.

[0064] 11) The technical solution of the present invention does overcome technical bias. Traditional medical image conversion technology, especially deep learning-based methods, usually relies on a large amount of paired labeled data for training, which is not only costly, but also has a sharp drop in performance when faced with new data with statistical characteristics significantly different from those of the training set in practical applications, limiting its application across different devices and patient groups. By introducing a controlled stochastic differential equation model, the present invention not only reduces the reliance on large-scale data sets, but also provides a dynamic modeling framework, enhances the generalization and adaptability of the model, and thus effectively overcomes the technical bias of traditional technologies in data dependence and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description, serve to explain the principles of the present disclosure;

[0066] Figure 1 is a flow chart of a proton density weighted and T1 weighted MRI image conversion method provided by an embodiment of the present invention;

[0067] Figure 2 It is a principle diagram of a proton density weighted and T1 weighted MRI image conversion method provided in an embodiment of the present invention;

[0068] Figure 3 is a schematic diagram of a data preprocessing module unit provided in an embodiment of the present invention;

[0069] Figure 4 It is a schematic diagram of a controlled stochastic differential equation model training method provided by an embodiment of the present invention;

[0070] Figure 5 is a flow chart of an image conversion execution method provided by an embodiment of the present invention;

[0071] Figure 6 is a schematic diagram of a cSDE module unit provided by an embodiment of the present invention;

[0072] Figure 7 It is a schematic diagram of a post-processing module unit provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific implementation disclosed below.

[0074] The innovation of the present invention is:

[0075] 1. Modular design: By combining four main modules (preprocessing module, controlled stochastic differential equation model training module, image conversion execution module and post-processing module), the embodiment of the present invention not only improves the quality of image conversion, but also enhances the flexibility and scalability of the system.

[0076] 2. Controlled Stochastic Differential Equation (cSDE) Model: The present invention uses the cSDE model for image conversion, which uses stochastic differential equations to dynamically adjust the image generation process, significantly improving the accuracy and generalization ability of the converted image.

[0077] 3. Real-time control and high interpretability: Another innovation of the cSDE model is that it allows operators to adjust model parameters, such as diffusion intensity, in real time during the image conversion process. This not only provides higher controllability and flexibility, but also increases the interpretability of the model, enabling clinicians to customize the image conversion process according to specific medical needs.

[0078] 4. Innovation of pre-processing and post-processing modules:

[0079] The preprocessing module uses a variety of techniques (such as denoising, contrast enhancement, and normalization) to optimize the input image, providing a higher quality starting point for subsequent transformations.

[0080] The post-processing module uses techniques such as sharpening and contrast adjustment to further enhance the quality of the converted images, ensuring the clinical suitability and diagnostic value of the output images.

[0081] These innovations reflect the technological leadership and practicality of the present invention, and effectively solve some key limitations of the existing technology, such as high data dependence, limited generalization ability, large consumption of computing resources and lack of real-time control mechanism.

[0082] In view of the problems of high data dependence, limited generalization ability, high computing resource cost and low application flexibility in the current PD-weighted image and T1-weighted image conversion methods, the present invention proposes a PD-weighted image and T1-weighted image conversion method which combines a preprocessing module, a controlled stochastic differential equation model training module, an image conversion execution module and a post-processing module.

[0083] Among them, the preprocessing module uses a denoising module to denoise the original PD-weighted image and T1-weighted image data, filter out interference noise, and keep the cleanliness of the source data; the contrast enhancement module is used to adjust the bright and dark areas of the denoised image to highlight the pathological features and important structures in the image; the normalization module is used to standardize the scale of the enhanced image data, reduce training deviation, and improve system stability while maintaining high performance.

[0084] The controlled stochastic differential equation model training module uses the cSDE variable assignment module to digitize the training method so that it can perform parameter optimization through subsequent training and other processes; the forward stage noise dilution module is used to simulate the noise in the image acquisition process in the real world, which is used to restore the original image data distribution in the subsequent inverse process; the backward stage image recovery module is used to restore the valid data and learn the distribution; finally, the cSDE model training module is used to verify the convergence of the entire model, thereby ensuring that the model has high-precision data recovery capabilities and strong robustness and generalization capabilities when facing different data.

[0085] The image conversion execution module uses the guidance information assignment module to define the guidance information, increase the model condition attributes, and enable the model to be trained according to the established supervision information; the computing resource dependency reduction setting module is used to make adjustable statistical settings for the module information, and the model parameters and related functions in the formula are set in a targeted manner to maintain the model performance while reducing the training computing power requirements; the prediction correction sampling module is used to perform multiple supervision corrections during the sampling process to improve the accuracy of the converted image; the image conversion training module is used to compare the predicted value with the true value to enhance the conversion effect.

[0086] Finally, the post-processing module uses a sharpening processing module to enhance the edge and structural clarity of the lesion area for doctors to identify and diagnose; and uses a contrast adjustment module to adjust the contrast between different tissues in the image to make them easier to distinguish.

