A method for determining a two-phase CT image bidirectional synthesis model

Abstract

The application discloses a determination method of a two-phase CT image bidirectional synthesis model, and relates to the field of medical image synthesis. The method comprises the following steps: constructing a data set, including multiple groups of original arterial phase CT images and original portal venous phase CT images of the same patient; preprocessing each group of images to obtain a training data set, and generating an anchor point coordinate with a lesion area as the center for each image; constructing a two-phase CT image bidirectional synthesis model; and training the model based on the training data set through a composite loss function composed of an adversarial loss, a cycle consistency loss and a pairwise term constraint loss to obtain a trained model. The model comprises a first trained generator and a second trained generator; the first trained generator is used for converting a clinical arterial phase CT image into a predicted portal venous phase CT image; and the second trained generator is used for converting a clinical portal venous phase CT image into a predicted arterial phase CT image. The application can synthesize a high-quality other-phase CT image based on a single-phase CT image.

Description

Technical Field

[0001] This application relates to the field of medical image synthesis, and in particular to a method for determining a bidirectional synthesis model of dual-phase CT images. Background Technology

[0002] In clinical practice, dual-phase CT imaging (arterial phase CT images and portal venous phase CT images) is a key imaging tool for evaluating rectal cancer. Arterial phase CT images can clearly show tumor neovascularization, while portal venous phase CT images can better demonstrate tumor parenchymal enhancement and its relationship with surrounding tissues. The two types of images provide complementary diagnostic information. However, existing dual-phase CT imaging technology has the following significant drawbacks: 1) Patients need to undergo two scans, increasing the potential risks of radiation exposure and iodine contrast agents; 2) Physiological movements such as breathing and heartbeat between the two scans make image registration difficult, seriously affecting the accuracy of subsequent fusion analysis; 3) The scanning process is time-consuming and requires a high level of medical resources and patient cooperation.

[0003] Medical image synthesis has emerged as a promising alternative, with the core objective of synthesizing images from one phase to another. However, existing methods based on Generative Adversarial Networks (GANs) primarily focus on maintaining the overall visual realism or cyclic consistency of images, lacking explicit modeling and constraints on the anatomical correspondences crucial for medical diagnosis. This results in synthesized images that, while visually realistic, exhibit structural distortions or spatial misalignments in key diagnostic areas such as tumor margins and microvessels, failing to meet the stringent requirements of anatomical accuracy in clinical diagnosis. Furthermore, the inherent noise and artifacts in medical images, along with the significant anatomical variability among patients, also challenge the robustness and generalization capabilities of existing models. Summary of the Invention

[0004] The purpose of this application is to provide a method for determining a bidirectional synthetic model of two-phase CT images, which can synthesize another phase CT image with high quality based on a single-phase CT image. This significantly reduces the risks of clinical radiation and contrast agents, while improving the structural fidelity of the synthesized image in key anatomical areas such as lesions and the robustness of the model to image noise and individual differences.

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for determining a bidirectional synthetic model of dual-phase CT images, including: Construct a dataset. The dataset includes multiple sets of dual-phase CT images; each set of dual-phase CT images includes one original arterial phase CT image and one original portal venous phase CT image of the same patient.

[0006] Each set of dual-phase CT images in the dataset is preprocessed to obtain a training dataset, and at least one anchor point coordinate centered on the lesion area is generated for each image.

[0007] Construct a two-way synthesis model for dual-phase CT images.

[0008] Based on the training dataset, the bidirectional synthesis model of dual-phase CT images is trained using a composite loss function to obtain a trained bidirectional synthesis model of dual-phase CT images. The composite loss function consists of adversarial loss, cycle consistency loss, and pairwise constraint loss. The trained bidirectional synthesis model of dual-phase CT images includes a trained first generator and a trained second generator. The trained first generator is used to convert clinical arterial phase CT images into predicted portal venous phase CT images. The trained second generator is used to convert clinical portal venous phase CT images into predicted arterial phase CT images.

[0009] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method for determining a bidirectional synthesis model of dual-phase CT images. The trained bidirectional synthesis model of dual-phase CT images can directly generate another phase CT image based on a single-phase CT image (arterial phase or portal venous phase), thus eliminating the need to perform two scans on the same patient and avoiding problems such as radiation exposure, contrast agent risks, difficulty in motion registration, and consumption of medical resources.

