Method and system for automatic lesion segmentation and scoring of brainstem ischemic stroke

Abstract

This invention provides an automatic segmentation and scoring method and system for lesions in ischemic stroke of the brainstem, including preprocessing brain CT images; automatic segmentation of the brainstem; and acquisition of mirror brainstem images and a brainstem atlas containing the midbrain, pons, and medulla oblongata regions. stem The method involves region segmentation of brainstem and mirror brainstem images; construction of a brainstem infarction lesion detection and segmentation network model; the brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder, and six difference calculation modules. Each encoder consists of three convolutional layers, each decoder consists of three deconvolutional layers, and the difference calculation module includes multi-scale pyramid convolutions and feature fusion modules at each scale; prediction scoring is performed based on the segmented lesions. The lesion detection and segmentation method of this invention can quickly, accurately, and objectively detect and segment lesions.

Description

Technical Field

[0001] This invention relates to the technical field of medical image processing, specifically to a method for automatic segmentation and scoring of lesions in brainstem ischemic stroke based on deep learning, and more particularly to a method for automatic segmentation and scoring of lesions in brainstem ischemic stroke based on deep learning. Background Technology

[0002] Brainstem ischemic stroke refers to cerebral infarction caused by acute vertebrobasilar artery occlusion, accounting for approximately 10% of all acute ischemic strokes. Brainstem infarction has a very poor prognosis, leading to severe neurological impairment and a very high mortality rate. Early diagnosis and intervention are crucial for improving prognosis; however, the presence of pseudostrokes in clinical practice makes diagnosis challenging. To determine the optimal reperfusion method, lesion detection and brainstem scoring are necessary.

[0003] For ischemic stroke, diffusion-weighted magnetic resonance imaging (MRI), digital subtraction angiography (DSA), and CT angiography (CTA) can be used to detect brainstem infarcts. However, MRI equipment and examination costs are high, its availability is low, and the examination time is long. DSA and CTA are invasive examinations and also have long examination times, none of which can meet the requirements for early detection, early diagnosis, and early treatment. Compared with the above imaging techniques, non-contrast computerized CT (NCCT) is non-invasive, fast, and widely available. It can quickly image infarct lesions and is the preferred examination method in the routine treatment of acute stroke.

[0004] Despite its numerous advantages, identifying stroke lesions in NCCT images remains challenging in the early stages of stroke symptoms, especially brainstem infarction. This is because, in the early stages of infarction, changes in density and volume of the infarct area on the image are not significant, and the brainstem region is easily affected by skull artifacts. Artificial intelligence (AI) is a cutting-edge discipline in current scientific and technological development. Its powerful post-processing capabilities and further learning and analysis abilities are gradually gaining recognition in the medical field and are widely used. With the development of AI technology, stroke lesions that are difficult for the human eye to identify can be trained and automatically detected using convolutional neural networks (CNNs). However, current lesion detection research suffers from problems such as inaccurate training set labeling, excessive time consumption, and the invalidation of manually labeled standards due to data acquisition delays.

[0005] A patent document with publication number CN113808191A discloses an automatic quantification and three-dimensional modeling method for lesion areas in acute ischemic stroke, comprising: desquamation of brain images from DWI sequences; desquamation of brain images from ADC sequences; creating new magnetic resonance diffusion-weighted imaging sequences based on the desquamated DWI and ADC brain images; segmenting the lesion area; desquamating the calculated T1 sequence brain images; partitioning the T1 brain images based on the desquamated T1 brain images into two types of brain anatomical structures; mapping the identified lesion areas to the two types of brain anatomical structure partitions, calculating the lesion volume and proportion of each brain anatomical structure based on the anatomical structure; and performing three-dimensional modeling based on the mapping results.

[0006] Therefore, a new technical solution is needed to improve the above-mentioned technical problems. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the purpose of this invention is to provide an automatic segmentation and scoring method and system for lesions in brainstem ischemic stroke.

[0008] The present invention provides an automatic segmentation and scoring method for lesions in ischemic stroke of the brainstem, the method comprising the following steps:

[0009] Step S1: Obtain plain CT images of the brain, obtain standard brain atlases, and perform registration and preprocessing on the brain CT images;

[0010] Step S2: Automatic segmentation of the brainstem is performed using a 3D U-net network model. The 3D U-net network model includes an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers. The loss function is the standard Dice score.

[0011] Step S3: Acquire mirror images of the brainstem and a brainstem atlas containing the midbrain, pons, and medulla oblongata regions. stem Regions were divided into brainstem and mirror brainstem images;

[0012] Step S4: Construct a brainstem infarction lesion detection and segmentation network model; The brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale.

