Medical image cross-modal conversion method

The multimodal medical image registration module and the non-alignment error detection module extract prior information and construct a prior loss term, which solves the problem of relying on pixel-level alignment data in cross-modal conversion of medical images, significantly improving the generator performance and Reliability of image generation.

CN119027303APending Publication Date: 2024-11-26SOUTHEAST UNIV +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411116177.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-14
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

Existing medical image cross-modal conversion methods rely on pixel-level alignment data, which is difficult to obtain and the reliability of the unsupervised method cannot be guaranteed, resulting in insufficient generator performance.

Method used

A prior extraction network consisting of multimodal medical image registration module and non-alignment error detection module is adopted to extract pixel-level prior information from non-alignment training data, build a prior loss term, and limit the solution space of the unsupervised GAN model. Improve generator performance.

Benefits of technology

In the absence of pixel-level alignment data, the generator performance of medical images is significantly improved, ensuring the accuracy and reliability of generated images, and is better than the other six most advanced cross-modal conversion methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027303A_ABST
    Figure CN119027303A_ABST
Patent Text Reader

Abstract

The invention discloses a medical image cross-modal conversion method, which comprises the following steps of: extracting available pixel-level prior information from non-aligned training data by using a prior extraction network consisting of a multi-modal medical image registration module and a multi-modal non-aligned error detection module; and then, constructing a prior loss item by using the extracted prior information to limit the solution space of the unsupervised GAN model based on the cyclic consistency constraint, thereby improving the performance of the generator. Based on multi-modal MRI (Magnetic Resonance Imaging) data, the method disclosed by the invention is compared with other six different most advanced cross-modal conversion methods. Qualitative and quantitative results show that the method has obvious superiority when non-alignment errors exist in training data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of medical image computer processing, and in particular to a method for realizing reliable cross-modal conversion of medical images without relying on pixel-level alignment data. Background Art

[0002] Cross-modal conversion of medical images refers to the technology of converting one medical image into another. This conversion can be achieved based on computer vision and deep learning technology. Cross-modal conversion is widely used in the medical field, aiming to improve the accuracy and efficiency of diagnosis. In addition, cross-modal medical image synthesis technology is also constantly developing, aiming to improve the performance of task models through synthetic images. For example, in the imaging diagnosis of Alzheimer's disease, the simultaneous use of MRI and PET is considered to be an effective means to achieve better performance, but many samples do not collect PET images. Through cross-modal image synthesis technology, missing data can be supplemented and all samples can be used to train multimodal diagnostic models. The application scenarios of this technology include image conversion, data augmentation, data unification, privacy protection, etc.

[0003] There is a contradiction in the current mainstream cross-modal conversion methods of medical images. The supervised methods with excellent performance rely on pixel-level aligned training data to constrain model optimization, but pixel-level aligned multimodal medical images are difficult to obtain in medical scenarios. Unsupervised methods can be trained without paired data, but their reliability cannot be guaranteed. Summary of the invention

[0004] The purpose of the present invention is to address the deficiencies of the prior art and provide a method for achieving reliable cross-modal conversion of medical images without relying on pixel-level aligned data. A prior extraction network composed of a multimodal medical image registration module and a multimodal non-aligned error detection module extracts usable pixel-level prior information from non-aligned training data; then, the extracted prior information is used to construct a prior loss term to limit the solution space of an unsupervised GAN model based on cycle consistency constraints, thereby improving the performance of the generator.

[0005] In order to solve the problems existing in the prior art, the present invention proposes a method for cross-modal conversion of medical images, comprising the following steps:

[0006] (1) Collect medical images of two modalities and construct the medical images of two modalities into a dataset. to represent the non-aligned multimodal medical image dataset, where and They come from mode X and mode Y respectively. “~” indicates that there is a misalignment error between them. express The area containing correct prior information, Represents the remaining areas containing erroneous prior information; introduces Used to indicate a relationship that exists only in theory. Pixel-level aligned modality Y images;

[0007] (2) The multimodal registration module MReg is introduced to extract available pixel-level prior information from non-aligned training data and eliminate The misalignment error in the register is Can provide more available pixel-level prior information;

[0008] (3) Introducing the multimodal misalignment error detection module MDet, for regions that are not detected as having obvious misalignment errors by the multimodal error detection model D, MDet will reserve weight values ​​for them in the confidence matrix W, so that the correct prior information contained in these regions can pass the optimization of the prior regularization term constraint model normally;

[0009] (4) After obtaining the pre-trained MReg and MDet modules, the extracted prior information is used to construct a priori loss terms and added to the cycle consistency module Cycle to limit the solution space of the unsupervised GAN model based on cycle consistency constraints.