[0087] Examples, such as Figure 1As shown, the proton density weighted and T1 weighted MRI image conversion method provided by the embodiment of the present invention includes the following steps:

[0088] S1, data preprocessing: for PD-weighted images and corresponding T1-weighted images, the data preprocessing module is used to standardize the data;

[0089] S2, cSDE model training: The cSDE model is predicted based on the deep learning network and dynamically adjusted according to the difference between the current predicted image data and the theoretical restored image;

[0090] S3, image conversion execution: providing an interface to allow the operator to adjust the model parameters in real time according to specific clinical needs to adapt to specific clinical needs or to cope with the performance of the model in specific situations;

[0091] S4, post-processing: The converted T1-weighted image is optimized through a post-processing module, which includes a sharpening processing module and a contrast adjustment module, to obtain a new T1-weighted image.

[0092] In step S1, the input of the data preprocessing module is the system input Image data pairs, where is the original T1-weighted image with noise, is with The corresponding original PD-weighted image, Indicates The output of the data preprocessing module is also a paired data pair for training the controlled stochastic differential equation model. ,in is the T1-weighted image after preprocessing. It is the PD-weighted image after preprocessing, which serves as the controlled condition of the cSDE model input.

[0093] like Figure 2 As shown, the data preprocessing provided by the embodiment of the present invention includes:

[0094] S101: First, use the trained small-core CNN denoising module to denoise the original data pair. The variable with the Ω symbol indicates that the denoising operation has been performed. For the T1-weighted image, the denoising module uses the original image with noise and the corresponding noise-free image pair. Conduct training, including is the original T1-weighted image with noise, is the corresponding noise-free image. For the PD weighted image, the denoising module uses the original image with noise and the corresponding noise-free image Conduct training, including is the original PD-weighted image with noise, is the corresponding noise-free PD weighted image. Through training, the model can effectively learn how to remove noise, that is, Convert to noise-free .

[0095] S102: Contrast enhancement module For each pixel value in ,in is the input pixel value, is the output pixel value, is the gamma value, is an optional scaling factor. After the data preprocessing contrast enhancement module, Convert to ,in Indicates the contrast after being processed by the contrast enhancement module. , Indicates the contrast after being processed by the contrast enhancement module. .

[0096] Among them, the formula The working principle of is to multiply each pixel value in the input image by the gamma value, effectively adjusting the brightness distribution of the image pixels. The main effect of applying this formula is to enhance the image contrast. Contrast refers to the difference in brightness between different areas of the image. By multiplying each pixel value by the gamma value, the formula amplifies the brightness difference between pixels, making the overall image look clearer and more eye-catching. The gamma value in the formula plays a vital role in controlling the strength of the contrast enhancement. Gamma values ​​greater than 1 increase contrast, while gamma values ​​less than 1 reduce contrast. The optional scaling factor A allows for further adjustment of the overall brightness level of the image. By manipulating the brightness distribution of pixels, the formula effectively improves visual clarity, feature enhancement, depth perception, and noise suppression. The gamma value and scaling factor provide additional control over the strength and overall brightness of the enhanced image.

[0097] S103: Use the minimum and maximum normalization method to process the data. This module uses the normalization formula Process each pixel value, where is the original pixel value, is the minimum pixel value, is the maximum pixel value, is the normalized pixel value. That is, after the data preprocessing normalization module, Convert to ,in After being processed by the normalization module , After normalization .

[0098] Among them, the formula Used to normalize image pixel values ​​to a specific range. The formula works by subtracting the minimum pixel value from each pixel value in the input image and then dividing it by the difference between the maximum and minimum pixel values. This transformation effectively scales the pixel values ​​to a defined range, usually between 0 and 1, which ensures that all pixels in the image have a uniform representation within the defined range. By bringing all pixel values ​​into a common range, the impact of extreme values ​​is reduced, resulting in more consistent processing and analysis. Feature extraction can be improved by ensuring that all features have similar amplitudes, preventing features with larger amplitudes from dominating the analysis. This can lead to more accurate and reliable feature detection and classification. Computational costs in machine learning algorithms can be reduced by ensuring that all input data falls within a similar range, usually between 0 and 1. This can improve the convergence and stability of machine learning models. Normalization plays a key role in various image processing and computer vision applications by rescaling, standardizing, and improving feature extraction, especially in the fields of machine learning, image enhancement, feature extraction, and medical imaging.

[0099] After the data preprocessing normalization module Convert to The specific process is:

[0100] T1-weighted images and PD-weighted images Each pixel value is scaled to between 0 and 1 based on the distance between the maximum and minimum values, removing the impact of instantaneous gradient changes caused by spike noise. That is, after the data preprocessing normalization module, Convert to ,in After being processed by the normalization module , After normalization .

[0101] To simplify subsequent variable representation and integrated module design, the output of the data preprocessing normalization module Defined as ,in Indicates that after being processed by the preprocessing module , Indicates that after being processed by the preprocessing module .

[0102] In the embodiment of the present invention, for example, Figure 3As shown in the figure, 2 small kernel convolution layers, activation functions, 2 convolution layers, batch normalization and activation functions, 1 convolution layer and activation function and 1 convolution layer are configured; data type check, gamma correction function and output range adjustment are configured; minimum value acquisition, maximum value acquisition and normalization functions are configured.

[0103] The cSDE model consists of two parts. The first part is the design and training of the stochastic differential equation model, which is used for high-precision image restoration and enables the model to have strong restoration generation performance. The second part is the design of a controlled guidance mechanism, which is used to guide the model to generate the required sampling results. For ease of understanding, this part is divided into S2 and S3 for separate explanation. In actual operation, these two modules will be integrated into one module, namely the cSDE module.

