A non-contrast-enhanced multi-modal MRI image-based acute ischemic stroke perfusion image generation system

Abstract

The application discloses an acute ischemic stroke perfusion image generation system based on non-contrast enhancement multi-modal images, which comprises four parts, an information acquisition module, a pretreatment module, an image generation module and a model evaluation module; the information acquisition module is used for collecting data information of patients to be diagnosed; the pretreatment module is used for standardizing and pretreating multi-modal images, that is, all images are processed through a standardized automatic pretreatment process; the image generation module is used for generating clinically available perfusion Tmax images by using a constructed adversarial generation network; and the model evaluation module is used for completing accurate segmentation and quantitative evaluation of key low perfusion areas by using a pre-trained segmentation network. The system disclosed by the application effectively expands the clinical application range of perfusion Tmax images, reduces medical costs, and reduces model deviation, so that the system can maintain stable performance on different devices and different population data.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, specifically to a system for generating perfusion images of acute ischemic stroke based on non-contrast enhanced multimodal magnetic resonance imaging (MRI). Background Technology

[0002] Acute ischemic stroke (AIS) is a common critical illness in clinical practice. Rapid and accurate assessment of the ischemic state of the patient's brain tissue is crucial for treatment planning and improving neurological prognosis. Currently, perfusion Tmax, as the most important parameter reflecting the delay in cerebral blood flow perfusion, is closely related to the pathophysiological process of AIS, but it must be obtained through perfusion imaging. MRI has the advantages of multi-sequence and multi-parameter capabilities, and its sensitivity in detecting the infarct core area of ​​AIS is significantly better than that of computed tomography (CT). Therefore, perfusion-weighted imaging (PWI) is currently the best imaging method for evaluating the ischemic state of brain tissue in AIS patients. Combined with ADC maps, the infarct core area (ADC ≤ 620 mm) can be effectively assessed. 2 Quantitative analysis of Tmax images (Tmax > 6s) and key hypoperfusion areas (Tmax > 6s) was performed to identify and quantify ischemic penumbra brain tissue at the edge of irreversible hypoperfusion. Furthermore, different threshold analyses of Tmax images (Tmax > 4s / 6s / 8s / 10s) were closely related to collateral circulation status, hemorrhagic transformation risk, and neurological prognosis in AIS patients.

[0003] Although MRI perfusion imaging is currently the best method for assessing the ischemic state of brain tissue in AIS, it still has the following limitations in actual clinical applications: contrast agent dependence: it requires the injection of gadolinium contrast agent, which carries the risk of allergies and is not suitable for patients with renal insufficiency; high equipment requirements: it is difficult to popularize in primary hospitals; long imaging time: some AIS patients cannot tolerate it or have large motion artifacts.

[0004] Recently, artificial intelligence technology has gradually penetrated into the imaging diagnosis and treatment decision-making process of acute ischemic stroke (AIS), showing significant potential, especially in medical image analysis and synthesis. Deep learning models, represented by Generative Adversarial Networks (GANs), have made progress in various medical image generation tasks. Previous studies have confirmed that DWI sequences simultaneously contain information on the diffusion and movement of water molecules between cells and perfusion information of peripheral blood vessels. Therefore, this patent discloses a system capable of generating clinically usable brain perfusion Tmax images based on non-contrast-enhanced conventional DWI sequences, ADC maps, and infarct core masks. Summary of the Invention

[0005] The purpose of this invention is to provide a perfusion image generation system for acute ischemic stroke based on non-contrast enhanced multimodal MRI images, in order to solve the technical problems existing in the prior art.

[0006] This invention proposes an acute ischemic stroke perfusion image generation system based on non-contrast enhanced multimodal MRI images, comprising four parts: an information acquisition module, a preprocessing module, an image generation module, and a model evaluation module. The information acquisition module is used to collect data information of the patient to be diagnosed, including head MRI-DWI, ADC data, real perfusion Tmax image and key low perfusion area mask (Tmax>6s). The preprocessing module is used to perform standardized preprocessing on multimodal images. That is, all images are processed through a standardized automated preprocessing workflow. The image generation module is used to generate clinically usable perfusion Tmax images using a constructed generative adversarial network; The image generation module is specifically used for: (1) Construct a ternary improved generative adversarial network framework that integrates a generator, a discriminator, and a segmentation network; (2) During the training process, the image generation module optimizes the model performance through a triple loss function; The model evaluation module is used by the pre-trained segmentation network to accurately segment and quantitatively evaluate low-perfusion regions.

[0007] Preferably, the information acquisition module is specifically used for: (1) Acquisition included DWI imaging, ADC maps, and dynamic magnetic susceptibility contrast PWI imaging based on gadolinium contrast agent; (2) The PWI raw data were automatically post-processed using the clinically validated RAPID software, and the peak time of whole brain voxels was calculated by the deconvolution algorithm to generate the real Tmax map. (3) Data screening is based on screening criteria, which are divided into data with inclusion criteria and data with exclusion criteria.

