Prognosis prediction model construction method and device, electronic equipment and storage medium

Abstract

The invention relates to a prognosis prediction model construction method and device, electronic equipment and a storage medium. The prognosis prediction model construction method comprises the following steps: determining key distinguishing features capable of distinguishing a normal region from a focus region based on a brain medical image; determining a hidden damaged tissue boundary around the focus area according to the key distinguishing features; wherein the brain tissue between the focus area and the damaged tissue boundary is a hidden abnormal blood flow tissue area which is difficult to directly detect based on a brain medical image; and constructing the prognosis prediction model based on the brain tissue characteristics in the damaged boundary. The embodiment of the invention can improve the prediction level.

Description

Technical Field

[0001] This disclosure relates to medical image processing technology, and in particular to a method for constructing prognostic prediction models. Background Technology

[0002] In related technologies, there are generally two methods for assessing the prognosis of patients with brain diseases, especially stroke. The first method assesses the patient's prognosis within three months based on clinically recorded basic information, past medical history, clinical NIHSS score, mRS score, and treatment methods. However, this method is limited by the accuracy and completeness of clinical information; in the absence of information, the accuracy of assessment and prediction is difficult to guarantee. The second method is based on intelligent processing technology of clinical images, such as direct processing of DWI, PWI, CT sequences, or analysis of the size and location of lesions to determine the patient's prognosis. However, this method focuses more on assessing the lesion area directly obtained clinically. Due to the complexity of brain tissue, an abnormality in one area will inevitably alter the blood flow of the surrounding tissues. The blood flow in this hidden tissue area can also affect the patient's recovery. Therefore, it is necessary to consider this often overlooked tissue area surrounding the lesion to fully utilize in predicting the patient's prognosis and thus improve the accuracy of the prediction. Summary of the Invention

[0003] This disclosure proposes a method, apparatus, electronic device, and storage medium for constructing a prognostic prediction model, which addresses the problem that existing prognostic prediction methods focus more on clinical information and lesion area, resulting in insufficient information and low prediction accuracy.

[0004] According to one aspect of this disclosure, a method for constructing a prognostic prediction model is provided, comprising: determining key distinguishing features that can differentiate normal regions from lesion regions based on brain medical images; determining the boundary of hidden damaged tissue around the lesion region based on the key distinguishing features; wherein the brain tissue between the lesion region and the boundary of the damaged tissue is a hidden abnormal blood flow tissue region that is difficult to detect directly based on brain medical images; and constructing the prognostic prediction model based on the brain tissue features within the damaged boundary.

[0005] In some possible implementations, constructing the prognostic prediction model based on brain tissue features within the damaged boundary includes at least one of the following methods: constructing the prognostic prediction model based on key distinguishing features within the damaged boundary; performing weighted processing on a first key distinguishing feature within the lesion region and a second key distinguishing feature between the lesion region and the damaged boundary, and constructing the prognostic prediction model using the weighted features; and constructing the prognostic prediction model based on brain tissue images within the damaged boundary.

[0006] In some possible implementations, determining the boundary of the hidden damaged tissue around the lesion area based on the key distinguishing features includes: using a surrounding tissue division strategy to divide the surrounding tissue around the lesion area into multiple levels; and determining the boundary of the damaged tissue around the lesion area based on the key distinguishing features within the multiple levels of surrounding tissue.

[0007] In some possible implementations, the step of using a surrounding tissue segmentation strategy to perform multi-level surrounding tissue segmentation around the lesion region includes: expanding outward from the lesion region according to preset voxels, obtaining a first-level surrounding region for each preset voxel expansion; obtaining the multi-level surrounding regions based on multiple preset voxel expansions; and / or The step of expanding outward according to a preset voxel includes: using the voxel coordinates on the outer boundary of the lesion area as a reference, expanding outward to the adjacent outer side according to a preset number of voxels, and the expanded area forms a primary surrounding area.

[0008] In some possible implementations, the step of using a surrounding tissue division strategy to divide the lesion area into multiple levels of surrounding tissue includes: based on the lesion area, expanding outward at a preset distance to form multiple preset shapes; the boundary of the lesion area and the first preset shape, as well as the adjacent preset shapes, constitute the multi-level surrounding area.

[0009] In some possible implementations, determining the boundary of damaged tissue around the lesion area based on key distinguishing features within the multi-level surrounding tissues includes: sequentially determining whether key distinguishing features in each of the surrounding areas meet preset conditions in order of distance from the lesion area; if the preset conditions are met, the corresponding surrounding area is identified as damaged tissue, and the key blood flow features of the next surrounding area are determined to meet preset conditions, until the boundary of the corresponding surrounding area is identified as the boundary of damaged tissue if the preset conditions are not met.

[0010] In some possible implementations, the step of sequentially determining whether the key distinguishing features in each of the surrounding areas meet preset conditions in order of distance from the lesion region includes at least one of the following methods: determining whether the similarity between the key distinguishing features in each of the surrounding areas and the key distinguishing features in the lesion region is greater than a first threshold; identifying the proportion of abnormal tissue in the surrounding areas based on the key distinguishing features in the surrounding areas, and determining whether the proportion is greater than a second threshold; determining whether the distance between the key distinguishing features in each of the surrounding areas and the key distinguishing features in the lesion region is less than a third threshold.

