Construction and prediction method of primary brain tumor non-enhanced sequence synthesis model

Abstract

The invention relates to the technical field of medical image data, in particular to a method for constructing and predicting a primary brain tumor non-enhanced sequence synthesis model, which comprises the following steps of: receiving non-enhanced magnetic resonance image data acquired and output by magnetic resonance imaging equipment; performing rigid registration, resampling, intensity cutting and normalization processing on the image data; inputting the input data into a tumor awareness generative adversarial network (TA-GAN); according to the method, the tumor perception generative adversarial network (TA-GAN) is constructed, so that the synthesis of the cerebral blood volume diagram under the condition of non-enhanced magnetic resonance image data is realized; the problems of safety risk, high scanning cost, crowd limitation and the like caused by dependence on a gadolinium-based contrast agent in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of medical imaging data technology, and in particular to a method for constructing and predicting non-enhanced sequence synthesis models of primary brain tumors. Background Technology

[0002] Current techniques for imaging assessment of primary intracranial tumors typically employ dynamic susceptibility contrast-perfusion-weighted imaging (DSC-PWI) to acquire cerebral blood volume (CBV) maps, reflecting the angiogenesis characteristics of the tumor. However, while this method can accurately characterize hemodynamic changes in brain tissue, it still suffers from numerous technical limitations and shortcomings in practical applications.

[0003] First, current DSC-PWI technology relies on intravenous injection of gadolinium-based contrast agents (GBCAs). While gadolinium contrast agents can enhance vascular signals and enable the quantitative calculation of perfusion parameters, their use may pose health risks, including adverse reactions such as nephrogenic systemic fibrosis (NSF) and allergic reactions. Long-term or repeated use of gadolinium contrast agents may also cause gadolinium ion deposition in brain tissue, leading to potential neurological safety hazards. Therefore, existing contrast agent-dependent brain perfusion imaging methods pose significant safety risks.

[0004] Secondly, this type of imaging method is costly, time-consuming, and relies on complex acquisition and post-processing procedures. DSC-PWI requires the continuous and rapid acquisition of multiple frames of signals during contrast agent injection to track its passage through brain tissue, which places high demands on the device's timing control, signal sampling rate, and data reconstruction capabilities. The complex scanning process not only increases examination costs and operational difficulty but also limits its promotion and application in primary healthcare institutions and multi-center imaging research.

[0005] Furthermore, existing technologies have significant limitations in terms of applicable populations. Due to the contraindications of gadolinium-based contrast agents, pregnant women, patients with renal insufficiency, and certain children are often unable to undergo such contrast-enhanced scans. This means that the technology cannot cover all clinically relevant populations and cannot achieve widespread, universal application.

[0006] Finally, DSC-PWI also has shortcomings in imaging stability. Because its signal intensity is greatly affected by changes in contrast agent concentration and magnetic field sensitivity, image quality is easily affected by motion artifacts, magnetic susceptibility artifacts, and noise interference, which leads to unstable or even distorted perfusion parameter calculations, affecting the repeatability and quantification reliability of CBV maps.

[0007] In summary, existing DSC-PWI brain perfusion imaging methods have significant shortcomings in terms of safety, cost-effectiveness, applicability, and imaging stability. Their reliance on gadolinium contrast agents not only introduces potential physiological safety risks but also increases scanning complexity and cost, limiting their application to specific populations. Therefore, a novel brain blood volume map generation method that does not require gadolinium-based contrast agents, possesses high imaging consistency, and exhibits good image representation capabilities is needed to overcome these technical deficiencies and improve the safety and universality of brain perfusion imaging. Summary of the Invention

[0008] To address the problems of existing technologies, such as reliance on gadolinium-based contrast agents, high scanning costs, low imaging efficiency, and severe artifacts, this invention proposes a method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors.

[0009] This method constructs a Tumor-Aware Generative Adversarial Network (TA-GAN), using conventional unenhanced magnetic resonance sequences as input data. Through collaborative training of the generator, discriminator, and Tumor-Aware Module (TAM), a nonlinear mapping relationship is established from conventional structural images to brain blood volume maps (CBV maps), enabling contrast-free synthesis and high-fidelity reconstruction of brain perfusion features.

[0010] The proposed solution can generate a composite CBV map with blood perfusion characteristics without any contrast agent intervention, thereby significantly reducing scanning time and cost while maintaining imaging safety, and has good universality and scalability.

[0011] This invention is achieved through the following technical solution: The method for constructing and predicting non-enhanced sequence synthesis models of primary brain tumors includes the following steps: S1. Receive non-enhanced magnetic resonance image data acquired and output by a magnetic resonance imaging device, wherein the image data includes a T2-weighted sequence, a diffusion-weighted sequence and its apparent diffusion coefficient map; S2. Perform rigid registration, resampling, intensity cropping and normalization on the image data to obtain spatially aligned and numerically consistent input data. S3. Input the input data into the tumor perception generative adversarial network (TA-GAN), wherein the TA-GAN includes a generator, a discriminator, and a tumor perception module; The generator includes an encoder, a mapping module, and a decoder. The encoder extracts image features, the mapping module generates style vectors and adjusts the convolution weights of the decoder to control image details, and the generator outputs a synthesized image of brain blood volume. The discriminator uses a multi-scale convolutional network to distinguish between the brain blood volume image output by the generator and the real brain blood volume map; The tumor perception module receives the combined features of the apparent diffusion coefficient map and the cerebral blood volume image output by the generator, generates a full perfusion mask and a high perfusion mask, and updates the generator weights through backpropagation to enhance the feature representation of high perfusion structural regions. S4, obtained through adversarial training, corresponds to the actual brain blood volume. Figure 1 Highly contagious synthetic cerebral blood volume map.