[0104] In step S2, the cSDE model simulates the time state from 0 to The original data distribution is gradually transformed into a Gaussian distribution in the process. In the reverse phase, the cSDE model transforms the simulation time state from The reverse process to 0.

[0105] like Figure 2 As shown, the controlled stochastic differential equation model training module provided by the embodiment of the present invention includes:

[0106] S201: Indicates at time Image in state , Indicates at time Image in state The probability density function of Representing images From time arrive The state transition probability of is also called Gaussian perturbation kernel when it obeys Gaussian distribution, where , It's the deadline. is the original distribution of the data, is the initial probability distribution of the system state, represents the probability distribution of the system state after adding noise. In the cSDE framework, as time goes from 0 to The system gradually evolved from Towards change, Represents a time The state of the system after noise and other disturbances are added.

[0107] S202: When calculating the diffusion process of the cSDE model, the main requirement is to solve ,in Representation variables Compliance interval The uniform distribution on Representation interval Uniform distribution on . is the drift term that controls how the image data evolves over time. is the diffusion term, controlling the randomness in the process. Indicates a small change in the state of the image data, Represents the differential of Brownian motion, which is used to simulate the change of random fluctuations in a small time interval. At this stage, the model simulates the noise in the image acquisition process in the real world by adding randomness. It is used to restore the original image data distribution in the subsequent inverse process.

[0108] Among them, the deterministic evolution function Describes the state variables Over time The deterministic evolution path of . In MRI image transformations, this can represent physical or biological properties that change over time, such as the magnetic susceptibility of tissue or the systematic change of water molecule concentration. By accurately simulating this process, the prediction accuracy of the state of a specific biological tissue can be improved, making the MRI image more realistically reflect the actual state of the tissue. Represents state variables In real biological processes, such as MRI imaging, many system variables are subject to unknown perturbations and inherent randomness. This term enables the model to capture this randomness and thus more fully simulate real-world conditions. and Being able to capture the deterministic and stochastic changes of the system separately, the model can more accurately describe and predict the dynamics of complex biological tissues. For example, changes in tissue properties, such as tumor growth or regression, can be better understood and predicted when performing long scans. In the process of converting proton density weighted images to T1-weighted images, this model can help the model be designed into a more sophisticated image processing algorithm, thereby achieving higher quality image conversion. By considering time changes and random perturbations, the algorithm can better adjust image properties. Using this advanced model, doctors and diagnostic experts can obtain more accurate image data to help them better understand and analyze disease states. Especially when diagnosing early-stage or subtle changes in diseases that are difficult to observe, it not only provides a dynamic and adaptable method for the proton density and T1-weighted conversion of MRI images, but also improves the authenticity and accuracy of the model's simulation of complex biological processes by introducing random factors.

[0109] S203: When calculating the inverse diffusion process of the cSDE model, the main requirement is to solve , used to simulate from any time to 0, where represents the inverse Brownian motion, It is the gradient term, which indicates the trend of the system state evolution over time and is used to find the current probability density function of the given data. A more appropriate evolution direction in the system, guiding the system to follow the current probability density function The logarithmic gradient of is used to adjust the image data. Represents the current probability density function Take the logarithm of the value.

[0110] in, The introduction of the gradient of the probability density into the drift term enhances the responsiveness of the model to the probability distribution of the system state. By taking into account the change in the probability distribution, this term can help adjust the drift rate more accurately to adapt it to the actual distribution of the system state. Since the terms related to the probability density gradient are added to the drift term, the model can adaptively adjust its path to be closer to the distribution of the actual data. This is very useful for MRI image conversion, which can more naturally simulate the changes in tissues, thereby improving the accuracy and naturalness of image conversion. In the conversion between proton density weighted and T1-weighted images, this model can more effectively process and predict subtle changes in images, especially when distinguishing different types of tissues or lesions. For example, different brain tissues and lesions appear differently in T1-weighted images due to subtle differences in their physical properties, and this advanced model can be used to distinguish them more accurately. This model can ensure that the converted images remain of high quality and increase the accuracy and reliability of diagnosis by providing a more accurate and adaptable image conversion method.

[0111] S204: In order to minimize the difference between the image predicted by the model and the image restored by theory, the objective function is designed as ,in Express expectations, Indicates at time The expected value of Representing images The expected value of the initial time, Representing images From time Return to the expected value of the initial time, is a weighted function greater than 0, , because in the process of solving real problems Difficult to calculate, requires extremely large computing resources, so neural networks are used To approximate ,in are the parameters of the neural network, Representing images From time status The gradient of the probability density function when restored to the initial time state 0 indicates the optimal direction of recovery, which is the theoretically optimal path. The task of the model is to approximate this gradient and then complete the high-precision image restoration process. Represents the probability density distribution in the original state of the data.