[0013] Step S5: Perform a predictive score based on the segmented lesions.

[0014] Preferably, step S1 includes the following steps:

[0015] Step S1.1: Obtain the deformation field from the plain brain CT image to the standard brain atlas;

[0016] Step S1.2: Register brain CT images using a deformation field;

[0017] Step S1.3: Remove the skull from the image obtained in step S1.2;

[0018] Step S1.4: Normalize the grayscale values ​​of the image obtained in step S1.3 to 0-255;

[0019] Step S1.5: Perform feature annotation on the image obtained in step S1.4.

[0020] Preferably, step S2 includes the following steps:

[0021] Step S2.1: Construct a 4-layer 3D U-net network model;

[0022] Step S2.2: Use the dataset processed in step S1.5 to train the 3D U-net network model and obtain the loss function and segmentation results;

[0023] Step S2.3: Adjust and optimize the network model parameters based on the loss function and segmentation results;

[0024] Step S2.4: Generate and save the optimized model after parameter tuning;

[0025] Step S2.5: Save the model segmentation results to obtain the brainstem image Stem. origin .

[0026] Preferably, step S3 includes the following steps:

[0027] Step S3.1: Process brainstem images. origin Perform a mirroring operation to obtain the mirror brainstem. mirror ;

[0028] Step S3.2: Based on the brainstem atlas Atlas stem to brainstem images Stem origin The mapping divides the brainstem regions;

[0029] Step S3.3: Based on the brainstem atlas Atlas stem To the mirror brainstem Stem mirror The mapping divides the brainstem regions.

[0030] Preferably, step S4 includes the following steps:

[0031] Step S4.1: Construct a brainstem infarction lesion detection and segmentation network model;

[0032] Step S4.2: The brainstem image Stem obtained in step S3.4... origin Mirror brainstem Stem mirror and the obtained brainstem atlas Atlas stem Input model;

[0033] Step S4.3: The three input datasets are encoded by the three encoders in the model to obtain feature maps;

[0034] Step S4.4: The brainstem image Stem obtained in step S4.2... origin The feature map is input into the decoder for decoding;

[0035] Step S4.5: Input the feature maps obtained in step S4.3 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas;

[0036] Step S4.6: In the difference calculation module, the two input feature maps M1 and M2 are first subtracted element-wise to obtain the difference feature map S. Then, pyramid convolution is performed at multiple scales. Assume that the difference feature map at the i-th scale is S. i The convolution result is obtained.

[0037]

[0038]

[0039] Where W1 represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction, w i This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution;

[0040] Step S4.7: Input the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, use average pooling to extract dense features F1(X) in its c-th channel. c ), and extract dense features F2(X) using the maximized output. c ):

[0041]

[0042]

[0043] Where H, W, Z, and C represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively, and h, w, z, and c represent the coordinates of a specific input during the calculation.

[0044] Step S4.8: First, combine the two dense features F1(X) c ) and F2(X c The activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and

[0045]

[0046] in, Represents the convolution operation. The parameters represent the dimensionality reduction fully connected layer, and μ represents the ReLU function. ω represents the parameters of the upgraded fully connected layer, and ω represents the sigmoid function.

[0047] Step S4.9: Element-wise summation of the two densely activated features yields the summed feature S. sum ;

[0048]

[0049] Where W2 represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation;

[0050] Step S4.10: Utilize the summation characteristic S sum The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features S. mix ;

[0051] S mix =S sum ⊙X

[0052] Where ⊙ represents channel-level multiplication operation;

[0053] Step S4.11: Use a skip connection to connect S mix Connecting the features of the corresponding resolution in the decoder obtained in step S4.4 yields the connection feature S. skip The connection features were obtained through a total of three connections. and

[0054] Step S4.12: For Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution

[0055]

[0056]

[0057] Among them, W 1′ and W 2′ They are respectively the corresponding The parameters of the deconvolution layer, Represents the convolution operation;

[0058] Step S4.13: For and Element-wise addition is performed using the ReLU activation function, and a multi-resolution fusion coefficient map S is generated using a 1×1×1 convolutional layer and the sigmoid function. map ;

[0059]

[0060] in, This represents element-wise addition, where μ represents the ReLU function. ω represents the convolution operation, W3 represents the parameters of a 1×1×1 convolutional layer, and ω represents the sigmoid function.