[0012] Furthermore, the encoder of the generator includes seven convolutional blocks, each of which includes two 3×3 convolutional layers. The first convolutional layer maintains spatial resolution, and the second convolutional layer performs double downsampling with stride convolution to obtain multi-level feature embeddings.

[0013] Furthermore, the mapping module includes eight fully connected layers for converting latent vectors into style vectors, which control the weight adjustment of convolutional kernels during the decoding stage.

[0014] Furthermore, the decoder uses bilinear interpolation for double upsampling and combines convolution operations after upsampling at each layer to recover spatial details. The parameters of the convolutional layers are dynamically adjusted by the style vector output by the mapping module.

[0015] Furthermore, the tumor sensing module includes an encoder and a decoder structure, both of which have five convolutional layers with a double sampling rate, and the ends include two 1×1 convolutional layers and a channel attention mechanism.

[0016] Furthermore, the loss function of the tumor perception module includes a pixel difference constraint term and a structural consistency constraint term. The gradient is calculated through backpropagation and the generator parameters are updated so that the generator focuses on enhancing the texture features of high-perfusion areas during training.

[0017] Furthermore, the discriminator consists of six hierarchical 3×3 convolutional blocks, which extract multi-scale features and output true / false classification signals through a fully connected layer.

[0018] Furthermore, the training process of TA-GAN adopts a mini-maximum game structure, where the generator minimizes the comprehensive loss function, the discriminator maximizes the discrimination accuracy between real and fake samples, and TAM participates in the optimization of generator parameters through its back gradient.

[0019] Furthermore, the input features of the tumor perception module are obtained by splicing the structural features of the apparent diffusion coefficient map with the perfusion features of the generated image. After channel-weighted fusion, the features are input into its convolutional network for the extraction of regional features of high-perfusion structures.

[0020] The beneficial effects of this invention are: This invention achieves brain blood volume (CBV) synthesis using non-contrast-enhanced magnetic resonance imaging (MRI) data by constructing a tumor-sensing generative adversarial network (TA-GAN). This overcomes the safety risks, high scanning costs, and limited patient population associated with existing technologies that rely on gadolinium-based contrast agents. The method uses T2-weighted sequences and DWI-ADC sequences as inputs, and leverages a generator, discriminator, and tumor-sensing module for collaborative optimization to reconstruct CBV data that closely resembles the real CBV without contrast agents. Figure 1 Highly consistent synthetic images. The tumor perception module guides the generator to focus on high-perfusion region features through joint feature input and inverse gradient update, effectively improving the structural consistency and texture fidelity of the images. The overall solution has high imaging safety, low cost, and strong stability, and can maintain good generalization performance under multi-center conditions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating the construction and prediction method of the non-enhanced sequence synthesis model for primary brain tumors proposed in this invention. Figure 2 This is a schematic diagram of the tumor perception generative adversarial network framework for the construction and prediction method of the non-enhanced sequence synthesis model of primary brain tumors proposed in this invention. Figure 3 This is a schematic diagram of the terminal device for the construction and prediction method of the non-enhanced sequence synthesis model of primary brain tumors proposed in this invention; Figure 4 This is a schematic diagram of a readable storage medium for the construction and prediction method of the non-enhanced sequence synthesis model of primary brain tumors proposed in this invention; In the diagram, 200 is the terminal device, 210 is the memory, 211 is the RAM, 212 is the cache, 213 is the ROM, 214 is the program / utility, 215 is the program module, 220 is the processor, 230 is the bus, 240 is the external device, 250 is the I / O interface, 260 is the network adapter, and 300 is the program product. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0024] Example 1 refer to Figure 1 This embodiment discloses a method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors. The core of this method is a tumor-sensing generative adversarial network (TA-GAN). This network generates synthetic images with high consistency with real cerebral blood volume (CBV) maps based solely on non-enhanced magnetic resonance imaging sequences without using gadolinium-based contrast agents, thereby achieving perfusion feature reconstruction of the brain tumor region. This method reduces imaging costs and operational risks while ensuring image structural consistency, making it suitable for primary intracranial tumor image analysis scenarios.

[0025] The specific steps include: 1. Receive non-enhanced magnetic resonance imaging data acquired and output by magnetic resonance imaging equipment. The system receives non-contrast magnetic resonance imaging (MRI) data acquired by an MRI machine. This data includes T2-weighted sequences, diffusion-weighted sequences, and their apparent diffusion coefficient maps. The model input consists of these two types of conventional non-contrast MRI sequences, which provide structural and diffusion feature information for subsequent feature learning by the deep network. Input data is derived from the MRI system and input into the computation module, serving as the foundational data for model training and inference.

[0026] 2. Perform resampling, rigid registration, and normalization on the image data. This embodiment performs standardized preprocessing on the input image data to ensure spatial alignment and numerical consistency.

[0027] First, rigid registration is performed using the T2WI sequence as a fixed spatial reference, and the coordinate transformation relationship between floating images such as DWI and ADC and the reference image is calculated. This process is achieved by constructing a linear transformation model containing rotation matrices and translation vectors. The rotation matrix is ​​constrained within a special three-dimensional orthogonal group, satisfying the orthogonality condition and having a determinant of one, ensuring that the transformation process only includes rigid body rotation and translation, and strictly prohibiting the introduction of scaling or deformation, thereby maintaining the geometric authenticity of the anatomical structure.

[0028] Next, a resampling operation is performed. Based on the transformation parameters determined in the registration stage, a high-order interpolation algorithm is used to remap all heterogeneous modal image data to the same spatial resolution and physical dimension as the T2WI sequence, achieving strict voxel-scale alignment. Subsequently, intensity cropping is performed. By setting high and low percentile thresholds, the resampled images are truncated to remove outlier signal values ​​generated by imaging noise or interpolation operations.

