Method and product for training a model for segmenting a sequence of brain perfusion images

Abstract

The application discloses a method and product for training a brain perfusion image sequence segmentation model, which comprises the following steps: marking the boundary of ischemic and infarcted areas in a brain perfusion image sequence to obtain marking information; inputting the image sequence into a difference calculation module to perform difference calculation and obtain a difference feature map; obtaining a brain basic feature map; inputting the brain basic feature map into a wavelet transform module to perform wavelet transform and obtain a wavelet transform feature map; inputting the brain basic feature map into a local information extraction module to perform information extraction and obtain a local information feature map; combining the difference feature map, the wavelet transform feature map and the local information feature map and then inputting them into a segmentation output module to perform regional segmentation and output a segmentation result; and updating the parameters of the brain perfusion image sequence segmentation model based on the marking information and the segmentation result to realize the training. According to the scheme, the segmentation accuracy of ischemic and infarcted areas can be improved.

Description

Technical Field

[0001] This application generally relates to the field of computer vision technology. More specifically, this application relates to a method, apparatus, and computer-readable storage medium for training a brain perfusion image sequence segmentation model. Furthermore, this application also relates to an apparatus and computer-readable storage medium for segmenting brain perfusion image sequences. Background Technology

[0002] Acute ischemic stroke (AIS), a serious disease threatening human life and health, is characterized by high disability and mortality rates due to the sudden interruption of blood flow to the brain. Rapid and accurate assessment of the ischemic and infarct areas is crucial for developing effective treatment strategies and improving patient prognosis. The advent of computed tomography perfusion (CTP) technology can visually present hemodynamic information of brain tissue, providing important evidence for the diagnosis and treatment of AIS. However, traditional manual assessment methods rely heavily on the professional experience of radiologists, revealing several drawbacks in practice. On the one hand, the assessment process is highly subjective, with differing judgment criteria among doctors leading to inconsistent results; on the other hand, manual analysis is time-consuming, failing to meet the urgent need for "time is life" in the emergency clinical treatment of AIS, potentially delaying the patient's optimal treatment window.

[0003] In recent years, deep learning has achieved significant breakthroughs in the field of medical image analysis, particularly in computer-aided diagnosis and medical image segmentation. Deep learning-based automatic segmentation methods, with their powerful feature extraction and analysis capabilities, can effectively improve the accuracy of ischemic region identification, reduce human error interference, and significantly shorten image analysis time. However, existing deep learning methods still suffer from limitations in processing CTP images of acute ischemic stroke, including insufficient capture of detailed features in ischemic and infarcted areas and weak multi-dimensional information fusion capabilities. These limitations result in segmentation accuracy and stability that cannot meet the stringent clinical requirements for precise assessment.

[0004] Therefore, there is an urgent need to provide a scheme for training a brain perfusion image sequence segmentation model so that the trained brain perfusion image sequence segmentation model can accurately segment ischemic and infarcted areas, providing clinicians with a more scientific and objective basis for decision-making. Summary of the Invention

[0005] In order to at least address one or more of the technical problems mentioned above, this application proposes a scheme for segmenting brain perfusion image sequences in the following aspects.

[0006] In a first aspect, this application provides a method for training a brain perfusion image sequence segmentation model, the brain perfusion image sequence segmentation model including a difference calculation module, a basic feature extraction module, a wavelet transform module, a local information extraction module, and a segmentation output module; the method includes: acquiring a brain perfusion image sequence, and in each image of the brain perfusion image sequence, annotating the boundaries of ischemic and infarcted regions to obtain annotation information; inputting the brain perfusion image sequence into the difference calculation module to perform difference calculation to obtain a difference feature map; and inputting the brain perfusion image sequence into the basic feature extraction module to perform feature extraction. The brain perfusion image sequence segmentation model is trained by: obtaining a basic brain feature map; inputting the basic brain feature map into the wavelet transform module for wavelet transform to obtain a wavelet transform feature map; inputting the basic brain feature map into the local information extraction module for information extraction to obtain a local information feature map; combining the difference feature map, the wavelet transform feature map, and the local information feature map to construct a combined feature map, and inputting the combined feature map into the segmentation output module for region segmentation to output the segmentation result; and updating the parameters of the brain perfusion image sequence segmentation model based on the annotation information and the segmentation result to train it.

[0007] In some embodiments, performing difference calculation in the difference calculation module to obtain a difference feature map includes: calculating the image difference between a first image and a second image at adjacent time points based on images at different time points in the brain perfusion image sequence as a difference feature map; and denoising the difference feature map to obtain a denoised difference feature map.

[0008] In some embodiments, performing wavelet transform in the wavelet transform module to obtain a wavelet transform feature map includes: convolving the brain baseline feature map through a convolutional layer using a preset convolutional kernel to obtain a processed brain baseline feature map; performing wavelet transform on the processed brain baseline feature map to obtain low-frequency information and high-frequency information, wherein the high-frequency information includes horizontal information, vertical information, and diagonal information; combining the horizontal information, the vertical information, and the diagonal information, and then processing them through a convolutional layer using a 1×1 convolutional kernel and a batch normalization layer to obtain processed high-frequency information; processing the low-frequency information through a convolutional layer using a 1×1 convolutional kernel to obtain processed low-frequency information; and combining the processed high-frequency information with the processed low-frequency information to obtain the wavelet transform feature map.