[0112] in, It is used to evaluate the performance of the model in MRI image conversion by calculating the state predicted by the model. The square of the Euclidean distance between the logarithmic gradient of the true probability density can be used to evaluate the accuracy of the model. It provides a quantitative indicator for model adjustment and optimization, helping to understand the performance of the model under different conditions and find areas for improvement. , the parameter settings or state transfer functions of the model can be improved to more accurately reflect the real system dynamics. This is especially important for the conversion of MRI images, because the image quality directly affects the reliability of the diagnosis results. The model is able to more accurately simulate real biological processes, such as the conversion of proton density and T1-weighted characteristics in MRI images. This makes the image conversion more precise, reduces errors, and improves the reliability of diagnosis. Reducing model prediction errors can provide more accurate image analysis tools, especially in medical image processing and disease monitoring. For example, with more accurate image conversion, different types of tissues and lesions can be better distinguished, helping doctors to perform more effective disease assessment and treatment planning. It provides an important performance evaluation and optimization tool for MRI image conversion models. By accurately measuring model performance, it helps to improve the prediction accuracy of the model, optimize the effect of medical image processing, and thus enhance the accuracy and efficiency of medical diagnosis.

[0113] For example, Figure 4 It is a schematic diagram of a controlled stochastic differential equation model training method provided by an embodiment of the present invention;

[0114] In step S3, the controlled stochastic differential equation model is used The image data is trained for supervised learning. The prediction-correction sampling method is introduced, combining the basic prediction step and multiple correction steps to gradually reversely reconstruct the process from the final noise state to the initial true state. During the conversion process, an interface is provided to allow the operator to adjust the model parameters according to specific clinical needs, such as diffusion intensity. The parameters can be adjusted in real time to adapt to specific clinical needs or to respond to the performance of the model in specific situations.

[0115] like Figure 2 As shown, the image conversion execution module provided by the embodiment of the present invention includes:

[0116] S301: In order to achieve the desired sampling output, the sampling model introduces guidance information in the forward and reverse training phases of diffusion. In this way, the model is supervised throughout the training process, ensuring the guidance and effectiveness of the learning process. This strategy helps the model to more accurately capture and reproduce the key features of the data distribution, thereby improving the quality and consistency of the sampling results.

[0117] S302: Using PD weighted image As guidance information, it supervises the forward and reverse diffusion processes. Specifically, we set , using Gaussian noise at different scales Dilute the original data distribution and model the data, that is, express yes About time status function. For images In Gaussian noise The probability density function of image The probability density function of For images In Gaussian noise The probability density function of

[0118] Given ,Right now, express The square root of the derivative of the rate of change of , then , by calculating The time derivative of the diffusion rate can be obtained , because the guidance information is added during the training process ,network Updated to Therefore, when conducting guided training, the present invention will solve the equation in the positive diffusion process of S202 Simplified to , the solution in the S203 reverse diffusion process is Simplified to , given , which means the mean is , the covariance is The multivariate normal distribution of is the identity matrix, then , so the objective function is .

[0119] in, Indicates the state variable The probability distribution of Close to initial state This indicates that the state of the system does not change much in a short period of time and still maintains distribution characteristics close to the initial state. In the context of MRI image conversion, this means that the basic features of the image do not change much in the early stage, which helps to maintain the continuity and stability of the image, especially during dynamic imaging. It shows that as time goes by State variables The probability distribution of is close to the standard normal distribution This means that over time, the state of the system tends to diffuse evenly, without obvious directional preference. In MRI image conversion, this indicates that after a long time, the image features tend to an equilibrium state, which is conducive to removing specific biases or noise from the initial image and achieving a more uniform or standardized image expression. The diffusion coefficient is defined Over time and the randomness coefficient The diffusion coefficient increases exponentially over time. This exponential diffusion enhancement model is particularly important for simulating biological tissue properties that change over time in MRI images, such as lesions that exhibit different diffusion behaviors. The diffusion coefficient is defined For time-varying The square root of the rate of change of diffusion. It describes how the diffusion rate changes dynamically over time. In MRI image transformation, this allows the model to more precisely control image properties that vary over time, such as the motion of water molecules in diffusion-weighted imaging. Further refinement of the randomness coefficient The time-dependent form of This complex time dependence allows the model to more realistically reproduce the dynamic changes of various complex physiological processes in MRI images, such as tumor growth or tissue degradation.

[0120] And will Simplified to It emphasizes that the variation is completely dominated by random perturbations. In MRI image processing, this pure random perturbation model is very important for simulating and analyzing noise caused by scanning equipment instability or random biological activities in vivo. This helps to develop more robust image processing algorithms, especially when processing or identifying image data with significant noise effects. Simplified to , by combining a time-dependent drift term and an adjusted random term to adjust the dynamics of the state variables. In particular The term represents the gradient influence of the probability density function and provides a mechanism to adjust the drift speed according to the gradient of the state distribution. In MRI image conversion applications, this can be used to dynamically adjust the changes in image characteristics, especially when simulating the diffusion behavior of complex tissue structures such as tumor regions. This method helps to more accurately capture and reproduce the behavioral characteristics of biological tissues under real-world conditions. and Given an initial state Under the condition, the state The conditional probability density of and the gradient of this density. This shows how the diffusion path of the state variable evolves from its initial state over time. In medical imaging analysis, the use of such a model can greatly enhance the understanding of the dynamic changes in images, such as tracking the progression of a disease or monitoring the effectiveness of treatment. It provides a quantitative method for diagnosis, comparing the statistical properties of images before and after treatment. is a performance evaluation formula that measures the deviation between the model output and the actual observed value. , It also allows errors at different time points to have different weights, thereby more finely adjusting the model evaluation process. This evaluation method is crucial for optimizing MRI image conversion algorithms, and can help researchers and clinicians understand the performance of the model in practical applications, and adjust model parameters or select the most suitable image processing strategy accordingly to improve the diagnostic value and clinical relevance of the image.