[0011] According to a second aspect of this disclosure, a prognostic prediction device is provided, which is constructed using the prognostic prediction model construction method described in any one of the first aspects, comprising: a first determining module, configured to determine key distinguishing features that can differentiate normal regions from lesion regions based on brain medical images; a second determining module, configured to determine the boundary of hidden damaged tissue around the lesion region based on the key distinguishing features; wherein the brain tissue between the lesion region and the boundary of the damaged tissue is a hidden abnormal blood flow tissue region that is difficult to detect directly based on brain medical images; and a construction module, configured to construct the prognostic prediction model based on the combined blood flow features within the damaged boundary.

[0012] According to a third aspect of this disclosure, an electronic device is provided, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to perform the method described in any one of the first aspects.

[0013] According to a fourth aspect of this disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, characterized in that, when the computer program instructions are executed by a processor, they implement the method described in any one of the first aspects.

[0014] Based on the above configuration, this embodiment first identifies key distinguishing features between normal and lesion regions using brain medical imaging; it then determines the hidden boundaries of damaged tissue surrounding the lesion region; and finally, it constructs the prognostic prediction model using brain tissue features within the damaged boundaries. This embodiment combines the lesion region and the potentially surrounding damaged tissue to construct the prognostic prediction model. This method surpasses traditional methods, considers more comprehensively, and helps improve model accuracy.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure.

[0016] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the specification, serve to illustrate the technical solutions of this disclosure.

[0018] Figure 1 This is a flowchart of a method for constructing a prognostic prediction model according to an embodiment of this disclosure. Detailed Implementation

[0019] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0020] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0021] In this document, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0022] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0023] The method provided in this disclosure can be executed by an image processing device. For example, the method can be executed by a terminal device, a server, or other processing device. The terminal device can be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, an in-vehicle device, a wearable device, etc. In some possible implementations, the method can be implemented by a processor calling computer-readable instructions stored in memory.

[0024] It is understood that the various method embodiments mentioned above in this disclosure can be combined with each other to form combined embodiments without violating the principle and logic. Due to space limitations, this disclosure will not elaborate further.

[0025] Figure 1 This is a flowchart of the prognostic prediction model construction method in the embodiments of this disclosure; as follows: Figure 1 As shown, the method for constructing the prognostic prediction model includes: S10: Identify key distinguishing features that differentiate normal and lesion areas based on brain medical imaging; In some possible implementations, embodiments of this disclosure can acquire brain medical images, which can be acquired in real time, such as by real-time acquisition of images from CT or MR imaging devices, or by receiving transmitted brain medical images via wired or wireless means, or by reading brain medical images from local or cloud server devices or other storage devices. This disclosure does not limit this. Furthermore, the brain medical images in the embodiments of this disclosure can be perfusion sequences, such as CTP, PWI, or other time-varying four-dimensional brain images. However, this disclosure does not specifically limit this.

[0026] In some possible implementations, key distinguishing features can be identified by utilizing the differences between the tissues of the normal group region and the lesion region. These key distinguishing features can characterize abnormalities in the vascular morphology and blood flow of brain tissue.

[0027] S20: Based on the key distinguishing features, determine the hidden damaged tissue boundary around the lesion area; wherein, the brain tissue between the lesion area and the damaged tissue boundary is a hidden abnormal blood flow tissue area that is difficult to detect directly based on brain medical imaging; In some possible implementations, key distinguishing features that can differentiate whether blood flow is abnormal can be used to identify whether there is damage to the surrounding tissues at various levels around the lesion area. If damage is found, the tissue in that area is identified as damaged tissue. The process continues to determine whether there is damage to the surrounding tissues at the outermost level, and so on. In this way, the damage to the surrounding tissues of the lesion area can be identified sequentially until the first-level surrounding area is determined to be undamaged. At this point, the boundary of the damaged tissue can be determined based on that area.

[0028] S30: Construct the prognostic prediction model based on the combined blood flow characteristics within the damaged boundary.

[0029] In some possible implementations, feature information related to blood flow in brain tissue within the damaged boundary can be extracted, joint blood flow features can be constructed, and the model can be trained using the joint blood flow features to obtain a prognostic prediction model.

[0030] Based on the above configuration, this embodiment of the present disclosure can further expand the traditional method of segmenting lesion tissue to obtain brain tissue regions that may be damaged, that is, to deeply analyze the boundary of damaged tissue and use the combined blood flow characteristics within the boundary of damaged tissue to construct a prognostic prediction model, so as to achieve accurate prognostic analysis of cerebrovascular diseases such as stroke.

[0031] The embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The embodiments of this disclosure can first acquire brain medical images. These images can be dynamic perfusion images with time information, such as CTP, PWI, and other sequence images, or three-dimensional images, such as CT, ADC, DWI, T1, T2, and other images. This disclosure does not specifically limit these; any image that can be used to identify abnormal blood flow or tissue damage in brain tissue can be used as a brain image in the embodiments of this disclosure.

[0032] Specifically, the acquisition of brain medical images can include at least one of the following methods: ① Directly acquire images using medical imaging equipment (CT, MR); ② Transmit and receive brain medical images via electronic devices; In this embodiment, images transmitted by other electronic devices can be received via communication, which may include wired and / or wireless communication, and this disclosure does not impose specific limitations. ③ Read brain medical images stored in a database; In this embodiment, image sequences stored locally or on a server can be read according to received data reading instructions, and this disclosure does not impose specific limitations.

