Non-paired cross-modal medical image conversion method based on contour guide path regularization

Abstract

The invention belongs to the technical field of medical images and deep learning, and particularly relates to a non-paired cross-modal medical image conversion method based on contour guide path regularization, which comprises the following specific steps: constructing a conversion path from a source domain to a target domain; a path regularization method is applied to smooth path transition and ensure the stability in the image conversion process; according to a path regularization technology based on contour guidance, the model can be focused on a key anatomical region, the structural consistency and accuracy of the generated CT image are further improved, and the training time can be remarkably shortened by reducing additional training steps. The invention aims to simplify the unsupervised training normal form and overcome the defects of the traditional GAN-based single-step generation method in the aspect of capturing the complex anatomical structure of the medical image. While the training time and the sampling time are considered, the problem that the quality of the generated medical image is limited is overcome.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging and deep learning technology, and specifically to an unpaired cross-modal medical image conversion method based on contour-guided path regularization. Background Art

[0002] Unpaired cross-modal medical image conversion refers to the conversion of one medical image modality to another, and the conversion process does not rely on paired image data for training. This conversion method is of great significance in medical image processing because it can make full use of a large number of unpaired data sets for training, thereby improving the accuracy and generalization ability of the conversion. The implementation of unpaired cross-modal medical image conversion mainly relies on deep learning technology, especially generative adversarial networks (GANs). By training the GANs model, the mapping relationship of the target modality image can be learned from the source modality image, thereby realizing the image conversion. This method can not only complete specific unidirectional conversion tasks, but also realize bidirectional conversion, ensuring that the pathological information of the source image does not change.

[0003] However, the current methods of unpaired cross-modal medical image conversion have the following problems:

[0004] 1. Additional training process complexity: Methods such as SynDiff require additional training processes, which include first using an unsupervised model to generate paired samples and then using these samples for supervised training. This approach not only increases the complexity of model training, but may also limit the final effect, because the supervised training part of the diffusion model may be affected by the pre-training process.

[0005] 2. Limited generation quality: Although the GAN-based method is concise, it may not be able to fully capture the complex anatomical structure features in medical images in the one-step generation process, resulting in limited generation quality.

[0006] 3. The training time of the diffusion model is long: Although the diffusion model-based method performs well in the quality and diversity of generated images, its unsupervised training mode and long sampling time pose significant challenges to the efficiency and practicality of practical applications, especially in medical cross-modal translation.

[0007] Therefore, an unpaired cross-modal medical image conversion method based on contour-guided path regularization is invented, aiming to simplify the unsupervised training process and avoid repeated redundant training processes, while overcoming the shortcomings of traditional GAN-based single-step generation methods in capturing the complex anatomical structure of medical images, while taking into account the training and sampling time, and overcoming the problem of limited quality of generated medical images. Summary of the invention

[0008] In view of the above problems and / or the problems existing in the existing unpaired cross-modal medical image conversion method based on contour-guided path regularization, the present invention is proposed.

[0009] Therefore, the purpose of the present invention is to provide an unpaired cross-modal medical image conversion method based on contour-guided path regularization, which can solve the above-mentioned existing problems.

[0010] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0011] The unpaired cross-modal medical image conversion method based on contour-guided path regularization includes the following specific steps:

[0012] S1, path establishment: constructing a conversion path from the source domain to the target domain;

[0013] S2, path regularization: The path regularization method is applied to smooth the path transition and ensure the stability during the image conversion process;

[0014] S3, contour guidance: The path regularization technology based on contour guidance enables the model to focus on key anatomical areas, further improving the structural consistency and accuracy of the generated CT images.