[0121] S303: First, from the prior distribution Take a sample from , and then numerically solve .when When it is large, the mean of the prior distribution is close to 0, and the prior distribution is approximated as The predictor of the prediction-correction sampling method adopts a fixed discretization strategy, using a small increment replace , and use Gaussian noise Alternative ,therefore ,in ; The corrector uses annealed Langevin dynamics and sets the Langevin step size ,in is the signal-to-noise ratio, According to the annealing Langevin kinetic formula Take samples. represents the multivariate normal distribution from 0 to the identity matrix, represents Gaussian noise, Indicates time The Gaussian noise function is For a simplified inverse Brownian motion, Represents the probability density distribution of the original state of the data; For neural networks, The Gaussian noise function is represents the state variable at time i, Gaussian noise at time i;

[0122] in, The evolution of the state variables is expressed by deterministic drift terms and random disturbance terms. indicates that the drift strength increases exponentially with time, allowing the model to adapt to rapidly changing dynamics; Functions depend on state , control parameters and time , thereby regulating the evolution path of the state. In MRI image conversion, the application of this model can help describe, for example, the diffusion process of water molecules in biological tissues, especially when these processes are significantly affected by external conditions such as temperature or pressure changes. This equation helps to more accurately simulate dynamic changes in images, such as changes in tissue structure during disease progression. Indicates time The state distribution at time , assuming that the mean is 0 and the variance is This means that over time, the state of the system tends to a uniform distribution enhanced by random fluctuations. For MRI image transformation, this means that after long-term observation, the image features show more extensive diffusion characteristics, which helps to reveal hidden biomarkers or pathological features, thereby improving the diagnostic value of the image. replace It is used to discretize the time variables in the continuous-time model for easy implementation in numerical simulation. This provides a methodological simplification for the time series analysis of MRI images, making it possible to more accurately track changes in image characteristics through discrete time steps, which is very important for understanding rapidly changing biological processes. Alternative Provides an explicit statistical model for the random disturbance term, where is a normally distributed random variable with time-dependent variance, reflecting the uncertainty spread over time. When processing MRI images, considering the randomness spread over time can help model and compensate for noise in the image acquisition process, such as noise introduced by patient movement or slight vibrations of the scanning equipment. The state at the previous time point is estimated by adding the reverse influence of drift adjustment and random perturbations. This is a time reversal technique that can be used to verify or optimize past state estimates. In MRI image analysis, this reverse simulation technique can be used for error correction, especially in dynamic imaging and time series analysis, to help improve the accuracy and reliability of image reconstruction. The signal-to-noise ratio in the Langevin kinetic model is defined as , which is based on the random variable The norm and system state function The norm of . This factor is used to balance the effects of random perturbations and system drift, ensuring the proper proportion of these two forces during the model evolution process. In MRI image transformation, the appropriate The α value can help control the speed and direction of the evolution of image properties such as proton density and T1-weighted signal, thereby reducing the effects of noise while preserving image detail.

[0123] S304: The cSDE model will be trained repeatedly until the loss function converges. The operator can adjust the model parameters according to the specific clinical needs to adjust the model optimization direction in real time and generate the required results, that is, according to the input PD-weighted image Generate the corresponding T1-weighted image .

[0124] For example, Figure 5 is a flow chart of an image conversion execution method provided by an embodiment of the present invention;

[0125] In the embodiment of the present invention, Figure 6 As shown, 2 encoders, pooling layers, 2 upsampling, merging and decoding are configured; 2 data comparisons, discretization functions, annealing Langevin functions and parameter updates are configured.

[0126] In step S4, the converted T1-weighted image is enhanced via a post-processing module.

[0127] S401: Use a high-pass filter to enhance high-frequency components to highlight the details and boundaries in the image. That is, the input of the sharpening processing module is the T1-weighted image output by S304. , the output of the sharpening module is , which is also the input of the post-processing contrast adjustment module S402.

[0128] S402: Adaptive histogram equalization technology is used. The adaptive histogram equalization technology can locally adjust the contrast without excessively enhancing the noise of the entire image. First, Perform fast partition processing. Represents a weighted image after sharpening; divided into , then for each independent block Contrast enhancement is performed to ensure uniformity and naturalness of the enhancement. By setting the window size parameters of adaptive histogram equalization, it can adapt to different image areas and diagnostic requirements. Reorganize and output the final desired image.

[0129] in, The image is first segmented into multiple small blocks or sub-regions. This segmentation makes it possible to adjust the contrast of each block independently, optimizing it according to its own characteristics and needs. Histogram equalization is applied to each independent small block. This step increases the local contrast by redistributing the brightness of the pixels and expanding the common brightness range. As a result, areas that originally appeared dark or had unclear details in the original image can become clearer. Especially in medical imaging such as MRI, subtle differences in tissue structure can be seen more clearly, which helps in the early detection and diagnosis of diseases. This technology enhances image quality by locally adjusting the contrast, rather than simply increasing the brightness of the entire image, so as to more effectively highlight important features and reduce noise interference.

[0130] In the embodiment of the present invention, Figure 7 As shown, two high-pass filter kernels, consistent convolution and sharpening functions are configured; two adaptive histogram equalization (CLAHE) functions and image display are configured.