[0033] With access to brain medical images, the lesion regions of brain diseases can be segmented using these images. These brain diseases can include stroke, brain tumors, etc., but this disclosure does not specifically limit the scope of these diseases. Taking stroke as an example, with access to brain medical images, ischemic areas of the brain can be segmented from the images.

[0034] In some embodiments, the method for segmenting ischemic brain regions in this disclosure may include at least one of the following: ① Processing perfusion sequences such as CTP, PWI, etc., firstly obtaining the temporal density curve of each pixel in the selected layer, and then processing it through a mathematical model to obtain: cerebral blood volume (CBV), cerebral blood flow (CBF), mean transit time (MTT), time to peak (TTP), and time to top (Tmax). Tmax represents the time it takes for the brain tissue's blood storage function to reach its maximum value, and is a sensitive indicator of changes in tissue perfusion and cerebral infarction. Brain tissue regions with Tmax greater than 6 seconds and less than 10 seconds can be defined as penumbras, i.e., ischemic brain regions. Correspondingly, the core infarction region is defined as brain tissue with Tmax greater than 10 seconds and a decrease in CBV. ② Processing brain medical images using a deep learning segmentation model to obtain ischemic brain regions. The segmentation model is a trained model that can be used for ischemic brain region segmentation. The specific training process will not be elaborated here. Those skilled in the art can use any of the following methods—strong supervision, semi-supervision, weak supervision, and unsupervision—to obtain the segmentation model used in the embodiments of this disclosure.

[0035] In this embodiment of the disclosure, the processing method of the ischemic tissue segmentation model includes: performing frequency feature extraction processing on the input image to obtain a high-frequency feature group and a low-frequency feature group; performing a first fusion processing on the high-frequency feature group to obtain a first fusion feature, and performing a second fusion processing on the low-frequency feature group to obtain a second fusion feature; obtaining a third fusion feature based on the first fusion feature and the second fusion feature; and using the third fusion feature to detect lesion region objects in the input image.

[0036] In some possible implementations, frequency feature extraction processing is performed on the input brain medical image to obtain a high-frequency feature group and a low-frequency feature group. The frequency feature extraction processing in this embodiment may include wavelet transform (DWT) processing; in other embodiments, Fourier transform and other methods may also be used, and this disclosure does not specifically limit this. After performing frequency feature extraction processing on the input image, a high-frequency feature group and a low-frequency feature group are obtained, wherein the high-frequency feature group includes at least two high-frequency feature maps, and the low-frequency feature group includes at least two low-frequency feature maps. That is, the high-frequency feature group and low-frequency feature group obtained in this embodiment each include at least two feature maps, representing the high-frequency and low-frequency features of the original input brain image. The high-frequency feature maps in the high-frequency feature group represent feature information at different high frequencies, and correspondingly, the low-frequency feature maps in the low-frequency feature group represent feature information at different low frequencies. Through the above configuration, feature information under different frequency conditions can be collected, and the richness of features can improve the accuracy of target detection.

[0037] In some possible implementations, performing a second fusion process on the low-frequency feature group to obtain a second fused feature includes: performing connection processing in the height and width directions on the low-frequency feature maps in the low-frequency feature group to obtain a first low-frequency connection feature and a second low-frequency connection feature; performing feature encoding processing on the first low-frequency feature and the second low-frequency feature to obtain a first encoded feature and a second encoded feature; and fusing the first encoded feature and the second encoded feature to obtain the second fused feature.

[0038] Specifically, the low-frequency feature maps within the low-frequency feature group are connected along the height direction to obtain the first low-frequency connected feature, and the low-frequency feature maps within the low-frequency feature group are connected along the width direction to obtain the second low-frequency connected feature. Given the first and second low-frequency connected features, a feature encoding module performs encoding processing to obtain encoding results under different feature connection modes, thus achieving multi-dimensional feature processing.

[0039] The feature encoding process includes: performing a two-branch process on the input connection features (a first low-frequency connection feature and a second low-frequency connection feature); wherein, the first branch process includes: performing a linear transformation and activation process on the connection features; the second branch process includes: performing a linear transformation, feature dimensionality reduction, activation process and state space transformation process on the connection features; and performing a product process and a linear transformation on the two features obtained from the two-branch process to obtain the encoded features corresponding to the connection features.

[0040] Taking the encoding process of the first low-frequency connection feature as an example, the processing of the second low-frequency feature is the same and will not be repeated. The encoding process of the first low-frequency connection feature includes: using the first branch to perform a linear transformation and activation processing (sigmoid) on the first low-frequency connection feature to obtain the first activation feature. Using the second branch to perform a linear transformation, feature dimensionality reduction (such as convolution), activation processing (sigmoid), and state space transformation on the first low-frequency connection feature to obtain the state space feature; performing a product processing and linear transformation on the first activation feature and the state space feature to obtain the encoded feature corresponding to the connection feature. The weight parameters used in the linear transformation of the first branch and the second branch are different; the specific parameters are learned during training. The convolution processing of the second branch uses a 1*1 kernel size, or it can also be 3*3; this disclosure does not specifically limit this. Furthermore, the model used in the state space transformation is the SSM model, and the state equation and output equation used are set by those skilled in the art according to their needs and are not specifically limited here. State space transformation can be used to achieve accurate representation of features over long distances, thereby improving model performance.

