Method and product for training brain perfusion image sequence segmentation model

Abstract

The invention discloses a method and product for training a brain perfusion image sequence segmentation model, and the method comprises the steps: carrying out the marking of the boundary of an ischemia and infarction region in a brain perfusion image sequence, and obtaining the marking information; the image sequence is input into a difference calculation module for difference calculation, and a difference feature map is obtained; obtaining a brain basic feature map; inputting the brain basic feature map into a wavelet transform module, and performing wavelet transform to obtain a wavelet transform feature map; inputting the brain basic feature map into a local information extraction module for information extraction to obtain a local information feature map; the difference value feature map, the wavelet transform feature map and the local information feature map are combined and then input into a segmentation output module for region segmentation, and a segmentation result is output; and on the basis of the labeling information and the segmentation result, updating parameters of the brain perfusion image sequence segmentation model, and training the brain perfusion image sequence segmentation model. By utilizing the scheme of the invention, the segmentation precision of the ischemia and infarction regions can be improved.

Description

Technical Field

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

[0002] Acute ischemic stroke (AIS), a serious threat to human health and life, is caused by the sudden interruption of cerebral blood flow and is characterized by high rates of disability and mortality. Rapid and accurate assessment of the extent of ischemia and infarcted areas is crucial for developing effective treatment strategies and improving patient outcomes. The emergence of computed tomography perfusion (CTP) technology, which can visually visualize brain hemodynamics, provides a crucial basis for the diagnosis and treatment of AIS. However, traditional manual assessment methods, which rely primarily on the professional experience of radiologists, have numerous drawbacks in practice. Firstly, the assessment process is highly subjective, and judgment criteria vary between physicians, leading to inconsistent results. Secondly, manual analysis is time-consuming, making it difficult to meet the urgent need for emergency treatment of AIS, where time is of the essence, and potentially delaying optimal treatment opportunities.

[0003] In recent years, deep learning has achieved significant breakthroughs in medical image analysis, particularly in computer-aided diagnosis and medical image segmentation. Automatic segmentation methods based on deep learning, with their powerful feature extraction and analysis capabilities, can effectively improve the accuracy of ischemic region identification, reduce human error, and significantly shorten image analysis time. However, existing deep learning methods still suffer from insufficient capture of detailed features of ischemic and infarcted regions and weak multi-dimensional information fusion capabilities when processing CTP images of acute ischemic stroke. As a result, segmentation accuracy and stability remain to be improved, making it difficult to meet the stringent clinical requirements for accurate assessment.

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

[0005] In order to at least solve one or more of the above-mentioned technical problems, the present application proposes a solution for segmenting a brain perfusion image sequence in the following aspects.

[0006] In a first aspect, the present application provides 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: acquiring a brain perfusion image sequence, and in each image of the brain perfusion image sequence, marking the boundaries of the ischemic area and the infarct area to obtain marking information; inputting the brain perfusion image sequence into the difference calculation module to perform difference calculation to obtain a difference feature map; inputting the brain perfusion image sequence into the basic feature extraction module to perform feature extraction to obtain a basic brain feature map; input the basic brain feature map into the wavelet transform module for wavelet transform to obtain a wavelet transform feature map; input the basic brain feature map into the local information extraction module for information extraction to obtain a local information feature map; combine the difference feature map, the wavelet transform feature map and the local information feature map to construct a combined feature map, and input the combined feature map into the segmentation output module for regional segmentation to output a segmentation result; based on the annotation information and the segmentation result, update the parameters of the brain perfusion image sequence segmentation model to achieve training thereof.