[0131] This paper selects the medical imaging dataset IXIBrain Development Dataset collected and publicly used for non-commercial research and educational purposes by multiple hospitals in London, UK (including Hammersmith Hospital, Guy's Hospital and Institute of Psychiatry) as the training set and test set, uses SSIM and PSNR as evaluation indicators, selects CNN, WGAN_GP, conditional DDPM and ODE as comparison algorithms, and uses the UNet architecture to build the network to reduce the impact of different model architectures. The Adam optimizer is used, and the learning rate is set to , beta is set to 0.9 and 0.99, epsilon is set to 1e-8, epochs is set to 200, and batchsize is set to 2. All experiments are performed using PyTorch2.3 on an Nvidia RTX 3090 GPU with 24G of video memory.

[0132] The experimental comparison results are shown in Table 1. The experimental results show that the proton density weighted and T1-weighted MRI image conversion method and system of the present invention achieve optimal values ​​for the SSIM value and PSNR value of the converted T1-weighted image, and compared with the suboptimal values, they are improved by 0.01 and 0.04 respectively, which is better than other algorithms.

[0133] Table 1 Comparison results of different algorithms

[0134]

[0135] In addition, the present invention can also be applied to the following specific scenarios:

[0136] 1. High-resolution medical imaging: In the field of medical imaging, especially neuroimaging, the clarity and details of images are extremely high, because this is directly related to the accuracy of lesion diagnosis. By generating high-quality images, the present invention can help radiologists and neurologists more accurately identify tiny lesions such as microvascular abnormalities, early tumors or microhemorrhages. This technology improves the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) of the image, both of which are important parameters for evaluating image quality. High values ​​indicate that the image quality is closer to the original or ideal.

[0137] 2. Multimodal fusion technology: One of the common challenges in medical imaging is how to effectively fuse different types of imaging data (such as T1-weighted, T2-weighted, proton density-weighted, etc.) to provide more comprehensive diagnostic information. The application of the present invention in this scenario can optimize this fusion process, by enhancing image quality and consistency, providing a more coherent and detailed view, thereby assisting doctors in making more comprehensive analysis and judgments.

[0138] 3. Real-time imaging and rapid diagnosis: In emergency medicine and surgical navigation, image data needs to be quickly acquired and processed to guide emergency medical decisions or surgical procedures. The efficient algorithm of the present invention can output high-quality images in a shorter time, which is particularly important for real-time imaging systems. By improving imaging speed and quality, doctors can make faster and more accurate diagnoses in emergency situations.

[0139] 4. Long-term disease monitoring and evaluation: For chronic or long-term diseases such as multiple sclerosis, Alzheimer's disease, etc., it is very important to regularly monitor their progress. The improved MRI image quality using the present invention can more clearly show the progression of the lesion, helping doctors evaluate the treatment effect or disease changes, so as to carry out appropriate medical intervention.

[0140] 5. Research and development: In neuroscience and clinical research, high-quality images are essential, and they can be used to support a variety of research purposes such as pathology research and drug effect evaluation. By using this invention, researchers can obtain higher quality data, which will help improve the accuracy and reliability of research.

[0141] These scenarios demonstrate the potential application and importance of the present invention in improving the quality of medical imaging, reflecting its wide applicability in practical medical treatment and research.

[0142] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the protection scope of the present invention.

Claims

1. A method for converting proton density weighted and T1 weighted MRI images, characterized in that: The method includes: S1, data preprocessing: for the initial PD-weighted image and the corresponding T1-weighted image, the data preprocessing module is used to standardize the data to obtain a training image pair; S2, cSDE model training: The preprocessed training image pairs are input into the cSDE model. Based on the prediction of the deep learning network, the conversion relationship from PD-weighted image to T1-weighted image is established according to the forward statistical feature dilution process and the dynamic adjustment process of the difference between the current predicted image data and the theoretical restored image, and the cSDE image conversion model that directly converts the PD-weighted image to the corresponding T1-weighted image is obtained; S3, image conversion execution: receiving a PD-weighted image to be converted into a corresponding T1-weighted image, and outputting a T1-weighted image generated under the guidance of the PD-weighted image through a cSDE image conversion model; The image conversion execution specifically includes: Guidance information setting: Add guidance information in the forward and inverse training process of diffusion, use the PD weighted image c as guidance information, and perform network training in a supervised manner; Computational resource reduction settings: Set f(x, t) = 0, use Gaussian noise σ1, σ2…σ at different scales T Dilute the original data distribution and model the data, then g=f(t→σ)=σ T Denote that g is a function of σ with respect to the time state t; given g(t) is the square root of the derivative of the rate of change of σ(t), σ(0) = 0, then By calculating σ 2t The time derivative of the diffusion rate σ(t) is obtained; due to the addition of guidance information c during the training process, the neural network s θ (x, t) is updated to s θ (x(t), c, t); during guided training, the solution formula dx = f(x, t)·dt+g(t)·dω in the positive diffusion process is simplified to dx = σ t ·dω, the solution in the inverse diffusion process Simplified to Given p 0t (x(t)|x(0))=N(x(t);x(0),σ(t)·I) represents a multivariate normal distribution with mean x(0) and covariance σ(t)·I, where I is the identity matrix. Then So the objective function is Prediction correction sampling: From the prior distribution p T ~N(x(0),σ(T)·I) to extract a sample x T , and then numerically solve When σ(T) is large, the mean of the prior distribution is close to 0, approximating the prior distribution to p T ~N(0, σ(T)·I); The predictor of the prediction correction sampling method adopts a fixed discretization strategy, using a small increment Δt instead of dt and using Gaussian noise z~N(0, Δt·I) instead therefore where z t ~N(0,I); the corrector uses annealed Langevin dynamics and sets the Langevin step size Where r is the signal-to-noise ratio, z~N(0,I), according to the annealing Langevin kinetic formula Sampling, x t is the state variable at time t, x t-1 is the state variable at time t-1, is the disturbance term associated with the standard normal distributed random variable z, Adjust the amplitude of random perturbations; Image conversion training: The cSDE model is repeatedly trained until the loss function converges; the diffusion control term g(t) adjusts the model parameters according to the operator's specific clinical needs to adjust the model optimization direction in real time, and generates the corresponding T1-weighted image x based on the input PD-weighted image c; S4, post-processing: The converted T1-weighted image is optimized through a post-processing module, which includes a sharpening processing module and a contrast adjustment module, to obtain a new T1-weighted image.