[0041] Given the first and second encoded features, they are fused to obtain the second fused feature. This can be achieved by performing dimensionality transformations on both the first and second encoded features to restore them to the original input image dimensions, then performing channel-wise concatenation to obtain the second fused feature based on the concatenation features. Alternatively, the concatenation features can be convolved and then summed with the low-frequency feature maps within the low-frequency feature group to obtain the second fused feature.

[0042] Based on the above configuration, low-frequency feature fusion processing can be achieved, resulting in accurate representation of low-frequency features over long distances, reducing feature loss and effectively improving feature expression.

[0043] In some possible implementations, performing a first fusion process on the high-frequency feature group to obtain a first fused feature includes: performing connection fusion processing on the high-frequency feature map of the high-frequency feature group in the height direction, width direction, and diagonal direction respectively to obtain high-frequency connection features; performing attention processing on the high-frequency connection features to obtain high-frequency attention features; and performing feature transformation processing on the high-frequency attention features to obtain the first fused feature.

[0044] The high-frequency feature maps of the high-frequency feature group are subjected to connection and fusion processing in the height, width, and diagonal directions to obtain high-frequency connection features. This includes: performing connection processing in the height, width, and diagonal directions on the high-frequency feature maps of the high-frequency feature group to obtain a first high-frequency connection feature, a second high-frequency connection feature, and a third high-frequency connection feature; and performing channel direction connection processing on the first, second, and third high-frequency connection features to obtain high-frequency connection features. This configuration can effectively integrate feature information within different high-frequency feature maps.

[0045] In some possible implementations, after obtaining the high-frequency connection features, the embodiments of this disclosure can further perform fusion optimization on the first fusion feature and the high-frequency connection features to obtain optimized high-frequency connection features. The fusion optimization method includes any of the following: ① performing channel-direction connection processing on the first fusion feature and the high-frequency connection features to obtain optimized high-frequency connection features; ② calculating the distance between the first fusion feature and the high-frequency fusion connection features; if the distance is less than a distance threshold, using the first fusion feature as the optimized high-frequency feature; if the distance is greater than the distance threshold, performing channel-direction connection processing on the first fusion feature and the high-frequency connection features to obtain optimized high-frequency connection features. With the optimized high-frequency connection features obtained, the optimized high-frequency connection features are used as high-frequency connection features for subsequent attention processing.

[0046] The high-frequency connection features are subjected to attention processing to obtain high-frequency attention features, including: performing recombination processing on the high-frequency connection features to obtain recombination features; performing three-branch processing on the recombination features to obtain first-branch features, second-branch features and third-branch features respectively; using the product of the second-branch features and the third-branch features to obtain first-branch product features; and using the product of the first-branch features and the first-branch product features to obtain high-frequency attention features.

[0047] The high-frequency connection features are recombined to obtain recombined features. This includes: performing block processing on the high-frequency connection features in the width and height directions to obtain multiple block features; and performing connection processing on the block features in the height direction to obtain the recombined features. The number of blocks can be set according to requirements, such as 2*2, or other numbers; this disclosure does not specifically limit this. By configuring the recombined features, feature information can be recombined, establishing associations and distinctions between feature information at different locations, thereby improving model performance.

[0048] The reconstructed features are processed in three branches to obtain first-branch features, second-branch features, and third-branch features. This includes: using the first and second high-frequency branches to sequentially perform feature centering, square root transformation, channel orientation standardization, and width and height plane standardization on the reconstructed features. Feature centering involves subtracting the mean from the reconstructed features to obtain features with a mean of 0 and a standard deviation of 1, following a standard normal distribution. Square root transformation applies the square root to the features, ensuring they conform to a normal distribution. Channel orientation standardization and width and height plane standardization normalize the features from different dimensions. The output of the third high-frequency branch is the reconstructed feature, which is directly used as the output of the third-branch feature for subsequent processing.

[0049] Furthermore, the product of the second branch features and the third branch features is used to obtain the first branch product features; the product of the first branch features and the first branch product features is used to obtain the high-frequency attention features. Given the high-frequency attention features, feature transformation processing is performed on them to obtain the first fused features. The feature transformation includes transforming the feature dimensions; through block processing in the width and height directions, the features are restored to the dimensions of the high-frequency features within the high-frequency feature group, thus obtaining the first fused features.

[0050] Based on the above configuration, high-frequency features are fused to obtain effective high-frequency features from different perspectives, thereby optimizing the algorithm accuracy.

[0051] Having obtained the first and second fusion features, a third fusion feature is further obtained by connecting the first and second fusion features along the channel direction. The location of the lesion region object can then be obtained through the third fusion feature. Detecting the lesion region object in the input image using the third fusion feature includes: directly performing activation processing on the third fusion feature to obtain the location information of the lesion region object, wherein the location of the lesion region object is where the feature value is greater than 0.5 in the activated feature. Alternatively, the location of the lesion region object can be obtained by performing convolution processing on the third fusion feature followed by activation processing. This disclosure does not specifically limit this approach.