[0010] In the bidirectional synthesis model of dual-phase CT images, a context discarding module is integrated into the bottleneck layer of the generator. This module generates a local discard mask on the feature map output by the encoder based on the anchor point coordinates and performs random masking. Simultaneously, the composite loss function includes a pairwise constraint loss. Through these two mechanisms, the model is forced to focus on the feature correspondence between the lesion region and its surroundings during training, and explicitly penalizes inconsistencies in corresponding positions between the two-phase images, thereby improving the structural accuracy of the synthesized image in key diagnostic regions. Furthermore, the context discarding module simulates information loss by randomly masking the feature map, enabling the model to adapt to local variations in the input image and enhancing its ability to handle different image qualities and anatomical differences. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1An application environment diagram for a method for determining a bidirectional synthetic model of dual-phase CT images provided in an embodiment of this application; Figure 2 A flowchart illustrating a method for determining a bidirectional synthetic model of dual-phase CT images according to an embodiment of this application; Figure 3 This diagram illustrates the data flow and supervision relationship in the bidirectional synthesis of two-phase CT images. Figure 4 A schematic diagram of the specific architecture of the bidirectional synthesis model for dual-stage CT images; Figure 5 This is a schematic diagram of the structure of a generator provided in one embodiment of this application; Figure 6 This is a schematic diagram of the structure of a discriminator provided in one embodiment of this application; Figure 7 This is a schematic diagram of the context discarding module structure provided in another embodiment of this application; Figure 8 Qualitative results of image segmentation using different generative networks provided in one embodiment of this application; Figure 9 A visualization comparison of the quantitative performance of the bidirectional synthesis model of dual-phase CT images provided in one embodiment of this application and existing models on the dual-phase CT image synthesis task; Figure 10 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] The method for determining the bidirectional synthesis model of dual-phase CT images provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be set up independently, integrated into server 104, or placed in the cloud or on another server. Terminal 102 can send a dataset to server 104. After receiving the dataset, server 104 preprocesses each group of dual-phase CT images in the dataset to obtain a training dataset and generates at least one anchor point coordinate centered on the lesion region for each image. A bidirectional synthesis model of dual-phase CT images is constructed. Based on the training dataset, the bidirectional synthesis model of dual-phase CT images is trained using a composite loss function to obtain a trained bidirectional synthesis model of dual-phase CT images. The composite loss function consists of adversarial loss (based on Wasserstein loss and gradient penalty), cycle consistency loss, and pairwise constraint loss. The trained bidirectional synthesis model of dual-phase CT images includes a trained first generator and a trained second generator. The trained first generator is used to convert clinical arterial phase CT images into predicted portal venous phase CT images. The trained second generator is used to convert clinical portal venous phase CT images into predicted arterial phase CT images. Server 104 can feed back the obtained trained bidirectional synthesis model of dual-phase CT images to terminal 102. Furthermore, in some embodiments, the method for determining the bidirectional synthesis model of dual-phase CT images can also be implemented separately by server 104 or terminal 102. For example, terminal 102 can directly process the dataset, or server 104 can obtain the dataset from the data storage system and process it.

[0016] The terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, and smart in-vehicle devices. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. The server 104 can be implemented using a standalone server or a server cluster composed of multiple servers, or it can be a cloud server.

[0017] In one exemplary embodiment, such as Figure 2 As shown, a method for determining a bidirectional synthetic model of dual-phase CT images is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 201 to 208. Wherein: Step 201: Construct a dataset; the dataset includes multiple sets of dual-phase CT images; each set of dual-phase CT images includes one original arterial phase CT image and one original portal venous phase CT image of the same patient.

[0018] Step 202: Preprocess each set of dual-phase CT images in the dataset to obtain a training dataset, and generate at least one anchor point coordinate centered on the lesion area for each image.

[0019] Step 203: Construct a bidirectional synthesis model of dual-stage CT images (DPD-GAN architecture).

[0020] Step 204: Based on the training dataset, the bidirectional synthesis model of dual-phase CT images is trained using a composite loss function to obtain a trained bidirectional synthesis model of dual-phase CT images. The composite loss function consists of adversarial loss, cycle consistency loss, and pairwise constraint loss. The trained bidirectional synthesis model of dual-phase CT images includes a trained first generator and a trained second generator. The trained first generator is used to convert clinical arterial phase CT images into predicted portal venous phase CT images. The trained second generator is used to convert clinical portal venous phase CT images into predicted arterial phase CT images.

[0021] By implementing steps 201 to 204 above, and integrating the context discarding module and paired-term constraint loss, the model can synthesize another phase image based on a single-phase CT image with high quality. This significantly improves the structural fidelity and robustness of key areas such as lesions while avoiding the risks of secondary scanning radiation and contrast agents.