[0008] Preferably, the preprocessing module is specifically used for: (1) extraction and fusion optimization of the infarct core region; (2) extraction and spatial registration of brain tissue regions; and (3) preprocessing of negative data.

[0009] The preferred method for extracting and optimizing the infarct core region specifically includes: Based on its integrated VB-Net segmentation model, the high-signal region of acute infarction is automatically segmented from the original DWI image, denoted as DWI-Core; Based on clinical standard ADC thresholds, diffusion-restricted regions are automatically segmented from the ADC map and denoted as Rapid-Core. Perform a logical OR operation between DWI-Core and Rapid-Core to generate a fused infarct core mask.

[0010] Preferably, the construction of the ternary improved adversarial generative network integrating the generator, discriminator, and segmentation network specifically includes: The generator uses ResNet-9blocks as the backbone network and constructs an encoder-decoder structure, which includes a diffusion-weighted imaging map, an apparent diffusion coefficient map, and an optimized infarct core region mask. The discriminator adopts the PatchGAN architecture, which does not make a single judgment on the entire image, but performs real and fake identification on the local receptive field of the image through convolution operation and outputs a discrimination matrix; The segmentation network is based on the VNet architecture, which specializes in 3D medical images, and was pre-trained on the processed real Tmax base map and its corresponding real mask.

[0011] Preferably, the model evaluation module is specifically used for: (1) segmentation of low-perfusion areas; (2) Quantitative evaluation of model performance, specifically including: First, the baseline clinical and imaging characteristics of the training set, validation set, internal test set, and external validation set were statistically described and compared to assess the balance of data distribution. Second, to objectively evaluate the pixel-level and structural similarity between the Tmax image synthesized by the generator and the real Tmax image; Third, calculate the Spearman rank correlation coefficient and intraclass correlation coefficient (ICC) of the low perfusion volume measured by the synthetic mask and the real mask, and plot the Bland-Altman plot. Fourth, multiple thresholds are applied to the synthetic Tmax and the real Tmax plots, and pairwise consistency analysis is performed on the volume under each threshold. The ICC value is calculated and the Bland-Altman plot is plotted.

[0012] Preferably, the objective evaluation of the similarity between the Tmax image synthesized by the generator and the real Tmax image at the pixel level and in terms of structure specifically adopts the following three image quality evaluation indicators: A. Mean Square Error: Measures the average squared value of the difference in pixel intensity between images. The calculation formula is: MSE=(1 / N)*Σ(Syn_i-Real_i)^2 Where N is the total number of pixels, and the lower the MSE value, the smaller the pixel-level error; B. Peak Signal-to-Noise Ratio (PSNR): Calculated based on MSE, it reflects the ratio of signal to noise in an image. The calculation formula is: PSNR = 10 * log10(MAX^2 / MSE) Where MAX is the maximum possible value of the image pixels, and the higher the PSNR value, the better the image quality; C. Structural Similarity Index: This index compares the similarity of images in terms of brightness, contrast, and structure. Its value ranges from -1 to 1. The closer it is to 1, the higher the structural similarity between the two images in visual perception.

[0013] Preferably, in calculating the Spearman rank correlation coefficient between the synthetic mask and the real mask for the low perfusion volume, the within-group correlation coefficient of a two-way random effects model is used to quantitatively assess the consistency strength of the two measurement results.

[0014] The method of the present invention has the following advantages: (1) By incorporating multicenter data and negative samples, the model bias was effectively reduced, enabling it to maintain stable performance on different devices and different population data.

[0015] (2) It breaks through the contrast agent dependence limitation of traditional perfusion imaging, is suitable for patients with renal insufficiency and allergies, and provides a new alternative for AIS patients who cannot tolerate perfusion examination, thus expanding the clinical application scope of perfusion Tmax images.

[0016] (3) It significantly reduces medical costs and simplifies the imaging process, which is beneficial for emergency scenarios and primary hospitals with limited imaging conditions. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the image system structure of the present invention; Figure 2 This is a schematic diagram of the multimodal image preprocessing process; Figure 3 This is a schematic diagram of a deep learning algorithm architecture; Figure 4 Scatter plot and Bland-Altman analysis plot showing the correlation between the volume consistency of the key low-perfusion mask in the synthetic (Syn Tmax-mask) and real (Real Tmax-mask) masks; Figure 5 Bland-Altman analysis plots of the synthesized and true Tmax volume estimates at different latency thresholds in the training set; Figure 6 Bland-Altman analysis plots of synthesized and real Tmax volume estimates under different delay thresholds in the internal test set; Figure 7 The image shows Bland-Altman analysis plots of the synthesized and real Tmax volume estimates under different delay thresholds in an external test set. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to specific embodiments. However, those skilled in the art should understand that the embodiments described below are only for illustrating the present invention and should not be regarded as limiting the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention provides a system for generating perfusion images of acute ischemic stroke based on non-contrast enhanced multimodal MRI images.