[0015] As a preferred solution of the unpaired cross-modal medical image conversion method based on contour-guided path regularization described in the present invention, the specific process of S1 is as follows:

[0016] S11: Let x0 and x1 represent samples in the source domain and target domain respectively, where x0 comes from the data distribution π0 of MRI images and x1 comes from the data distribution π1 of CT images;

[0017] S12: Given a basic data pair We assume that images from both domains are generated through the same decoder G* from a shared latent space Z and a continuous domain variable θ, which is specifically, z = E * (x0) = E * (x1), where E * represents the encoder function, which leads to x0 = G * (z, 0) and x1 = G * (z, 1), where G * is the decoder function, so the conversion from the source domain to the target domain is expressed as Specifically expressed as

[0018] As a preferred solution of the unpaired cross-modal medical image conversion method based on contour-guided path regularization described in the present invention, the specific process of S2 is as follows:

[0019] S21, calculate the path speed function: through the decoder function G dec (z, θ) calculates the velocity function V(x θ ,θ);

[0020] S22, perform path regularization optimization objectives, as follows:

[0021]

[0022] As a preferred solution of the unpaired cross-modal medical image conversion method based on contour-guided path regularization described in the present invention, wherein: the speed function v(x θ , θ) is calculated as follows:

[0023] v(x θ ,θ)=G dec (z, θ+Δθ)-G dec (z, θ)

[0024] Among them G dec (z, θ) represents the decoding result under the domain variable θ.

[0025] As a preferred solution of the unpaired cross-modal medical image conversion method based on contour-guided path regularization described in the present invention, wherein: the optimization objectives include reconstruction loss, adversarial loss, and multi-scale path loss;

[0026] The reconstruction loss is calculated as Among them, G(z,0) represents the output image of the generator in the starting domain, x0 represents the original image, and z is the encoding result of the original image;

[0027] The adversarial loss is calculated as in, is the output image of the generator in the destination domain, and D is the discriminator used to distinguish the generated image from the real image x1 in the target domain;

[0028] The multi-scale path loss is based on the path lengths of multiple feature layers l, which is calculated as f l =G l (z, 1)-G l (z, 0)-v l (x θ ,θ), where G l (z, 1) and G l (z, 0) represent the decoded features under the latent vector z and parameter θ respectively;

[0029] In order to measure the multi-scale loss, the path length loss is calculated as:

[0030]

[0031] Based on this, by minimizing the path length loss L path , ensuring the smoothness of transition paths at different scales, thus facilitating a coherent and accurate conversion from MRI to CT images.

[0032] As a preferred solution of the unpaired cross-modal medical image conversion method based on contour-guided path regularization described in the present invention, the specific process of S3 is as follows:

[0033] S31: Constructing the contour generation network G grad (x), and taking the MRI image as input, generate the corresponding CT image contour;

[0034] S32: Extract contours from CT images using the Sobel algorithm;

[0035] S33: Optimizing the contours extracted by the Sobel algorithm and the contour generation network G grad (x) The L1 norm between the generated contours to ensure that the contours extracted by the Sobel algorithm are consistent with the contour generation network G grad (x) The generated contours are aligned, which helps the model focus on detailed contour areas more effectively when generating CT images;

[0036] S34: G grad (x0) interpolates to f l The resolution of the matching layer is smoothed by applying bilateral filtering in the transition area near the contour, and then an attention weight W is assigned to the feature map on each layer l. l , and its modified path loss function combined with the attention weight is:

[0037]

[0038] Among them, α controls the importance of the attention mechanism and adjusts the feature map f of each layer l contribution in the calculation of loss;

[0039] In addition, the overall loss function of the model is:

[0040] L full =λ gan L gan +λ path L path +λ c L c +λ rec L rec .

[0041] As a preferred solution of the unpaired cross-modal medical image conversion method based on contour-guided path regularization described in the present invention, the loss function of the contour in S33 is calculated as follows:

[0042]

[0043] In order to achieve efficient and effective path regularization, contour constraints are only applied in the starting and ending domains, and the contour information is used as an additional attention map to correct the intermediate domain during the path transfer process.

[0044] Compared with existing technologies:

[0045] 1. Simplify the unsupervised training process: By reducing the additional training steps, the training time can be significantly shortened. Compared with the unsupervised method based on the diffusion model, the simplified training process not only improves efficiency, but also reduces the consumption of computing resources, allowing the model to converge quickly in a shorter time.

