Medical image post-processing method and system and medical image workstation
By generating TTP images from single-modal FLAIR magnetic resonance images using deep learning, the complexity and risks of traditional PWI technology are solved, enabling contrast agent-free cerebral hemodynamic assessment and providing rapid, automated quantitative assessment and reliable cerebral perfusion judgment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-19
- Publication Date
- 2026-03-31
AI Technical Summary
Traditional perfusion-weighted magnetic resonance imaging (PWI) is complex, time-consuming, and carries risks associated with contrast agent use, making it difficult to widely apply in primary hospitals. Furthermore, the post-processing results are highly subjective, making it difficult to achieve rapid and automated quantitative assessment.
Using a deep learning-based approach, TTP images are generated from single-modal FLAIR magnetic resonance images using a visual feature encoder and an image reconstruction decoder. Combined with a dynamic assessment model, contrast agent-free cerebral hemodynamic assessment is achieved.
To reduce imaging detection costs, avoid the risks associated with contrast agents, achieve rapid and automated quantitative assessment of cerebral hemodynamics, and provide reliable evidence for cerebral perfusion judgment.
Smart Images
Figure CN121767510A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image post-processing technology, specifically to a medical image post-processing method and system, and a medical image workstation. Background Technology
[0002] Perfusion-weighted imaging (PWI) is currently an important imaging technique for clinically assessing cerebral hemodynamics, identifying the ischemic penumbra, and guiding reperfusion therapy in acute stroke. PWI can obtain parameters such as cerebral blood flow (CBF), cerebral blood volume (CBV), mean transit time (MTT), and time-to-peak (TTP). TTP images directly reflect the delay in local blood flow arrival and are an important basis for assessing insufficient tissue perfusion. However, traditional PWI imaging relies on gadolinium contrast agent injection and dynamic acquisition sequences, making the imaging process complex and time-consuming, easily limited by equipment and operational conditions. Furthermore, the use of contrast agents may pose risks of allergies or kidney dysfunction, limiting its application in some patients and in primary care hospitals. In addition, PWI post-processing requires manual selection of arterial input functions and deconvolution analysis, resulting in highly subjective results and making rapid, automated quantitative assessment difficult. Summary of the Invention
[0003] The main technical problem addressed by this invention is how to reduce or replace the application of PWI imaging technology.
[0004] According to a first aspect, one embodiment provides a medical image post-processing method, comprising: Acquire magnetic resonance FLAIR sequence images of the head of the user to be examined; the magnetic resonance FLAIR sequence images are sequence images used to reflect the T2-weighted signal characteristics of the brain tissue of the user to be examined; The magnetic resonance FLAIR sequence image is input into a preset TTP image conversion mathematical model, and the predicted TTP image sequence corresponding to the user to be examined is obtained from the output of the TTP image conversion mathematical model. The TTP image conversion mathematical model includes a visual feature encoder, an image reconstruction decoder, and an activation function. The visual feature encoder is used to extract multi-scale global features and contextual information from the magnetic resonance FLAIR sequence images. The image reconstruction decoder is used to restore spatial resolution layer by layer through multi-level upsampling and convolution operations, and to fuse feature information from different levels through skip connections. The activation function is used as the output of the TTP image conversion mathematical model to output a single-channel predicted TTP image. The predicted TTP image is displayed and output in real time.
[0005] In one embodiment, the medical image post-processing method further includes: The predicted TTP image sequence is input into a preset dynamic evaluation mathematical model, and the cerebral hemodynamic evaluation table corresponding to the user under test is obtained from the output of the dynamic evaluation mathematical model; wherein, the cerebral hemodynamic evaluation table is used to identify the anatomical location, delay volume and / or multi-perfusion parameter correlation analysis results of the cerebral perfusion delay area; Output the cerebral hemodynamic assessment form.