[0020] This system proposes for the first time a multimodal input strategy that integrates DWI images, ADC maps, and infarct core masks, providing richer structural and functional information for the generative model. An attention mechanism is introduced into the discriminator, enabling the model to focus on feature discrimination of key regions at the time to peak perfusion function (Tmax), enhancing the clinical plausibility of the generated images. A multi-task optimization objective is designed, integrating image similarity loss, attention adversarial generative network (GAN) loss, and mask similarity loss, synergistically guiding the model to achieve comprehensive improvements in pixel-level fidelity, region discrimination ability, and semantic consistency. A ternary adversarial network architecture comprising a generator, discriminator, and a pre-trained segmentation network (Seg-Net) is constructed, using Seg-Net to achieve alignment constraints between generated and real images at the semantic mask level.

[0021] The system can rapidly and accurately generate clinically usable Tmax images and masks of key hypoperfusion areas (Tmax>6s) without contrast enhancement. This technology enables precise assessment of tissue ischemia in AIS patients in clinical settings with limited imaging conditions or contraindications to contrast agents, providing support for the diagnosis, assessment, and treatment decisions of AIS.

[0022] This invention discloses a perfusion image generation system for acute ischemic stroke based on non-contrast enhanced multimodal imaging, comprising four parts: an information acquisition module, a preprocessing module, an image generation module, and a model evaluation module, such as... Figure 1 As shown.

[0023] The information acquisition module is used to collect data information of the patient to be diagnosed, including head MRI-DWI, ADC data, real perfusion Tmax images and key low perfusion area masks.

[0024] This invention is based on a multicenter retrospective diagnostic study. Data were obtained from Shanghai Fourth People's Hospital (the main research center) and five other independent medical centers. All data were acquired using 1.5T or 3.0T MRI systems. The data were divided into a main dataset and external validation cohort data.

[0025] The main dataset consists of 967 cases from Shanghai Fourth People's Hospital between January 2018 and December 2024, which were randomly stratified in an 8:1:1 ratio into a training set (803 cases), a validation set (82 cases), and an internal test set (82 cases). The training set was used for deep learning and iterative optimization of model parameters, the validation set was used for hyperparameter tuning, monitoring overfitting, and implementing early stopping strategies, and the internal test set was used for preliminary independent evaluation after model training.

[0026] The external validation cohort consisted of 207 cases from four independent centers—Henan People's Hospital, Tianjin Huanhu Hospital, the Seventh Affiliated Hospital of Guangxi Medical University, and Zhejiang Hospital—during the period from March 2022 to January 2025. This cohort data was not involved in any stage of model development and was specifically used for the final, most rigorous external generalization test to evaluate the model's universality across different devices, scanning protocols, and patient populations.

[0027] The information acquisition module is specifically used for: (1) Acquisition included DWI imaging, ADC maps, and dynamic magnetic susceptibility contrast PWI imaging based on gadolinium contrast agent.

[0028] (2) The clinically validated RAPID software was used to automatically post-process the raw PWI data. The peak time of whole brain voxels was calculated by the deconvolution algorithm to generate the real Tmax map (Real-Tmax). At the same time, the threshold (Tmax>6 seconds) and small cluster filtering (<3mL) algorithms of the RAPID software were automatically applied to generate the corresponding real low perfusion mask (Real-Mask). Both served as the absolute reference standard for model training and evaluation.

[0029] (3) Data screening was conducted based on screening criteria, divided into data for inclusion criteria and data for exclusion criteria. Case screening criteria ensured data quality. Data for inclusion criteria included patients clinically suspected of having acute ischemic stroke who underwent brain MRI examinations within 24 hours of symptom onset, including DWI, ADC, and PWI sequences. Data for exclusion criteria included data where the quality of DWI, ADC, or PWI images did not meet the analysis requirements, or where Rapid processing results did not meet the analysis needs.

[0030] The preprocessing module is used to perform standardized preprocessing on multimodal images. That is, all images are processed through a standardized automated preprocessing workflow to ensure the quality and spatial consistency of the input data.

[0031] The preprocessing module is specifically used for: (1) Extraction and fusion optimization of the infarct core region DWI Core Extraction: Based on its integrated VB-Net segmentation model, the high-signal region of acute infarction is automatically segmented from the original DWI image and denoted as "DWI-Core".

[0032] ADC core extraction: based on clinical standard ADC threshold (≤620×10⁻⁶). -6 mm 2 The diffusion-restricted region is automatically segmented from the ADC image ( / s), and denoted as "Rapid-Core".

[0033] Core Fusion: A logical OR operation is performed between "DWI-Core" and "Rapid-Core" to generate a fused infarct core mask, complementing the advantages of the two segmentation methods, maximizing the capture of infarct tissue information, and providing more accurate spatial guidance for the generated network.