[0046] 2. Enhanced anatomical structure capture capability: The single-step generation method of traditional GAN ​​often fails to fully capture the details of anatomical structures when processing complex medical images. This invention introduces a multi-step path regularization mechanism to better preserve the complex anatomical information in MRI and CT images, so that the generated CT images are more realistic and clinically applicable.

[0047] 3. Balance training and sampling efficiency: While ensuring the quality of model generation, the present invention optimizes the training and sampling time. Through the improved generation strategy, the computational overhead caused by the increase in model complexity in traditional methods is avoided, thereby achieving efficient generation of medical images.

[0048] 4. Improve the quality of medical image generation: This invention overcomes the limitations of traditional generative models in terms of medical image quality, especially in terms of structural consistency and image details. The resulting CT images are not only more realistic visually, but also have higher accuracy and stability in medical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall model structure diagram of the present invention;

[0050] Figure 2 This is a schematic diagram of the path conversion of the present invention;

[0051] Figure 3 This is a schematic diagram of the application process of the model of the present invention;

[0052] Figure 4 This is a graph of the experimental results of the present invention. DETAILED DESCRIPTION

[0053] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0054] The present invention provides an unpaired cross-modal medical image conversion method based on contour-guided path regularization, please refer to Figure 1-Figure 4 , including the following specific steps:

[0055] S1, path establishment: constructing a conversion path from the source domain to the target domain;

[0056] The details of path establishment are as follows:

[0057] Let x0 and x1 represent samples in the source domain (MRI image) and target domain (CT image), respectively, where x0 comes from the data distribution π0 of MRI images and x1 comes from the data distribution π1 of CT images. In order to convert samples from π0 to π1, an effective method is needed to promote the transition path between the two domains.

[0058] Given a basic data pair Assume that images from both domains are passed through the same decoder G * , generated from a shared latent space Z and a continuous domain variable θ; specifically, z = E * (x0) = E * (x1), where E * represents the encoder function; this leads to x0 = G * (z, 0) and x1 = G * (z, 1), where G * is the decoder function; therefore, the transformation from the source domain to the target domain is expressed as It can be expressed as

[0059] In this paper, a unidirectional image translation strategy guided by domain variables θ is adopted instead of relying on the traditional cycle consistency structure; specifically, a path guided by θ is constructed in the latent space to achieve image translation from the source domain (MRI) to the target domain (CT); its goal is to generate an image in the target domain through the shortest path from θ = 0 to θ = 1 while preserving the structure and content of the source domain; this forms a path γ x :[0,1]→G(z x ,θ);

[0060] S2, path regularization: The path regularization method is applied to smooth the path transition and ensure the stability during the image conversion process;

[0061] The details of path regularization are as follows:

[0062] Single-step generation may not be able to fully capture the complex features of the image, resulting in limitations in the quality of generation; therefore, the entire path transfer process γ x :[0,1]→G(z x ,θ) is regularized, that is, the conversion process from source domain MRI to target domain CT to ensure that the consistency of anatomical structure and content is maintained as much as possible during the conversion process;

[0063] In order to achieve path regularization, the velocity field v is optimized to guide the sample to smoothly transition from the source domain to the target domain; the velocity field v describes the direction and speed of the sample during the conversion process; the goal is to minimize the square error between the velocity field and the sample difference to ensure the smoothness and accuracy of the domain δ conversion;

[0064] Through the decoder function G dec (z, θ) calculates the velocity function v(x θ ,θ); Specifically, the velocity function v(x θ , θ) is calculated as follows:

[0065] v(x θ ,θ)=G dec (z, θ+Δθ)-G dec (z, θ)

[0066] Among them G dec (z, θ) represents the decoding result under the domain variable θ, and the path regularization optimization objective is as follows

[0067]

[0068] This approach is able to directly measure the variation of the decoded representation with θ and optimize the velocity field accordingly, ensuring a smooth transition in the domain δ;

[0069] In order to standardize the path conversion process, we must first ensure that the starting domain can accurately reconstruct the original image, while ensuring that the image of the end domain conforms to the distribution of the target domain; during the path construction process, we gradually learn the mapping from the source domain to the target domain, ensuring a smooth transition while maximally retaining the content information of the source domain and successfully transferring it to the target domain image;