[0009] In some embodiments, information extraction in the local information extraction module to obtain a local information feature map includes: performing convolution processing on the basic brain feature map through a convolutional layer using a preset convolutional kernel to obtain a preliminary local feature map; performing convolution processing on the preliminary local feature map through a convolutional layer using the preset convolutional kernel to obtain a refined local feature map; performing feature extraction on the preliminary local feature map through a multi-scale feature extraction module to obtain a multi-scale local feature map; concatenating the refined local feature map with the multi-scale local feature map, and performing normalization processing using a batch normalization layer to obtain the local information feature map.

[0010] In some embodiments, the multi-scale feature extraction module includes a first convolutional block, a max-pooling layer, and a second convolutional block; the first convolutional block includes a batch normalization layer, a convolutional layer using a 1×1 convolutional kernel, and a convolutional layer using the preset convolutional kernel; the second convolutional block includes a convolutional layer using a 1×1 convolutional kernel and a convolutional layer using the preset convolutional kernel; the pooling kernels of the max-pooling layer include 5×5, 9×9, and 13×13.

[0011] In some embodiments, extracting features from the preliminary local feature map using a multi-scale feature extraction module to obtain a multi-scale local feature map includes: performing convolution processing on the preliminary local feature map using a first convolutional block to obtain a processed preliminary local feature map; performing pooling on the processed preliminary local feature map using a max pooling layer to obtain a first-scale local feature map, a second-scale local feature map, and a third-scale local feature map; concatenating the first-scale local feature map, the second-scale local feature map, and the third-scale local feature map to obtain a concatenated local feature map; and performing convolution processing on the concatenated local feature map using a second convolutional block to obtain the multi-scale local feature map.

[0012] In some embodiments, the segmentation output module includes a 1×1 convolutional block and at least one convolutional layer; and inputting the combined feature map into the segmentation output module for region segmentation to output a segmentation result includes: inputting the combined feature map into the 1×1 convolutional block for dimensionality reduction to obtain a dimensionality-reduced feature map; and inputting the dimensionality-reduced feature map into the at least one convolutional layer for convolution operation to output the segmentation result.

[0013] In a second aspect, this application provides an apparatus for training a brain perfusion image sequence segmentation model, comprising: a processor; and a memory storing program instructions for training the brain perfusion image sequence segmentation model, wherein when the program instructions are executed by the processor, the method described in the first aspect and the method described in several embodiments thereof are implemented.

[0014] In a third aspect, this application provides an apparatus for segmenting brain perfusion image sequences, comprising: a processor; and a memory storing program instructions for segmenting brain perfusion image sequences, wherein when the program instructions are executed by the processor, the apparatus performs the following operations: acquiring a brain perfusion image sequence of a patient; inputting the brain perfusion image sequence into a brain perfusion image sequence segmentation model trained according to the method described in the first aspect and its multiple embodiments for segmentation, so as to output a segmentation result.

[0015] In a fourth aspect, this application provides a computer-readable storage medium having stored program instructions thereon for training a brain perfusion image sequence segmentation model and / or for segmenting a brain perfusion image sequence, wherein when executed by a processor, the program instructions implement the operations implemented by the method and its multiple embodiments described in the first aspect and / or the device described in the third aspect.

[0016] The above-described scheme for training a brain perfusion image sequence segmentation model incorporates the temporal dynamics of the brain perfusion image sequence through a difference calculation module. This allows the model to capture differences in blood perfusion in brain tissue at different time points and quantify the dynamic changes in contrast agent distribution. Compared to traditional manual assessments that rely on static experience and existing methods that ignore temporal information, this mechanism can sensitively identify early micro-ischemic lesions, providing an objective temporal basis for accurate determination of the thrombolysis time window and avoiding treatment delays caused by lags in manual analysis. The wavelet transform module can perform multi-scale decomposition of the basic brain feature map, generating wavelet transform feature maps that include the overall brain region structure and refined edge information. This significantly enhances the boundary contrast between ischemic and infarcted areas and normal tissue, improving the model's segmentation accuracy for ischemic and infarcted areas. The local information extraction module, through pooling operations with pooling kernels of different sizes, can extract multi-scale local structural features, achieving fine-grained analysis of lesion areas and improving the model's ability to recognize complex morphological regions. The segmentation output module can fuse difference feature maps, wavelet transform feature maps, and local information features. Figure 3 Complementary features enable the model to learn multi-level information about time, space, and structure in brain perfusion image sequences, which can significantly improve the model's generalization performance.