[0006] In one embodiment, the method for obtaining the TTP image conversion mathematical model includes: Training set construction process; Acquire multiple sets of paired medical imaging data, each set of medical imaging data including magnetic resonance FLAIR sequence images and TTP images of the examined user's head; The paired medical image data for each pair are spatially registered to eliminate positional bias, and then subjected to intensity normalization, size standardization, and data augmentation to serve as standardized input data for training the TTP image conversion mathematical model. The data augmentation includes performing rotation, translation, scaling, brightness perturbation, and mirroring operations on the FLAIR images to improve the robustness of the TTP image conversion mathematical model to images with different orientations. The spatial registration is used in a mutual information-based image registration algorithm to align corresponding regions of brain structures in different modalities. The process of training and optimizing the model; A preset composite loss function is used to jointly optimize the TTP image conversion mathematical model, so as to enhance the learning weight of the TTP image conversion mathematical model for the TTP delay region and maintain the structural consistency of the generated TTP image. Model validation process; The newly acquired magnetic resonance FLAIR sequence image is input into the TTP image conversion mathematical model to obtain the predicted TTP image sequence output by the TTP image conversion mathematical model; Obtain the actual TTP image sequence of the user under investigation corresponding to the magnetic resonance FLAIR sequence image; The similarity index between the real TTP image sequence and the predicted TTP image sequence is evaluated based on the peak signal-to-noise ratio and / or structural similarity index calculated from the real TTP image sequence.
[0007] In one embodiment, the visual feature encoder adopts a PVTv2 structure, extracts feature maps in a hierarchical pyramid manner, and introduces an attention mechanism to achieve global context modeling; the image reconstruction decoder consists of multi-level deconvolutional layers and convolutional blocks, each level containing two convolutional units and a batch normalization layer, outputting features and corresponding features of the visual feature encoder, which are then fused through skip connections and processed by the ReLU function; the activation function is the Sigmoid activation function, which constrains pixel values within the range of [0,1] to express the TTP normalized perfusion time distribution.
[0008] In one embodiment, outputting the predicted TTP image includes: The predicted TTP image is uploaded to the hospital PACS system or medical imaging workstation to achieve real-time visualization and output of the predicted TTP image.
[0009] According to a second aspect, one embodiment provides a medical image post-processing system for applying the medical image post-processing method as described in the first aspect, the medical image post-processing system comprising: The image acquisition unit is used to acquire magnetic resonance FLAIR sequence images of the head of the user to be examined; the magnetic resonance FLAIR sequence images are sequence images that reflect the T2-weighted signal characteristics of the brain tissue of the user to be examined. An image prediction unit is used to input the magnetic resonance FLAIR sequence image into a preset TTP image conversion mathematical model and obtain the predicted TTP image sequence corresponding to the user to be examined, output by the TTP image conversion mathematical model. The TTP image conversion mathematical model includes a visual feature encoder, an image reconstruction decoder, and an activation function. The visual feature encoder is used to extract multi-scale global features and contextual information from the magnetic resonance FLAIR sequence image. The image reconstruction decoder is used to restore spatial resolution layer by layer through multi-level upsampling and convolution operations, and to fuse feature information from different levels through skip connections. The activation function serves as the output of the TTP image conversion mathematical model to output a single-channel predicted TTP image. The image analysis unit is used to input the predicted TTP image sequence into a preset dynamic evaluation mathematical model and obtain the cerebral hemodynamic evaluation table corresponding to the user to be examined, which is output by the dynamic evaluation mathematical model; wherein, the cerebral hemodynamic evaluation table is used to identify the anatomical location, delay volume and / or multi-perfusion parameter correlation analysis results of the cerebral perfusion delay area. The result output unit is used to graphically display the cerebral hemodynamic assessment form. According to a third aspect, one embodiment provides a medical imaging workstation including a computer-readable storage medium storing a computer program that can be executed by a processor to implement the medical image post-processing method as described in the first aspect.
[0010] According to a fourth aspect, one embodiment provides a computer program product including a computer program and / or instructions, characterized in that, when the computer program and / or instructions are executed by a processor, they implement the medical image post-processing method as described in the first aspect.