[0029] Finally, data normalization is performed to uniformly scale the intensity of the cropped image signal to a standard numerical range, thereby eliminating dimensional differences caused by different scanning devices or imaging parameters and providing input data with a consistent numerical distribution for subsequent deep neural networks.

[0030] Specifically, rigid registration uses the T2WI sequence as a spatial reference and calculates the geometric transformation relationship between floating images such as DWI and ADC relative to the reference image. This process is achieved by constructing a three-dimensional spatial coordinate transformation model, where the original image coordinates are... and target image coordinates The relationship between them is as follows: ;in It is a rotation matrix and belongs to a special orthogonal group. Satisfying orthogonality and This ensures that the transformation is purely rotational. This represents the translation amount.

[0031] In this embodiment, the image resampling setting is used to define the size of the original image. The target size is The spacing of the original pixels is The target pixel spacing is .

[0032] The formula for calculating the target pixel spacing is: ; The resampling operation formula is: ;in The original image is in position pixel values, It is the interpolation kernel function.

[0033] The spatial changes are as follows: .

[0034] This implementation also performs intensity cropping after image resampling, assuming the image intensity value is... and give the percentile range of the clipping. , indicating all below The prime values ​​will be cropped to All higher The pixel values ​​will be cropped to The cutting formula is as follows: .

[0035] Set the intensity range of the image input to: The normalization target range is [0, 255], and the intensity normalization formula is as follows: .

[0036] 3. Input the data into the tumor perception generative adversarial network (TA-GAN). The preprocessed input data is fed into the tumor perception generative adversarial network (TA-GAN). The TA-GAN consists of three parts: a generator, a discriminator, and a tumor perception module. These three parts are connected through an end-to-end structure to achieve feature mapping from non-enhanced MRI images to brain blood volume maps.

[0037] The generator comprises an encoder, a mapping module, and a decoder. The encoder extracts image features and contains seven convolutional blocks, each consisting of two 3×3 convolutional layers. The first convolutional layer maintains spatial resolution, while the second layer performs a 2x downsampling through stride convolution. After multiple convolutions, the input image is compressed into a high-dimensional feature representation. The mapping module is an eight-layer fully connected network that maps latent vectors to style vectors, used to control convolution weights during the decoding stage. The decoder employs a mirror-symmetric structure with the encoder, achieving a 2x upsampling through bilinear interpolation and recovering spatial details after each upsampling layer combined with convolution operations. The decoder's convolution weights are dynamically adjusted by the style vectors output from the mapping module to control the texture and contrast of the generated image, thereby outputting a brain blood volume image with the same dimensions as the original.

[0038] Specifically, the encoder is used to influence the preprocessed input. Multi-scale features are extracted from the encoder. The encoder contains... Several convolutional blocks, including the initial feature extraction layer. The first convolutional block directly acts on the input image. As shown in the formula: ; Then regarding the hierarchy to feature map of encoder We obtain this through the following recursive relationship: ;in This represents the feature map of the previous layer. This indicates a downsampling operation, which smooths the feature map using a low-pass filter before reducing the spatial resolution to suppress the introduction of high-frequency aliasing. It includes further feature convolution and activation operations, which not only compresses the spatial dimension, but also preserves the effective frequency band information that conforms to the sampling theorem.

[0039] The mapping module is responsible for fusing image content features with external condition information to generate a style vector for controlling the decoding process. .

[0040] First, the feature map output from the last layer of the encoder. Perform global average pooling (GAP) to extract latent vectors. : ; The mapping module then receives the potential vector. Modal category embedding and slice position embedding And map these joint conditional information into style vectors. : ;in Indicates modal category, Style vectors represent the position or depth information of a slice. During the decoding process, it is further processed into the scale of each layer. and offset parameter.

[0041] The decoder employs a mirror-symmetric structure, starting from the lowest resolution features. Initially, the original resolution is restored layer by layer. In each synthesis layer of the decoder, the core operation is modulation convolution, which modulates the input feature map. With style vectors By combining these methods, dynamic control of the feature channels can be achieved, and its mathematical expression is as follows: ;in This is the input feature map for the current decoding layer. For this layer generated by the mapping module The style vector, this operation is performed through Achieve channel-scale modulation of features, and then through Channel shifting is implemented to enable the generated feature map to dynamically adapt to changes in texture and contrast. The style of representation.

[0042] The discriminator distinguishes between the brain blood volume images output by the generator and real brain blood volume maps. The discriminator consists of six hierarchical 3×3 convolutional blocks. It extracts discriminative features through a multi-scale convolutional network. The features output by each convolutional block are flattened at the end and input into a fully connected layer, outputting a true / false classification signal to evaluate the realism of the input image. The generator and discriminator are trained in an adversarial structure. Synthetic images generated by the generator and real images are input into the discriminator. The discriminator improves its discrimination ability through multiple rounds of updates, while the generator continuously optimizes its parameters to reduce discrepancies, ultimately achieving consistency between the generated images and real brain blood volume maps in terms of structural features and statistical distribution.

[0043] Specifically, the discriminator is used to evaluate the authenticity of the input image. In this embodiment, a multi-level convolutional network structure is used to extract multi-scale features.

[0044] The forward propagation process of the discriminator consists of a series of cascaded convolutional layers, and its layer-by-layer convolution expression is as follows: First, input the image. As initial feature representation: ; Then the feature map go through Hierarchical convolutional block processing, in which : in, These are the output features of the previous layer; Representing the The convolutional operation of the layer uses spectral normalization applied to the convolutional kernel weights to stabilize the training process; This represents a non-linear activation function.