[0017] In the application phase of the brain perfusion image sequence segmentation model, the device used to segment brain perfusion image sequences, with the aid of a trained brain perfusion image sequence segmentation model, can efficiently and accurately segment ischemic and infarcted areas. The device acquires the patient's brain perfusion image sequence and inputs it into the trained model, quickly outputting segmentation results. This not only significantly shortens segmentation time and improves segmentation efficiency but also avoids the subjectivity and uncertainty of manual segmentation. Based on a model trained on a large amount of data, the segmentation results are more objective and accurate, providing reproducible objective data support for clinical practice. Attached Figure Description

[0018] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, and the same or corresponding reference numerals denote the same or corresponding parts, wherein:

[0019] Figure 1 An exemplary structural diagram of a brain perfusion image sequence segmentation model according to an embodiment of this application is shown;

[0020] Figure 2 An exemplary structural diagram of the wavelet transform module according to an embodiment of this application is shown;

[0021] Figure 3 An exemplary structural diagram of the local information extraction module according to an embodiment of this application is shown;

[0022] Figure 4 An exemplary flowchart of a method for training a brain perfusion image sequence segmentation model according to an embodiment of this application is shown;

[0023] Figure 5 An exemplary structural block diagram of an electronic device for training a brain perfusion image sequence segmentation model according to an embodiment of this application is shown;

[0024] Figure 6 An exemplary flowchart of a method for segmenting brain perfusion image sequences according to an embodiment of this application is shown. Detailed Implementation

[0025] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] It should be understood that the terms "comprising" and "including" used in the specification and claims of this application indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0027] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this specification and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this specification and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0028] As used in this specification and claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0029] The specific embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0030] Figure 1 An exemplary structural diagram of a brain perfusion image sequence segmentation model according to an embodiment of this application is shown. Figure 1 As shown, the brain perfusion image sequence segmentation model of this application may include a difference calculation module 101, a basic feature extraction module 102, a wavelet transform module 103, a local information extraction module 104, and a segmentation output module 105.

[0031] The difference calculation module 101 is used to calculate the image difference between the first and second images at adjacent time points in a brain perfusion image sequence, using these differences as difference feature maps. The difference feature maps record the dynamic changes in blood flow at different time points. In practical applications, the number of tomographic images in different brain perfusion image sequences varies, resulting in different numbers of difference feature maps. In practical applications, a brain perfusion image sequence containing N tomographic images can yield N-1 difference feature maps through calculation.

[0032] The basic feature extraction module 102 is used to extract basic features such as edges, textures, and basic structures from brain perfusion image sequences to obtain a basic brain feature map, providing "raw materials" for the subsequent wavelet transform module 103 and local information extraction module 104. The basic brain feature map is not a visually visible brain image, but rather a "coded map" that a computer can understand—where there are vascular abnormalities, where blood flow changes are significant, these are all converted into digital signals and recorded. In one embodiment, the basic feature extraction module 102 can use a classic convolutional neural network (CNN), such as a ResNet residual network or a DenseNet densely connected network, as the backbone network.

[0033] The aforementioned wavelet transform module 103 takes the aforementioned basic brain feature map as input and performs a wavelet transform on the basic brain feature map to obtain a wavelet transform feature map. The wavelet transform feature map contains the overall outline of the brain region and subtle edges in various directions. (See reference here.) Figure 2 The structure of the wavelet transform module 103 in this embodiment will be understood. For example... Figure 2 As shown, the wavelet transform module 103 may include convolutional layers using preset convolutional kernels (such as...). Figure 2 The diagram shows Conv k=3), and at least one convolutional layer using a 1×1 kernel (e.g., Figure 2 The Conv k=1 shown in the figure and the batch normalization layer (such as Figure 2 (BN shown in the figure). It is understood that those skilled in the art can select the specific size of the preset convolution kernel here and in the following places according to actual needs. For example, the size of the preset convolution kernel can be selected as 3×3, 5×5 or 7×7, etc., and the embodiments of this application do not make specific limitations. In addition, the splicing or combination operations in the following places correspond to Concat in the corresponding figures.

[0034] Based on this, in the wavelet transform module 103, performing a wavelet transform on the basic brain feature map to obtain a wavelet transformed feature map may include the following operations: convolving the basic brain feature map through a convolutional layer using a first convolutional kernel to obtain a processed basic brain feature map. Then, a wavelet transform can be performed on the processed basic brain feature map (corresponding to...). Figure 2The DWT (Digital Wavelength Wound) is used to enable multi-scale analysis of the processed basic brain feature map, obtaining low-frequency information (L) and high-frequency information (H / V / D). Specifically, high-frequency information can include horizontal information (H), vertical information (V), and diagonal information (D). Typically, low-frequency information (L) records the overall structure of the image (e.g., the approximate shape of the brain, the distribution of large blood vessels). Horizontal information (H) records horizontal edges in the image (e.g., the horizontal texture of the cerebral cortex, the horizontal orientation of blood vessels); vertical information (V) records vertical edges in the image (e.g., the vertical boundaries of sulci and gyri); and diagonal information (D) records diagonal edges in the image (e.g., oblique vascular branches, the oblique boundaries of irregular lesions).