[0011] According to the medical image post-processing method in the above embodiments, since it is possible to generate TTP perfusion images based on single-modal magnetic resonance FLAIR images, thereby achieving contrast agent-free assessment of cerebral hemodynamic status, it not only reduces imaging detection costs, but also eliminates the detection risks caused by contrast agent injection for the examinee. Attached Figure Description
[0012] Figure 1 This is a flowchart illustrating a medical image post-processing method in one embodiment. Figure 2 This is a functional block diagram of a medical image post-processing system in one embodiment. Figure 3 This is a schematic diagram of image comparison in one embodiment. Detailed Implementation
[0013] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.
[0014] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.
[0015] The serial numbers assigned to components in this document, such as "first" and "second," are used only to distinguish the described objects and have no sequential or technical meaning. The terms "connection" and "linkage" used in this application, unless otherwise specified, include both direct and indirect connections (linkages).
[0016] FLAIR (Fluid-Attenuated Inversion Recovery) magnetic resonance imaging is a routine, contrast-free scanning sequence widely used in imaging examinations for neurological diseases. Existing research has shown that the FLAIR hyperintense vessel (FHV) in FLAIR images can reflect perfusion delay areas to some extent, but its assessment relies on human visual judgment, which is highly subjective and has limited accuracy. In recent years, with the development of artificial intelligence technology, deep learning has shown great potential in image feature extraction and image-to-image prediction. However, existing research on predicting perfusion images from FLAIR images is mostly limited to shallow convolutional models, which lack spatial feature representation capabilities, struggle to capture multi-scale changes in cerebral blood flow, and have limited stability and clinical consistency of the generated results. Therefore, there is an urgent need for a deep learning-based method that can automatically generate high-fidelity TTP images from FLAIR images to achieve quantitative cerebral perfusion assessment under contrast-free conditions, thereby providing more comprehensive judgment criteria for rapid diagnosis and treatment decisions in cerebrovascular diseases. Example
[0017] Please refer to Figure 1 This is a flowchart illustrating a medical image post-processing method in one embodiment. The medical image post-processing method generates TTP perfusion images based on single-modal magnetic resonance FLAIR images, specifically including: Step 101: Obtain magnetic resonance images.
[0018] Magnetic resonance FLAIR sequence images of the head of the user under examination are acquired. These FLAIR sequence images are used to reflect the T2-weighted signal characteristics of the brain tissue of the user under examination.
[0019] Step 102: Generate TTP perfusion images.
[0020] The FLAIR magnetic resonance imaging (MRI) sequence images are input into a pre-defined TTP image conversion mathematical model, and the predicted TTP image sequence for the corresponding user is obtained from the output of the TTP image conversion mathematical model. The TTP image conversion mathematical model includes a visual feature encoder, an image reconstruction decoder, and an activation function. The visual feature encoder extracts multi-scale global features and contextual information from the FLAIR sequence images. The image reconstruction decoder restores spatial resolution layer by layer through multi-level upsampling and convolution operations, and fuses feature information from different levels through skip connections. The activation function serves as the output of the TTP image conversion mathematical model to output a single-channel predicted TTP image. In one embodiment, the visual feature encoder is a pyramid visual encoder with a PVTv2 structure, extracting feature maps in a hierarchical pyramid manner and introducing an attention mechanism to achieve global context modeling. In another embodiment, the image reconstruction decoder is a U-shaped decoder composed of multiple levels of deconvolutional layers and convolutional blocks. Each level contains two convolutional units and a batch normalization layer. The output features are fused with the features of the corresponding pyramid visual encoder through skip connections and then processed by the ReLU function. In one embodiment, the activation function is the Sigmoid activation function, which constrains the pixel values within the range of [0,1] to represent the TTP normalized perfusion time distribution.
[0021] Step 103: Output and display the predicted image.
[0022] Real-time visualization and output of predicted TTP images. In one embodiment, the predicted TTP images are uploaded to a hospital PACS system or medical imaging workstation to achieve real-time visualization and output of the predicted TTP images.
[0023] Step 104: Obtain the cerebral hemodynamic assessment form.
[0024] The predicted TTP image sequence is input into a pre-defined kinetic assessment mathematical model, and the cerebral hemodynamic assessment table for the corresponding user is obtained from the output of the kinetic assessment mathematical model. The cerebral hemodynamic assessment table is used to identify the anatomical location of the cerebral perfusion delay zone, the delay volume, and / or the results of multi-perfusion parameter correlation analysis.