[0007] In some embodiments, performing difference calculation in the difference calculation module to obtain a difference feature map includes: calculating the image difference between the first image and the second image at adjacent time points as a difference feature map based on the images at different time points in the cerebral perfusion image sequence; 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 basic brain feature map through a convolution layer using a preset convolution kernel to obtain a processed basic brain feature map; performing wavelet transform 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; combining the horizontal information, the vertical information and the diagonal information, and then processing them through a convolution layer using a 1×1 convolution kernel and a batch normalization layer to obtain the processed high-frequency information; processing the low-frequency information through a convolution layer using a 1×1 convolution kernel to obtain the processed low-frequency information; 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 is performed in the local information extraction module to obtain a local information feature map, including: convolution processing of the basic brain feature map through a convolution layer using a preset convolution kernel to obtain a preliminary local feature map; convolution processing of the preliminary local feature map through a convolution layer using the preset convolution kernel to obtain a deepened local feature map; feature extraction of the preliminary local feature map through a multi-scale feature extraction module to obtain a multi-scale local feature map; splicing the deepened local feature map with the multi-scale local feature map, and normalizing them 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 convolution block, a maximum pooling layer and a second convolution block; the first convolution block includes a batch normalization layer, a convolution layer using a 1×1 convolution kernel and a convolution layer using the preset convolution kernel; the second convolution block includes a convolution layer using a 1×1 convolution kernel and a convolution layer using the preset convolution kernel; the pooling kernels of the maximum pooling layer include 5×5, 9×9 and 13×13.

[0011] In some embodiments, subjecting the preliminary local feature map to feature extraction via a multi-scale feature extraction module to obtain a multi-scale local feature map includes: subjecting the preliminary local feature map to convolution processing via the first convolution block to obtain a processed preliminary local feature map; subjecting the processed preliminary local feature map to pooling via the maximum pooling layer to obtain a first-scale local feature map, a second-scale local feature map, and a third-scale local feature map; splicing the first-scale local feature map, the second-scale local feature map, and the third-scale local feature map to obtain a spliced ​​local feature map; subjecting the spliced ​​local feature map to convolution processing via the second convolution block to obtain the multi-scale local feature map.

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

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

[0014] In a third aspect, the present application provides a device for segmenting a cerebral perfusion image sequence, comprising: a processor; a memory storing program instructions for segmenting a cerebral perfusion image sequence, wherein when the program instructions are executed by the processor, the device performs the following operations: acquiring a patient's cerebral perfusion image sequence; inputting the cerebral perfusion image sequence into a cerebral perfusion image sequence segmentation model trained according to the method described in the first aspect and multiple embodiments thereof for segmentation, thereby outputting a segmentation result.

[0015] In a fourth aspect, the present application provides 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. When the program instructions are executed by a processor, the operations implemented by the method described in the first aspect and its multiple embodiments and / or the device described in the third aspect are implemented.

[0016] The scheme provided above for training the brain perfusion image sequence segmentation model introduces the temporal dynamic characteristics of the brain perfusion image sequence through the difference calculation module, so that the model can capture the differences in blood perfusion of brain tissue at different time points and quantify the dynamic changes in the distribution of contrast agents. Compared with the defects of traditional manual evaluation that relies on static empirical judgment and the existing methods that ignore temporal information, this mechanism can sensitively identify early small ischemic lesions, provide an objective basis in the time dimension for the accurate judgment of the thrombolysis time window, and avoid treatment delays caused by manual analysis lags. The wavelet transform module can perform multi-scale decomposition of the basic brain feature map and generate a wavelet transform feature map containing the overall structure of the brain region and refined edge information, significantly enhancing the boundary contrast between the ischemic and infarcted areas and normal tissues, and improving the segmentation accuracy of the model for ischemic and infarcted areas. The local information extraction module can extract multi-scale local structural features through pooling operations with pooling kernels of different sizes, realize fine-grained analysis of the lesion area, and improve the model's recognition ability for complex morphological areas. The segmentation output module can fuse the difference feature map, wavelet transform feature map and local information feature. Figure 3 The complementary features enable the model to deeply learn the multi-level information of time, space and structure in brain perfusion image sequences, which can significantly improve the generalization performance of the model.

[0017] During the application phase of the cerebral perfusion image sequence segmentation model, the device used to segment cerebral perfusion image sequences leverages the trained cerebral perfusion image sequence segmentation model to efficiently and accurately segment ischemic and infarcted areas. By acquiring the patient's cerebral perfusion image sequence and inputting it into the trained model, the device rapidly outputs segmentation results. This not only significantly shortens segmentation time and improves efficiency, but also avoids the subjectivity and uncertainty inherent in manual segmentation. Based on a model trained on a large amount of data, segmentation results are more objective and accurate, providing reproducible, objective data support for clinical practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The above and other objects, features and advantages of the exemplary embodiments of the present application will become readily understood by reading the detailed description below with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent 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 the present application is shown;

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

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

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

[0024] Figure 6 An exemplary flowchart of a method for segmenting a brain perfusion image sequence according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0025] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

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

[0027] It should also be understood that the terms used in this specification are for the purpose of describing specific embodiments only and are not intended to limit this 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 further understood that the term "and / or" as used in this specification and claims refers to any and all possible combinations of one or more of the associated listed items, including and including these combinations.