[0035] Next, the horizontal information H, vertical information V, and diagonal information D are combined and processed through a convolutional layer with a 1×1 kernel and a batch normalization layer to obtain the processed high-frequency information. Similarly, the low-frequency information L is processed through a convolutional layer with a 1×1 kernel to obtain the processed low-frequency information. Finally, by combining the processed high-frequency and low-frequency information, the wavelet transform feature map can be obtained.

[0036] The aforementioned local information extraction module 104, taking the aforementioned basic brain feature map as input, can extract multi-scale local information feature maps through pooling operations with pooling kernels of different sizes. These local information feature maps magnify local details of brain regions, especially the edges and surrounding structures of infarcted areas, enabling fine-grained analysis of lesion areas and improving the model's ability to recognize complex morphological regions. (See reference here.) Figure 3 The structure of the local information extraction module 104 in this application embodiment will be understood. For example... Figure 3 As shown, the local information extraction module 104 may include at least one convolutional layer using a preset convolutional kernel (such as...). Figure 3 The diagram shows Conv k=3), a multi-scale feature extraction module, and a batch normalization layer (e.g., ...). Figure 3 (BN shown).

[0037] Furthermore, the multi-scale feature extraction module may include a first convolutional block, a max pooling layer (such as...) Figure 3 The MaxPool shown in the diagram represents the first convolutional block, which may include a batch normalization layer (such as...). Figure 3 The diagram shows BN), and convolutional layers using 1×1 kernels (such as...). Figure 3 The diagram shows Conv k=1) and convolutional layers using preset kernels (such as...). Figure 3 In the diagram, Conv k = 3); the second convolutional block can include convolutional layers using 1×1 kernels (e.g., Figure 3 The diagram shows Conv k=1) and convolutional layers using preset kernels (such as...). Figure 3Conv k = 3 (as shown in the figure).

[0038] Figure 3 In this context, 5×5, 9×9, and 13×13 represent the sizes of the pooling kernels (sliding windows) in a max-pooling layer. Taking 5×5 as an example, it means sliding a 5x5 window across the feature map, taking the maximum value of pixels within the window as the output each time. The same logic applies to 9×9 and 13×13, only the number of rows and columns differs. Smaller pooling kernels (such as 5×5) can focus on smaller local regions, preserving detailed edge and texture information, and capturing fine features such as infarct edges. Larger pooling kernels (such as 13×13) can cover a larger area, extracting more macroscopic structural information and grasping the overall brain region layout. By combining pooling kernels of different sizes, the model can simultaneously focus on details and the overall picture, enriching feature representation and improving adaptability to complex brain structures.

[0039] Based on this, in the local information extraction module 104, extracting information from the basic brain feature map to obtain a local information feature map may include the following operations: convolving the basic brain feature map through a convolutional layer using a preset convolutional kernel to obtain a preliminary local feature map; convolving the preliminary local feature map through a convolutional layer using a preset convolutional kernel to obtain a refined local feature map; extracting features from the preliminary local feature map through a multi-scale feature extraction module to obtain a multi-scale local feature map; concatenating the refined local feature map with the multi-scale local feature map and normalizing it using a batch normalization layer to obtain a local information feature map.

[0040] Further, extracting features from the preliminary local feature map using a multi-scale feature extraction module to obtain a multi-scale local feature map may include the following operations: convolving the preliminary local feature map through a first convolutional block to obtain a processed preliminary local feature map; pooling the processed preliminary local feature map through the max pooling layer to obtain a first-scale local feature map, a second-scale local feature map, and a third-scale local feature map; concatenating the first-scale local feature map, the second-scale local feature map, and the third-scale local feature map to obtain a concatenated local feature map; and convolving the concatenated local feature map through a second convolutional block to obtain a multi-scale local feature map.

[0041] Continue to refer to Figure 1 The segmentation output module 105 in this embodiment may include a 1×1 convolutional block and at least one convolutional layer (e.g., ...). Figure 1The diagram shows Conv. The segmentation output module 105 takes the combined feature map constructed by combining the aforementioned difference feature map, wavelet transform feature map, and local information feature map as input, and outputs the segmentation result. In this embodiment, the 1×1 convolutional block can be composed of 1×1 convolution, batch normalization (BN) layer, and rectified linear unit (ReLU) activation function. The convolutional layer can extract more complex feature combinations (such as blood flow changes, edges in a specific direction, or local structural abnormalities at a certain location), and these combined features can more accurately locate ischemic and infarct areas.

[0042] Here, the 1×1 convolution in the 1×1 convolution block uses a 1×1 kernel, meaning that in each convolution operation, the kernel only operates on a single pixel and its corresponding channels in the input feature map. By using a 1×1 kernel for convolution on the feature map, high-dimensional feature maps can be mapped to a low-dimensional space, reducing the number of channels and removing redundant information, thereby reducing model complexity. Batch Normalization (BN) layers normalize the input data, making the data distribution more stable and helping to accelerate model convergence. BN layers also have a regularization effect, improving the model's generalization ability and reducing the risk of overfitting. The ReLU (Rectified Linear Unit) activation function introduces non-linearity to the model. After 1×1 convolution and batch normalization, the ReLU function sets values ​​less than 0 to 0 and keeps values ​​greater than 0 unchanged, thus increasing the model's non-linear expressive power. This allows the model to learn more complex feature patterns, improving the performance of tasks such as classification and regression.