[0025] In one embodiment, the mathematical model for kinetic assessment can be a model combining the Arterial Input Function (AIF) with a deconvolution algorithm. The following describes the process of outputting a cerebral hemodynamic assessment table using a Gamma variable hemodynamic model as an example, specifically including: The Gamma variable hemodynamic model, based on TTP (time to peak) image sequences, calculates key perfusion parameters by quantifying the time-to-concentration curve characteristics of cerebral blood flow perfusion. Changes in contrast agent concentration within arteries are used as AIF (Area Intensity Forward), and changes in contrast agent concentration within brain tissue are used as the tissue time-to-concentration curve (TTC). A deconvolution algorithm is used to decompose the dynamic processes such as "blood flow perfusion" and "extravascular space distribution" within the TTC, and perfusion parameters are obtained by fitting with a Gamma function. The input data for the Gamma variable hemodynamic model is a TTP image sequence, including time-dimension images of the entire brain of the user (typically 10-20 frames, frame interval 1-2 seconds); auxiliary data includes images of the user's brain anatomy (such as MRI T1-weighted images) and AIF extraction regions (usually the middle cerebral artery supply area). The calculation of core parameters (corresponding to the core indicators of the evaluation table) includes: 1. Anatomical location of the delayed cerebral perfusion zone.
[0026] By registering TTP image sequences with anatomical images and combining them with the "TTP delay threshold (usually > normal mean + 2SD)", the location of the delayed area in the brain lobe and sulci (e.g., "around the calcarine sulcus of the right parietal lobe") is marked.
[0027] 2. Delay volume.
[0028] Pixel counts are performed on the marked delay region, and the actual volume (unit: ml) is calculated by combining the pixel spacing of the image.
[0029] 3. Correlation analysis of multiple perfusion parameters.
[0030] The model fits and outputs hemodynamic parameters (cerebral blood flow CBF, cerebral blood volume CBV, and mean transit time MTT), delay-related parameters (maximum TTP, standard deviation of TTP, and delay severity grading), and correlation analysis parameters (CBF / CBV ratio and correlation coefficient between MTT and TTP).
[0031] The output of the Gamma variable hemodynamic model is mapped to the evaluation table, which includes, for example, anatomical location data (three-dimensional coordinate range of the delayed region, corresponding anatomical structure or matching label based on brain atlas), volume data (total number of pixels in the delayed region → converted to actual volume, such as "8.2 ml") and parameter matrix (the CBF, CBV, MTT, TTP and other parameter values of each brain voxel, and the results of parameter correlation analysis).
[0032] Based on the Gamma variable hemodynamic model described above, a cerebral hemodynamic assessment table can be obtained by predicting TTP image sequences. Alternatively, the volume can be calculated based on simple threshold post-processing of the TTP map (e.g., TTP ≥ 6 seconds).
[0033] Step 105: Output and display the evaluation table.
[0034] Visualize or print the cerebral hemodynamic assessment form.
[0035] In one embodiment, the method for obtaining the TTP image conversion mathematical model includes a training set construction process, a model training and optimization process, and a model validation process.
[0036] The training set construction process includes: Multiple sets of paired medical image data were acquired, each set including FLAIR MRI images and TTP images of the examined user's head. Spatial registration was performed on each pair of medical image data to eliminate positional bias, followed by intensity normalization, size standardization, and data augmentation to serve as standardized input data for training the TTP image conversion mathematical model. Data augmentation included rotation, translation, scaling, brightness perturbation, and mirroring operations on the FLAIR images to improve the robustness of the TTP image conversion mathematical model to images with different orientations. Spatial registration was used in a mutual information-based image registration algorithm to align corresponding brain structural regions across different modalities.
[0037] The process of training and optimizing the model includes: A pre-defined composite loss function is used to jointly optimize the TTP image conversion mathematical model, thereby enhancing the model's learning weights for TTP delay regions and maintaining the structural consistency of the generated TTP images. In one embodiment, a composite loss function is composed of a focal regression loss function (focal regression loss) and a structural similarity loss function (SSIM loss). The focal regression loss function is used to enhance the model's learning weights for TTP delay regions, while the structural similarity loss function, calculated based on brightness, contrast, and structural components, maintains the structural consistency of the generated TTP images.