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

[0029] The specific implementation of the present application will be described in detail below with reference to the accompanying drawings.

[0030] Figure 1 FIG. 4 shows an exemplary structural diagram of a brain perfusion image sequence segmentation model according to an embodiment of the present application. Figure 1 As shown, the brain perfusion image sequence segmentation model of the present application may include a difference calculation module 101 , a basic feature extraction module 102 , a wavelet transformation 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 image and the second image at adjacent time points based on the images at different time points in the cerebral perfusion image sequence as a difference feature map. The difference feature map records the dynamic changes in blood flow at different time points. In actual applications, the number of calculated difference feature maps will also vary depending on the number of tomographic images in different cerebral perfusion image sequences. In actual applications, for a cerebral perfusion image sequence containing N tomographic images, N-1 difference feature maps can be obtained through calculation.

[0032] The basic feature extraction module 102 is used to extract basic features such as edges, textures and basic structures from the brain perfusion image sequence to obtain a basic brain feature map, which provides "raw materials" for the subsequent wavelet transform module 103 and the local information extraction module 104. The basic brain feature map is not a brain image visible to the naked eye, but a "code map" that can be understood by the computer - where there are vascular abnormalities and where there are obvious changes in blood flow are 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 brain basic feature map as input and performs wavelet transform on the brain basic feature map to obtain a wavelet transform feature map. The wavelet transform feature map contains the overall outline of the brain area and the subtle edges in all directions. Figure 2 Understand the structure of the wavelet transform module 103 in the embodiment of the present application. Figure 2 As shown, the wavelet transform module 103 may include a convolution layer (such as Figure 2 Conv k=3 shown in ), at least one convolutional layer using a 1×1 convolution kernel (such as Figure 2 Conv k=1) and batch normalization layers (as shown in Figure 2 ). It is understood that those skilled in the art can select the specific size of the preset convolution kernel here and in the following according to actual needs. For example, the preset convolution size can be selected as 3×3, 5×5, or 7×7, etc., and this embodiment of the application does not specifically limit this. In addition, the splicing or combination operations in the following correspond to the Concat in the corresponding figures.

[0034] Based on this, in the wavelet transform module 103, performing wavelet transform on the basic brain feature map to obtain a wavelet transform feature map may include the following operations: performing convolution processing on the basic brain feature map through a convolution layer using a first convolution kernel to obtain a processed basic brain feature map. Then, the processed basic brain feature map may be subjected to wavelet transform (corresponding to Figure 2The DWT in the image is used to perform multi-scale analysis of the processed basic brain feature map, obtaining low-frequency information L and high-frequency information (H / V / D). The high-frequency information can specifically include horizontal information H, vertical information V, and diagonal information D. Generally, the low-frequency information L can record the overall structure of the image (such as the general shape of the brain and the distribution of large blood vessels). The horizontal information H can record the horizontal edges in the image (such as the horizontal texture of the cerebral cortex and the horizontal direction of the blood vessels); the vertical information V can record the vertical edges in the image (such as the vertical boundaries of sulci and gyri); and the diagonal information D can record the diagonal edges in the image (such as oblique blood vessel branches and the oblique boundaries of irregular lesions).

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

[0036] The local information extraction module 104 takes the aforementioned brain basic feature map as input and extracts multi-scale local information feature maps through pooling operations with different-sized pooling kernels. The local information feature map amplifies the local details of the brain area, especially the edges and surrounding structures of the infarct area, realizes fine-grained analysis of the lesion area, and improves the model's recognition ability for complex morphological areas. Figure 3 Understand the structure of the local information extraction module 104 in the embodiment of the present application. Figure 3 As shown, the local information extraction module 104 may include at least one convolution layer using a preset convolution kernel (e.g. Figure 3 Conv k=3 shown in ), multi-scale feature extraction module and batch normalization layer (as shown in Figure 3 BN shown in the figure).