[0043] Based on the above structure of the segmentation output module 105, when performing region segmentation in the segmentation output module 105 to output the segmentation result, the following operations can be performed: inputting the combined feature map into a 1×1 convolutional block for dimensionality reduction to obtain a dimensionality-reduced feature map; inputting the dimensionality-reduced feature map into at least one convolutional layer for convolution operation to output the segmentation result.

[0044] In practice, the output of at least one convolutional layer is not the final segmentation result. The following steps can be taken to output the segmentation result: Based on the output of the convolutional layer, the softmax function can be used to calculate the probability that each pixel belongs to an ischemic region, an infarcted region, or normal tissue. The category with the highest probability is the final judgment result. Next, based on the final judgment result, a segmentation map can be output as the segmentation result. In the segmentation map, different colors can be used to represent different regions to clearly show the location of ischemia and infarction.

[0045] Figure 4An exemplary flowchart of a method 400 for training a brain perfusion image sequence segmentation model according to an embodiment of this application is shown. It is understood that method 400 can be executed by any suitable device with data processing capabilities, such as, but not limited to, terminal devices, processors, and servers.

[0046] Based on this, such as Figure 4 As shown, in step S401, method 400 can acquire a brain perfusion image sequence and annotate the boundaries of the ischemic and infarcted regions in each image of the brain perfusion image sequence to obtain annotation information. In step S402, method 400 can input the brain perfusion image sequence into the difference calculation module for difference calculation to obtain a difference feature map. In step S403, method 400 can input the brain perfusion image sequence into the basic feature extraction module for feature extraction to obtain a basic brain feature map.

[0047] Next, in step S404, method 400 can input the basic brain feature map into the wavelet transform module to perform wavelet transform and obtain a wavelet transform feature map. In step S405, method 400 can input the basic brain feature map into the local information extraction module for information extraction and obtain a local information feature map. In step S406, method 400 can combine the difference feature map, wavelet transform feature map, and local information feature map to construct a combined feature map, and input the combined feature map into the segmentation output module for region segmentation to output the segmentation result. Finally, in step S407, method 400 can update the parameters of the brain perfusion image sequence segmentation model based on the annotation information and the segmentation result to train it.

[0048] In step S401 above, brain perfusion image sequences of patients with acute ischemic stroke can be retrospectively collected. Brain perfusion image sequences refer to multiple CT scans of the same region of interest within a short period, acquiring dynamic imaging data of that region after contrast agent injection. These data reflect the flow and distribution of the contrast agent in brain tissue. Furthermore, when annotating the boundaries between ischemic and infarcted areas, a radiologist can manually annotate the images, clearly marking the precise boundaries of the ischemic and infarcted regions. Additionally or optionally, another radiologist can review the annotations to ensure their accuracy.

[0049] In this embodiment, to improve the training efficiency and save time of the brain perfusion image sequence segmentation model, the aforementioned brain perfusion image sequence can be preprocessed by denoising, standardizing, and normalizing before being used for model training. This preprocessed brain perfusion image sequence is then used for model training. This reduces unnecessary computation while ensuring sufficient information, allowing for more efficient use of computing resources and accelerating the entire training process.

[0050] During the acquisition of raw brain perfusion image sequences, a significant amount of noise is introduced due to factors such as electronic noise from the equipment itself, minor patient movements during scanning, and uneven contrast agent distribution. This noise negatively impacts feature extraction in subsequent deep learning models, causing them to misidentify noise as features and thus affecting segmentation accuracy. Therefore, methods such as median filtering, Gaussian filtering, or nonlocal mean filtering can be used to filter brain perfusion image sequences to remove noise while better preserving edge and detail information. This is particularly important for protecting the complex brain tissue structures in brain perfusion image sequences.

[0051] Furthermore, the imaging parameters (such as tube voltage, tube current, and slice thickness) of different CT devices vary, and the scanning conditions of the same device may also fluctuate at different times. This can lead to significant differences in grayscale range and contrast in the acquired brain perfusion image sequences. For example, two CT devices from different manufacturers scanning the same patient's brain may produce brain perfusion image sequences with completely different grayscale distributions. Standardization aims to eliminate the impact of these device differences, giving the images a uniform format and characteristics. Common standardization methods include mapping the image's grayscale values ​​to a fixed range, such as adjusting the grayscale values ​​of all pixels to the [0,1] interval, or converting the grayscale values ​​of all pixels to a data distribution with a mean of 0 and a standard deviation of 1. This ensures that brain perfusion image sequences from different sources are comparable in subsequent processing, providing stable and consistent input for model training.