[0045] In the final layer of feature extraction, the discriminator outputs a log-probability map of true and false classifications, implementing a discriminative structure similar to PatchGAN. The final output of the discriminator... From the last convolutional layer Given: This output As a multidimensional feature map, each element corresponds to the authenticity score of a region in the input image, rather than a single global true / false signal.

[0046] Generator With discriminator In adversarial training, the discriminator aims to maximize its ability to distinguish between real and synthetic samples, while the generator aims to minimize the probability that the discriminator will classify its output as fake.

[0047] This embodiment uses an adversarial loss based on the Softplus function as the objective function, and the total loss of the discriminator is... Loss from real samples and synthetic sample loss constitute.

[0048] For data distribution from real data samples The loss items are: ; For generators From potential space The generated synthetic sample The loss items are: ; Total loss of the discriminator The sum of the two items above: ,in .

[0049] This loss function is designed to train the discriminator to accurately distinguish between real and fake images. and composite images .

[0050] refer to Figure 2 The Tumor-Aware Module (TAAM) is embedded in the TA-GAN network to guide the generator to focus on tumor regions and high-perfusion structures during training. TAM employs an encoder-decoder structure; the encoder contains five convolutional layers with 2x downsampling, and the decoder mirrorively contains five convolutional layers with 2x upsampling, ending with two 1×1 convolutional layers and a channel attention mechanism to output a feature mask. TAM receives dual inputs: an apparent diffusion coefficient map from the original image and a brain blood volume image from the generator output. The module fuses the joint features of these two inputs to output a full-tumor mask and a high-perfusion mask. Its training process is based on a tumor-aware loss function. ; in, This represents the pixel difference constraint term for the entire tumor mask. For reconstruction constraints in high-perfusion regions, For high-injection boundary constraints, The structural consistency constraint term is used; TAM calculates the gradient through backpropagation and passes the result to the generator, so that the generator's parameters are updated in each iteration in the direction of enhancing the details of the tumor region, thereby achieving structural enhancement of key regions.

[0051] The generator, discriminator, and TAM module are jointly optimized in an end-to-end manner. The generator minimizes the overall loss to improve image quality, the discriminator maximizes the discrimination accuracy to enhance the model's robustness, and the tumor perception module guides the generator update through inverse gradient descent, achieving targeted optimization of specific region features. Through multiple rounds of adversarial training, the model gradually converges, and the generator is able to generate images that are highly consistent with real brain blood volume maps in terms of spatial structure and texture features.

[0052] 4. Through adversarial training, the actual brain blood volume was obtained. Figure 1 Highly contagious synthetic cerebral blood volume map After adversarial training, the TA-GAN model enters the inference phase. Standardized T2WI and DWI-ADC images are input into the trained generator, which directly outputs a synthetic cerebral blood volume map (CBVsyn) through encoding, mapping, and decoding. The spatial resolution of the output image is consistent with the input, and image details are enhanced in high-perfusion areas. Because the TAM module has optimized the generator parameters through loss feedback during training, the generator maintains high responsiveness to tumor regions and the ability to reconstruct structural details during inference.

[0053] The generated cerebral blood volume images can be directly used in medical image analysis systems for quantitative analysis and image assessment of brain tumor regions. This model can synthesize cerebral blood volume maps without contrast agents, and the resulting images maintain consistency with real CBV maps in terms of structural hierarchy, boundary continuity, and local texture features. The TA-GAN's structural design and training process ensure stable mapping capabilities under non-enhanced MRI input conditions, enabling the model to extract latent features related to perfusion patterns from input images, achieving accurate reconstruction of cerebral blood volume distribution. The model can be integrated into MRI mainframes, image post-processing systems, or standalone AI workstations to achieve automated inference and image output.

[0054] Example 2 Based on the tumor perception generative adversarial network method described in Example 1, this example proposes a preferred implementation method to further improve the robustness, data adaptability, and biological interpretability of the model under different clinical conditions.

[0055] 1. Sample augmentation methods with multi-center and population expansion To improve the model's generalization performance, this embodiment expands the original single-center training data into a multi-center dataset, including non-enhanced MRI images from different equipment manufacturers, imaging protocols, and scanning parameters. Specifically, a unified preprocessing standard is used to normalize the data from different centers in terms of spatial resolution, signal dynamic range, and grayscale distribution, ensuring consistency in numerical scale. During the training phase, samples from different sources are alternately input into the network, and a batch randomization strategy is used to prevent the model from overfitting to features from a single center.

[0056] Furthermore, this embodiment includes cases of primary brain tumors in children. Because the brain tissue structure and imaging signal characteristics differ between children and adults, a stratified sampling strategy is employed during model training. This ensures that children's samples constitute a fixed proportion of the total population in each training cycle, allowing the network to simultaneously learn the image feature distributions of different age groups. If children's samples are still insufficient, a transfer learning strategy can be used. This involves using a model trained on adult samples as initial parameters, followed by fine-tuning with children's samples, thereby achieving cross-population adaptation. Through these steps, the model's stability and generalization ability across multiple institutions, devices, and age groups are significantly improved.

[0057] Specifically, to achieve numerical consistency of multi-center data, this embodiment performs a unified preprocessing procedure on all sources of non-enhanced MRI image data, including spatial resolution normalization, grayscale cropping, and grayscale normalization.

[0058] All input images Resampled to a uniform spatial resolution, such as target height and target width To ensure spatial alignment at the voxel scale: ; Resampled images Perform grayscale cropping to remove extreme values ​​(such as noise or artifacts) from the data. The cropping is based on a set percentile threshold. and ,in and It can be determined by statistically analyzing the grayscale distribution of all samples.