[0038] The model validation process includes: First, the newly acquired FLAIR magnetic resonance image sequence is input into the TTP image conversion mathematical model to obtain the predicted TTP image sequence output by the TTP image conversion mathematical model; then, the actual TTP image sequence of the user to be examined corresponding to the FLAIR magnetic resonance image sequence is obtained; finally, the similarity index between the actual TTP image sequence and the predicted TTP image sequence is evaluated based on the peak signal-to-noise ratio and / or structural similarity index calculated from the actual TTP image sequence.
[0039] Please refer to Figure 2This is a functional block diagram of a medical image post-processing system in one embodiment. The medical image post-processing system is used to apply the medical image post-processing method described above, and includes an image acquisition unit 10, an image prediction unit 20, an image analysis unit 30, and a result output unit 40. The image acquisition unit 10 is used to acquire magnetic resonance FLAIR sequence images of the head of the user to be examined. The magnetic resonance FLAIR sequence images are sequence images reflecting the T2-weighted signal characteristics of the brain tissue of the user to be examined. The image prediction unit 20 is used to input the magnetic resonance FLAIR sequence images into a preset TTP image conversion mathematical model, and acquire the predicted TTP image sequence of the user to be examined output by the TTP image conversion mathematical model. The TTP image conversion mathematical model includes a visual feature encoder, an image reconstruction decoder, and an activation function. The visual feature encoder is used to extract multi-scale global features and contextual information from the magnetic resonance FLAIR sequence images. The image reconstruction decoder is used to recover spatial resolution layer by layer through multi-level upsampling and convolution operations, and to fuse feature information from different levels through skip connections. The activation function serves as the output of the TTP image conversion mathematical model to output a single-channel predicted TTP image. The image analysis unit 30 is used to input the predicted TTP image sequence into a preset dynamic evaluation mathematical model and obtain the cerebral hemodynamic evaluation table of the corresponding user output by the dynamic evaluation mathematical model. The cerebral hemodynamic evaluation table is used to identify the anatomical location of the cerebral perfusion delay area, the delay volume, and / or the results of multi-perfusion parameter correlation analysis. The result output unit 40 is used to display the cerebral hemodynamic evaluation table graphically.
[0040] This application aims to propose a deep learning-based medical image post-processing method that can automatically generate TTP perfusion images from single-modal FLAIR magnetic resonance images, enabling contrast-free assessment of cerebral hemodynamic status, reducing imaging costs and patient detection risks. To facilitate understanding of the application scenarios and methods of the medical image post-processing method disclosed in this application, specific embodiments are described below, including: The process of acquiring a TTP image conversion mathematical model can be briefly summarized as follows: data acquisition and preprocessing, model building, loss function design and optimization strategy, and model training and validation. The data acquisition and preprocessing process includes the following steps: 1) Collect paired image data of patients that simultaneously contain magnetic resonance FLAIR (Fluid-Attenuated Inversion Recovery) sequences and TTP (Time-to-Peak) perfusion images; 2) Perform spatial registration processing on the images to align the brain anatomical structures of the two modalities in the same spatial coordinates, so as to eliminate the displacement error between the modalities; 3) Perform intensity normalization on the registered image to standardize the pixel intensity range to a uniform interval; 4) Unify all image sizes to a preset size (preferably 512×512 pixels) to ensure consistency of network input; 5) Perform data augmentation on the training samples, including random rotation, translation, and scaling operations, to improve the model's generalization performance under different scanning conditions.
[0041] The model building process includes: The model employs a combined structure of a pyramidal visual Transformer and a U-shaped decoder (PVTv2-UNet). The encoder utilizes a multi-scale pyramidal Transformer module to extract multi-level semantic features from the input FLAIR image, combining this with a self-attention mechanism to model global contextual information. The decoder progressively restores spatial resolution through multi-level deconvolution and convolution operations, fusing high-level semantic features with low-level spatial details via skip connections. The model output uses a 1×1 convolutional layer to generate a single-channel TTP prediction image, and the pixel values are constrained to the range [0,1] using a sigmoid activation function to obtain a continuous perfusion temporal distribution map.