[0037] Furthermore, the multi-scale feature extraction module may include a first convolution block, a maximum pooling layer (such as Figure 3 The first convolution block may include a batch normalization layer (such as Figure 3 BN shown in the figure), convolutional layers using 1×1 convolution kernels (such as Figure 3 Conv k=1) and convolutional layers using preset convolution kernels (such as Figure 3 Conv k=3 in the figure); the second convolution block may include a convolution layer using a 1×1 convolution kernel (such as Figure 3 Conv k=1) and convolutional layers using preset convolution kernels (such as Figure 3Conv k=3) shown in .

[0038] Figure 3 The 5×5, 9×9, and 13×13 in the figure represent the sizes of the pooling kernels (sliding windows) in the maximum pooling layer. Taking 5×5 as an example, it means sliding a window of 5 rows and 5 columns across the feature map, taking the maximum value of the pixels within the window each time as the output. The same applies to 9×9 and 13×13, except that the number of windows is different. Small pooling kernels (such as 5×5) can focus on smaller local areas, retaining detailed information such as fine edges and textures, and capturing fine features such as the edges of infarcted areas. Large pooling kernels (such as 13×13) can cover a larger area, extract more macroscopic structural information, and grasp the overall layout of the brain. By combining pooling kernels of different sizes, the model can simultaneously focus on details and the overall picture, enriching feature expression and improving adaptability to complex brain tissue structures.

[0039] Based on this, in the local information extraction module 104, information extraction is performed on the basic feature map of the brain to obtain a local information feature map, which may include the following operations: convolution processing is performed on the basic feature map of the brain through a convolution layer using a preset convolution kernel to obtain a preliminary local feature map; convolution processing is performed on the preliminary local feature map through a convolution layer using a preset convolution kernel to obtain a deepened local feature map; feature extraction is performed on the preliminary local feature map through a multi-scale feature extraction module to obtain a multi-scale local feature map; the deepened local feature map and the multi-scale local feature map are spliced ​​and normalized using a batch normalization layer to obtain a local information feature map.

[0040] Furthermore, subjecting the preliminary local feature map to feature extraction via a multi-scale feature extraction module to obtain a multi-scale local feature map may include the following operations: subjecting the preliminary local feature map to convolution processing via a first convolution block to obtain a processed preliminary local feature map; subjecting the processed preliminary local feature map to pooling via the maximum pooling layer to obtain a first-scale local feature map, a second-scale local feature map, and a third-scale local feature map; splicing the first-scale local feature map, the second-scale local feature map, and the third-scale local feature map to obtain a spliced ​​local feature map; subjecting the spliced ​​local feature map to convolution processing via a second convolution block to obtain a multi-scale local feature map.

[0041] Continue to refer Figure 1 , the segmentation output module 105 of the embodiment of the present application may include a 1×1 convolution block and at least one convolution layer (such as Figure 1Conv is shown in ). 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 an embodiment of the present application, a 1×1 convolution block can be composed of a 1*1 convolution, a batch normalization layer (Batch Normalization, BN) and an activation function (RectifiedLinear Unit, ReLu). The convolution layer can extract more complex feature combinations (such as whether a certain location has both blood flow changes and edges in a specific direction, or local structural abnormalities). 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 size. This means that in each convolution operation, the kernel operates on only one pixel in the input feature map and all its corresponding channels. By using a 1×1 convolution kernel on the feature map, the high-dimensional feature map can be mapped to a low-dimensional space, reducing the number of channels and removing redundant information, thereby reducing model complexity. The Batch Normalization layer normalizes the input data, making the data distribution more stable and helping to accelerate model convergence. The Batch Normalization layer also provides a certain degree of regularization, improving the model's generalization ability and reducing the risk of overfitting. The ReLU (Rectified Linear Unit) activation function introduces nonlinearity into the model. After the 1×1 convolution and batch normalization, the ReLU function sets values ​​less than 0 to 0 and leaves values ​​greater than 0 unchanged, thereby increasing the model's nonlinear representation capabilities. This enables the model to learn more complex feature patterns, improving performance in 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 and outputting the segmentation result, the following operations can be performed: the combined feature map is input into a 1×1 convolution block for dimensionality reduction to obtain a feature map after dimensionality reduction; the feature map after dimensionality reduction is input into at least one convolution 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 operations can be performed to output the segmentation result: Based on the output of the convolutional layer, a softmax function can be used to calculate the probability of each pixel belonging to an ischemic area, an infarcted area, 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 areas to clearly show the location of ischemia and infarction.