[0052] Based on the above configuration, the embodiments of this disclosure can accurately segment the cerebral ischemia lesion area, providing a good foundation for subsequent processing and helping to improve the accuracy of the algorithm.

[0053] Having obtained the lesion area, a control normal area can be further obtained from the lesion area. This embodiment utilizes the principle of brain symmetry, using the brain symmetry line as a reference to obtain the normal area on the symmetrical side of the brain lesion area. The method for determining the brain symmetry line includes first determining the image centroid and the image deflection angle (which can be obtained simultaneously during image acquisition) based on the brain image. The line passing through the centroid with the deflection angle as its direction is the brain symmetry line. Alternatively, the center line of the brain image can also be used as the symmetry line; this disclosure does not specifically limit this approach.

[0054] Having obtained the lesion region and the corresponding normal region, key distinguishing features that can differentiate between the normal region and the lesion region can be further obtained, including: acquiring a first brain image set; segmenting cerebral blood vessels and the lesion region from the first brain image set; extracting vascular morphological features and dynamic blood flow features based on the lesion region, cerebral blood vessels, and the normal region on the symmetrical side of the lesion region, respectively; and determining key distinguishing features that can differentiate between the normal region and the lesion region based on the vascular morphological features and the dynamic blood flow features.

[0055] This disclosure embodiment utilizes brain symmetry mechanisms and feature statistics to obtain key distinguishing features. Specifically, a first brain image set can be acquired. The image type of the first brain image set can be the same as the brain medical image types described in the previous embodiments, such as PWI or CTP images. The acquisition method is also the same and will not be repeated here. The first brain image set includes multiple brain medical images; the more images, the more accurate the subsequent analysis results. This disclosure embodiment includes at least 20 examples, but this is only for illustrative purposes and no specific requirement is made.

[0056] Given a set of first-order brain images, cerebral blood vessels and lesion regions can be segmented from these images. This segmentation can be achieved using a network structure. In practice, the network model can be trained to segment these regions. Once the cerebral blood vessels and lesion regions are obtained, the principle of brain symmetry can be used to identify the normal region on the symmetrical side of the lesion region, based on the line of symmetry. Then, vascular morphology and dynamic blood flow features can be extracted from the lesion region, the cerebral blood vessels, and the normal region on the symmetrical side of the lesion region. The method for determining the line of symmetry includes first determining the image centroid and the image deflection angle based on the brain images; the line passing through the centroid with the deflection angle as its direction is the line of symmetry.

[0057] In addition, in this embodiment of the disclosure, vascular morphological features are extracted based on the lesion region, cerebral blood vessels, and the normal region on the symmetrical side of the lesion region, including: calculating morphological parameters such as vessel length, volume, cross-sectional area, and curvature of the cerebral blood vessels in the lesion region and the normal region to obtain the vascular morphological features; correspondingly, the vascular morphological features may include information such as vessel length, volume, cross-sectional area, curvature, and curvature. Other morphological features may also be included in other embodiments, and this disclosure does not specifically limit them.

[0058] Based on the lesion region, cerebral blood vessels, and the normal region symmetrical to the lesion region, dynamic blood flow features are extracted by performing dynamic radiomics feature extraction and multi-dimensional feature screening on the brain tissue of the lesion region and the normal region to obtain the dynamic blood flow features. The images in this embodiment are 4D images, composed of 3D images scanned at multiple time points. Therefore, this embodiment can extract radiomics features from the lesion region and the normal region of each 3D image separately, obtaining dynamic radiomics features at various time points within a continuous time period. In other words, the dynamic radiomics features are radiomics features of the lesion region and the normal region at multiple time points. Based on this, dynamic radiomics features of the lesion region and dynamic radiomics features of the normal region can be obtained.

[0059] Having obtained the dynamic radiomics features of both the lesion and normal regions, a multi-dimensional feature selection process is employed to identify dynamic blood flow features that can distinguish between the normal and lesion regions. This multi-dimensional feature selection process includes: using a T-test to select significant features, for example, selecting features with a significance greater than 0.05; and using the Lasso algorithm to select dynamic blood flow features from these significant features. In other embodiments, clustering algorithms, analysis of variance, or other feature selection methods can also be used to obtain dynamic blood flow features; this disclosure does not specifically limit this approach.

[0060] Based on the above process, the vascular morphological features and dynamic blood flow features of normal and lesion areas can be obtained. Combining these two types of features constitutes the key distinguishing features. The key distinguishing features obtained in this embodiment can effectively differentiate the vascular morphological features and dynamic blood flow features of normal and abnormal areas, taking into account the possibility of tissue abnormalities from both static and dynamic perspectives, thus improving the accuracy of tissue differentiation.

[0061] Having obtained key distinguishing features that can differentiate between normal and abnormal tissues, the hidden damaged tissue boundary around the lesion area is determined based on the key distinguishing features, including: using a surrounding tissue division strategy to divide the surrounding tissue around the lesion area into multiple levels; and determining the damaged tissue boundary around the lesion area based on the key distinguishing features within the multiple levels of surrounding tissue.