[0010] A further technical solution of the present invention is the method for cross-modal conversion of medical images described above, wherein the multimodal registration module MReg adopts a coarse-to-fine cascade registration method;

[0011] Coarse Registration Model In mutual information loss Train under the constraint of to learn the affine deformation field ,make and The mutual information value between them reaches the maximum; fine registration model By minimizing the error results output by the multimodal error detector D in the MDet module, the parameters are continuously optimized to generate a more accurate deformation field. , the image obtained after rough registration is The slight but irregular elastic deformation error in the repair; the mutual information loss Described as: .

[0012] In the aforementioned method for cross-modal conversion of medical images, the MReg module will obtain a complete deformation vector field , by Apply to Get a new image pair , can be repaired The misalignment error that can be registered in There are regions with misalignment errors that can be registered In, with The area where there is a misalignment error that can be registered is recorded as ,and The area where there is non-registration error is recorded as , then ; Therefore, the corresponding prior regularization term after the introduction of the MReg module Can be described as:

[0013]

[0014] When the registration is completed, The area containing the correct prior information is replaced by the original Expand to , the area containing the wrong prior information is replaced by the original Reduce to , so the introduction of MReg adds correct pixel-level prior information and facilitates the training process.

[0015] In the aforementioned method for cross-modal conversion of medical images, the multimodal non-alignment error detection module MDet uses a multimodal error detection model D to The residual non-alignment error in the image is detected, and the activation function Act is used to activate the detection result to generate a confidence matrix with the same shape as the image and a value range of [0,1]. ; For regions detected by D to have obvious misalignment errors, MDet will assign them very low weight values ​​in the confidence matrix W to prevent the erroneous prior information contained in these regions from participating in the prior regularization term middle.

[0016] The aforementioned method of cross-modal conversion of medical images, the corresponding prior regularization term after the introduction of the MDet module Can be described as:

[0017] ;

[0018] Ideally, the area where W is zero should be The area where W has a value of 1 should correspond exactly to Completely corresponds; ideally the corresponding prior regularization term Can be described as:

[0019] ;

[0020] The prior extraction network composed of the MReg module and the MDet module has the ability to extract The ability to extract correct prior information and eliminate erroneous prior information as much as possible.

[0021] The aforementioned method of cross-modal conversion of medical images, after obtaining the pre-trained MReg and MDet modules, uses the prior regularization loss Add to the Cycle module, where , ; Prior regularization loss Described as: ;

[0022] The complete optimization object of this method can be expressed as: ;

[0023] in is the cycle consistency loss, is the discriminative loss, G and F are generators, and is the discriminator.

[0024] In the aforementioned method for cross-modal conversion of medical images, the cycle consistency loss is described as: ;

[0025] The discriminative loss is described as:

[0026] .

[0027] Compared with the prior art, the method of the present invention uses a prior extraction network composed of a multimodal medical image registration module and a multimodal misalignment error detection module to extract usable pixel-level prior information from non-aligned training data; then, the extracted prior information is used to construct a prior loss term to restrict the solution space of the unsupervised GAN model based on cycle consistency constraints, thereby improving the performance of the generator. Based on multimodal MRI data, the method of the present invention is compared with six other state-of-the-art cross-modal conversion methods. Both qualitative and quantitative results show that the method of the present application has obvious superiority when there are misalignment errors in the training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is the network framework structure of this embodiment.

[0029] Figure 2 Network framework diagram for the training process of the multimodal misalignment error detector.

[0030] Figure 3 The process of converting the error map output by D into the confidence matrix W using the activation function Act.

[0031] Figure 4Qualitative comparison of different methods on datasets built based on (a) BraTS2020 and (b) PDGM.

[0032] Figure 5 shows a qualitative comparison of ablation experiments performed on datasets constructed based on (a) BraTS2020 and (b) PDGM.