[0042] The process of designing and optimizing loss functions includes: A composite loss function is used for optimization during training, employing both focal regression loss and structural similarity loss. Focal regression loss is used to enhance the model's sensitivity to regions of delayed blood perfusion; its formula is as follows: ; in, This is the actual TTP value. These are the model's predicted values. and To adjust the parameters, specifically, As a focusing factor, This is the scaling parameter. It is adjusted by... and The value is used to balance the weights of the overall error and the local salient regions.
[0043] Structural similarity loss (SSIM Loss) is used to maintain consistency between the predicted image and the true TTP image in terms of brightness, contrast, and structure. The SSIM Loss is calculated based on the brightness, contrast, and structure components, and is defined by the following formula: ; SSIM is used to measure the consistency of the predicted image with the real image in terms of local structure.
[0044] The total loss function is a weighted combination of the focus regression loss and the structural similarity loss mentioned above. The formula for the weighted combination of the two losses is as follows: ; in, and These are the weighting coefficients.
[0045] The optimizer for the optimization strategy uses the AdamW algorithm with a learning rate of 1×10⁻⁶. -4 The weight decay coefficient is 1×10 -5 In one embodiment, the training batch size is 8, the image size is 512×512 pixels, and the training epochs are 100. The optimizer updates gradients steadily and prevents overfitting.
[0046] The model training and validation process includes: (1) Input the preprocessed FLAIR-TTP paired data into the TTP image conversion mathematical model and perform iterative training with a batch size of 8; (2) The total number of training rounds is set to 100 rounds, and the model weights and some generated samples are automatically saved after each training cycle; (3) During the validation phase, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) are used to quantitatively evaluate the predicted TTP images in order to verify the generation quality and structural fidelity of the model. During the model training process, the weight file and generated samples are saved regularly, and the average training loss, peak signal-to-noise ratio and structural similarity index are calculated as performance monitoring indicators. (4) After the model training is completed, output the best parameter combination and save the final network weights.
[0047] In one embodiment of this application, the trained model is first applied to a new magnetic resonance FLAIR image input; then, the corresponding TTP prediction image is automatically generated and compared with the real PWI-TTP image; then, the Dice coefficient and crossover ratio (IoU) of the perfusion delay zone (TTP≥6 seconds) are calculated to evaluate the accuracy of the delay zone prediction; finally, the results are output and visualized in the form of a triptych, including the input FLAIR image, the predicted TTP image and the real TTP image, to achieve intuitive display of cerebral blood flow perfusion status and clinical auxiliary judgment.
[0048] The medical image post-processing system disclosed in this application can be embedded in a hospital PACS or neuro-emergency imaging workstation to achieve automatic recognition of FLAIR sequences, generation of TTP images, and real-time visualization output. This medical image post-processing system can generate TTP images solely based on FLAIR sequences, reducing patient risk (no intravenous perfusion required) and enabling quantitative assessment (accurate quantification of perfusion delay based on the output continuous pixel-level perfusion parameter map).
[0049] Please refer to Figure 3 This is a schematic diagram of image comparison in one embodiment. In actual testing, the predicted TTP image is highly consistent with the actual PWI results in terms of PSNR and SSIM. The medical image post-processing method disclosed in this application embodiment has strong application scalability and can be extended to other neuroimaging prediction tasks (such as CBF / CBV generation, etc.).
[0050] The medical image post-processing method disclosed in this application first acquires FLAIR magnetic resonance imaging (MRI) sequence images; then, it inputs these FLAIR sequence images into a TTP image conversion mathematical model to obtain a predicted TTP image sequence. The TTP image conversion mathematical model includes a visual feature encoder, an image reconstruction decoder, and an activation function. The visual feature encoder extracts multi-scale global features and contextual information from the FLAIR sequence images. The image reconstruction decoder restores spatial resolution layer by layer through multi-level upsampling and convolution operations, and fuses feature information from different levels through skip connections. The activation function outputs a single-channel predicted TTP image. Finally, the predicted TTP image is displayed and output. Because it can generate TTP perfusion images from single-modal FLAIR MRI images, it enables contrast-free assessment of cerebral hemodynamics, reducing imaging costs and eliminating the risks associated with contrast agent injection for the patient.