[0045] Figure 4An exemplary flow chart of a method 400 for training a cerebral perfusion image sequence segmentation model according to an embodiment of the present application is shown. It is understood that the method 400 can be executed by any appropriate device with data processing capabilities, including but not limited to a terminal device, a processor, and a server.

[0046] Based on this, Figure 4 As shown, at step S401, method 400 can acquire a cerebral perfusion image sequence and mark the boundaries of the ischemic area and the infarct area in each image of the cerebral perfusion image sequence to obtain marking information. At step S402, method 400 can input the cerebral perfusion image sequence into a difference calculation module to perform difference calculation to obtain a difference feature map. At step S403, method 400 can input the cerebral perfusion image sequence into a basic feature extraction module to perform feature extraction to obtain a basic brain feature map.

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

[0048] At the aforementioned step S401, a brain perfusion image sequence of patients with acute ischemic stroke can be retrospectively collected. A brain perfusion image sequence refers to multiple CT scans of the same area of ​​interest within a short period of time to obtain dynamic image data of the area after contrast agent injection. These data reflect the flow and distribution of the contrast agent in the brain tissue. Furthermore, when marking the boundaries of the ischemic area and the infarct area, a radiologist can manually mark the image, that is, mark the exact boundaries of the cerebral ischemic and cerebral infarction areas. Additionally or optionally, another radiologist can also review the image to ensure the accuracy of the marking results.

[0049] In an embodiment of the present application, to improve the training efficiency of the brain perfusion image sequence segmentation model and save time and cost, the aforementioned brain perfusion image sequence can be pre-processed by denoising, standardizing, and normalizing the image sequence before using it for model training, thereby obtaining a pre-processed brain perfusion image sequence for model training. This can reduce unnecessary computational effort while ensuring sufficient information, enabling more efficient use of computing resources and accelerating the entire training process.

[0050] During the acquisition process of raw brain perfusion image sequences, a large amount of noise is introduced due to factors such as electronic noise from the equipment itself, slight patient movement during scanning, and uneven distribution of the contrast agent. This noise can negatively impact feature extraction in subsequent deep learning models, causing the model to mistakenly identify noise as features, thereby affecting segmentation accuracy. Based on this, methods such as median filtering, Gaussian filtering, or non-local mean filtering can be used to filter brain perfusion image sequences. This can remove noise while better preserving image edges and detail information, which is particularly important for preserving the complex brain tissue structure in brain perfusion image sequences.

[0051] Furthermore, imaging parameters (such as tube voltage, tube current, and scanning slice thickness) vary between CT devices, and scanning conditions on the same device may fluctuate at different times. This can lead to significant differences in grayscale range, contrast, and other aspects of the acquired brain perfusion image sequences. For example, two CT devices from different manufacturers scanning the same patient's brain may produce completely different grayscale value distributions in the resulting brain perfusion image sequences. Standardization aims to eliminate the impact of these device differences and give the images a uniform format and characteristics. Common standardization methods map the image's grayscale values ​​to a fixed range, such as adjusting the grayscale values ​​of all pixels to the interval [0, 1]. Another example is converting the grayscale values ​​of all pixels into 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]) so that the data has a similar scale. During the training process of deep learning models, if the scale differences of the input data are too large, the model convergence speed will be slowed down, and even problems such as gradient vanishing or gradient explosion will occur. For example, the CT values ​​of different tissues in a brain perfusion image sequence span a large range. Without normalization, the model may over-focus on areas with large CT values ​​(such as bones) during the learning process, while ignoring brain tissue areas with smaller CT values ​​but critical for stroke diagnosis. By compressing all pixel values ​​to a unified scale through normalization, the model training speed 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 many noise points (such as machine noise and interference caused by slight patient movement), which can interfere with the model's interpretation of actual blood flow changes. Therefore, the difference image map can be denoised to eliminate these noise points and retain the actual blood flow change areas.