[0033] in:

[0034] Figure 2 (a) Domain generalization method for simulating different modal distributions; (b) Training process of multimodal misaligned error detector D. DETAILED DESCRIPTION

[0035] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the embodiments.

[0036] This embodiment discloses a method for realizing reliable cross-modal conversion of medical images without relying on pixel-level alignment data. To facilitate the introduction of the scheme of this embodiment, some variables that need to be used are first defined: to represent the non-aligned multimodal medical image dataset, where and They come from mode X and mode Y respectively, and “~” indicates that there is a misalignment error between them. express The area containing correct prior information, Represents the remaining areas containing erroneous prior information. , used to indicate a phenomenon that exists only in theory Pixel-level aligned modality Y images. Under the conditions of The core of the method in this embodiment is to use a prior extraction network composed of a multimodal registration module MReg and a multimodal misalignment error detection module MDet to extract the prior information from the non-aligned image pair. The available pixel-level prior information is extracted, and the extracted prior information is used to construct a priori loss term to constrain the optimization of the model. The overall structure of the network model of this embodiment is as follows Figure 1 shown.

[0037] First, the multimodal registration module MReg is as follows Figure 1 (a) is shown, the purpose of which is to eliminate The misalignment error in the register is It can provide more available pixel-level prior information. In order to obtain the best possible registration effect, this embodiment adopts a coarse-to-fine cascade registration method. In mutual information loss Train under the constraint of (Eq.1) to learn an affine deformation field ,make and The mutual information value between them reaches the maximum. The coarse registration model can overall repair the misalignment errors caused by large but relatively regular affine deformations, thereby reducing the workload of the fine registration model.

[0038]

[0039] Fine Registration Model By minimizing the error results of the multimodal error detector D in the MDet module, the parameters (Eq. 2) are continuously optimized to generate a more accurate deformation field. , the image obtained after rough registration is In addition, to ensure It can generate a smooth deformation field, and also introduces a diffusion regularization operator in the gradient direction of the deformation vector field. Constraints are applied (Eq.3). Fine registration model The overall optimization object can be expressed as Eq.4:

[0040]

[0041]

[0042]

[0043] Finally, the MReg module will obtain a complete deformation vector field , by Apply to Get a new image pair , can be repaired There is a misalignment error in the registration. There are regions with misalignment errors that can be registered In, with The area where there is a misalignment error that can be registered is recorded as ,and The area where there is non-registration error is recorded as , then Therefore, after the introduction of the MReg module, the corresponding prior regularization term It can be described as Eq.5:

[0044]

[0045] It can be seen that with the completion of the registration, The area containing the correct prior information is replaced by the original Expand to , the area containing the wrong prior information is replaced by the original Reduce to , therefore, the introduction of MReg can increase the correct pixel-level prior information and facilitate the training process.

[0046] because The non-registration errors in the region cannot be fixed by registration. The erroneous prior information contained in the region will still interfere with the optimization of the model. Therefore, after introducing the MReg module, the multimodal misalignment error detection module MDet is introduced, as shown in Figure 1 (c) The multimodal misalignment error detection module MDet uses a multimodal error detection model D to The residual non-alignment error in the image is detected, and the activation function Act is used to activate the detection result to generate a confidence matrix with the same shape as the image and a value range of [0,1]. For regions detected by D to have obvious misalignment errors, MDet will assign them very low weight values ​​in the confidence matrix W to prevent the erroneous prior information contained in these regions from participating in the prior regularization term. For regions that are not detected by D as having obvious misalignment errors, MDet will reserve a certain weight value for them in W, so that the correct prior information contained in these regions can pass the prior regularization term normally. Optimization of the constraint model. The corresponding prior regularization term after the introduction of the MDet module It can be described as Eq.6:

[0047]

[0048] Ideally, the area where W is 0 should be The area where W has a value of 1 should correspond exactly to Completely corresponds. Ideally, the corresponding prior regularization term It can be described as Eq.7:

[0049]

[0050] At this point, the prior extraction network composed of the MReg module and the MDet module has the ability to extract the data from the non-aligned dataset. The ability to extract correct prior information and eliminate erroneous prior information as much as possible.