[0051] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.
[0052] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.
Claims
1. A medical image post-processing method, characterized in that, The method comprises: obtaining a magnetic resonance FLAIR sequence image of a head of a user to be detected; the magnetic resonance FLAIR sequence image is a sequence image reflecting T2-weighted signal characteristics of brain tissue of the user to be detected; inputting the magnetic resonance FLAIR sequence image into a preset TTP image conversion mathematical model, and obtaining a predicted TTP image sequence corresponding to the user to be detected output by the TTP image conversion mathematical model; wherein the TTP image conversion mathematical model comprises a visual feature encoder, an image reconstruction decoder, and an activation function; the visual feature encoder is used to extract multi-scale global features and context information of the magnetic resonance FLAIR sequence image; the image reconstruction decoder is used to recover spatial resolution layer by layer through multi-level upsampling and convolution operation, and fuse feature information of different levels through a skip connection; and the activation function is used as an output end of the TTP image conversion mathematical model to output a single-channel predicted TTP image; real-time visualizing displaying and outputting the predicted TTP image.
2. The medical image post-processing method of claim 1, wherein, The method further comprises: inputting the predicted TTP image sequence into a preset dynamic evaluation mathematical model, and obtaining a cerebral blood flow dynamic evaluation table corresponding to the user to be detected output by the dynamic evaluation mathematical model; wherein the cerebral blood flow dynamic evaluation table is used to identify an anatomic location of a cerebral perfusion delay area, a delay volume, and / or a multi-perfusion parameter correlation analysis result; outputting the cerebral blood flow dynamic evaluation table.
3. The medical image post-processing method of claim 1, wherein, The method for obtaining the TTP image conversion mathematical model comprises: a training set construction process; obtaining a plurality of groups of paired medical image data, each group of the medical image data comprising a magnetic resonance FLAIR sequence image of a head of a user who has been detected and a TTP image; performing spatial registration on each group of paired medical image data to eliminate positional deviation, and performing intensity normalization, size standardization, and data enhancement processing to serve as standardized input data for training of the TTP image conversion mathematical model; wherein the data enhancement processing comprises performing rotation, translation, scaling, brightness disturbance, and mirror operation on the FLAIR image to improve robustness of the TTP image conversion mathematical model to different orientation images; and the spatial registration is based on a mutual information-based image registration algorithm to align corresponding regions of brain structures under different modalities.
4. The medical image post-processing method of claim 3, wherein, The method for obtaining the TTP image conversion mathematical model further comprises: a model training and optimization process; using a preset composite loss function to jointly optimize the TTP image conversion mathematical model to enhance learning weight of the TTP image conversion mathematical model to a TTP delay area and maintain structural consistency of a generated TTP image.
5. The medical image post-processing method of claim 4, wherein, The method for obtaining the TTP image conversion mathematical model further comprises: a model verification process; inputting the newly obtained magnetic resonance FLAIR sequence image into the TTP image conversion mathematical model to obtain a predicted TTP image sequence output by the TTP image conversion mathematical model; obtaining a real TTP image sequence of the user to be detected corresponding to the magnetic resonance FLAIR sequence image; The peak signal-to-noise ratio and / or the structural similarity index calculated according to the real TTP image sequence are used to evaluate the similarity index of the predicted TTP image sequence.
6. The medical image post-processing method of claim 1, wherein, The visual feature encoder is a pyramid visual encoder adopting a PVTv2 structure, which extracts multi-scale global features in a hierarchical pyramid manner and introduces an attention mechanism to realize global context modeling; the image reconstruction decoder is a U-shaped decoder composed of multiple deconvolution layers and convolution blocks, each level containing two convolution units and a batch normalization layer, the output features are fused with the features of the visual feature encoder through a jump connection and then processed by a ReLU function; the activation function is a Sigmoid activation function, which constrains the pixel value in the range of [0, 1] and is used to express the TTP normalized perfusion time distribution.