[0070] Its optimization objectives include reconstruction loss, adversarial loss, and multi-scale path loss;

[0071] Reconstruction loss: To ensure that the starting domain can accurately reconstruct the original image, the reconstruction loss is optimized and is defined as follows: Among them, G(z,0) represents the output image of the generator in the starting domain, x0 represents the original image, and z is the encoding result of the original image;

[0072] Adversarial loss: In order to make the images generated by the destination domain better match the style of the target domain, adversarial loss is introduced, which is defined as follows: in is the output image of the generator in the destination domain, and D is the discriminator used to distinguish the generated image from the real image x1 in the target domain; this loss ensures that the generated image is visually close to the distribution of the target domain;

[0073] Multi-scale path loss: To ensure the smoothness of multi-scale paths, the path lengths of multiple feature layers l are considered, defined as: f l =G l (z, 1)-G l (z, 0)-v l (x θ ,θ) where G l (z, 1) and G l (z, 0) represent the decoded features under the latent vector z and parameter θ respectively;

[0074] To measure the multi-scale loss, the path length loss is defined as:

[0075]

[0076] By minimizing the path length loss L path , ensuring the smoothness of transition paths at different scales, thus facilitating a coherent and accurate conversion from MRI to CT images;

[0077] In summary, by adopting path regularization, we not only ensure structural consistency in image translation, but also solve potential problems related to path smoothness and content preservation; this approach helps to accurately and reliably generate target domain images from source domain MRI images;

[0078] S3, contour guidance: The path regularization technology based on contour guidance enables the model to focus on key anatomical areas, further improving the structural consistency and accuracy of the generated CT images;

[0079] The details about contour guidance are as follows:

[0080] The contour area in CT images is of great significance for capturing key bone feature areas. In the unsupervised paradigm, considering unpaired CT images, a contour generation network G is constructed. grad (x), the contour generation network G grad (x) Taking MRI images as input, the corresponding CT image contours are generated; this method enables the model to be indirectly guided to focus on important contour features without paired CT images;

[0081] Specifically, the contour generation network Ggrad (x) generates contours from MRI images, and another generator G(z x , 1) then generate paired CT images; then use the Sobel algorithm to extract contours from these generated CT images; in order to ensure that the generated contours are consistent with G grad (x) Align the generated contours and optimize the contours extracted by Sobel with G grad (x) The L1 norm between the generated contours; this optimization helps the model focus on detailed contour regions more effectively when generating CT images;

[0082] Its contour loss is defined as follows:

[0083]

[0084] In order to achieve efficient and effective path regularization, contour constraints are only applied to the starting and ending domains, and the contour information is used as an additional attention map to correct the intermediate domain during the path transfer process; this method avoids imposing strict contour invariance constraints on each point of the path, thereby reducing computational overhead; the contour generation network does not participate in back propagation in this process, which significantly improves model performance and reduces computational burden; this strategy ensures a balance between effectiveness and efficiency;

[0085] G grad (x0) interpolates to f l The resolution of the matching layer is adjusted, and bilateral filtering is applied to the transition area near the contour for smoothing; then an attention weight W is assigned to the feature map on each layer l. l ; The modified path loss function combined with the attention weight is:

[0086]

[0087] Among them, α controls the importance of the attention mechanism and adjusts the contribution of each layer of feature map fl in the loss calculation;

[0088] The overall loss function of the model is defined as:

[0089] L full =λ gan L ga n+λ path L path +λ c L c +λ rec L rec .

[0090] like Figure 1As shown in Figure 2, the entire network consists of a backbone network and a contour generator. The backbone network uses an encoder-decoder structure to encode the source domain image into a latent code. The domain variable θ adjusts the affine factor of the AdaIN layer to control the decoding result and facilitate the entire path conversion process (e.g. Figure 2 The contour generator generates latent CT contours from the source MRI images and uses them as attention maps to guide the path conversion process of the decoder. The paired CT images generated by the backbone network are contour extracted by the Sobel operator and optimized with the latent CT contours using the L1 loss.