[0061] Step S4.14: Analyze the fusion coefficient map S map and the last level connection features Element-wise multiplication is used to obtain the segmentation feature S. seg ;

[0062]

[0063] in, Represents element-wise multiplication;

[0064] Step S4.15: Based on segmentation feature S seg The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

[0065] Preferably, step S5 includes the following steps:

[0066] Step S5.1: Based on the brainstem lesion mask, perform a step on the brainstem image. origin The region of interest (ROI) is obtained from the brainstem image and then placed within the brainstem image. origin By outlining the contours, a brainstem image containing the lesion region is obtained. focus ;

[0067] Step S5.2: Based on the brainstem atlas Atlas stemStem images of the brainstem containing lesions focus The brain was divided and located in the midbrain, pons, and medulla oblongata regions.

[0068] Step S5.3: Calculate the volume of lesions in each region of the midbrain, pons, and medulla oblongata using voxels;

[0069] Step S5.4: Calculate the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel;

[0070] Step S5.5: Score the lesions in each region separately. Add 2 points if the volume of the lesion is greater than 50% of the volume of the same site on one side, add 1 point if it is less than 50% of the volume of the same site on one side, and add no points if there are no lesions.

[0071] The present invention also provides an automatic lesion segmentation and scoring system for ischemic stroke in the brainstem, the system comprising the following modules:

[0072] Module M1: Acquire plain CT images of the brain, acquire standard brain atlases, and perform registration and preprocessing on the brain CT images;

[0073] Module M2: Automatic brainstem segmentation is performed using a 3D U-net network model. The 3D U-net network model includes an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers. The loss function is the standard Dice score.

[0074] Module M3: Acquires mirror images of the brainstem and brainstem atlases containing the midbrain, pons, and medulla oblongata regions. stem Regions were divided into brainstem and mirror brainstem images;

[0075] Module M4: Constructs a brainstem infarction lesion detection and segmentation network model; The brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale.

[0076] Module M5: Predicts and scores based on the segmented lesions.

[0077] Preferably, module M1 includes the following modules:

[0078] Module M1.1: Acquires the deformation field from plain brain CT images to standard brain atlases;

[0079] Module M1.2: Registering brain CT images using a deformation field;

[0080] Module M1.3: Performs skull removal on the image obtained from module M1.2;

[0081] Module M1.4: Normalizes the grayscale values ​​of the image obtained from module M1.3 to 0-255;

[0082] Module M1.5: Performs feature annotation on the image obtained from module M1.4;

[0083] Module M2 includes the following modules:

[0084] Module M2.1: Constructs a 4-layer 3D U-net network model;

[0085] Module M2.2: Uses the dataset processed by Module M1.5 to train the 3D U-net network model, and obtains the loss function and segmentation results;

[0086] Module M2.3: Adjusts and optimizes network model parameters based on loss function and segmentation results;

[0087] Module M2.4: Generates and saves the optimized model after parameter tuning;

[0088] Module M2.5: Saves the brainstem image Stem from the model segmentation results. origin .

[0089] Preferably, module M3 includes the following modules:

[0090] Module M3.1: For brainstem images... origin Perform a mirroring operation to obtain the mirror brainstem. mirror ;

[0091] Module M3.2: Based on brainstem atlas Atlas stem to brainstem images Stem origin The mapping divides the brainstem regions;

[0092] Module M3.3: Based on brainstem atlas Atlas stem To the mirror brainstem Stem mirror The brainstem region is mapped and divided; module M4 includes the following modules:

[0093] Module M4.1: Constructing a network model for detecting and segmenting brainstem infarcts;

[0094] Module M4.2: Module M3.4 generates brainstem images. origin Mirror brainstem Stem mirror and the obtained brainstem atlas Atlas stem Input model;

[0095] Module M4.3: The three input datasets are encoded into feature maps by the three encoders in the model, respectively;

[0096] Module M4.4: Module M4.2 generates brainstem images. origin The feature map is input into the decoder for decoding;

[0097] Module M4.5: Input the feature maps obtained from module M4.3 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas;

[0098] Module M4.6: In the difference calculation module, the two input feature maps M1 and M2 are first subtracted element-wise to obtain the difference feature map S. Then, pyramid convolution is performed at multiple scales. Assume that the difference feature map at the i-th scale is S. i The convolution result is obtained.

[0099]

[0100]

[0101] Where W1 represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction, w i This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution;

[0102] Module M4.7: Inputs the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, average pooling is used in its c-th channel to extract dense features F1(X). c ), and extract dense features F2(X) using the maximized output. c ):

[0103]

[0104]

[0105] Where H, W, Z, and C represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively, and h, w, z, and c represent the coordinates of a specific input during the calculation.