[0051] After obtaining the pre-trained MReg and MDet modules, the following prior regularization loss can be (Eq.8) is added to the cycle consistency module Cycle, such as Figure 1 (b), where , .

[0052]

[0053] At this point, the complete optimization object of this method can be expressed as:

[0054]

[0055] in is the cycle consistency loss (Eq.10), is the discriminative loss (Eq.11), G and F are generators, and is the discriminator.

[0056]

[0057] In order to effectively detect the results of MReg The remaining non-registration errors in the present embodiment are as follows: Figure 2 (b) is the training process to train the detector D. First, the input modality X image Apply random affine deformation and elastic deformation to obtain Modality X images with unimodal misalignment between them . Then Use as Figure 2 The domain generalization method shown in (a) is specifically, this method randomly divides the image distribution into parts, which are then randomly shuffled and remapped in the shuffled order to simulate the image distribution of different modalities, thereby obtaining Pixel-level alignment, but with Images with multimodal misalignment between them Direct quantification It is very difficult to correct the multimodal misalignment errors between The unimodal non-aligned error between them can be expressed as the residual graph between them It is easy to quantify. Therefore, consider the standardized As training labels, to train the detector D The multimodal misalignment error between Through this training, the detector D can quantify the multimodal misalignment error and give an error map with a range of [0,1]. The complete optimization object of the detector D can be expressed as Eq.12:

[0058]

[0059] However, the detection results of the detector D cannot be directly used as the confidence matrix W, so some post-processing is still required, such as Figure 3 When D is used to process the image after MReg module When performing detection, if the residual value of a certain area in the detection result of D is greater than a threshold th, it can be determined The area in There is a significant misalignment error between them, so the weight value of the corresponding area in W can be directly set to 0. Since the size of the unimodal residual value is also related to the pixel value of the image itself, for the area with a residual value less than th, it is still necessary to appropriately reduce its weight according to the size of its residual value. Taking the above into consideration, the final confidence matrix W and activation function Act are shown in Eq.13, where Indicates the mask where the residual value is less than the th region.

[0060]

[0061] Specific experiments in this example

[0062] 1. Data Preparation

[0063] In the process of constructing the training set, first The first slice and modal Y volume data The image pairs are then randomly added Degrees of rotation, The translation and The final training set contains unregisterable misalignment errors and significant registerable misalignment errors. Based on the T1-T2 modality images of the BraTS2020 dataset, a non-aligned training set was constructed using 12,000 pairs of images according to the above method, and a test set was constructed using 1,000 pairs of highly aligned images. Based on the T2-FLAIR images of the PDGM dataset, a non-aligned training set was constructed using 8,000 pairs of images according to the above method, and a test set was constructed using 800 pairs of highly aligned images.

[0064] 2. Implementation and Training Details

[0065] In the MReg module, the coarse registration model By five The convolutional layer and two fully connected layers are composed of two The first convolution operation has a stride of 2, and the rest have a stride of 1. Each convolution operation is followed by Batch Normalization and Leaky-ReLU activation. The number of filters in the five convolutional layers and the last two fully connected layers is set as follows: [32, 64, 64, 64, 64, 32, 4]. Finally, Will output a set of parameters representing rotation, scaling and translation ,according to Able to obtain affine deformation field . Fine Registration Model Based on the UNet implementation, the number of filters in the downsampling layer is set as follows: [32, 64, 64, 64, 64, 64], and the number of filters in the upsampling layer is set as follows: [64, 64, 64, 64, 64, 64, 32]. The convolutional layer directly outputs a deformation field with 2 channels. The multimodal error detector D in the MDet module and the generators G and F in the Cycle module are consistent with the generative model containing 9 layers of Res blocks in CycleGAN[*]. The discriminator in the Cycle module and It is implemented based on PatchGAN. The threshold th of the activation function Act in MDet is set to 0.1.

[0066] All models in this embodiment are trained with the same training set, except that the multimodal misalignment error detector D in MDet is only trained with BraTS2020 T1 modality data and PDGM T2 modality data in the training set. All network models are implemented based on the PyTorch framework and run on a computer equipped with an Nvidia GeForce RTX 4090 graphics card. The batch size is set to 1, the number of training rounds for the two pre-training modules MReg and MDet is set to 80, and the number of training rounds for MisT is set to 60. All loss functions are calculated by and the Adam optimizer with a learning rate of 1e-4. The weights of the loss function are set to , , .