[0052] The main purpose of the aforementioned normalization is to adjust the numerical range of image data to a specific interval (usually [0, 1] or [-1, 1]), making the data have a similar scale. During deep learning model training, if the scale difference of the input data is too large, it can lead to slower model convergence and even problems such as vanishing or exploding gradients. For example, the CT values ​​of different tissues in brain perfusion image sequences vary greatly. Without normalization, the model may overemphasize areas with large CT values ​​(such as bones) during learning, while ignoring brain tissue areas with smaller CT values ​​that are crucial for stroke diagnosis. By normalizing and compressing all pixel values ​​to a uniform scale, the training speed of the model can be accelerated, the stability and training efficiency of the model can be improved, and the model can more effectively learn the characteristics of ischemic and infarcted areas.

[0053] In step S402, the difference feature map obtained by direct subtraction may contain a lot of noise (such as machine noise and interference from slight patient movements), which can interfere with the model's judgment of real blood flow changes. Therefore, the difference image can be denoised to eliminate these noises and preserve the real blood flow change area.

[0054] In step S407, method 400 may perform the following operations to update the parameters of the brain perfusion image sequence segmentation model based on the annotation information and segmentation results: determining the loss value using the Dice loss function and the cross-entropy loss function based on the annotation information and segmentation results; and updating the parameters of the brain perfusion image sequence segmentation model using a preset optimization algorithm based on the loss value. Here, the preset optimization algorithm may be Adaptive Moment Estimation (Adam) or Stochastic Gradient Descent (SGD), etc.

[0055] In cerebral ischemia and infarction segmentation, the lesion region (positive sample) typically accounts for only 5%-15% of the entire image, and traditional loss functions are easily dominated by normal tissue (negative sample). The Dice loss function, by maximizing the overlap between the predicted and real regions, forces the model to focus on small targets. The cross-entropy loss function penalizes the classification error of each pixel individually, making it suitable for refining boundary details (such as the blurred transition area between the ischemic penumbra and normal tissue). For example, if a pixel is actually an ischemic area and the model predicts a probability of 0.6 (correct classification but low confidence), the CE loss will force an increase in the prediction probability of that pixel through gradient backpropagation. In the embodiments of this application, by combining the Dice loss function and the cross-entropy loss function, the model can simultaneously satisfy both region integrity and pixel accuracy.

[0056] In some implementation scenarios, cross-validation can be used to evaluate the model's generalization ability. Specifically, when implementing cross-validation, the labeled dataset can be divided into k non-overlapping subsets. Each time, k-1 subsets are used for training, and one subset is used for validation, repeated k times. This ensures that every sample is used for both training and validation, allowing the model to adapt to the diversity of real-world clinical data and improving its generalization ability. Furthermore, in scenarios where medical data annotation is costly (such as in this application where each tomographic image requires manual annotation by experts), cross-validation can maximize the use of limited data, reduce the sample consumption of independent test sets, and improve data utilization.

[0057] In the embodiments of this application, when implementing cross-validation, the Dice coefficient or Hausdorff distance can be used as evaluation metrics to quantitatively evaluate the segmentation results. The Dice coefficient is used to measure the degree of overlap between the model's predicted region and the actual labeled region. It can reflect the model's overall ability to capture lesion regions and is particularly suitable for evaluating the segmentation effect of small targets (such as small ischemic lesions). The Hausdorff distance is used to evaluate the maximum error between the boundary of the predicted region and the actual labeled region, that is, the distance between the farthest points on their contours. It can detect subtle deviations in the model's boundary localization (such as missing or extra labeled edge pixels).

[0058] In this embodiment, to improve the stability of the brain perfusion image sequence segmentation model and prevent overfitting, data augmentation operations such as rotation, flipping, scaling, translation, brightness / contrast adjustment, and adding Gaussian noise are performed on the original brain perfusion image sequence. This allows the training dataset to be artificially expanded without relying on additional clinical data. For example, randomly rotating or scaling the labeled ischemic / infarct area tomographic images can simulate real-world clinical scenarios such as different patient head positions and CT scan slice thickness differences. This enables the model to learn more general anatomical features, avoiding overfitting caused by limited training data, and thus maintaining stable performance in brain perfusion image sequence segmentation for unknown patients.

[0059] The above combination Figure 4 This paper describes a method for training a segmentation model of brain perfusion image sequences. By introducing a difference calculation module, the temporal dynamics of the brain perfusion image sequences (i.e., difference feature maps) are incorporated, enabling the model to capture differences in blood perfusion in brain tissue at different time points and quantify the dynamic changes in contrast agent distribution. Compared to traditional manual assessments that rely on static experience and existing methods that ignore temporal information, this mechanism can sensitively identify early micro-ischemic lesions (such as ischemic penumbras with slow but not completely interrupted blood flow), providing an objective temporal basis for accurate determination of the thrombolysis time window and avoiding treatment delays caused by the lag in manual analysis.

[0060] Furthermore, the wavelet transform module can perform multi-scale decomposition (low-frequency contour + high-frequency edge) on the basic brain feature map, generating wavelet transform feature maps that contain overall brain region structure and refined edge information. This significantly enhances the boundary contrast between ischemic and infarcted areas and normal tissue, improving the model's segmentation accuracy for ischemic and infarcted areas. The local information extraction module, through pooling operations with pooling kernels of different sizes, can extract multi-scale local structural features, achieving fine-grained analysis of lesion areas and improving the model's ability to recognize complex morphological regions. The segmentation output module can fuse difference feature maps, wavelet transform feature maps, and local information features. Figure 3 Complementary features enable the model to learn multi-level information about time, space, and structure in brain perfusion image sequences, which can significantly improve the model's generalization performance.