[0091] like Figure 2 As shown, the goal is to translate along the trajectory rather than the path shown by the dashed line. The end point of the dashed path may result in loss of content and anatomical structure, thus maintaining structural consistency of key anatomical regions during medical image translation.

[0092] like Figure 3 As shown in the figure, when the trained model is applied, the input MRI image will enter the contour generator module and the path transfer module (parallel processing) respectively. Among them, the contour generator module generates potential CT contours and interpolates and adjusts them to match the scale in the path transfer module, thereby achieving an effect similar to the attention map, making the model pay more attention to the complex features of the contour area. Finally, the path transfer module outputs the CT image corresponding to the MRI image.

[0093] like Figure 4 As shown, from left to right are the source domain MRI image, the real paired CT image, and the CT image generated by the model. This method ensures that the generated images not only have a high degree of anatomical accuracy, but also retain the key details required for medical diagnosis, radiotherapy planning, and surgical operations.

[0094] The experimental process of the present invention is as follows:

[0095] The model was implemented using the PyTorch framework and trained on an Nvidia RTX 4090 GPU. The Adam optimizer was used during training, with β1 = 0.5 and β2 = 0.9. For the hyperparameter setting, the step size of Δθ was fixed to 0.05 and the number of cross-validations was set to 100. The learning rate was configured to 2.0×10 -4 , the contour loss weight and path loss weight are set to 5.0 and 0.1 respectively, and the reconstruction loss weight is set to 5.0. During the evaluation process, each slice generates a target image from the source image, and the model performance is evaluated by the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learning-perceptual image patch similarity (LPIPS) indicators.

[0096] Two different datasets are used to verify the effectiveness of the model: a multimodal pelvic MRI-CT dataset and a head MRI-CT dataset provided by the First Affiliated Hospital of Sun Yat-sen University. The evaluation results are shown in Tables 1 and 2. Through experimental verification, the effect of this method on the test set is better than other classic methods.

[0097] Table 1: Comparison results of head MRI-CT datasets

[0098]

[0099]

[0100] Table 2: Comparison results of multimodal pelvic MRI-CT datasets

[0101]

[0102] In addition, the present invention includes but is not limited to the following cases:

[0103] Step 1: Data preprocessing

[0104] First, the MR image volume was preprocessed, including denoising, N4 bias field correction, and registration with the corresponding CT image. Then, 150 axial slices containing skull structures were extracted from each patient's image volume, and the image matrix was reconstructed to a resolution of 512×512 pixels for subsequent calculations. Finally, the dataset was divided into training, test, and validation sets in a ratio of 7:2:1.

[0105] Step 2: Construct a transfer path from the source domain (MRI image) to the target domain (CT image)

[0106] Step 2-1

[0107] The source domain (MRI) image is first mapped to a shared latent space Z through an encoder. The encoder structure includes a 7x7 convolution layer, two 3x3 convolution downsampling layers, a batch normalization layer, and a ReLU activation function, which can compress the input image layer by layer and extract its key structural and semantic information, and finally generate a low-dimensional feature representation.

[0108] Step 2-2

[0109] The domain variable θ is first encoded through an embedding layer, and the embedding result is then used to control the adaptive instance normalization (AdaIN) layer of the decoder. In this process, the embedding result of the domain variable θ provides the mean and standard deviation parameters required by the AdaIN layer to adjust the features in the latent space Z. When the domain variable θ = 0, the decoder maintains the source domain features through the AdaIN layer and reconstructs the MRI image of the source domain. When the domain variable θ = 1, the AdaIN layer adjusts the features in the Z space according to the statistical characteristics of the target domain (CT) to generate the corresponding CT image. This forms a transfer path γ x :[0,1]→G(z x , θ).