[0059] Cropped image The calculation formula is as follows: ; Finally, the cropped image is normalized to grayscale, mapping its signal dynamic range to the [0, 1] interval to eliminate the influence of scanning parameters or device differences on grayscale distribution. The normalization formula adopts a Min-Max scaling form: ;in and It also serves as the minimum / maximum value for both the cropping threshold and scaling.

[0060] During the model training phase, to improve the model's generalization ability across different age groups (adults and children), this embodiment employs a stratified sampling strategy. In each training batch, the system uses a fixed proportion of pre-set children's samples. To calculate the number of children to be sampled. .

[0061] Subsequently from the children subset Select One sample, and select from the remaining adult samples. Each sample is used to form a training batch.

[0062] This strategy ensures that the proportion of children's samples in the overall training remains stable and fixed, preventing the dilution of characteristics from small sample groups. Furthermore, by employing a batch randomization strategy, different centers... The source samples are randomly mixed and input into the network to prevent the model from overfitting to single-center-specific imaging features.

[0063] If the sample size of pediatric primary brain tumor cases is insufficient, a transfer learning strategy can be employed to achieve cross-population adaptation, thereby fully utilizing the common imaging features contained in the adult sample data. This strategy first pre-trains the model parameters $\theta$ on the adult dataset (to obtain...). Then use the children's dataset Fine-tune the model.

[0064] The fine-tuning process uses a small learning rate. To update the model parameters and enable knowledge transfer from adults to children, the parameter update steps are as follows: ;in For the model in children's samples The loss function calculated above gradient, These are the fine-tuned model parameters. Through this step, the model can effectively adapt to the differences in children's brain tissue structure and imaging signals.

[0065] 1. Robust Optimization Strategy Based on Artifact Awareness To address the vulnerability of diffusion-weighted imagery (DWI-ADC) to motion artifacts and magnetic susceptibility signals, this embodiment introduces an artifact detection mechanism during model training to enhance robustness. Specifically, during the training data preparation phase, a random perturbation algorithm is used to simulate common artifact types in a subset of the input images, including mild motion blur, gradient nonlinear signal shift, and random noise interference. Then, the original image and the image containing the artifacts are mixed and input into the generator. During training, the model learns to distinguish between artifacts and real tissue signals through convolutional feature extraction, thereby acquiring the ability to automatically suppress artifact features.

[0066] In terms of model structure, to enhance the generator's noise resistance, a feature normalization layer is added during the encoding stage to suppress local abnormal highlighting or signal drift regions, making the feature extraction process more stable. During training, the discriminator's discrimination samples simultaneously include the synthesized results of artifact input and the real image, enabling the generator to learn to maintain image structural consistency and detail integrity even under artifact interference conditions through multiple iterations. This method does not require modification of the network backbone structure; it achieves adaptive learning of artifacts solely through input perturbations and feature constraints, thereby significantly improving the model's stable performance in practical applications facing different scan qualities.

[0067] Specifically, to enhance the model's robustness to common artifacts in actual scanning (such as motion blur and nonlinear magnetic field gradients), this embodiment introduces an artifact detection mechanism during the training data preparation phase. This mechanism uses a random perturbation algorithm to detect artifacts in the original image. Simulate various artifact types and generate artifact samples. .

[0068] Generate artifact samples The generation is the cumulative result of applying the following perturbations in sequence: Motion blurring, achieved through random motion convolution kernels For the original image Perform convolution operations (simulating mild motion artifacts): ;in, It is a randomly defined motion convolution kernel.

[0069] Gradient nonlinear migration for images with motion blur Applying spatially dependent nonlinear signal offset Simulate signal drift caused by non-uniform magnetic fields: ; Random noise in images with gradient offset The superposition follows a Gaussian distribution random noise Simulate thermal noise interference, and finally obtain : .

[0070] During the model training phase, the original images Compared with the artifact images generated above Randomly mixed to form the final input batch. This mixed input strategy ensures that the model is exposed to data with artifacts in each training epoch, thereby improving generalization ability. The definition is as follows: ;in The probability of introducing artifacts is preset to a fixed value.

[0071] (3) Validation method for the correspondence between model output and histological angiogenesis This embodiment proposes a verification process based on imaging-pathology comparisons to validate the biological rationale of the generated images. The specific implementation steps are as follows: After model training and inference are completed, a sample set with postoperative histological analysis data is selected to perform spatial registration and region labeling on the synthetic cerebral blood volume map (CBVsyn) output by the model. By delineating high-perfusion areas in the tumor lesion region, quantitative parameters such as the average signal intensity of the region are extracted; at the same time, pathological section results of the corresponding tissue samples are collected, such as microvessel density (MVD) and angiogenesis marker expression levels.

[0072] During validation, the distribution of high-perfusion regions in the generated images was spatially compared with pathological angiogenesis regions to assess their consistency in location and extent. Statistical analysis methods (such as consistency ratio and interval matching degree) were used to quantify this correspondence. If the high-perfusion regions output by the model correspond well with the actual tissue areas with high vascular density, it indicates that the CBVsyn images generated by the model have interpretability and credibility at the biological level. This validation method is independent of the network training process and is used to verify the actual clinical relevance of the model's prediction results, providing a basis for the reliability of the model in subsequent clinical applications.

[0073] This embodiment aims to verify the synthetic cerebral blood volume map output by the generator through imaging-pathological comparison. The biological plausibility of the characterized high-perfusion area with the actual histological angiogenesis area.