[0022] In another exemplary embodiment of this application, such as Figure 3 As shown, the bidirectional synthesis model for dual-phase CT images includes a first generator, a second generator, a first discriminator, and a second discriminator. The dual discriminators employ a WGAN-GP architecture to separately distinguish between genuine and fake images in the arterial and portal venous phases, achieving a refined evaluation of the local texture and structural realism of the images. The first output of the first generator is connected to the second input of the second generator.

[0023] The second output of the first generator is connected to the input of the first discriminator.

[0024] The first output of the second generator is connected to the second input of the first generator.

[0025] The second output of the second generator is connected to the input of the second discriminator.

[0026] Both the first generator and the second generator employ an encoder-bottleneck layer-decoder network architecture, and both the first generator and the second generator integrate a context discarding module in their bottleneck layers. The context discarding module generates a local discard mask on the feature map output by the encoder based on the anchor point coordinates, and randomly masks the feature map output by the encoder. Specifically, this includes randomly setting the regions within the local discard mask to zero or discarding them according to probability.

[0027] Figure 3 In the diagram, dashed arrows represent cyclic consistency constraints used to enforce interphase anatomical consistency. Blue / green solid arrows represent data flow. ART' is the generated predicted arterial phase CT image, PV' is the generated predicted portal venous phase CT image, ART'' is the cyclically reconstructed arterial phase CT image, and PV'' is the cyclically reconstructed portal venous phase CT image.

[0028] In another exemplary embodiment of this application, such as Figure 3 As shown, step 204 above specifically includes: The original arterial phase CT image is input into the first generator to generate a predicted portal venous phase CT image.

[0029] The original portal venous phase CT image is input into the second generator to generate a predicted arterial phase CT image.

[0030] The predicted portal venous phase CT image is input into the second generator to generate a cyclically restored arterial phase CT image.

[0031] The predicted arterial phase CT image is input into the first generator to generate a cyclically restored portal venous phase CT image.

[0032] Calculate the first cycle consistency loss between the original arterial phase CT image and the cyclically reconstructed arterial phase CT image.

[0033] Calculate the second cycle consistency loss between the original portal venous phase CT image and the cyclically restored portal venous phase CT image.

[0034] Calculate the first pairwise constraint loss between the predicted portal venous phase CT image and the original portal venous phase CT image.

[0035] Calculate the second pairwise constraint loss between the predicted arterial phase CT image and the original arterial phase CT image.

[0036] The original portal venous phase CT image and the predicted portal venous phase CT image are input into a first discriminator to obtain a first discrimination result; a first adversarial loss is calculated based on the first discrimination result.

[0037] The original arterial phase CT image and the predicted arterial phase CT image are input into the second discriminator to obtain the second discrimination result; the second adversarial loss is calculated based on the second discrimination result.

[0038] Based on the first cycle consistency loss, the second cycle consistency loss, the first pairwise constraint loss, the second pairwise constraint loss, the first adversarial loss, and the second adversarial loss, the composite loss function value is calculated, and it is determined whether the training termination condition is met, thus obtaining the first judgment result.

[0039] If the first judgment result is negative, the network parameters of the bidirectional synthesis model of dual-phase CT images are adjusted according to the composite loss function value, and the training process is returned. If the first judgment result is positive, the bidirectional synthesis model of dual-phase CT images trained in this session is determined as the trained bidirectional synthesis model of dual-phase CT images.

[0040] In another exemplary embodiment of this application, the optimization of the bidirectional synthesis model of dual-phase CT images adopts an alternating training strategy. Specifically, in each round of iterative training, the network parameters of the first generator and the second generator are first fixed, and the network parameters of the first discriminator and the second discriminator are updated according to the first adversarial loss and the second adversarial loss; then, the updated network parameters of the first discriminator and the second discriminator are fixed, and the network parameters of the first generator and the second generator are updated according to the composite loss function value. The above process is repeated until the composite loss function value meets the preset training termination condition or reaches the preset number of iterations, that is, the model is determined to have converged.

[0041] This alternating training strategy ensures that the discriminator can continuously provide effective adversarial gradients to the generator, while avoiding imbalance between the generator and the discriminator during training, thereby further improving the structural fidelity of the synthesized image in key areas such as lesions and the model's convergence stability.

[0042] In another exemplary embodiment of this application, such as Figure 4 As shown, the output of the encoder of the first generator is connected to the input of the bottleneck layer of the first generator, and the output of the bottleneck layer of the first generator is connected to the input of the decoder of the first generator. The encoder of the first generator includes multiple downsampling units connected in series, each of which consists of a convolutional block and a downsampling layer connected in series. The decoder of the first generator includes a plurality of upsampling units connected in series, each of which consists of a convolutional block and an upsampling layer connected in series.