[0062] Having obtained the lesion area, this embodiment further performs multi-level tissue division around the lesion area. That is, multiple levels of surrounding tissue areas are sequentially obtained around the lesion area in order from near to far. In some possible implementations, the multi-level surrounding tissue division around the lesion area using a surrounding tissue division strategy includes: expanding outwards from the lesion area according to a preset voxel count, obtaining a first-level surrounding area for each preset voxel expansion; and obtaining the multi-level surrounding areas based on multiple preset voxel expansions. Specifically, expanding outwards according to the preset voxel count includes: using the voxel coordinates on the outer boundary of the lesion area as a reference, expanding outwards to adjacent outer regions according to a preset number of voxels, with the expanded regions forming the first-level surrounding areas.

[0063] The innermost region is the obtained lesion region. Starting from the outer edge of the lesion region, a preset voxel is extended to the adjacent peripheral region. For example, the number of preset voxels in this embodiment can be a region of [1, 10], but this is not a specific limitation of this disclosure. In this embodiment, the number of preset voxels can be 4. That is, by extending 4 voxels outward from the lesion region, a first-level peripheral region can be obtained. Then, starting from the outer periphery of the first-level peripheral region, preset voxels are extended to the adjacent outer side to obtain a second-level peripheral region; and so on, multiple levels of peripheral regions can be obtained successively. In other embodiments, at least two levels of peripheral regions can be obtained, and this disclosure does not specifically limit this. In addition, the layer expansion method in this embodiment can include an expansion method, and those skilled in the art can choose other methods to expand, and this disclosure does not limit this.

[0064] In other embodiments, the method of using a surrounding tissue division strategy to divide the lesion area into multiple levels of surrounding tissue includes: based on the lesion area, expanding outward at a preset distance to form multiple preset shapes; the boundary of the lesion area and the first preset shape, as well as the adjacent preset shapes, constitute the multi-level surrounding area.

[0065] In other words, in this embodiment of the disclosure, the area can be expanded outward according to the shape of a preset graphic to obtain multiple levels of surrounding areas. The preset shape may include shapes such as circles, ellipses, and squares, expanding outward from the lesion area as the center at a preset distance. A schematic diagram of a multi-level surrounding tissue division result according to another embodiment of the present disclosure is shown. Taking an ellipse as an example, the shape parameters of the ellipse can be set according to requirements. For example, the semi-minor axis of the ellipse can be set to a preset number of voxels, and the semi-major axis can be set to twice the preset number of voxels, where the preset number of voxels can be 4 voxels or other set values. In this way, multiple preset shapes can be successively expanded around the lesion area, and the distance between each preset shape can be a preset number of voxels or other set values. In other embodiments of the present disclosure, other shapes can also be used for expansion, and the present disclosure does not specifically limit this.

[0066] In addition, the embodiments of this disclosure may also set termination conditions for the division of the surrounding area, which may include at least one of the following: (1) When expanding the surrounding area outward, the expansion operation shall be terminated if the expanded area reaches or exceeds the image boundary; (2) The expansion operation is terminated when the number of expanded areas reaches the preset requirement. The preset requirement can be a number of expansions greater than 2, which can be set by those skilled in the art according to their needs.

[0067] Based on the above configuration, embodiments of this disclosure can obtain multiple surrounding tissue regions around the lesion region. With these surrounding tissue regions obtained, the boundary of damaged tissue around the lesion region can be determined based on key distinguishing features within the multi-level surrounding tissue regions. The brain tissue between the lesion region and the boundary of the damaged tissue is a region of abnormal blood flow that is difficult to observe based on brain medical imaging; the key distinguishing features can be used to differentiate between normal tissue and diseased tissue.

[0068] In some possible implementations, the boundary of damaged tissue around the lesion area can be determined based on key distinguishing features within the multi-level surrounding tissues, including: sequentially judging whether the key distinguishing features in each level of the surrounding area meet preset conditions in order of distance from the lesion area; if the preset conditions are met, the corresponding surrounding area is determined as damaged tissue, and the key distinguishing features of the next surrounding area are judged to meet the preset conditions, until the boundary of the corresponding surrounding area is determined as the boundary of damaged tissue if the preset conditions are not met.

[0069] The step of judging whether the key distinguishing features in the surrounding areas at each level meet the preset conditions in order of distance from the lesion area, from near to far, includes at least one of the following methods: (1) Determine whether the similarity between the key distinguishing features in each of the surrounding areas and the key distinguishing features in the lesion area is greater than a first threshold; In some possible implementations, key distinguishing features of the surrounding tissue at each level can be sequentially acquired from near to far of the lesion region. The similarity between the key distinguishing features of the surrounding tissue at each level and the key distinguishing features of the lesion region can then be obtained. If the similarity is greater than a first threshold, the tissue in that surrounding region is determined to be damaged tissue, and the outer boundary of the corresponding surrounding region is the boundary of the damaged tissue. If the similarity is less than or equal to the first threshold, the key distinguishing features of the surrounding tissue at the next level are acquired and compared with the key distinguishing features of the lesion region until the damaged boundary is found. The first threshold can be a value greater than or equal to 0.5, which can be set by those skilled in the art according to their needs.

[0070] In some possible implementations, the similarity is determined by calculating at least one or more of the following: cosine similarity, adjusted cosine similarity, Pearson correlation coefficient, Jaccard similarity coefficient, log-likelihood similarity, and Tanimoto coefficient between key distinguishing features in the surrounding region and key distinguishing features in the lesion region. When multiple similarities are calculated, the mean of all similarities is used as the final similarity. The above is merely illustrative and does not constitute a specific limitation.