2. The proton density weighted and T1 weighted MRI image conversion method according to claim 1, characterized in that: In step S1, the input of the data preprocessing module is the system input (x i , c i ) image data pair, where x i is the original T1-weighted image with noise, c i For x i The corresponding original PD weighted image, i is the i-th sample pair; The output of the data preprocessing module is the paired data pairs (x i , c i ), where x i is the preprocessed T1-weighted image, c i is the preprocessed PD-weighted image, which serves as the controlled condition of the cSDE model input.

3. The proton density weighted and T1 weighted MRI image conversion method according to claim 2, characterized in that: Data preprocessing, including: Data denoising: The trained small-core CNN denoising module is used to denoise the original data pair. The variables with the Ω symbol represent the denoising operation. For T1-weighted images, the denoising module uses the original image with noise and the corresponding noise-free image pair. Training is performed, where x i is the original T1-weighted image with noise, is the corresponding noise-free T1-weighted image; For PD weighted images, the denoising module uses the original noisy image and the corresponding noise-free image pair. Training is performed, where c i is the original PD-weighted image with noise, is the corresponding noise-free PD weighted image; through training, the model converts the original noisy (x i ,c i ) into a noise-free Data contrast enhancement: The contrast enhancement module receives the denoised data. Apply nonlinear transformation to enhance dark or bright areas of the image; contrast enhancement module For each pixel value in Among them, V out is the output pixel value, A is the scaling factor, γ is the gamma value, V in is the input pixel value; After the data contrast enhancement module, Convert to in, After being processed by the contrast enhancement module After being processed by the contrast enhancement module Data normalization: When processing PD-weighted images and T1-weighted image data, the data normalization module uses the minimum and maximum normalization method to process the data. This module uses the normalization formula Process each pixel value, where ρ is the original pixel value, ρ min is the minimum pixel value, ρ max is the maximum pixel value, ρ' is the normalized pixel value; After the data preprocessing normalization module, Convert to in, After being processed by the normalization module After normalization The output of the data preprocessing normalization module Defined as (x i ,c i ), where x i is x after being processed by the preprocessing module i , c i is c after being processed by the preprocessing module i .

4. The proton density weighted and T1 weighted MRI image conversion method according to claim 1, characterized in that: In step S2, the cSDE model training is divided into a forward diffusion stage and a reverse diffusion stage; in the forward diffusion stage, the cSDE model gradually transforms the original data distribution into a Gaussian distribution while simulating the time state from 0 to t, and in the reverse diffusion stage, the cSDE model simulates the inverse process of the time state from t to 0.

5. The proton density weighted and T1 weighted MRI image conversion method according to claim 4, characterized in that: The forward diffusion stage and the reverse diffusion stage include: Variable setting: x(t) is the image x at time t, p t (x) is the probability density function of image x at time t, p st (x(t)|x(s)) is the state transition probability of image x from time s to time t. When it obeys Gaussian distribution, 0≤s<t<T, T is the cutoff time, p0 is the original distribution of the data, and p is the starting probability distribution of the system state. T is the probability distribution of the system state after a series of processes; in the cSDE framework, as time evolves from 0 to T, the system gradually changes from p0 to p T Transformation, p T is the state of the system after adding noise and other disturbances at a certain time T; Noise dilution in the diffusion stage: When calculating the diffusion process of the cSDE model, solve dx = f(x, t)·dt+g(t)·dω, where t~U[0, T] means that the variable t obeys a uniform distribution on the interval [0, T]; f(x, t) is a drift term that controls how the image data evolves over time; g(t) is a diffusion term that controls the randomness of the process; dx is a small change in the state of the image data, and dω is the differential of the Brownian motion, which is used to simulate the amount of change in random fluctuations within a small time interval; in this stage, the model simulates the noise in the image acquisition process in the real world by adding randomness, which is used to restore the original image data distribution in the subsequent inverse process; Image restoration in the inverse diffusion stage: When calculating the inverse diffusion process of the cSDE model, solve Used to simulate the process from any time t to 0, where represents the inverse Brownian motion, is the gradient term, which represents the trend of the system state evolution over time and is used to find the probability flow p of given data t (x) is a more appropriate evolution direction, guiding the system to follow the current probability density function p t (x) log gradient adjusted image data; cSDE model training optimization: The cSDE model design objective function is Where E is the expectation, λ(t) is a weighting function greater than 0, used to adjust the importance; t~U[0,T], x(0)~p0(x), using the neural network s θ (x t , t) approximation Among them, θ is the parameter of the neural network, It is the gradient of the probability density function when the image x is restored from time state t to the initial time state 0, indicating the optimal direction of recovery. The task of the model is to perform a high-precision image restoration process by approximating the gradient.