[0043] The output of the encoder of the second generator is connected to the input of the bottleneck layer of the second generator, and the output of the bottleneck layer of the second generator is connected to the input of the decoder of the second generator; the encoder of the second generator includes multiple downsampling units connected in series, and each downsampling unit is composed of a convolutional block and a downsampling layer connected in series. The decoder of the second generator includes multiple upsampling units connected in series, each of which consists of a convolutional block and an upsampling layer connected in series.

[0044] In another exemplary embodiment of this application, such as Figure 5 As shown, the first generator and the second generator each further include multiple skip connections to fuse shallow detail information and high-level semantic information; except for the deepest downsampling unit, each skip connection is used to fuse the first feature map output by any downsampling unit in the encoder with the second feature map output by the upsampling unit of the corresponding level in the decoder to obtain a fused feature map, and the fused feature map is used as the output of the upsampling unit of the corresponding level in the decoder.

[0045] In another exemplary embodiment of this application, such as Figure 4 As shown, the structure of the first discriminator is the same as that of the second discriminator; the first discriminator includes multiple downsampling units connected in series and a convolutional unit; each downsampling unit consists of a convolutional block and a downsampling layer connected in series; the convolutional unit includes two convolutional layers connected in series, and the convolutional unit is connected to the output of the last downsampling unit. Figure 6 As shown, taking the first discriminator as an example, its network structure is as follows: First, a convolution block (CB) with a kernel size of 7×7 and 32 output channels is used to receive the input image or feature map. Then, three downsampling units for feature extraction are sequentially connected in series. The first downsampling unit consists of a convolution block and a downsampling layer connected in series, where the kernel size of the convolution block is 3×3 and the number of output channels is expanded to 64; the kernel size of the convolution block of the second downsampling unit is 3×3 and the number of output channels is expanded to 128; the kernel size of the convolution block of the third downsampling unit is also 3×3, and the number of output channels is expanded to 256. At the output of the last downsampling unit, two convolutional layers are connected in series for further feature processing. Finally, a sigmoid activation function is used to output the final discrimination result (e.g., the probability of judging the input image as a real image).

[0046] In another exemplary embodiment of this application, such as Figure 7As shown, the first generator includes a first context discarding module; when the first context discarding module receives the first feature map output by the encoder of the first generator, it performs the following operations: The first context discarding module is used to: receive the anchor point coordinates when receiving the first feature map output by the encoder of the first generator; The region to be processed on the first feature map is determined based on the anchor point coordinates; Channel-level pooling is performed on the first feature map of the region to be processed to obtain a pooled first feature map, and a first operation is performed on the pooled feature map to obtain a first weight map; The first feature map is multiplied element-wise with the first weight map to obtain the first discarded feature map, and the discarded feature map is input into the decoder of the first generator.

[0047] The second generator is completely symmetrical and identical to the first generator in terms of network structure, core modules, loss participation methods, and training strategies.

[0048] In another exemplary embodiment of this application, the first operation in the above embodiments specifically includes: The pooled first feature map is input into the activation function to obtain the first importance map; Simultaneously, the pooled first feature map is processed to obtain the first region feature value. The first region feature value is compared with the randomly generated drop probability threshold to obtain the comparison result. Based on the comparison result, a drop decision is generated for each local region, and a drop mask is constructed based on all drop decisions.

[0049] The first importance graph is multiplied element-wise with the discard mask to obtain the first weight graph.

[0050] In another exemplary embodiment of this application, the formula for calculating the composite loss function value is as follows: ; Where λ is the cycle consistency loss coefficient; This represents the second wave of losses in the confrontation; The first generator converts the original arterial phase CT image into a predicted portal venous phase CT image; The first discriminator; These are the original arterial phase CT images; These are the original portal venous phase CT images; This represents the first instance of combat losses; The second generator converts the original portal venous phase CT image into a predicted arterial phase CT image. This is the second discriminator; This is the sum of the consistency loss from the first cycle and the consistency loss from the second cycle. The loss coefficient for paired-term constraints; It is the sum of the first pairwise constraint loss and the second pairwise constraint loss.

[0051] In another exemplary embodiment of this application, the adversarial loss in the composite loss function employs a Wasserstein loss with gradient penalty. This loss measures the difference in data distribution between the generated predicted image and the real original image, forcing the generator to synthesize an output that is difficult for the discriminator to distinguish from the real image through adversarial learning. Its expression is: ; in, For adversarial loss function; These are CT images from the portal venous phase. These are CT images taken during the arterial phase. For the second expectation operator; , All are real data jointly distributed; This represents the response of the second discriminator to the real image pairs; It is the first expectation operator; This represents the response of the first discriminator to the real image pair.