[0074] Specifically, regarding the model output After spatial registration of the images to a unified anatomical coordinate system, the tumor region is first... Delineate high-irrigation areas Its definition Signal strength exceeds preset threshold The set of all voxels: ; Based on the extracted high-perfusion areas Calculate the following quantitative parameters: (1) High-perfusion average signal : Average within the region Signal intensity, used to characterize the average blood vessel density in the region: ;in The number of voxels in the high-perfusion region.

[0075] (2) High infusion volume High-irrigation areas The total volume of space, i.e. Number of voxels included: .

[0076] Will Spatial comparison was performed with the pathological section results of the corresponding tissue samples. First, a mask of the pathological angiogenesis region was constructed based on pathological analysis indicators (such as microvessel density, MVD). And register it to the image space.

[0077] High-perfusion areas predicted using the Dice coefficient quantification model With histological angiogenesis region Consistency in location and extent serves as an indicator of regional matching: ;in This indicates the number of voxels that overlap between two regions. and Each represents the number of voxels in its respective region.

[0078] If the calculated Dice coefficient reaches the preset threshold, it indicates that the generated model output... The high-perfusion areas indicated by the imaging corresponded well with the actual angiogenesis areas in histopathology. This result can serve as a basis for verifying the biological rationality and clinical relevance of the model's prediction results.

[0079] (4) Multimodal functional image extended input method Building upon the existing T2WI and DWI-ADC sequences as inputs, this embodiment proposes introducing advanced functional magnetic resonance imaging (fMRI) sequences to enrich the model's input feature space. Specifically, the pseudo-diffusion coefficient and perfusion fraction extracted from inverse diffusion-weighted imaging (IVIM) sequences are used as new input channels, forming a multimodal input group together with T2WI and DWI-ADC images. During the input phase, unified voxel resampling and intensity normalization ensure consistency in spatial location and signal range across the multimodal images. The generator's input layer integrates features from various sequences through channel stitching, while the encoder simultaneously extracts multimodal texture, diffusion, and microcirculation features during convolution, achieving feature-level fusion.

[0080] During the decoding stage, the generator dynamically adjusts the weights of different modal channels based on the style vector output by the mapping module, enabling the model to automatically balance the contributions of different input sequences to the final image features when generating cerebral blood volume maps. In this way, the model can capture information related to microcirculation perfusion from IVIM sequences, thereby further improving the accuracy and detail of blood flow distribution in synthetic images without contrast agents.

[0081] Specifically, to enrich the model input feature space, this embodiment introduces the pseudo-diffusion coefficient extracted from the inverse diffusion-weighted imaging (IVIM) sequence, based on the T2WI and ADC sequences. and infusion fraction .

[0082] These four modalities of imagery constitute the multimodal input group, specifically including: , pseudo-diffusion coefficient and infusion fraction .

[0083] Before the input generator, all modes undergo uniform voxel resampling and intensity normalization (represented as...). This ensures consistency in spatial location and signal range. The generator's input layer integrates the features of each sequence through channel splicing to form a multi-channel input. : ;in This indicates that the input tensor is concatenated along the channel dimension. have Each channel has a unified spatial dimension. .

[0084] During feature extraction by the encoder, the network simultaneously processes multimodal features from four channels through convolutional operations, achieving feature-level fusion. The encoder operates at different levels... Extracted feature map It not only includes texture and diffusion information, but also microcirculation features, which are further fused to achieve information complementarity: ;in These are the features extracted at the current level; Operations (such as channel concatenation, element-wise addition, or attention mechanisms) will By fusing features from higher or other modalities, feature information from different sources and different levels of abstraction can be aggregated, thereby enhancing the ability to characterize the complex biological properties of tumor regions.

[0085] Example 3 This embodiment provides a medical image processing system based on Tumor Perception Generative Adversarial Network (TA-GAN) and its application. The system can be integrated into magnetic resonance imaging equipment (MRI main unit), hospital image archiving and communication system (PACS), or a standalone artificial intelligence image analysis workstation. Combining the construction and prediction method for the non-enhanced sequence synthesis model of primary brain tumors described in Embodiment 1, this system can automatically generate cerebral blood volume maps (CBVsyn) from conventional non-enhanced MRI sequences without the use of gadolinium-based contrast agents. This data can then be further used for image-assisted diagnosis, grading, and molecular subtype prediction of primary intracranial tumors (such as gliomas).

[0086] The overall system architecture includes: The system comprises an input acquisition unit, a data preprocessing unit, a TA-GAN intelligent synthesis unit, a result analysis unit, and a visualization output unit. These units work collaboratively to complete the entire process of image input, model inference, feature analysis, and result output.

[0087] The system's hardware platform includes a central processing unit (CPU), a graphics processing unit (GPU), a storage module, and a communication interface module. The system can connect to the MRI host's data acquisition server via a high-speed data bus, and can also retrieve non-contrast MRI image sequences from the PACS system via the DICOM protocol. The system software platform includes an operating system, an image reading module, a model inference engine, an image post-processing module, and a user interface module. The model inference engine integrates the aforementioned TA-GAN network architecture, supporting multi-threaded parallel computing and GPU acceleration to achieve efficient image synthesis and feature extraction.

[0088] The system's input acquisition unit receives non-contrast MRI image sequences acquired by the magnetic resonance scanning equipment. Input data includes at least T2-weighted imaging (T2WI) sequences and diffusion-weighted imaging (DWI) sequences along with their apparent diffusion coefficient maps (ADCs). This unit can automatically identify DICOM files of different modalities according to clinical scanning protocols and import them into the system's working directory. To ensure data consistency under different scanning conditions, the system automatically invokes the data preprocessing unit to perform standardized operations during the input phase.