[0034] (2) Brain tissue region extraction and spatial registration Skull removal and ventricle removal: The ADC images were processed using the Uniskull algorithm (UniSkull is a general segmentation component developed by United Imaging Intelligence, which uses a series of segmentation networks trained with multimodal methods (T1W, ADC, etc.). For details, please refer to the following reference (Lin Teng, Zihao Zhao, Yulin Wang, Jiameng Liu, Feng Shi, Dinggang Shen, "BrainSMM: Lifespan Brain Segmentation Model with Metadata-Driven Prompt Learning", IEEE Transactions on Medical Imaging, 2025.) to remove the skull. Subsequently, the ventricle region was automatically identified and removed to eliminate the interference of cerebrospinal fluid signals on subsequent analysis.

[0035] Generate input mask: Perform strict affine transformation registration between the fused core mask and the processed ADC image to ensure that the two are perfectly aligned in space, generating the final "Merged-Core". For negative cases, this mask is an all-zero matrix.

[0036] Preprocessing and registration operations: The ADC and DWI modal images are resampled into the space of the PWI modality while maintaining the same resolution; the ADC, DWI, and PWI images are truncated and normalized, and the pixel values ​​below 1% and above 99% are truncated and mapped to [-1,1]; rigid registration technology is used to spatially align the ADC and DWI modal images and the fusion core mask "Merged-Core" with the PWI modality.

[0037] (3) Preprocessing of negative data For stroke negative data, the same processing procedures and parameter settings as for positive data are used, with only the infarct core region segmentation result being an empty mask. This ensures data structure consistency, reduces the risk of false positives in the model, and enhances its ability to recognize normal brain tissue patterns.

[0038] The image generation module is used to generate clinically usable perfusion Tmax images using a constructed generative adversarial network (GAN).

[0039] The image generation module is specifically used for: (1) Construct a three-element improved generative adversarial network (GAN) that integrates generator, discriminator and segmentation network (Seg-Net).

[0040] The generator uses ResNet-9blocks as its backbone network to construct an encoder-decoder structure. The input is a stack of multimodal images that have been strictly spatially aligned, specifically including a surface diffusion-weighted imaging map (DWI), an observation diffusion coefficient map (ADC), and an optimized infarct core region mask (Merged-Mask). The network extracts deep features through multi-layer convolutional downsampling (supplementing the specific network structure, parameters of each layer, and extracted features), and then gradually restores spatial details through transposed convolutional upsampling (supplementing the specific network structure and parameters of each layer), finally outputting a high-fidelity synthetic Tmax image (Syn-Tmax).

[0041] The specific network structure of the generator is as follows: The input layer consists of a 7×7×7 convolution (3 padding, 64 channels), a batch normalization layer, and a ReLU activation layer.

[0042] The two downsampling operations are 3×3×3 convolutions with a stride of 2, and the number of channels is doubled each time (64→128→256).

[0043] Each of the nine residual blocks contains two layers of 3×3×3 convolutions, using skip connections to maintain 256 channels.

[0044] The two upsampling operations are 3×3×3 transposed convolutions with a stride of 2, and the number of channels is halved each time (256→128→64).

[0045] The output layer is a 7×7×7 convolutional layer that outputs the target channel, followed by a Tanh activation layer. The final output is a high-fidelity synthetic Tmax image (Syn-Tmax).

[0046] The discriminator adopts the PatchGAN architecture, which does not make a single judgment on the entire image, but performs real and fake identification on the local receptive field of the image through convolution operation, and outputs a discrimination matrix to accurately constrain the authenticity of the generated image in terms of texture details. It is specifically responsible for distinguishing between the real Tmax image from the gold standard (Real-Tmax) and the Tmax image synthesized by the generator (Syn-Tmax).

[0047] The specific network structure of the discriminator is as follows: The input layer consists of a 3D convolutional layer (4×4×4 convolutional kernel, stride 2, padding 1) and a LeakyReLU activation layer.

[0048] The intermediate layers consist of three convolutional blocks, each containing a 3D convolutional layer (4×4×4 convolutional kernel, stride 2, padding 1), a batch normalization layer, and a LeakyReLU activation layer.

[0049] The output layer is a 3D convolutional layer (4×4×4 convolutional kernel, stride 1, padding 1), outputting a single-channel result.

[0050] The segmentation network (Seg-Net) is based on the VNet architecture, which specializes in 3D medical imaging, and was pre-trained on the real Tmax base map (Real-Tmax) processed by RAPID software and its corresponding real mask (Real-Mask, i.e., the region where Tmax>6s). During the main training of the GAN, its network parameters are frozen, acting as a stable "semantic supervisor" responsible for converting the synthesized Tmax image (Syn-Tmax) and the real Tmax image (Real-Tmax) output by the generator into binary masks of low-perfusion regions with smooth, physiologically reasonable boundaries (Syn Tmax-mask and Real Tmax-mask, respectively). This design bypasses the problem of sharp boundaries and noise in the original RAPID mask, providing higher quality and more clinically meaningful supervisory signals for generative adversarial training.