[0061] Next, combine Figure 5 An exemplary description is provided of an electronic device 500 for training a brain perfusion image sequence segmentation model, as provided in embodiments of this application. Figure 5 As shown, the electronic device 500 in this application embodiment may include a processor 501, a memory 502, and a communication bus 503.

[0062] In specific embodiments, the processor 501 described above can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), CPU, controller, microcontroller, and microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and this embodiment does not specifically limit it.

[0063] In this embodiment, the communication bus 503 is used to establish communication between the processor 501 and the memory 502; the memory 502 stores program instructions for training a brain perfusion image sequence segmentation model; when the processor 501 executes the program instructions stored in the memory 502, it implements the combination of this application. Figure 4 The method described is for training a brain perfusion image sequence segmentation model.

[0064] The above combination Figure 5An electronic device for training a brain perfusion image sequence segmentation model is described. It should be understood that the device structure or architecture described herein is merely exemplary, and the implementation methods and entities described herein are not limited thereto, but can be modified without departing from the spirit of this application.

[0065] It is understood that the description of the various embodiments in this disclosure emphasizes the differences between the various embodiments, and the similarities or corresponding parts can be referred to each other. For the purpose of brevity, this disclosure will not go into detail.

[0066] Furthermore, this application also provides a device for segmenting brain perfusion image sequences. The structure or architecture of this device, in conjunction with the foregoing description... Figure 5 The electronic device 500 described for training a brain perfusion image sequence segmentation model has the same structure or framework. For the sake of brevity, this disclosure will not elaborate further. In the device for segmenting brain perfusion image sequences, a memory stores program instructions for segmenting the brain perfusion image sequences, which, when executed by a processor, enable the device to achieve the following combination. Figure 6 The method 600 described is for segmenting brain perfusion image sequences.

[0067] Figure 6 An exemplary flowchart of a method 600 for segmenting a brain perfusion image sequence according to an embodiment of this application is shown. It is understood that method 600 can be executed by any suitable device with data processing capabilities, such as, but not limited to, processors, terminal devices, and servers.

[0068] like Figure 6 As shown, in step S601, method 600 can acquire the patient's brain perfusion image sequence. Next, in step S602, method 600 can input the brain perfusion image sequence into a trained brain perfusion image sequence segmentation model for segmentation, and output the segmentation result. It is understood that the trained brain perfusion image sequence segmentation model here refers to the model using the method described in this application. Figure 4 The method described herein for training a brain perfusion image sequence segmentation model has been successfully trained.

[0069] It is understandable that, in step S602 above, before inputting the brain perfusion image sequence into the segmentation model, preprocessing operations (such as denoising, standardization, and normalization) that are completely consistent with those in the training data acquisition process are usually required to ensure the accuracy of the segmentation results.

[0070] The above combination Figure 6A method for segmenting brain perfusion image sequences is described. By acquiring brain perfusion image sequences from patients and inputting them into a trained model, segmentation results can be output quickly. This not only significantly reduces segmentation time and improves segmentation efficiency but also avoids the subjectivity and uncertainty of manual segmentation. Based on a model trained on a large amount of data, the segmentation results are more objective and accurate, providing reproducible objective data support for clinical practice.

[0071] Based on the foregoing description in conjunction with the accompanying drawings, those skilled in the art will understand that the embodiments of this application can also be implemented by software programs. Therefore, this application also provides a computer-readable storage medium. This computer-readable storage medium stores program instructions for training a brain perfusion image sequence segmentation model and / or for segmenting brain perfusion image sequences, which can be used to implement the embodiments of this application. Figure 4 The described method or combination for training a brain perfusion image sequence segmentation model Figure 6 The method described is for segmenting brain perfusion image sequences.

[0072] It should be noted that although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. On the contrary, the steps depicted in the flowchart can be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0073] While numerous embodiments of this application have been shown and described herein, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will arise for those skilled in the art without departing from the spirit and intent of this application. It should be understood that various alternatives to the embodiments of this application described herein may be employed in the practice of this application. The appended claims are intended to define the scope of protection of this application and therefore cover equivalents or alternatives within the scope of these claims.

[0074] The collection and acquisition of various data in this application comply with relevant laws and regulations and are authorized by the data providers. Any organization or individual that needs to obtain external data shall obtain authorization in accordance with the law and ensure data security, and shall not illegally collect, use, process, or transmit unauthorized or unprotected data, nor shall it illegally buy, sell, provide, or disclose unauthorized or unprotected data.