[0089] The data preprocessing unit includes a resampling module, a registration module, and a normalization module. The resampling module performs voxel interpolation on DWI and ADC images based on the spatial resolution of the T2WI sequence, ensuring consistency of multimodal images in three-dimensional space. The registration module employs a rigid registration algorithm to precisely align each image sequence in anatomical space. The normalization module uses an intensity normalization method to linearly scale the image signal range to the [-1,1] interval, ensuring a uniform dynamic range of the input data. After the above preprocessing, spatially aligned and numerically consistent input tensors are output, providing standardized input for model inference.

[0090] The intelligent synthesis unit is the core component of the system, which contains a generator, a discriminator, and a tumor sensing module (TAM) to complete the synthesis and structural enhancement of non-enhanced images into brain blood volume maps.

[0091] Generator module The generator consists of an encoder, a mapping module, and a decoder. The encoder contains seven convolutional blocks, each consisting of two 3×3 convolutional layers: the first convolutional layer maintains spatial resolution, and the second convolutional layer performs a 2x downsampling with stride convolution, thus obtaining multi-level feature embeddings. The mapping module consists of eight fully connected layers that map latent vectors to style vectors, used to control the adjustment of convolutional kernel weights during the decoding stage. The decoder uses bilinear interpolation for progressive upsampling and performs convolution operations after each upsampling layer to restore spatial details. The parameters of the decoder's convolutional layers are dynamically adjusted by the style vectors output by the mapping module during each decoding iteration, enabling fine-grained control over image texture, perfusion features, and local contrast.

[0092] Discriminator module The discriminator employs a multi-scale convolutional structure, comprising six hierarchical 3×3 convolutional blocks, to distinguish generated images from real CBV images through multi-layer feature extraction. The final features are processed by a fully connected layer to output a true / false signal. This module, together with the generator, forms a minimization-maximization game structure: the generator minimizes the overall loss function to improve synthesis quality, while the discriminator maximizes its discrimination accuracy, thus achieving balanced training through alternating optimization.

[0093] Tumor Sensing Module (TAM) The tumor perception module employs an encoder-decoder symmetric structure. Both the encoder and decoder contain five convolutional layers with a double sampling rate, and the terminal layers include two 1×1 convolutional layers and a channel attention mechanism. The TAM module receives joint features from the apparent diffusion coefficient map and the generator's output cerebral blood volume image to generate full-perfusion and high-perfusion masks. During model training, the TAM calculates a loss function that includes pixel difference constraints and structural consistency constraints, updating the generator weights through backpropagation to enhance the generator's texture representation in high-perfusion regions. During inference, the TAM module can directly output perfusion masks for clinical visualization or subsequent quantitative analysis.

[0094] Model training and inference process Before system deployment, the TA-GAN model had completed its training phase. The training process employed an adversarial game structure, with the generator and discriminator alternately optimizing the model, and the TAM module participating in generator parameter updates. After training, the model weights were stored in the system's model library. In clinical application, the system directly loaded the trained model for forward inference. After the input underwent feature extraction by the encoder, style vector generation by the mapping module, and image reconstruction by the decoder, the system automatically output a synthesized cerebral blood volume image (CBVsyn) and simultaneously generated perfusion mask results.

[0095] The results analysis unit processes the CBVsyn images and masking results generated by the model, outputting image display, quantitative parameter calculation, and region annotation functions. The system can simultaneously display the original T2WI, DWI-ADC, and synthesized CBV images in a 3D interface. Users can manually or automatically select tumor regions on the interface to calculate quantitative indicators such as mean cerebral blood volume, standard deviation, and the proportion of high-perfusion areas within the region. The system supports outputting results as color overlay images, allowing real-time comparison between the synthesized images and the original structural images, assisting physicians in determining tumor perfusion distribution characteristics.

[0096] In AI workstation applications, the system can transmit inference results back to the PACS system in DICOM format, allowing radiologists to view them directly in their regular image reading systems. For MRI host-integrated versions, the system can automatically initiate the image inference process after scanning, generating synthetic images within minutes, achieving seamless clinical integration.

[0097] This system can not only generate synthetic perfusion images that are highly consistent with real CBV maps, but also play a role in the multi-level diagnosis and prediction of primary intracranial tumors.

[0098] At the diagnostic level, the system visually displays the perfusion distribution in the tumor area through CBVsyn imaging, helping doctors identify tumor angiogenesis characteristics and distinguish between high-perfusion malignant tumors and low-perfusion benign lesions.

[0099] At the grading level, the proportion of high-perfusion areas and the signal intensity within the regions extracted by the system can be used as perfusion quantitative indicators to provide auxiliary judgment for distinguishing between low-grade and high-grade gliomas.

[0100] At the molecular subtype prediction level, the image features and tumor perfusion heterogeneity output by the system can be used to infer the status of molecular markers, such as isocitrate dehydrogenase (IDH1) mutation or O6-methylguanine-DNA methyltransferase (MGMT) methylation status. Combined with the image feature analysis module, radiomics-assisted prediction can be achieved.

[0101] The system eliminates the need for contrast agent injections, avoiding the risks of renal systemic fibrosis and gadolinium deposition, while significantly shortening the scanning process and reducing costs. The generated images exhibit high signal stability and repeatability, providing quantitative references for follow-up or radiotherapy response assessment. Through integration with PACS or clinical information systems, this system can serve as part of a hospital's intelligent image analysis module, achieving a fully automated closed-loop workflow from data acquisition and model inference to result feedback.

[0102] Example 4 refer to Figure 3 Based on Example 1, this example proposes a terminal device for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors. The terminal device 200 includes at least one memory 210, at least one processor 220, and a bus 230 connecting different platform systems.

[0103] The memory 210 may include a readable medium in the form of volatile memory, such as RAM 211 and / or cache memory 212, and may further include ROM 213.