[0054] At step S407, method 400 may perform the following operations to update the parameters of the cerebral perfusion image sequence segmentation model based on the annotation information and the segmentation results: determining a loss value based on the annotation information and the segmentation results using the Dice loss function and the cross-entropy loss function; and updating the parameters of the cerebral perfusion image sequence segmentation model based on the loss value using a preset optimization algorithm. The preset optimization algorithm may be, for example, an adaptive moment estimation (Adam) algorithm or a stochastic gradient descent (SGD) algorithm.

[0055] In cerebral ischemia and infarction segmentation, the lesion area (positive sample) usually only accounts for 5%-15% of the whole image, and the traditional loss function is easily dominated by normal tissue (negative sample). The Dice loss function can force the model to focus on small targets by maximizing the overlap between the predicted area and the true area. The cross entropy loss function penalizes the classification error of each pixel individually, which is suitable for refining boundary details (such as the fuzzy transition area between the ischemic penumbra and normal tissue). For example, if a pixel is actually an ischemic area, the model predicts a probability of 0.6 (correct classification but low confidence), and the CE loss will force the prediction probability of the pixel to be increased through gradient back propagation. In this embodiment of the application, the Dice loss function and the cross entropy loss function are combined to enable the model to simultaneously meet regional integrity and pixel accuracy.

[0056] In some implementation scenarios, cross-validation can be used to evaluate the generalization ability of the model. Specifically, when implementing cross-validation, the labeled data set can be divided into k non-overlapping subsets, and k-1 subsets are used for training each time, and 1 subset is used for verification, and repeated k times to ensure that each sample is used for training and verification, so that the model can adapt to the data diversity in real clinical practice and improve the generalization ability of the model. In addition, in scenarios where the cost of medical data annotation is high (such as each tomographic image in this application needs to be manually annotated 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 the present application, when implementing cross-validation, the Dice coefficient or Hausdorff distance can be used as an evaluation indicator to quantitatively evaluate the segmentation results. The Dice coefficient is used to measure the degree of overlap between the model's predicted area and the true annotated area. It can reflect the model's overall ability to capture the lesion area 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 area and the true annotated area, that is, the distance between the farthest points on their contours. It can detect subtle deviations in the model's boundary positioning (such as missing or over-labeled edge pixels).

[0058] In an embodiment of the present application, in order to improve the stability of the brain perfusion image sequence segmentation model and prevent overfitting, data enhancement operations such as rotation, flipping, scaling, translation, brightness / contrast adjustment, and adding Gaussian noise are performed on the original brain perfusion image sequence, so that the size of the training data set can be artificially expanded without relying on additional clinical data. For example, random angle rotation or scale scaling of the annotated ischemic / infarct area tomographic images can simulate actual clinical scenarios such as different patient head postures and CT scan layer thickness differences, enabling the model to learn more universal anatomical structural features, avoiding overfitting problems caused by a single training data set, and thus maintaining stable performance in the segmentation of brain perfusion image sequences of unknown patients.

[0059] Combination of the above Figure 4 This paper describes a method for training a segmentation model for cerebral perfusion image sequences. By incorporating temporal dynamic features of cerebral perfusion image sequences (i.e., difference feature maps) through a difference calculation module, the model can capture differences in brain perfusion at different time points and quantify the dynamic changes in contrast agent distribution. Compared to traditional manual assessment methods that rely on static empirical judgment and ignore temporal information, this mechanism can sensitively identify early, small ischemic lesions (such as ischemic penumbras where blood flow is slow but not yet completely interrupted), providing an objective temporal basis for accurately determining the thrombolytic time window and avoiding treatment delays caused by lags in manual analysis.

[0060] Furthermore, the wavelet transform module can perform multi-scale decomposition of the basic feature map of the brain (low-frequency contour + high-frequency edge), and can generate a wavelet transform feature map containing the overall structure of the brain area and refined edge information, which significantly enhances the boundary contrast between the ischemic and infarcted areas and normal tissues, and improves the segmentation accuracy of the model for ischemic and infarcted areas. The local information extraction module can extract multi-scale local structural features through pooling operations of pooling kernels of different sizes, realize fine-grained analysis of the lesion area, and improve the model's recognition ability for complex morphological areas. The segmentation output module can fuse the difference feature map, wavelet transform feature map and local information feature. Figure 3 The complementary features enable the model to deeply learn the multi-level information of time, space and structure in brain perfusion image sequences, which can significantly improve the generalization performance of the model.