6. The proton density weighted and T1 weighted MRI image conversion method according to claim 1, characterized in that: In step S3, the cSDE model uses (x i , c i ) image data for supervised learning training; introduce a prediction correction sampling method, combine basic prediction steps and multiple correction steps to gradually reversely reconstruct the process from the final noise state to the initial true state; the process includes continuous time simulation, solving the cSDE model by numerical method, and generating T1-weighted images; During the conversion process, an interface is provided to allow the operator to adjust the model parameters according to specific clinical needs. The model parameters include diffusion intensity. The model parameters can be adjusted in real time to adapt to specific clinical needs or respond to the performance of the model in specific situations.

7. The proton density weighted and T1 weighted MRI image conversion method according to claim 1, characterized in that: In step S4, the converted T1-weighted image is enhanced via a post-processing module.

8. The proton density weighted and T1 weighted MRI image conversion method according to claim 1, characterized in that: In step S4, post-processing includes: Sharpening: Design a sharpening module, use a high-pass filter, enhance the high-frequency components to highlight the details and boundaries in the image, the input of the sharpening module is the T1-weighted image x output from step S3, and the output of the sharpening module is x ⊥ , which is also the input of the contrast adjustment module; Contrast adjustment: Design a contrast adjustment module and use adaptive histogram equalization technology to adjust x ⊥ Perform partition block processing and divide it into x ⊥1 ,x ⊥2 …x ⊥n , and then for each independent block x ⊥i Contrast enhancement is performed; the window size parameters of adaptive histogram equalization are set to adapt to different image areas and diagnostic requirements; finally, the enhanced independent block x ⊥i Reorganize and output the final desired image.

9. A proton density weighted and T1 weighted MRI image conversion system, characterized in that: The system is used to control the proton density weighted and T1-weighted MRI image conversion method according to any one of claims 1 to 8, and the system comprises: A data preprocessing module, used for standardizing the data of the PD-weighted image and the corresponding T1-weighted image using the data preprocessing module, including denoising, contrast enhancement and normalization; cSDE model training module, cSDE sets drift term and diffusion term to control the smooth transition path of image data and the introduced random disturbance respectively; based on the deep learning network prediction cSDE model is dynamically adjusted according to the difference between the current predicted image data and the theoretical restored image; The image conversion execution module is used to establish the conversion relationship from PD-weighted images to T1-weighted images by training the cSDE model. During the conversion process, an interface is provided to allow the operator to adjust the model parameters in real time according to specific clinical needs to adapt to specific clinical needs or to cope with the performance of the model in specific situations. The image conversion execution specifically includes: Guidance information setting: Add guidance information in the forward and inverse training process of diffusion, use the PD weighted image c as guidance information, and perform network training in a supervised manner; Computational resource reduction settings: Set f(x, t) = 0, use Gaussian noise σ1, σ2…σ at different scales T Dilute the original data distribution and model the data, then g=f(t→σ)=σ T Denote that g is a function of σ with respect to the time state t; given g(t) is the square root of the derivative of the rate of change of σ(t), σ(0) = 0, then By calculating σ 2t The time derivative of the diffusion rate σ(t) is obtained; due to the addition of guidance information c during the training process, the neural network s θ (x, t) is updated to s θ (x(t), c, t); during guided training, the solution formula dx = f(x, t)·dt+g(t)·dω in the positive diffusion process is simplified to dx = σ t ·dω, the solution in the inverse diffusion process Simplified to Given p 0t (x(t)|x(0))=N(x(t);x(0),σ(t)·I) represents a multivariate normal distribution with mean x(0) and covariance σ(t)·I, where I is the identity matrix. Then So the objective function is Prediction correction sampling: From the prior distribution p T ~N(x(0),σ(T)·I) to extract a sample x T , and then numerically solve When σ(T) is large, the mean of the prior distribution is close to 0, approximating the prior distribution to p T ~N(0, σ(T)·I); The predictor of the prediction correction sampling method adopts a fixed discretization strategy, using a small increment Δt instead of dt and using Gaussian noise z~N(0, Δt·I) instead therefore where z t ~N(0,I); the corrector uses annealed Langevin dynamics and sets the Langevin step size Where r is the signal-to-noise ratio, z~N(0,I), according to the annealing Langevin kinetic formula Sampling, x t is the state variable at time t, x t-1 is the state variable at time t-1, is the disturbance term associated with the standard normal distributed random variable z, Adjust the amplitude of random perturbations; Image conversion training: The cSDE model is repeatedly trained until the loss function converges; the diffusion control term g(t) adjusts the model parameters according to the operator's specific clinical needs to adjust the model optimization direction in real time, and generates the corresponding T1-weighted image x based on the input PD-weighted image c; The post-processing module is used to optimize the converted T1-weighted image through the post-processing module, and the post-processing module includes a post-processing sharpening processing module and a post-processing contrast adjustment module.

Images