[0067] 3. Experimental Results

[0068] This embodiment uses 7 different methods including supervised GAN methods (Pix2Pix, RegGAN), unsupervised GAN methods (CycleGAN, UNIT, MUNIT), diffusion model (SynDiff) and the method of the present application scheme to train on the two constructed data sets, and uses two performance indicators, structural similarity (SSIM) and peak signal-to-noise ratio (PSNR), to quantitatively evaluate the performance of all models on the test set. The quantitative results are listed in Table 1, and the image results of different methods and their residual maps with the reference image are shown in Figure 4 Among the results of supervised GAN methods, Pix2Pix's generated images have obvious errors and its quantization results are poor. This is in line with expectations, because Pix2Pix is ​​highly dependent on aligned data and cannot avoid any non-alignment errors from interfering with model optimization. RegGAN's score is still not ideal, and its image result quality is not significantly improved compared to Pix2Pix, which indicates that the registration network in RegGAN cannot properly handle the non-registration errors. The quantization results of unsupervised GAN methods are generally better than those of supervised GAN methods, because they do not use the pixel-level prior information in the training data to constrain model optimization. However, the lack of guidance from pixel-level prior information also leads to poor fidelity of these methods to image content, especially some detailed structures. This defect is most obvious in the results of MUNIT, which may be due to the loss of information in the process of decoupling and reorganizing content and style in MUNIT. SynDiff can better preserve the overall structure of some tissues in the image, and its performance indicators are also relatively good among unsupervised methods, all of which are due to the excellent performance of the diffusion model in generating high-quality images. However, SynDiff still cannot guarantee the correctness of the image content. Figure 4 The enlarged area of ​​the SynDiff result image has lower contrast than other methods, and the area indicated by the green arrow accidentally generates erroneous content similar to tissue that does not actually exist. It can be seen that due to the lack of pixel-level prior constraints, even unsupervised methods based on powerful diffusion models still have an unclear solution space and cannot generate stable and reliable translation results. From the perspective of PSNR and SSIM, MITIA achieved the highest score. In terms of the quality of the generated image, compared with other methods, MITIA can more accurately depict the image content and some detailed structures.

[0069] Table 1. Quantitative comparison results of different methods

[0070]

[0071] 4. Ablation Experiment

[0072] In order to verify the effectiveness of each module in MisT, four experiments were conducted based on two datasets constructed using BraTS2020 and PDGM respectively. The experimental settings are as follows:

[0073] V1: CycleGAN model. This baseline model does not include the MReg module and the MDet module, nor does it introduce any pixel-level prior loss. and Unsupervised training is performed under the constraints of .

[0074] V2: Introduced the MReg module in V1 and used , and pixel-level prior loss (Eq.14) constrains the model optimization, where .

[0075]

[0076] V3: Introduced the MDet module in V1 and used , and pixel-level prior loss (Eq.15) constrains the model optimization.

[0077]

[0078] V4: The complete MisT model, including the three modules MReg, MDet and Cycle. It uses the complete optimization object shown in Eq.9 during training.

[0079]

[0080] Table 2. Quantitative results of ablation experiments

[0081]

[0082] As shown in Table 2, it can be seen that the PSNR and SSIM scores have significantly improved from V1 to V4. From V1 to V2, PSNR increased by 1.64dB and 1.38dB, respectively, and SSIM increased by 2.01% and 3.47%, respectively. This result shows that the pixel-level prior information extracted after the introduction of the MReg module to eliminate the repairable misalignment error is very helpful in improving the performance of the generator. The PSNR and SSIM scores in V3 have also been greatly improved compared to V1, which shows that non-aligned data can directly provide some correct pixel-level prior information to guide the training of the model without registration processing. Since there is no registration process in V3, V2 should have more sufficient correct prior information than V3, but its quantization results are similar to V3, which means that the erroneous prior information caused by the residual misalignment error in V2 has caused relatively serious interference to the optimization of the model. The evaluation index values ​​of V4 are higher than those of V2 and V3, which indicates that with the combination of MReg and MDet modules, V4 can eliminate the interference of residual non-alignment errors in V2 and obtain more correct prior information than V3, thus providing more reliable guidance for model optimization. Figure 5 From the enlarged area in (a), V1 has a messy texture, while V2 and V3 are slightly better, which shows that the introduction of the two modules MReg and MDet is helpful to improve the quality of the generated image. However, compared with V4, which can generate an image result that is highly similar to the reference image, V2 is overall darker and the boundary of V3 is a bit blurred. In addition, Figure 5 In the area pointed by the arrow in (b), V2 and V3 images both have the problem of losing detailed structures, but V4 can well preserve the detailed structures in the image. These differences once again prove that each module in MITIA is indispensable, and making them work together is very important to improve the performance of the generator.