Claims

1. A method for training a brain perfusion image sequence segmentation model, wherein the brain perfusion image sequence segmentation model includes a difference calculation module, a basic feature extraction module, a wavelet transform module, a local information extraction module, and a segmentation output module; the method includes: A brain perfusion image sequence is acquired, and the boundaries between the ischemic region and the infarct region are marked in each image of the brain perfusion image sequence to obtain annotation information; The brain perfusion image sequence is input into the difference calculation module for difference calculation to obtain a difference feature map; the difference feature map is a feature map composed of the image differences between adjacent time points in the brain perfusion image sequence; The brain perfusion image sequence is input into the basic feature extraction module for feature extraction to obtain a basic brain feature map; The basic brain feature map is input into the wavelet transform module to perform wavelet transform, so as to obtain the wavelet transform feature map; The basic brain feature map is input into the local information extraction module for information extraction to obtain a local information feature map; The difference feature map, the wavelet transform feature map, and the local information feature map are combined to construct a combined feature map, and the combined feature map is input into the segmentation output module for region segmentation to output the segmentation result; Based on the annotation information and the segmentation results, the parameters of the brain perfusion image sequence segmentation model are updated to enable training. Performing wavelet transform in the wavelet transform module to obtain wavelet transform feature maps includes: The brain basic feature map is processed by convolutional layer using a preset convolutional kernel to obtain the processed brain basic feature map. Wavelet transform is performed on the processed basic brain feature map to obtain low-frequency information and high-frequency information, wherein the high-frequency information includes horizontal information, vertical information and diagonal information; The horizontal information, the vertical information, and the diagonal information are combined, and then processed using 1 The convolutional layers with 1 convolutional kernel and the batch normalization layer are used to process the high-frequency information after processing. The low-frequency information is transmitted via 1 The convolutional layer with 1 convolutional kernel is used to process the low-frequency information after processing. The processed high-frequency information is combined with the processed low-frequency information to obtain the wavelet transform feature map; The difference calculation module performs difference calculations to obtain a difference feature map, including: Based on the images at different time points in the brain perfusion image sequence, the image difference between the first and second images at adjacent time points is calculated as a difference feature map; The difference feature map is denoised to obtain a denoised difference feature map.

2. The method according to claim 1, wherein, The local information extraction module performs information extraction to obtain a local information feature map, including: The basic brain feature map is processed by convolutional layer using a preset convolutional kernel to obtain a preliminary local feature map; The preliminary local feature map is convolved through a convolutional layer using the preset convolutional kernel to obtain a deeper local feature map; The preliminary local feature map is processed by a multi-scale feature extraction module to extract features, thereby obtaining a multi-scale local feature map. The enhanced local feature map is concatenated with the multi-scale local feature map, and then normalized using a batch normalization layer to obtain the local information feature map.

3. The method according to claim 2, wherein, The multi-scale feature extraction module includes a first convolutional block, a max-pooling layer, and a second convolutional block; the first convolutional block includes a batch normalization layer, a max-pooling layer, and a second convolutional block. A convolutional layer with one convolutional kernel and a convolutional layer using the preset convolutional kernel; the second convolutional block includes a convolutional layer using one convolutional kernel. A convolutional layer with 1 convolutional kernel and a convolutional layer using the preset convolutional kernel; the pooling kernel of the max pooling layer includes 5... 5, 9 9 and 13 13.

4. The method according to claim 3, wherein, The preliminary local feature map is processed by a multi-scale feature extraction module to extract features, thereby obtaining a multi-scale local feature map, including: The preliminary local feature map is convolved through the first convolutional block to obtain the processed preliminary local feature map. The processed preliminary local feature map is pooled through the max pooling layer to obtain a first-scale local feature map, a second-scale local feature map, and a third-scale local feature map. The first-scale local feature map, the second-scale local feature map, and the third-scale local feature map are stitched together to obtain a stitched local feature map. The stitched local feature map is convolved through the second convolution block to obtain the multi-scale local feature map.

5. The method according to claim 1, wherein, The segmentation output module includes 1 Convolutional blocks and at least one convolutional layer; The combined feature map is input into the segmentation output module for region segmentation, and the segmentation result is output, including: The combined feature map is input to the 1 Dimensionality reduction is performed in the convolutional block to obtain the dimensionality-reduced feature map; The reduced feature map is input into the at least one convolutional layer for convolution operation to output the segmentation result.

6. An apparatus for training a brain perfusion image sequence segmentation model, characterized in that, include: processor; as well as A memory storing program instructions for training a brain perfusion image sequence segmentation model, which, when executed by a processor, cause the method according to any one of claims 1-5 to be implemented.

7. An apparatus for segmenting brain perfusion image sequences, characterized in that, include: processor; A memory storing program instructions for segmenting brain perfusion image sequences, which, when executed by the processor, cause the device to perform the following operations: Obtain brain perfusion image sequences from the patient; The brain perfusion image sequence is input into a brain perfusion image sequence segmentation model trained according to any one of claims 1-5 for segmentation, so as to output the segmentation result.

8. A computer-readable storage medium having stored thereon program instructions for training a brain perfusion image sequence segmentation model and / or for segmenting a brain perfusion image sequence, wherein the program instructions, when executed by a processor, implement the operation implemented by the method according to any one of claims 1-5 and / or the device according to claim 6.

Images