[0071] (2) Based on the key distinguishing features in the surrounding area, identify the proportion of abnormal tissue in the surrounding area and determine whether the proportion is greater than the second threshold.

[0072] In some possible implementations, an abnormal tissue proportion judgment model is first obtained. This model is then used to determine key distinguishing features within the surrounding area, identifying the proportion of abnormal tissue. If this proportion is greater than a second threshold, the surrounding area is determined to be damaged tissue, and its outer boundary is identified as the damaged tissue boundary. If the proportion is less than or equal to the second threshold, the key distinguishing features of the surrounding area at the next level are obtained, and the abnormal tissue proportion is judged again, until the damaged boundary is found. The second threshold can be a value greater than or equal to 0.6, which can be set by those skilled in the art according to their needs.

[0073] The abnormal tissue proportion judgment model can be a trained model capable of identifying abnormal proportions within tissues; this model can be a machine learning model or a deep learning model. This disclosure does not impose specific limitations on it. Generally, because different brain regions have different structures, if the proportion obtained by this abnormal tissue proportion model exceeds a second threshold, it can be determined that the brain tissue region is damaged.

[0074] (3) In some possible implementations, it is determined whether the distance between the key distinguishing features in each of the surrounding areas and the key distinguishing features in the lesion area is less than a third threshold; In some possible implementations, key distinguishing features of the surrounding tissue at each level can be sequentially acquired from near to far of the lesion region. The distance between the key distinguishing features of the surrounding tissue at each level and the key distinguishing features of the lesion region can then be obtained. If this distance is less than a third threshold, the tissue in that surrounding region is determined to be damaged tissue, and the outer boundary of the corresponding surrounding region is the boundary of the damaged tissue. If the similarity is greater than or equal to the third threshold, the key distinguishing features of the surrounding tissue at the next level are acquired and compared with the key distinguishing features within the lesion region until the damaged boundary is found. The third threshold can be a value greater than or equal to 0.5, which can be set by those skilled in the art according to their needs.

[0075] In some possible implementations, the distance is determined by calculating at least one or more of the following: Euclidean distance, normalized Euclidean distance, Mahalanobis distance, Manhattan distance, Chebyshev distance, Minkowski distance, and Hamming distance between key distinguishing features in the surrounding region and key distinguishing features in the lesion region. When multiple distances are calculated, the mean of all distances is used as the final distance. The above is merely illustrative and does not constitute a specific limitation herein.

[0076] Having obtained the boundary of the damaged tissue, the prognostic prediction model can be further constructed using the brain tissue features within the damaged tissue boundary. It should be noted that the prognostic prediction model constructed in this embodiment can determine the recovery level of stroke patients within three months after treatment, for example, it can predict the mRS score. This score can be one of seven levels from 0 to 6, with a lower score indicating a better recovery level. There are seven levels in total: 0: Completely asymptomatic; 1: Symptoms present, but no obvious disability, able to complete all regularly performed work and activities; 2: Mild disability, unable to complete all work and activities, but able to handle personal affairs without assistance; 3: Moderate disability, requires assistance, but can walk without assistance; 4: Severe disability, unable to walk without assistance, unable to care for oneself; 5: Severe disability, bedridden, incontinent, requiring continuous care, requiring multiple 24-hour care; 6: Death. This embodiment can obtain a training dataset in advance, which includes brain images and corresponding prognostic scores. Based on this, the prognostic prediction model can be trained. However, the feature information required for training the model can be of various types, and this disclosure does not specifically limit it. The method for constructing the prediction model according to this disclosure includes at least one of the following: ① Construct the prognostic prediction model based on the key distinguishing features within the damaged boundary; In some possible implementations, vascular morphology and dynamic blood flow characteristics within brain tissue regions within the lesion boundary (including the lesion area) can be utilized; these are key distinguishing features within the lesion boundary. A prognostic prediction model is constructed by inputting these key distinguishing features within the lesion boundary into the model and training it according to the mRS score.

[0077] ② The first key distinguishing feature within the lesion area and the second key distinguishing feature between the lesion area and the damaged boundary are weighted and processed to construct the prognostic prediction model using the weighted features; In some possible implementations, a weighted average can be applied to the first key distinguishing feature within the lesion region and the second key distinguishing feature of the hidden damaged tissue (brain tissue between the lesion region and the damaged boundary). For example, the weight of the first key distinguishing feature can be a, and the weight of the second key distinguishing feature can be b. The corresponding weights are multiplied by the key distinguishing features and then summed to obtain the weighted features. The weighted features are then input into the model and trained according to the mRS score to construct a prognostic prediction model.

[0078] ③ Based on the brain tissue images within the damaged boundary, the prognostic prediction model is constructed.

[0079] In some possible implementations, after obtaining the damaged boundary, images of the damaged tissue within the damaged boundary (including the lesion area) in brain medical imaging can be obtained. These damaged tissue images are then input into the model, trained according to the mRS score, to construct a prognostic prediction model. Alternatively, in other implementations, images of the lesion area and the tissue area between the lesion area and the damaged tissue boundary can be obtained separately. These two images can be connected along their respective channel directions to obtain a connected image. This connected image is then input into the model, trained according to the mRS score, to construct a prognostic prediction model.