[0083] It should be particularly noted that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A method for cross-modal conversion of medical images, characterized in that The following steps are involved: (1) Collect medical images of two modalities and construct the medical images of two modalities into a dataset. to represent the non-aligned multimodal medical image dataset, where and They come from mode X and mode Y respectively. "~" indicates that there is a misalignment error between them. express The area containing correct prior information, Represents the remaining areas containing erroneous prior information; introduces Used to indicate a relationship that exists only in theory. Pixel-level aligned modality Y images; (2) The multimodal registration module MReg is introduced to extract available pixel-level prior information from non-aligned training data and eliminate The misalignment error in the register is Can provide more available pixel-level prior information; (3) For the regions that are not detected by the multimodal error detection model D as having obvious misalignment errors, the multimodal misalignment error detection module MDet is introduced, and MDet will reserve weight values ​​for them in the confidence matrix W; (4) After obtaining the pre-trained MReg and MDet modules, the extracted prior information is used to construct a priori loss terms and added to the cycle consistency module Cycle to limit the solution space of the unsupervised GAN model based on cycle consistency constraints.

2. The method for cross-modal conversion of medical images according to claim 1, characterized in that: The multimodal registration module MReg adopts a coarse-to-fine cascade registration method; Coarse Registration Model In mutual information loss Train under the constraint of to learn the affine deformation field ,make and The mutual information value between them reaches the maximum; fine registration model By minimizing the error results output by the multimodal error detector D in the MDet module, the parameters are continuously optimized to generate a more accurate deformation field. , the image obtained after rough registration is Corrects the slight but irregular elastic deformation errors in the image.

3. The method for cross-modal conversion of medical images according to claim 2, characterized in that: The mutual information loss Described as: .

4. The method for cross-modal conversion of medical images according to claim 2, characterized in that: The MReg module will obtain a complete deformation vector field , by Apply to Get a new image pair , can be repaired The misalignment error that can be registered in There are regions with misalignment errors that can be registered In, with The area where there is a misalignment error that can be registered is recorded as ,and The area where there is non-registration error is recorded as , then ; Therefore, the corresponding prior regularization term after the introduction of the MReg module Can be described as: ; With the registration completed, The area containing the correct prior information is replaced by the original Expand to , the area containing the wrong prior information is replaced by the original Reduce to , so the introduction of MReg adds correct pixel-level prior information and facilitates the training process.

5. The method for cross-modal conversion of medical images according to claim 1, characterized in that: The multimodal misalignment error detection module MDet uses a multimodal error detection model D to The residual non-alignment error in the image is detected, and the activation function Act is used to activate the detection result to generate a confidence matrix with the same shape as the image and a value range of [0,1]. .

6. The method for cross-modal conversion of medical images according to claim 5, characterized in that: The corresponding prior regularization term after the introduction of the MDet module Can be described as: ; Ideally, the area where W is zero should be The area where W has a value of 1 should correspond exactly to Completely corresponds; ideally the corresponding prior regularization term Can be described as: ; The prior extraction network composed of the MReg module and the MDet module has the ability to extract The ability to extract correct prior information and eliminate erroneous prior information as much as possible.

7. The method for cross-modal conversion of medical images according to claim 1, characterized in that: After obtaining the pre-trained MReg and MDet modules, the prior regularization loss Add to the Cycle module, where , ; ; At this point, the complete optimization object of this method can be expressed as: ; in is the cycle consistency loss, is the discrimination loss.

8. The method for cross-modal conversion of medical images according to claim 7, characterized in that: The cycle consistency loss is described as: ; The discriminative loss is described as: ; Where G and F are generators, and is the discriminator.