[0051] The network structure of Seg-Net is as follows: The input layer is a 3×3×3 convolution + BN + ReLU.

[0052] There are 3 downsampling layers, each containing 2 3×3×3 residual convolutions, 2×2×2 convolutions (stride 2) + BN + ReLU.

[0053] The intermediate layer consists of three 3×3×3 residual convolutions, a 3D convolution downsampling layer (3×3×3 convolution kernel, stride 2, padding 1), an encoding feature extraction block (double 3×3×3 convolution + PReLU + residual connection), a bottleneck module (double 3×3×3 convolution + PReLU + residual connection), and a 3D transposed convolution upsampling layer (2×2×2 convolution kernel, stride 2) and a decoding feature extraction block (double 3×3×3 convolution + PReLU + residual connection).

[0054] The output layer is a 3D convolutional layer (1×1×1 convolutional kernel, stride 1) that maps the number of channels to the number of target segmentation categories and generates a 3D segmentation probability map through the Softmax activation function.

[0055] (2) During the training process, the image generation module optimizes the model performance through a triple loss function.

[0056] Image similarity loss: L1 loss is used to constrain the pixel-level similarity between the synthesized Tmax image and the real Tmax image.

[0057] Attention-enhanced GAN loss: An attention mechanism is introduced into the discriminator. The binary mask of the lesion is added with 5% random perturbation and then fed into the discriminator as two channels along with the image for training. For negative data, an all-zero image with perturbation is used as the attention channel, so that the discriminator focuses on the perfusion abnormal area and enhances the model's ability to distinguish pathological features.

[0058] Mask similarity loss: L1 loss is used to measure the consistency between the synthesized mask and the real mask. The semantic-level supervision provided by the segmentation network improves the physiological rationality and smoothness of the infusion boundary of the generated image. The segmentation network parameters are fixed during the training phase to stabilize the training process.

[0059] The image size was set to 128×128, the batch size to 32, the learning rate to 0.0004, the optimization frequency ratio of the generator to the discriminator to 1:1, and a balanced ratio of positive and negative samples was maintained in the data loader to ensure stable training convergence and capture the complex mapping relationship between the infarct core and the low-perfusion region.

[0060] The model evaluation module is used by the pre-trained segmentation network (Seg-Net) to complete the accurate segmentation and quantitative evaluation of key low-perfusion regions.

[0061] The model evaluation module is specifically used for: (1) Segmentation of key low-irrigation areas Using a pre-trained Seg-Net with parameter freezing, the synthesized Tmax image (Syn-Tmax) and the real Tmax image (Real-Tmax) output by the generator are processed to generate corresponding binary masks for key low-perfusion regions, namely the synthesized mask (Syn Tmax-mask) and the real mask (Real Tmax-mask).

[0062] (2) Quantitative evaluation of model performance First, the baseline clinical and imaging characteristics of the training set, validation set, internal test set, and external validation set were statistically described and compared to assess the balance of data distribution.

[0063] Continuous variables were described using median and interquartile range. The Mann-Whitney U test (a non-parametric test) was used for comparisons between two independent samples. Categorical variables were described using frequency and percentage. Chi-square tests were used for comparisons between groups; Fisher's exact test was used if the expected count in any cell was less than 10. Significance level: All statistical analyses were two-tailed tests, and a p-value < 0.05 was defined as statistically significant.

[0064] Second, to objectively evaluate the pixel-level and structural similarity between the generator-synthesized Tmax image (Syn-Tmax) and the real Tmax image (Real-Tmax).

[0065] The following three classic image quality evaluation metrics are used: A. Mean Square Error: Measures the average squared value of the difference in pixel intensity between images. The calculation formula is: MSE=(1 / N)*Σ(Syn_i-Real_i)^2 Where N is the total number of pixels. The lower the MSE value, the smaller the pixel-level error.

[0066] B. Peak Signal-to-Noise Ratio (PSNR): Calculated based on MSE, it reflects the ratio of signal to noise (error) in an image. The calculation formula is: PSNR = 10 * log10(MAX^2 / MSE) MAX represents the maximum possible value of the image pixels. A higher PSNR value indicates better image quality.

[0067] C. Structural Similarity Index (SSIM): This index compares image similarity based on brightness, contrast, and structure. Its value ranges from -1 to 1, and the calculation formula is relatively complex, usually implemented using standard library functions. The closer the SSIM value is to 1, the higher the visual structural similarity between the two images. This evaluation is performed in a Python environment.

[0068] Third, calculate the Spearman rank correlation coefficient between the low perfusion volume measured by the synthetic Tmax-mask and the real Tmax-mask.