[0052] In another exemplary embodiment of this application, the cycle consistency loss in the composite loss function is used to ensure that the image can be restored to the original image after bidirectional conversion, maintaining structural stability, and its expression is: in, Let this be the cycle consistency loss function; The edge distribution of arterial phase CT images; The edge distribution of CT images during the portal venous phase; It is the third expectation operator; It is the fourth expectation operator; The generated portal venous phase image; For the generated arterial phase image; Arterial phase images for circulatory reconstruction; Portal venous phase image for circulatory reconstruction; It is an L1 norm.

[0053] In another exemplary embodiment of this application, the pairwise constrained loss in the composite loss function explicitly calculates the feature differences between the predicted image and its paired original image at corresponding spatial locations, maintaining the consistency of anatomical structures across phase images, with particular attention to the correspondence between lesions and surrounding key tissues. Its expression is: ; in, Pairwise constraint loss function; The edge distribution of arterial phase CT images; The edge distribution of CT images during the portal venous phase; It is the third expectation operator; It is the fourth expectation operator; The generated portal venous phase image; For the generated arterial phase image; It is an L1 norm.

[0054] In another exemplary embodiment of this application, a specific example of step 202 is given: Each set of dual-phase CT images in the dataset is preprocessed to obtain a training dataset, and at least one anchor point coordinate centered on the lesion region is generated for each image. This preprocessing specifically includes: spatial normalization, which unifies the voxel spacing of all images to a preset standard resolution (e.g., 0.98 mm × 0.98 mm × 3.0 mm) through resampling to eliminate spatial scale differences caused by different scanners and acquisition parameters; and intensity normalization, which maps the pixel values ​​of the images to a preset numerical range (e.g., [-1, 1]) through linear or piecewise linear transformation to reduce inconsistencies in intensity distribution caused by factors such as contrast agent dosage and patient constitution. The anchor point coordinates, i.e., the lesion guidance anchor points, are used to generate masks in the subsequent guidance context discarding module; the anchor point coordinates can be resampled randomly in each training iteration to enhance the model's robustness to changes in lesion location.

[0055] Compared with existing technologies, the method for determining the bidirectional synthetic model of dual-phase CT images provided in this application has the following significant advantages: (1) The anatomical fidelity of the synthesized images is significantly improved. On the LNMCR public dataset, the normalized cross-correlation (NCC) value between the synthesized image and the target real image achieved by our method reaches 0.9953, significantly outperforming existing comparative methods in quantitative evaluation. Simultaneously, visual evaluation also confirms that our method can more accurately reproduce the morphological changes of tissues surrounding the pelvis and the complex structural details of image edges. This advantage mainly stems from the synergistic effect of pairwise constraints and the cycle consistency loss function.

[0056] (2) The model’s robustness to noise and data variation is significantly enhanced. Ablation experiments show that after introducing the context discarding module designed in this application, the model's performance on the key evaluation metric mean squared error (MSE) is improved by up to 2.83%, indicating that the model can still maintain stable and high-quality synthetic output even when there is interference or missing information in the input image. This advantage stems directly from the context discarding module mechanism based on lesion anchor points. This mechanism generates a local discard mask by using randomly selected lesion center coordinates as anchor points, achieving a kind of "guided stochastic regularization": it forces the network to learn more generalizable robust features by actively discarding local context information, and ensures the focus and stability of the reconstruction of the core pathological structure because the anchor points surround the key diagnostic area (lesion).

[0057] (3) The clinical applicability of the synthesized images has been effectively verified. In downstream lesion segmentation tasks, the dual-channel U-Net model trained on images synthesized using this method achieved optimal segmentation accuracy. In lymph node metastasis classification tasks, the ResNet50 model trained on images synthesized using this method also achieved the highest area under the receiver operating characteristic (AUC) value. This combined advantage reflects the optimization of the aforementioned two techniques: high anatomical fidelity provides accurate spatial prior information for downstream tasks, while strong robustness ensures the consistency of input data quality. The combination of these two aspects enables the synthesized images to not only possess visual realism but also practical clinical value in supporting accurate assisted diagnosis.