[0080] The model in this embodiment can be a ResNet model, a UNet model, or other deep learning models, and this disclosure does not specifically limit it. The model can also be a network framework, as long as the output is a classification value.

[0081] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0082] In addition, this disclosure also provides a prognostic prediction device, electronic device, computer-readable storage medium, and program, all of which can be used to implement any of the prognostic prediction model construction methods provided in this disclosure. The corresponding technical solutions and descriptions are described in the corresponding records in the method section and will not be repeated here.

[0083] This disclosure also provides a prognostic prediction device, constructed using the prognostic prediction model construction method described in the above embodiments, comprising: a first determining module, used to determine key distinguishing features that can differentiate normal regions from lesion regions based on brain medical images; a second determining module, used to determine the hidden damaged tissue boundary around the lesion region based on the key distinguishing features; wherein the brain tissue between the lesion region and the damaged tissue boundary is a hidden abnormal blood flow tissue region that is difficult to detect directly based on brain medical images; and a construction module, used to construct the prognostic prediction model based on the combined blood flow features within the damaged boundary.

[0084] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0085] This disclosure also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method. The computer-readable storage medium may be a non-volatile computer-readable storage medium.

[0086] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technical improvements to the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for constructing a prognostic prediction model, characterized in that, include: Based on brain medical imaging, key distinguishing features can be identified to differentiate between normal and lesion areas; Based on the key distinguishing features, the boundary of the hidden damaged tissue surrounding the lesion area is determined; wherein, the brain tissue between the lesion area and the boundary of the damaged tissue is a hidden blood flow abnormality tissue area that is difficult to detect directly based on brain medical imaging. The prognostic prediction model is constructed based on the brain tissue characteristics within the damaged boundary.

2. The method according to claim 1, characterized in that, The construction of the prognostic prediction model based on brain tissue features within the damaged boundary includes at least one of the following methods: The prognostic prediction model is constructed based on the key distinguishing features within the damaged boundary. The first key distinguishing feature within the lesion area and the second key distinguishing feature between the lesion area and the damaged boundary are weighted and processed to construct the prognostic prediction model using the weighted features. The prognostic prediction model is constructed based on images of brain tissue within the damaged boundary.

3. The method according to claim 1, characterized in that, The step of determining the boundary of the hidden damaged tissue surrounding the lesion area based on the key distinguishing features includes: The surrounding tissue is divided into multiple levels using a surrounding tissue division strategy; Based on the key distinguishing features within the surrounding multi-level tissues, the boundaries of the damaged tissues surrounding the lesion area are determined.

4. The method according to claim 3, characterized in that, The method of using a surrounding tissue delineation strategy to perform multi-level surrounding tissue delineation around the lesion area includes: Based on the lesion area, expand outward according to a preset voxel, and obtain a first-level surrounding area for each preset voxel expansion; Based on the expansion of multiple preset voxels, the surrounding regions of the multi-level are obtained; and / or The expansion outward according to a preset voxel includes: Based on the voxel coordinates on the outer boundary of the lesion area, an extended region is formed to the adjacent outer side according to a preset number of voxels, and the extended region forms a primary surrounding area.

5. The method according to claim 3, characterized in that, The method of using a surrounding tissue delineation strategy to perform multi-level surrounding tissue delineation around the lesion area includes: Based on the lesion area, multiple preset patterns are formed by expanding outward at preset distances; The boundary of the lesion area and the first preset pattern, as well as the adjacent preset patterns, constitute the surrounding area of ​​the multi-level system.

6. The method according to claim 3, characterized in that, The determination of the damaged tissue boundary surrounding the lesion region based on key distinguishing features within the multi-level surrounding tissues includes: According to the order of distance from the lesion area to the farthest, determine whether the key distinguishing features in each of the surrounding areas meet the preset conditions; If the preset conditions are met, the corresponding surrounding area is identified as damaged tissue, and the key blood flow characteristics of the next surrounding area are judged to meet the preset conditions until the preset conditions are not met, at which point the boundary of the corresponding surrounding area is identified as the boundary of damaged tissue.

7. The method according to claim 6, characterized in that, The step of sequentially determining whether key distinguishing features in each surrounding area meet preset conditions in order of distance from the lesion region to the farthest region includes at least one of the following methods: Determine whether the similarity between key distinguishing features in each of the surrounding areas and key distinguishing features in the lesion area is greater than a first threshold; Based on key distinguishing features within the surrounding area, the proportion of abnormal tissue within the surrounding area is identified, and it is determined whether the proportion is greater than a second threshold. Determine whether the distance between the key distinguishing features in each of the surrounding areas and the key distinguishing features in the lesion area is less than a third threshold.

8. A prognostic prediction device, characterized in that, The model is constructed using the prognostic prediction model construction method according to any one of claims 1-7, comprising: The first determination module is used to determine key distinguishing features that can differentiate normal areas from lesion areas based on brain medical images; The second determining module is used to determine the boundary of hidden damaged tissue around the lesion area based on the key distinguishing features; wherein, the brain tissue between the lesion area and the boundary of the damaged tissue is a hidden blood flow abnormality tissue area that is difficult to detect directly based on brain medical imaging. A construction module is used to construct the prognostic prediction model based on the combined blood flow characteristics within the damaged boundary.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images