[0069] The within-class correlation coefficient (ICC, model "absolutely consistent") of a two-way random effects model was used to quantitatively assess the strength of agreement between the two measurements. A Bland-Altman plot was plotted to visually display the distribution of the differences between the two measurement methods relative to the mean, and the 95% agreement threshold was calculated to assess the presence of systematic or proportional bias.

[0070] Fourth, on the synthetic Tmax (Syn-Tmax) and the real Tmax (Real-Tmax) plots, thresholds of Tmax>4s,>6s,>8s, and>10s are applied respectively. Paired consistency analysis is performed on the volume under each threshold, and the ICC value is calculated and the Bland-Altman plot is drawn.

[0071] ICC is calculated using SPSS (Statistical Package for the Social Sciences) (two-way mixing, absolute consistency): Among them, MS R Mean square among subjects, MS E Mean squared residuals, MS C : Inter-rater mean square, k: Number of measurements, n: Number of subjects.

[0072] Bland-Altman diagram as shown Figure 4-7 As shown: The horizontal axis typically represents the average of two measurement results. That is, (Result from Method A + Result from Method B) / 2, which represents the best estimate of the "true value" for each sample (usually assumed to be the average of the gold standard and the predicted value).

[0073] Vertical axis: The difference between the two measurement results. That is, the result of method A minus the result of method B (the direction must be clearly stated, e.g., predicted value - gold standard). It directly shows the degree of inconsistency between the two methods on each sample.

[0074] The three key horizontal reference lines (parallel dashed lines) in the diagram are: Center line: The mean of all differences. It represents the systematic bias between the two methods. If this line is close to 0, it indicates no significant systematic bias; if it deviates significantly from 0 (and passes a statistical test), it indicates that one method systematically overestimates or underestimates the other.

[0075] Upper and lower bounds: Consistency limits. Typically, these are the difference between the means ± 1.96 × the standard deviation of the difference. Assuming the data follows a normal distribution, approximately 95% of the data points will fall between these two lines. This interval defines the "acceptable range" of the difference between the two methods.

[0076] Data point: Each point represents a sample (such as a patient).

[0077] The following analysis will examine the specific experimental process and results: 1. Baseline characteristics and balance analysis of the patient cohort A comprehensive comparison of baseline characteristics of the training set, validation set, internal test set, and external test set was conducted, as shown in Table 1. There were no statistically significant differences between the training set and the internal test set across all baseline variables, indicating that the internal partitioning was balanced. The external validation set showed significantly higher AIS proportions, critical low-perfusion volumes, and volumes meeting all Tmax thresholds than the main development cohort. This heterogeneity is crucial for evaluating the model's generalization ability, demonstrating the challenging nature of the external test set.

[0078] Table 1 IQR, interquartile range 2. Quantitative quality assessment results of the synthesized Tmax image To objectively evaluate the accuracy and reliability of the synthesized Tmax images generated by the method described in this invention, a series of recognized quantitative image quality evaluation indicators were adopted, including mean squared error, peak signal-to-noise ratio, and structural similarity index. The generated results were compared and analyzed pixel-by-pixel with the real Tmax images obtained based on standard perfusion imaging post-processing. Specific evaluation data are summarized in Table 2. High-fidelity image generation: During the model training and internal testing phases, the synthesized images already exhibited a high degree of similarity to the gold standard (SSIM>0.98, PSNR>51dB). This result verifies that the generative adversarial network framework constructed in this invention has the ability to effectively learn complex nonlinear mappings from non-enhanced MRI sequences to quantitative perfusion parameter maps.

[0079] Generalization performance and clinical reliability: On a completely independent external multicenter validation set, the quality metrics of the synthesized Tmax images remained at a high level: mean squared error was only 0.959, structural similarity index reached 0.979, and peak signal-to-noise ratio was 49.644 dB. These quantitative data indicate that the images generated by the method of this invention not only have extremely low errors in pixel intensity, but more importantly, they maintain a high degree of consistency with real clinical perfusion parameter maps in terms of the overall structure, local contrast, and texture details upon which visual perception depends.

[0080] Technical efficacy confirmation: The above results collectively demonstrate that this invention successfully achieved its technical objective, namely, generating Tmax parameter maps that are not inferior to standard PWI post-processing results in terms of both quantitative indicators and visual quality, without relying on contrast agents and dynamic perfusion scanning. This lays a solid and verifiable data foundation for subsequent clinical applications based on this synthesized image, such as low perfusion detection and volume quantification.

[0081] Table 2 3. Consistency analysis of quantitative measurement of critical low perfusion volume (1) Consistency analysis results of key low-perfusion volume prediction As shown in Table 3, the consistency strength is as follows: the intragroup correlation coefficients of all datasets are higher than 0.87, and their 95% confidence interval lower limits are all greater than 0.78. According to the ICC grading standard (>0.75 is "excellent"), this indicates that the model predictions and actual measurements have an excellent level of consistency.