[0058] In another exemplary embodiment of this application, an ablation study was conducted on the composite loss function and the context drop-out module. Table 1 shows that the model incorporating the context drop-out module significantly outperforms the baseline model using only the cycle consistency loss on all metrics, with a particularly significant improvement in mean squared error. The context drop-out module training strategy using a random selection mode improves the mean squared error by 2.83% compared to the baseline model; while the context drop-out strategy combined with an attention mechanism brings a 0.49% improvement in mean squared error. The introduction of the context drop-out module acts as a regularization mechanism, effectively preventing overfitting and enabling the model to have stronger generalization ability on unseen data. This mechanism forces the network to learn more robust feature representations, thereby reducing reconstruction error. Furthermore, by integrating an attention mechanism into the context drop-out mechanism framework, the model can focus on the most discriminative feature regions, further improving prediction accuracy.

[0059] Table 1. Results of ablation studies using loss function and context dropout module.

[0060] To further verify the superiority of the proposed method, we conducted downstream task research on biphasic images generated by different methods. We used a dual-channel U-Net model of the same size to segment lesions in various generated images, and the qualitative segmentation results are as follows: Figure 8 As shown in the figure, observation of the segmentation results leads to the conclusion that the images generated by our proposed method bring the most significant performance improvement in the segmentation task. This indicates that the anatomical correspondence established by the method effectively enhances the accuracy of lesion detection and contour delineation. Furthermore, the high-quality images generated by the model provide clearer and more detailed visualization of pathological features for clinical decision-making, especially in preoperative planning, where precise lesion localization is crucial. The method exhibits strong robustness to deformations caused by respiration and heartbeat, further demonstrating its applicability in diverse clinical scenarios and its ability to provide consistent and high-quality results for different patient groups. Figure 8 In the diagram, green indicates correctly segmented regions, while blue and red represent false negative and false positive segmentation results, respectively.

[0061] In another exemplary embodiment of this application, the overall performance of the complete solution of the present invention (bidirectional synthesis model of dual-stage CT images) on a standard dataset with optimal parameter configuration is verified.

[0062] 1.1 Test Methods Dataset: The publicly available LNMCR dual-phase CT dataset for rectal cancer was used, containing paired images from 117 patients. Five-fold cross-validation was employed. Images were resampled to (0.98, 0.98, 3.0) mm, and intensity was normalized to [-1, 1].

[0063] 1.2 Test Results The comparison methods include: M-GAN, FDE-Net, Cycle-GAN, MR-GAN, AMC-GAN, and bidirectional synthesis model of dual-phase CT images.

[0064] As shown in Table 2, the proposed bidirectional synthesis model for dual-phase CT images achieves the best results in terms of mean squared error (27.24), peak signal-to-noise ratio (27.856), and normalized cross-correlation (0.9953), indicating its significant advantages in the bidirectional synthesis task of dual-phase CT images. Considering all indicators, the bidirectional synthesis model for dual-phase CT images in this application demonstrates stronger robustness and overall effectiveness.

[0065] Table 2. Quantitative performance comparison between the two-stage CT image synthesis model and existing models on the two-stage CT image synthesis task.

[0066] Figure 9A visual comparison of the above results is provided. Each row corresponds to a different test image: the first column is the original portal venous phase image, the second column is the original arterial phase image, and subsequent columns show the differences between the synthesized results of each method and the original arterial phase images. It is evident that the proposed method not only excels in accurately restoring the pelvic periphery tissues and rectal deformation but also effectively preserves fine structural details at the image edges. Compared to other algorithms, the bidirectional synthesis model of dual-phase CT images is more accurate in maintaining these subtle anatomical features.

[0067] This application also provides an application scenario in which the above-described method for determining a bidirectional composite model of two-phase CT images is applied. Specifically, the method for determining a bidirectional composite model of two-phase CT images provided in this embodiment can be applied in the scenario of preoperative auxiliary diagnosis and surgical planning for colorectal cancer. The above scenario includes the image acquisition stage, the bidirectional image synthesis stage, and the lesion analysis and surgical planning stage. The method for determining a bidirectional composite model of two-phase CT images provided in this embodiment belongs to the bidirectional image synthesis stage.

[0068] In another exemplary embodiment of this application, the method for determining the bidirectional synthetic model of dual-phase CT images can be deployed in a dual-phase CT image synthesis system for rectal cancer. This system can be deployed on a server, workstation, or cloud, and its core consists of the following functional modules: (1) Data interface and preprocessing module: Receives external input single-phase contrast-enhanced CT images (original arterial phase CT images or original portal venous phase CT images) of rectal cancer patients, performs preprocessing operations, including: image resampling, size normalization, CT value intensity normalization, and format unification, to obtain standardized input images and eliminate the problem of inconsistent image distribution caused by differences in equipment and scanning parameters.