7. The medical image post-processing method of claim 1, wherein, The predicted TTP image is outputted, including: The predicted TTP image is uploaded to a hospital PACS system or a medical image workstation to realize real-time visual display and output of the predicted TTP image.
8. A medical image workstation, characterized by, The computer program and / or instructions are executed by the processor to realize the medical image post-processing method according to any one of claims 1 to 7.
9. A computer program product comprising computer programs and / or instructions, characterized in that, The computer program and / or instructions are executed by the processor to realize the medical image post-processing method according to any one of claims 1 to 7.
10. A medical image post-processing system, characterized in that, The medical image post-processing system for applying the medical image post-processing method according to any one of claims 1 to 7, comprising: An image acquisition unit is configured to acquire a magnetic resonance FLAIR sequence image of a head of a user to be examined; the magnetic resonance FLAIR sequence image is used to reflect a sequence image of T2-weighted signal characteristics of brain tissue of the user to be examined; An image prediction unit is configured to input the magnetic resonance FLAIR sequence image into a preset TTP image conversion mathematical model, and acquire a predicted TTP image sequence corresponding to the user to be examined output by the TTP image conversion mathematical model; wherein the TTP image conversion mathematical model comprises a visual feature encoder, an image reconstruction decoder, and an activation function; the visual feature encoder is configured to extract multi-scale global features and context information of the magnetic resonance FLAIR sequence image; the image reconstruction decoder is configured to recover spatial resolution layer by layer through multi-level upsampling and convolution operations, and fuse feature information of different levels through a jump connection; and the activation function is configured as an output end of the TTP image conversion mathematical model to output a single-channel predicted TTP image; An image analysis unit is configured to input the predicted TTP image sequence into a preset dynamics evaluation mathematical model, and acquire a cerebral blood flow dynamics evaluation table corresponding to the user to be examined output by the dynamics evaluation mathematical model; wherein the cerebral blood flow dynamics evaluation table is used to identify an anatomic location of a cerebral perfusion delay region, a delay volume, and / or a multi-perfusion parameter correlation analysis result; A result output unit is configured to image display the cerebral blood flow dynamics evaluation table. A computer readable storage medium having stored thereon a computer program executable by a processor to implement the medical image post-processing method according to any one of claims 1 to 7. The computer program and / or instructions are executed by the processor to realize the medical image post-processing method according to any one of claims 1 to 7. The medical image post-processing system for applying the medical image post-processing method according to any one of claims 1 to 7, comprising: An image acquisition unit is configured to acquire a magnetic resonance FLAIR sequence image of a head of a user to be examined; the magnetic resonance FLAIR sequence image is used to reflect a sequence image of T2-weighted signal characteristics of brain tissue of the user to be examined; An image prediction unit is configured to input the magnetic resonance FLAIR sequence image into a preset TTP image conversion mathematical model, and acquire a predicted TTP image sequence corresponding to the user to be examined output by the TTP image conversion mathematical model; wherein the TTP image conversion mathematical model comprises a visual feature encoder, an image reconstruction decoder, and an activation function; the visual feature encoder is configured to extract multi-scale global features and context information of the magnetic resonance FLAIR sequence image; the image reconstruction decoder is configured to recover spatial resolution layer by layer through multi-level upsampling and convolution operations, and fuse feature information of different levels through a jump connection; and the activation function is configured as an output end of the TTP image conversion mathematical model to output a single-channel predicted TTP image; An image analysis unit is configured to input the predicted TTP image sequence into a preset dynamics evaluation mathematical model, and acquire a cerebral blood flow dynamics evaluation table corresponding to the user to be examined output by the dynamics evaluation mathematical model; wherein the cerebral blood flow dynamics evaluation table is used to identify an anatomic location of a cerebral perfusion delay region, a delay volume, and / or a multi-perfusion parameter correlation analysis result; A result output unit is configured to image display the cerebral blood flow dynamics evaluation table.