[0082] Correlation verification: Spearman rank correlation coefficients were all above 0.88, which statistically confirmed a strong monotonic correlation between the predicted and actual values.

[0083] Statistical significance: The p-values ​​of all ICCs in the Spearman correlation test were less than 0.001, indicating extremely high statistical significance.

[0084] Generalization performance: as shown in the appendix Figure 4 As shown in Figure AC, the horizontal axis Hypopefusion_GT represents the critical low-perfusion segmentation volume of the real Tmax-mask in the real Tmax image, and the vertical axis Hypopefusion_Pred represents the critical low-perfusion segmentation volume of the synthetic Tmax-mask in the synthetic Tmax image. On a completely independent external validation set, the ICC value reaches 0.894 (95% CI: 0.811–0.941), and the Spearman correlation coefficient is 0.901, which is highly comparable to the performance on the training set and the internal test set. This demonstrates that the model of this invention has excellent cross-center and cross-device generalization ability, and its volume prediction performance is stable and reliable.

[0085] Table 3 (2) Bland-Altman Consistency Analysis To accurately assess the difference (bias) between predicted and actual values ​​and its distribution range, we plotted a Bland-Altman diagram (see attached diagram). Figure 4 (D-F diagram).

[0086] Systematic bias analysis: In the training set, the mean difference between predicted and true values ​​was 13.09 mL. In the internal and external validation sets, the mean differences were 13.09 mL and 19.18 mL, respectively. These relatively small systematic bias values ​​indicate a slight tendency for the model to overestimate or underestimate overall, but the magnitude of the bias is within clinically acceptable limits.

[0087] The Limits of Agreement (LoA) are clearly marked in the figure, indicating the 95% agreement interval (mean difference ± 1.96 standard deviations). The vast majority of data points (>95%) fall within this interval, demonstrating that the random difference between the model's predicted values ​​and the actual values ​​is controllable and the dispersion is within reasonable expectations.

[0088] Differential distribution pattern: In the Bland-Altman plot, the data points are uniformly distributed along the horizontal axis (the average of the predicted and actual values), with no obvious heteroscedasticity (i.e., the difference systematically increases or decreases as the measured average increases). This indicates that the predictive performance of the model of this invention remains stable at different sizes of low perfusion volume levels, demonstrating good applicability.

[0089] 4. Robustness assessment of volume measurements with different Tmax thresholds (1) Consistency analysis results of low perfusion volume prediction under different Tmax thresholds Table 4 evaluates the consistency and robustness of volume measurements in the synthetic Tmax maps generated by the model across different ischemia severity thresholds. Cross-threshold consistency: On the most challenging external validation set, the model maintains good to excellent volume consistency across all four Tmax thresholds (>4s, >6s, >8s, >10s), with ICC values ​​of 0.808, 0.782, 0.807, and 0.831, respectively.

[0090] Advantages in core ischemic regions: Notably, the ICC reaches a high of 0.831 in the core region, representing the most severe ischemia, i.e., at the threshold of Tmax > 10 seconds. This indicates that the method proposed in this invention provides the most accurate and robust delineation of the boundaries of the core severely ischemic region.

[0091] Performance stability: Under different thresholds, the ICC performance trends of the internal test set and the external validation set are consistent, and the model performance is generally better at more stringent thresholds (>8s, >10s). This proves that the model of this invention can not only identify areas of abnormal perfusion, but also accurately distinguish the severity of ischemia.

[0092] Table 4 (2) Bland-Altman analysis results Systematic Bias: The mean bias is close to zero across all datasets (training set, internal test set, external test set) and all four thresholds (>4s, >6s, >8s, >10s). Specifically, the bias values ​​are mostly distributed within a narrow range of ±15mL. For example, at the critical Tmax>6s threshold, the absolute value of the mean bias across the training set, internal test set, and external test set is less than 10mL. This indicates that the Syn-Tmax plot generated by this invention is not significantly different from the real Real-Tmax plot at the level of systematic error.

[0093] Limits of Agreement (LoA): The 95% agreement interval (LoA, i.e., mean difference ± 1.96 × standard deviation) remained reasonable and compact across all analyses. As the Tmax threshold increased from >4s to >10s (i.e., the defined lesion volume typically decreased), the width of the agreement interval did not exhibit abnormal expansion or contraction, indicating that the model maintains stable predictive accuracy for key hypoperfusion areas of varying severity and size.

[0094] Data Distribution and Heterogeneity: In all Bland-Altman plots, data points are uniformly and symmetrically distributed around the bias line, with the vast majority (>95%) falling within the 95% consistency interval. No obvious trend was observed, meaning the variance did not systematically increase or decrease with increasing average volume (no heteroscedasticity). This demonstrates that the consistency between Syn-Tmax and Real-Tmax is reliable at different volume levels.