[0104] The memory 210 also stores a computer program that can be executed by the processor 220, causing the processor 220 to perform the application of any of the above-described methods for constructing and predicting non-enhanced sequence synthesis models of primary brain tumors in this application embodiment. The specific implementation method and the achieved technical effects are consistent with those described in the above-described application embodiments, and some details will not be repeated here. The memory 210 may also include a program / utility 214 having a set (at least one) of program modules 215. Such program modules include, but are not limited to, an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0105] Accordingly, processor 220 can execute the aforementioned computer program, as well as executable program / utility 214.

[0106] Bus 230 can represent one or more of several types of bus structures, including a memory bus or memory controller, peripheral bus, graphics acceleration port, processor, or a local bus using any of the multiple bus structures.

[0107] Terminal device 200 can also communicate with one or more external devices 240, such as keyboards, pointing devices, Bluetooth devices, etc., and with one or more devices capable of interacting with it, and / or with any device that enables it to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication can be performed via I / O interface 250. Furthermore, terminal device 200 can communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) via network adapter 260. Network adapter 260 can communicate with other modules of terminal device 200 via bus 230. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with terminal device 200, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage platforms.

[0108] Example 5 This embodiment proposes a readable storage medium for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors. The computer-readable storage medium stores instructions that, when executed by a processor, implement any of the above-mentioned methods for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors. The specific implementation method is consistent with the implementation method and the technical effects achieved in the above-mentioned application embodiments, and some details will not be repeated.

[0109] Figure 4The present embodiment illustrates a program product 300 for implementing the above-described applications. This product may employ a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the program product 300 of the present invention is not limited thereto. In this embodiment, the readable storage medium may be any tangible medium containing or storing a program that may be used by or in conjunction with an instruction execution system, apparatus, or device. The program product 300 may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disc read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0110] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium, capable of sending, propagating, or transmitting a program for use by or in conjunction with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof. Program code for performing operations of the present invention may be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java, C++, etc., and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on a user computing device, partially on a user device, as a standalone software package, partially on a user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing devices can be connected to user computing devices via any type of network, including local area networks (LANs) or wide area networks (WANs), or they can be connected to external computing devices (e.g., via the Internet using an Internet service provider).

[0111] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for constructing and predicting a synthetic model of non-enhancing sequences of a primary brain tumor, characterized by, Includes the following steps: S1. Receive non-enhanced magnetic resonance image data acquired and output by a magnetic resonance imaging device, wherein the image data includes a T2-weighted sequence, a diffusion-weighted sequence and its apparent diffusion coefficient map; S2. Perform rigid registration, resampling, intensity cropping and normalization on the image data to obtain spatially aligned and numerically consistent input data. S3. Input the input data into the tumor perception generative adversarial network (TA-GAN), wherein the TA-GAN includes a generator, a discriminator, and a tumor perception module; The generator includes an encoder, a mapping module, and a decoder. The encoder extracts image features, the mapping module generates style vectors and adjusts the convolution weights of the decoder to control image details, and the generator outputs a synthesized image of brain blood volume. The discriminator uses a multi-scale convolutional network to distinguish between the brain blood volume image output by the generator and the real brain blood volume map; The tumor perception module receives the combined features of the apparent diffusion coefficient map and the cerebral blood volume image output by the generator, generates a full perfusion mask and a high perfusion mask, and updates the generator weights through backpropagation to enhance the feature representation of high perfusion structural regions. S4. A synthetic cerebral blood volume map with high consistency with the real cerebral blood volume map was obtained through adversarial training.

2. The method of constructing and predicting a primary brain tumor non-enhancing sequence synthetic model according to claim 1, wherein, The encoder of the generator includes seven convolutional blocks, each of which includes two 3×3 convolutional layers. The first convolutional layer maintains spatial resolution, and the second convolutional layer performs double downsampling with stride convolution to obtain multi-level feature embeddings.

3. The method of constructing and predicting a primary brain tumor non-enhancing sequence synthetic model of claim 1, wherein, The mapping module includes eight fully connected layers for converting latent vectors into style vectors, which control the weight adjustment of convolutional kernels during the decoding stage.

4. The method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors according to claim 1, characterized in that, The decoder uses bilinear interpolation for double upsampling and combines convolution operations after upsampling at each layer to restore spatial details. The parameters of the convolutional layers are dynamically adjusted by the style vector output by the mapping module.

5. The method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors according to claim 1, characterized in that, The tumor sensing module includes an encoder and a decoder structure. Both the encoder and the decoder have five convolutional layers with a double sampling rate, and the end contains two 1×1 convolutional layers and a channel attention mechanism.

6. The method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors according to claim 5, characterized in that, The loss function of the tumor perception module includes pixel difference constraints and structural consistency constraints. The gradient is calculated through backpropagation and the generator parameters are updated so that the generator focuses on enhancing the texture features of high-perfusion areas during training.

7. The method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors according to claim 1, characterized in that, The discriminator consists of six hierarchical 3×3 convolutional blocks, which extract multi-scale features and output true / false classification signals through a fully connected layer.

8. The method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors according to claim 1, characterized in that, The training process of TA-GAN adopts a mini-maximum game structure. The generator minimizes the comprehensive loss function, the discriminator maximizes the discrimination accuracy of distinguishing between real and fake samples, and TAM participates in the optimization of generator parameters through its back gradient.

9. The method for constructing and predicting a non-enhanced sequence synthesis model of primary brain tumors according to claim 1, characterized in that, The input features of the tumor perception module are obtained by splicing the structural features of the apparent diffusion coefficient map with the perfusion features of the generated image. After channel-weighted fusion, the features are input into its convolutional network for the extraction of regional features of high-perfusion structures.

Images