[0110] Step 3: Regularization path transfer process

[0111] Step 3-1, calculate the path speed function v(x θ ,θ)

[0112] Through the decoder function G dec (z, θ) calculates the velocity function v(x θ Specifically, the velocity function v(x θ , θ) is calculated as follows:

[0113] v(x θ ,θ)=G dec (z, θ+Δθ)-G dec (z, θ)

[0114] Among them G dec (z, θ) represents the decoding result under the domain variable θ;

[0115] Construct the overall optimization objective, i.e., the loss function, which includes path loss, reconstruction loss, and adversarial loss;

[0116] Step 3-2: Calculate multi-scale path loss

[0117] Considering the path length of multiple feature layers l, it is defined as: f l =G l (z, 1)-G l (z, 0)-v l (x θ ,θ) where G l (z, 1) and G l (z, 0) represents the decoded features under the latent vector z and parameter θ, respectively.

[0118] To measure the multi-scale loss, the path length loss is defined as:

[0119]

[0120] By minimizing the path length loss L path , ensuring the smoothness of the transition paths at different scales, thus facilitating a coherent and accurate conversion from MRI to CT images.

[0121] Step 3-3, calculate the reconstruction loss:

[0122] Reconstruction loss: To ensure that the starting domain can accurately reconstruct the original image, the reconstruction loss is optimized and is defined as follows: Among them, G(z, 0) represents the output image of the generator in the starting domain, x0 represents the original image, and z is the encoding result of the original image.

[0123] Step 3-4, calculate the adversarial loss:

[0124] In order to make the images generated by the destination domain better match the style of the target domain, an adversarial loss is introduced, which is defined as follows: in is the output image of the generator in the destination domain, and D is the discriminator used to distinguish the generated image from the real image x1 in the target domain. This loss ensures that the generated image is visually close to the distribution of the target domain.

[0125] Step 4: Contour guide

[0126] Step 4-1. First, build a contour generation network G grad (x), taking MRI images as input, it generates the corresponding CT image contours. In the absence of paired CT images, the network indirectly guides the model to focus on important areas in the CT image by capturing contour features.

[0127] Step 4-2: Use the Sobel algorithm to extract contours from the CT image generated above

[0128] Step 4-3: Calculate the contour loss function

[0129]

[0130] Step 4-4, G grad (x0) interpolates to f l The resolution of the image is matched, and bilateral filtering is applied to the transition area near the contour for smoothing. Then an attention weight W is assigned to the feature map on each layer l. l The modified path loss function combined with the attention weight is:

[0131]

[0132] Among them, α controls the importance of the attention mechanism and adjusts the feature map f of each layer l Contribution in loss calculation.

[0133] The overall loss function of the model is defined as:

[0134] L full =λ gan L gan +λ path L path +λ c L c +λ rec L rec

[0135] Step 5: Model training and evaluation

[0136] The model is trained using the loss function defined in step 4. The model is implemented using the PyTorch framework and is trained on an Nvidia RTX 4090 GPU. The Adam optimizer is used during training, with β1 = 0.5 and β2 = 0.9. For the hyperparameter setting, the step size of Δθ is fixed to 0.05 and the number of cross-validations is set to 100. The learning rate is configured to 2.0×10 -4 , the contour loss weight and path loss weight are set to 5.0 and 0.1 respectively, and the reconstruction loss weight is set to 5.0. During the evaluation process, each slice generates a target image from the source image, and the model performance is evaluated by the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and learning-perceptual image patch similarity (LPIPS) indicators.

[0137] Based on the above, the present invention realizes a smooth transition between the source domain (MRI) and the target domain (CT) by constructing a transition path in the latent space. Through path regularization, the present invention can effectively reduce the image drift and content loss problems that may occur in the direct generation process. The present invention also adopts adversarial training to ensure that the generated image conforms to the style characteristics of the target domain CT image. At the same time, the prior anatomical information extracted from the MRI image is introduced into the path transition process to further optimize the transition effect. The present invention not only reduces the risk of path loss, but also can focus on key anatomical areas to ensure the structural consistency and accuracy of the generated image.