[0069] (2) Lesion Anchor Point Generation Module: For each standardized input image, generate one or more anchor point coordinates centered on the lesion area; the anchor points are used to guide the mask generation of the subsequent context discarding module, so that the model focuses on feature learning of the lesion and key anatomical areas.

[0070] (3) Bidirectional synthesis network module for dual-phase CT images: The bidirectional synthesis network module for dual-phase CT images is the core synthesis unit of the system. Its architecture includes three parts: a bidirectional generator (first generator and second generator), a dual discriminator (first discriminator and second discriminator), and a composite loss calculation unit. The bottleneck layer of the bidirectional generator integrates a context discarding module, which is used to generate a local discarding mask on the feature map output by the encoder according to the anchor point coordinates, and randomly mask the feature map output by the encoder to improve the model's generalization ability and anti-interference ability. The two generators are used to convert clinical arterial phase CT images into predicted portal venous phase CT images and to convert clinical portal venous phase CT images into predicted arterial phase CT images, respectively, and are supervised by the first discriminator and the second discriminator. Cyclic consistency loss is used to force anatomical consistency between phases.

[0071] (4) Result post-processing and output module: The synthesized image is resampled back to its original spatial resolution and converted into a standard DICOM or image format for output to ensure compatibility with the original data; the output image can be directly connected to the clinical PACS system, image workstation or AI-assisted diagnostic system to achieve clinical compatibility.

[0072] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 10 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores centralized data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for determining a bidirectional synthetic model of dual-phase CT images.

[0073] Those skilled in the art will understand that Figure 10The structures shown are merely block diagrams of some structures related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. In an exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0074] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0075] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0076] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0077] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0078] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for determining a bidirectional synthesis model of dual-phase CT images, characterized in that, The determination method includes: Construct a dataset; the dataset includes multiple sets of dual-phase CT images; each set of dual-phase CT images includes one original arterial phase CT image and one original portal venous phase CT image of the same patient; Preprocess each set of dual-phase CT images in the dataset to obtain a training dataset, and generate at least one anchor point coordinate centered on the lesion area for each image. Construct a two-way synthesis model for dual-phase CT images; Based on the training dataset, the bidirectional synthesis model of dual-phase CT images is trained using a composite loss function to obtain a trained bidirectional synthesis model of dual-phase CT images. The composite loss function consists of adversarial loss, cycle consistency loss, and pairwise constraint loss. The trained bidirectional synthesis model of dual-phase CT images includes a trained first generator and a trained second generator. The trained first generator is used to convert clinical arterial phase CT images into predicted portal venous phase CT images. The trained second generator is used to convert clinical portal venous phase CT images into predicted arterial phase CT images.

2. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 1, characterized in that, The bidirectional synthesis model of dual-phase CT images includes a first generator, a second generator, a first discriminator, and a second discriminator; the first output of the first generator is connected to the second input of the second generator. The second output of the first generator is connected to the input of the first discriminator; The first output of the second generator is connected to the second input of the first generator; The second output of the second generator is connected to the input of the second discriminator; Both the first generator and the second generator adopt an encoder-bottleneck layer-decoder network architecture, and both the first generator and the second generator integrate a context discarding module in their bottleneck layers. The context discarding module is used to generate a local discard mask on the feature map output by the encoder based on the anchor point coordinates, and to randomly mask the feature map output by the encoder.

3. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 2, characterized in that, Based on the training dataset, the bidirectional synthesis model of dual-phase CT images is trained using a composite loss function to obtain a trained bidirectional synthesis model of dual-phase CT images, specifically including: The original arterial phase CT image is input into the first generator to generate a predicted portal venous phase CT image. The original portal venous phase CT image is input into the second generator to generate a predicted arterial phase CT image. The predicted portal venous phase CT image is input into the second generator to generate a cyclically restored arterial phase CT image. The predicted arterial phase CT image is input into the first generator to generate a cyclically restored portal venous phase CT image. Calculate the first cycle consistency loss between the original arterial phase CT image and the cyclically reconstructed arterial phase CT image; Calculate the second cycle consistency loss between the original portal venous phase CT image and the cyclically restored portal venous phase CT image; Calculate the first pairwise constraint loss between the predicted portal venous phase CT image and the original portal venous phase CT image; Calculate the second pairwise constraint loss between the predicted arterial phase CT image and the original arterial phase CT image; The original portal venous phase CT image and the predicted portal venous phase CT image are input into a first discriminator to obtain a first discrimination result; a first adversarial loss is calculated based on the first discrimination result. The original arterial phase CT image and the predicted arterial phase CT image are input into the second discriminator to obtain the second discrimination result; the second adversarial loss is calculated based on the second discrimination result. Based on the first cycle consistency loss, the second cycle consistency loss, the first pairwise constraint loss, the second pairwise constraint loss, the first adversarial loss, and the second adversarial loss, the composite loss function value is calculated, and it is determined whether the training termination condition is met, thus obtaining the first judgment result. If the first judgment result is negative, the network parameters of the bidirectional synthesis model of the dual-stage CT images are adjusted according to the composite loss function value, and the training process is returned. If the first judgment result is positive, the bidirectional synthesis model of the dual-stage CT images trained in this training is determined as the trained bidirectional synthesis model of the dual-stage CT images.

4. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 3, characterized in that, Adjusting the network parameters of the bidirectional synthesis model of dual-phase CT images based on the composite loss function value specifically includes: In each round of iterative training, the network parameters of the first generator and the second generator are first fixed, and the network parameters of the first discriminator and the second discriminator are updated according to the first adversarial loss and the second adversarial loss; then the updated network parameters of the first discriminator and the second discriminator are fixed, and the network parameters of the first generator and the second generator are updated according to the composite loss function value.

5. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 2, characterized in that, The output of the encoder of the first generator is connected to the input of the bottleneck layer of the first generator, and the output of the bottleneck layer of the first generator is connected to the input of the decoder of the first generator; the encoder of the first generator includes a plurality of downsampling units connected in series, each downsampling unit consisting of a convolutional block and a downsampling layer connected in series; the decoder of the first generator includes a plurality of upsampling units connected in series, each upsampling unit consisting of a convolutional block and an upsampling layer connected in series. The output of the encoder of the second generator is connected to the input of the bottleneck layer of the second generator, and the output of the bottleneck layer of the second generator is connected to the input of the decoder of the second generator; the encoder of the second generator includes a plurality of downsampling units connected in series, each downsampling unit consisting of a convolutional block and a downsampling layer connected in series; the decoder of the second generator includes a plurality of upsampling units connected in series, each upsampling unit consisting of a convolutional block and an upsampling layer connected in series.

6. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 5, characterized in that, The first generator and the second generator each further include multiple skip connections; except for the deepest downsampling unit, each skip connection is used to fuse the first feature map output by any downsampling unit in the encoder with the second feature map output by the upsampling unit of the corresponding layer in the decoder to obtain a fused feature map, and use the fused feature map as the output of the upsampling unit of the corresponding layer in the decoder.

7. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 1, characterized in that, The structure of the first discriminator is the same as that of the second discriminator; the first discriminator includes a plurality of downsampling units and a convolution unit connected in series; each downsampling unit is composed of a convolution block and a downsampling layer connected in series; the convolution unit includes two convolution layers connected in series, and the convolution unit is connected to the output of the last downsampling unit.

8. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 2, characterized in that, The first generator includes a first context discard module; The first context discard module is used for: When the first feature map output by the encoder of the first generator is received, the anchor point coordinates are received; The region to be processed on the first feature map is determined based on the anchor point coordinates; Channel-level pooling is performed on the first feature map of the region to be processed to obtain a pooled first feature map, and a first operation is performed on the pooled feature map to obtain a first weight map; The first feature map is multiplied element-wise with the first weight map to obtain the first discarded feature map, and the discarded feature map is input into the decoder of the first generator.

9. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 8, characterized in that, The first operation specifically includes: The pooled first feature map is input into the activation function to obtain the first importance map; Simultaneously, the pooled first feature map is processed to obtain the first region feature value. The first region feature value is compared with the randomly generated drop probability threshold to obtain the comparison result. Based on the comparison result, a drop decision is generated for each local region, and a drop mask is constructed based on all drop decisions. The first importance graph is multiplied element-wise with the discard mask to obtain the first weight graph.

10. The method for determining the bidirectional synthesis model of dual-phase CT images according to claim 3, characterized in that, The formula for calculating the value of the composite loss function is as follows: ； Where λ is the cycle consistency loss coefficient; This represents the second wave of losses in the confrontation; The first generator converts the original arterial phase CT image into a predicted portal venous phase CT image; The first discriminator; These are the original arterial phase CT images; These are the original portal venous phase CT images; This represents the first instance of combat losses; The second generator converts the original portal venous phase CT image into a predicted arterial phase CT image. This is the second discriminator; This is the sum of the consistency loss from the first cycle and the consistency loss from the second cycle. The loss coefficient for paired-term constraints; It is the sum of the first pairwise constraint loss and the second pairwise constraint loss.

Images