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

[0062] In a specific embodiment, the processor 501 may be at least one of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a CPU, a controller, a microcontroller, and a microprocessor. It is understood that for different devices, the electronic device used to implement the functions of the processor may also be other, and this embodiment does not specifically limit this.

[0063] In the embodiment of the present application, the communication bus 503 is used to realize the connection and communication between the processor 501 and the memory 502; the memory 502 stores program instructions for training the brain perfusion image sequence segmentation model; when the processor 501 executes the program instructions stored in the memory 502, the present application is realized. Figure 4 A method is described for training a segmentation model for brain perfusion image sequences.

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

[0065] It is understood that the description of each embodiment in this disclosure focuses on the differences between the embodiments, and the same or corresponding parts can be referenced to each other. For the purpose of brevity, this disclosure will not repeat them one by one.

[0066] In addition, the present application also provides a device for segmenting a brain perfusion image sequence. The structure or architecture of the device is combined with the above Figure 5 The structure or framework of the electronic device 500 for training the brain perfusion image sequence segmentation model is the same as that described above. For the purpose of brevity, the present disclosure will not describe them one by one. In the device for segmenting the brain perfusion image sequence, the memory stores program instructions for segmenting the brain perfusion image sequence. When the program instructions are executed by the processor, the device can implement the following combination Figure 6 A method 600 for segmenting a sequence of brain perfusion images is described.

[0067] Figure 6 An exemplary flow chart of a method 600 for segmenting a cerebral perfusion image sequence according to an embodiment of the present application is shown. It is understood that the method 600 can be executed by any appropriate device with data processing capabilities, including but not limited to a processor, a terminal device, and a server.

[0068] like Figure 6 As shown, at step S601, method 600 can obtain a patient's brain perfusion image sequence. Then, at step S602, method 600 can input the brain perfusion image sequence into the trained brain perfusion image sequence segmentation model for segmentation to output a segmentation result. It can be understood that the trained brain perfusion image sequence segmentation model here refers to the brain perfusion image sequence segmentation model used in combination with the present application. Figure 4 The described method for training a segmentation model for a brain perfusion image sequence trains a completed model.

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

[0070] Combination of the above Figure 6This paper describes a method for segmenting cerebral perfusion image sequences. By acquiring a patient's cerebral perfusion image sequence and inputting it into a trained model, segmentation results can be rapidly output. This method not only significantly shortens segmentation time and improves efficiency, but also avoids the subjectivity and uncertainty inherent in 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] According to the above description in conjunction with the accompanying drawings, those skilled in the art will also understand that the embodiments of the present application can also be implemented by a software program. Therefore, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores program instructions for training a brain perfusion image sequence segmentation model and / or for segmenting a brain perfusion image sequence. The program instructions can be used to implement the present application in conjunction with Figure 4 The method described for training a segmentation model for a brain perfusion image sequence or a combination thereof Figure 6 A method is described for segmenting a sequence of brain perfusion images.

[0072] It should be noted that although the operations of the present method are described in a particular order in the accompanying drawings, this does not require or imply that the operations must be performed in that particular order, or that all of the operations shown must be performed to achieve the desired results. Rather, the steps depicted in the flowcharts may be performed in a different order. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into a single step, and / or a single step may be broken down into multiple steps.

[0073] Although multiple embodiments of the present 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. Those skilled in the art can conceive of many changes, modifications, and alternatives without departing from the thought and spirit of the present application. It should be understood that in the process of practicing the present application, various alternatives to the embodiments of the present application described herein can be adopted. The accompanying claims are intended to define the scope of protection of the present application and therefore cover equivalents or alternatives within the scope of these claims.