[0138] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. Unpaired cross-modal medical image conversion method based on contour-guided path regularization, characterized in that: The specific steps are as follows: S1, path establishment: constructing a conversion path from the source domain to the target domain; S2, path regularization: The path regularization method is applied to smooth the path transition and ensure the stability during the image conversion process; S3, contour guidance: The path regularization technology based on contour guidance enables the model to focus on key anatomical areas, further improving the structural consistency and accuracy of the generated CT images.

2. The unpaired cross-modal medical image conversion method based on contour-guided path regularization according to claim 1, characterized in that: The specific process of S1 is as follows: S11: Let x0 and x1 represent samples in the source domain and target domain respectively, where x0 comes from the data distribution π0 of MRI images and x1 comes from the data distribution π1 of CT images; S12: Given a basic data pair And assume that images from both domains pass through the same decoder G * , generated from a shared latent space Z and a continuous domain variable θ, which is specifically, z = E * (x0) = E * (x1), where E * represents the encoder function, which leads to x0 = G * (z,0) and x1=G * (z,1), where G * is the decoder function, so the conversion from the source domain to the target domain is expressed as Specifically expressed as 3. The unpaired cross-modal medical image conversion method based on contour-guided path regularization according to claim 1, characterized in that: The specific process of S2 is as follows: S21, calculate the path speed function: through the decoder function G dec (z,θ) calculates the velocity function v(x θ ,θ); S22, perform path regularization optimization objectives, as follows:

4. The unpaired cross-modal medical image conversion method based on contour-guided path regularization according to claim 3, characterized in that: The velocity function v(x θ ,θ) is calculated as follows: v(x θ ,θ)=G dec (z,θ+Δθ)-G dec (z,θ) Among them G dec (z,θ) represents the decoding result under the domain variable θ.

5. The unpaired cross-modal medical image conversion method based on contour-guided path regularization according to claim 4, characterized in that: The optimization objectives include reconstruction loss, adversarial loss, and multi-scale path loss; The reconstruction loss is calculated as Among them, G(z,0) represents the output image of the generator in the starting domain, x0 represents the original image, and z is the encoding result of the original image; The adversarial loss is calculated as in, is the output image of the generator in the destination domain, and D is the discriminator used to distinguish the generated image from the real image x1 in the target domain; The multi-scale path loss is based on the path lengths of multiple feature layers l, which is calculated as f l =G l (z,1)-G l (z,0)-v l (x θ ,θ), where G l (z,1) and G l (z,0) represents the decoded features under the latent vector z and parameter θ respectively; In order to measure the multi-scale loss, the path length loss is calculated as: Based on this, by minimizing the path length loss L path , ensuring the smoothness of transition paths at different scales, thus facilitating a coherent and accurate conversion from MRI to CT images.

6. The unpaired cross-modal medical image conversion method based on contour-guided path regularization according to claim 1, characterized in that: The specific process of S3 is as follows: S31: Constructing the contour generation network G grad (x), and taking the MRI image as input, generate the corresponding CT image contour; S32: Extract contours from CT images using the Sobel algorithm; S33: Optimizing the contours extracted by the Sobel algorithm and the contour generation network G grad (x) The L1 norm between the generated contours to ensure that the contours extracted by the Sobel algorithm are consistent with the contour generation network G grad (x) The generated contours are aligned, which helps the model focus on detailed contour areas more effectively when generating CT images; S34: G grad (x0) interpolates to f l The resolution of the matching layer is smoothed by applying bilateral filtering in the transition area near the contour, and then an attention weight W is assigned to the feature map on each layer; l , and its modified path loss function combined with the attention weight is: Among them, α controls the importance of the attention mechanism and adjusts the feature map f of each layer l contribution in the calculation of loss; In addition, the overall loss function of the model is: L full =λ gan L gan +λ path L path +λ c L c +λ rec L rec 。 7. The unpaired cross-modal medical image conversion method based on contour-guided path regularization according to claim 6, characterized in that: The loss function calculation method of the contour in S33 is: In order to achieve efficient and effective path regularization, contour constraints are only applied in the starting and ending domains, and the contour information is used as an additional attention map to correct the intermediate domain during the path transfer process.

Images