Claims

1. An acute ischemic stroke perfusion image generation system based on non-contrast enhanced multimodal images, comprising four parts: an information acquisition module, a preprocessing module, an image generation module, and a model evaluation module; The information acquisition module is used to collect data information of the patient to be diagnosed, including head MRI-DWI, ADC data, real perfusion Tmax image and key low perfusion area mask; The preprocessing module is used to perform standardized preprocessing on multimodal images. That is, all images are processed through a standardized automated preprocessing workflow. The image generation module is used to generate clinically usable perfusion Tmax images using a constructed generative adversarial network; The image generation module is specifically used for: (1) Construct a ternary improved generative adversarial network framework that integrates a generator, a discriminator, and a segmentation network; (2) During the training process, the image generation module optimizes the model performance through a triple loss function; The model evaluation module is used by the pre-trained segmentation network to accurately segment and quantitatively evaluate key low-perfusion regions.

2. The image system according to claim 1, characterized in that, The information acquisition module is specifically used for: (1) Acquisition included DWI imaging, ADC maps, and dynamic magnetic susceptibility contrast PWI imaging based on gadolinium contrast agent; (2) The clinically validated RAPID software was used to automate the post-processing of the PWI raw data. The peak time of the whole brain voxels was calculated by the deconvolution algorithm to generate the real Tmax map and the key low perfusion area mask (Real Tmax-mask). (3) Data screening is based on screening criteria, which are divided into data with inclusion criteria and data with exclusion criteria.

3. The image system according to claim 1, characterized in that, The preprocessing module is specifically used for: (1) extraction and fusion optimization of the infarct core region; (2) extraction and spatial registration of brain tissue regions; and (3) preprocessing of negative data.

4. The image system according to claim 3, characterized in that, Extraction and fusion optimization of the infarct core region, specifically including: Based on its integrated VB-Net segmentation model, the high-signal region of acute infarction is automatically segmented from the original DWI image, denoted as DWI-Core; Based on clinical standard ADC thresholds, diffusion-restricted regions are automatically segmented from the ADC map and denoted as Rapid-Core. Perform a logical OR operation between DWI-Core and Rapid-Core to generate a fused infarct core mask.

5. The image system according to claim 1, characterized in that, The construction of the ternary improved adversarial generative network integrating a generator, discriminator, and segmentation network specifically includes: The generator uses ResNet-9blocks as the backbone network and constructs an encoder-decoder structure, which includes a diffusion-weighted imaging map, an apparent diffusion coefficient map, and an optimized infarct core region mask. The discriminator adopts the PatchGAN architecture, which does not make a single judgment on the entire image, but performs real and fake identification on the local receptive field of the image through convolution operation and outputs a discrimination matrix; The segmentation network is based on the VNet architecture, which specializes in 3D medical images, and was pre-trained on the processed real Tmax base map and its corresponding real mask.

6. The image system according to claim 1, characterized in that, The model evaluation module is specifically used for: (1) segmentation of key low-perfusion areas; (2) Quantitative evaluation of model performance, specifically including: First, the baseline clinical and imaging characteristics of the training set, validation set, internal test set, and external validation set were statistically described and compared to assess the balance of data distribution. Second, to objectively evaluate the pixel-level and structural similarity between the Tmax image synthesized by the generator and the real Tmax image; Third, calculate the Spearman rank correlation coefficient and intraclass correlation coefficient (ICC) of the low perfusion volume measured by the synthetic mask and the real mask, and plot the Bland-Altman plot. Fourth, multiple thresholds are applied to the synthetic Tmax and the real Tmax plots, and pairwise consistency analysis is performed on the volume under each threshold. The ICC value is calculated and the Bland-Altman plot is plotted.

7. The image system according to claim 6, characterized in that, To objectively evaluate the pixel-level and structural similarity between the Tmax image synthesized by the generator and the real Tmax image, the following three image quality evaluation metrics are specifically used: A. Mean Square Error: Measures the average squared value of the difference in pixel intensity between images. The calculation formula is: MSE=(1 / N)*Σ(Syn_i-Real_i)^2 Where N is the total number of pixels, and the lower the MSE value, the smaller the pixel-level error; B. Peak Signal-to-Noise Ratio (PSNR): Calculated based on MSE, it reflects the ratio of signal to noise in an image. The calculation formula is: PSNR = 10 * log10(MAX^2 / MSE) Where MAX is the maximum possible value of the image pixels, and the higher the PSNR value, the better the image quality; C. Structural Similarity Index: This index compares the similarity of images in terms of brightness, contrast, and structure. Its value ranges from -1 to 1. The closer it is to 1, the higher the structural similarity between the two images in visual perception.

8. The image system according to claim 6, characterized in that, In calculating the Spearman rank correlation coefficient between the synthetic mask and the real mask for low perfusion volume, the within-group correlation coefficient of the two-way random effects model is used to quantitatively assess the consistency strength of the two measurement results.

Images