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

Claims

1. A method for training a cerebral perfusion image sequence segmentation model, the cerebral perfusion image sequence segmentation model comprising 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 comprising: Acquiring a cerebral perfusion image sequence, and marking the boundaries of the ischemic area and the infarct area in each image of the cerebral perfusion image sequence to obtain marking information; Inputting the brain perfusion image sequence into the difference calculation module to perform difference calculation to obtain a difference feature map; Inputting the brain perfusion image sequence into the basic feature extraction module to perform feature extraction to obtain a brain basic feature map; Inputting the brain basic feature map into the wavelet transform module for wavelet transform to obtain a wavelet transform feature map; Inputting the brain basic feature map into the local information extraction module to extract information 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 to perform region segmentation and output a segmentation result; Based on the annotation information and the segmentation result, the parameters of the cerebral perfusion image sequence segmentation model are updated to achieve training thereof.

2. The method according to claim 1, wherein Performing difference calculation in the difference calculation module to obtain a difference feature map includes: Based on images at different time points in the cerebral perfusion image sequence, calculating an image difference between a first image and a second image at adjacent time points as a difference feature map; Denoising is performed on the difference feature map to obtain a denoised difference feature map.

3. The method according to claim 1 or 2, wherein: Performing wavelet transform in the wavelet transform module to obtain a wavelet transform feature map includes: Convolutionally processing the basic brain feature map through a convolution layer using a preset convolution kernel to obtain a processed basic brain feature map; performing a wavelet transform on the processed brain basic 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 processing the combined information through a convolution layer using a 1×1 convolution kernel and a batch normalization layer to obtain processed high-frequency information; Processing the low-frequency information through a convolution layer using a 1×1 convolution kernel to obtain processed low-frequency information; The processed high-frequency information is combined with the processed low-frequency information to obtain the wavelet transform feature map.

4. The method according to claim 1 or 2, wherein: Performing information extraction in the local information extraction module to obtain a local information feature map includes: Convolutionally processing the basic brain feature map through a convolution layer using a preset convolution kernel to obtain a preliminary local feature map; Convolutionally processing the preliminary local feature map through a convolution layer using the preset convolution kernel to obtain a deepened 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; The deepened local feature map is spliced ​​with the multi-scale local feature map, and a batch normalization layer is used to perform normalization processing to obtain the local information feature map.

5. The method according to claim 4, wherein The multi-scale feature extraction module includes a first convolution block, a maximum pooling layer and a second convolution block; the first convolution block includes a batch normalization layer, a convolution layer using a 1×1 convolution kernel and a convolution layer using the preset convolution kernel; the second convolution block includes a convolution layer using a 1×1 convolution kernel and a convolution layer using the preset convolution kernel; the pooling kernels of the maximum pooling layer include 5×5, 9×9 and 13×13.

6. The method according to claim 5, wherein: Performing feature extraction on the preliminary local feature map through a multi-scale feature extraction module to obtain a multi-scale local feature map includes: Performing convolution processing on the preliminary local feature map through the first convolution block to obtain a processed preliminary local feature map; Pooling the processed preliminary local feature map through the maximum pooling layer to obtain a first-scale local feature map, a second-scale local feature map, and a third-scale local feature map; splicing the first-scale local feature map, the second-scale local feature map, and the third-scale local feature map to obtain a spliced ​​local feature map; The spliced ​​local feature map is convolved through the second convolution block to obtain the multi-scale local feature map.

7. The method according to claim 1, wherein The segmentation output module includes a 1×1 convolution block and at least one convolution 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 convolution block for dimensionality reduction to obtain a feature map after dimensionality reduction; The feature map after dimensionality reduction is input into the at least one convolutional layer for convolution operation to output the segmentation result.

8. A device for training a segmentation model for a cerebral perfusion image sequence, characterized in that: include: processor; as well as A memory storing program instructions for training a brain perfusion image sequence segmentation model, wherein when the program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

9. A device for segmenting a brain perfusion image sequence, characterized in that: include: processor; A memory storing program instructions for segmenting a cerebral perfusion image sequence, wherein when the program instructions are executed by the processor, the device implements the following operations: Acquire a patient's brain perfusion image sequence; The brain perfusion image sequence is input into a brain perfusion image sequence segmentation model trained according to the method according to any one of claims 1 to 7 for segmentation, so as to output a segmentation result.

10. A computer-readable storage medium storing program instructions for training a cerebral perfusion image sequence segmentation model and / or for segmenting a cerebral perfusion image sequence, wherein when the program instructions are executed by a processor, the method according to any one of claims 1 to 7 and / or the operations implemented by the device according to claim 9 are implemented.

Images