[0106] Module M4.8: First, combine two dense features F1(X) c ) and F2(X c The activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and

[0107]

[0108] in, Represents the convolution operation. The parameters represent the dimensionality reduction fully connected layer, and μ represents the ReLU function. ω represents the parameters of the upgraded fully connected layer, and ω represents the sigmoid function.

[0109] Module M4.9: Element-wise addition of two densely activated features yields the summed feature S. sum ;

[0110]

[0111] Where W2 represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation;

[0112] Module M4.10: Utilizing the summation feature S sum The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features S. mix ;

[0113] S mix =S sum ⊙X

[0114] Where ⊙ represents channel-level multiplication operation;

[0115] Module M4.11: Using skip connections to connect S mix The connection feature S is obtained by connecting the corresponding resolution features in the decoder obtained from module M4.4. skip The connection features were obtained through a total of three connections. and

[0116] Module M4.12: for Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution

[0117]

[0118]

[0119] Among them, W 1′ and W 2′ They are respectively the corresponding The parameters of the deconvolution layer, Represents the convolution operation;

[0120] Module M4.13: For and Element-wise addition is performed using the ReLU activation function, and a multi-resolution fusion coefficient map S is generated using a 1×1×1 convolutional layer and the sigmoid function. map ;

[0121]

[0122] in, This represents element-wise addition, where μ represents the ReLU function. ω represents the convolution operation, W3 represents the parameters of a 1×1×1 convolutional layer, and ω represents the sigmoid function.

[0123] Module M4.14: For the fusion coefficient map S map and the last level connection features Element-wise multiplication is used to obtain the segmentation feature S. seg ;

[0124]

[0125] in, Represents element-wise multiplication;

[0126] Module M4.15: Based on segmentation feature S seg The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

[0127] Preferably, module M5 includes the following modules:

[0128] Module M5.1: Based on brainstem lesion masks in brainstem images origin The region of interest (ROI) is obtained from the brainstem image and then placed within the brainstem image. origin By outlining the contours, a brainstem image containing the lesion region is obtained. focus ;

[0129] Module M5.2: Based on the brainstem atlas Atlas stem Stem images of the brainstem containing lesions focus The brain was divided and located in the midbrain, pons, and medulla oblongata regions.

[0130] Module M5.3: Calculates the volume of lesions in the midbrain, pons, and medulla oblongata by voxel;

[0131] Module M5.4: Calculates the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel;

[0132] Module M5.5: Scores are given for lesions in each region. 2 points are added when the volume of the lesion is greater than 50% of the volume of the same region on one side, 1 point is added when it is less than 50% of the volume of the same region on one side, and no points are added when there are no lesions.

[0133] Compared with the prior art, the present invention has the following beneficial effects:

[0134] 1. This invention achieves a highly automated and integrated workflow through a series of deep convolutional neural network models, reducing the operational difficulty for doctors, improving their clinical diagnostic efficiency, and saving time from onset to treatment through rapid diagnosis, effectively reducing the occurrence of neurological damage and death in cases.

[0135] 2. This invention is based on non-contrast CT, which has a high penetration rate and low cost, to detect and segment brainstem ischemic stroke lesions with poor prognosis and the greatest diagnostic difficulty. It can quickly and accurately calculate the volume and proportion of lesions in each region of the brainstem, and provide objective and scientific scoring for the lesions, providing quantitative reference opinions for doctors' clinical diagnosis. Attached Figure Description

[0136] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0137] Figure 1 A flowchart is provided for an embodiment of the present invention to illustrate a method for lesion detection and brainstem scoring of ischemic stroke in the brainstem based on deep learning.

[0138] Figure 2 This is a schematic diagram of the structure of the model in an embodiment of the present invention. Detailed Implementation

[0139] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.

[0140] Example 1:

[0141] The present invention provides an automatic segmentation and scoring method for lesions in ischemic stroke of the brainstem, the method comprising the following steps:

[0142] Step S1: Obtain plain CT images of the brain, obtain standard brain atlases, and perform registration and preprocessing on the brain CT images;

[0143] Step S1.1: Obtain the deformation field from the plain brain CT image to the standard brain atlas;

[0144] Step S1.2: Register brain CT images using a deformation field;

[0145] Step S1.3: Remove the skull from the image obtained in step S1.2;

[0146] Step S1.4: Normalize the grayscale values ​​of the image obtained in step S1.3 to 0-255;

[0147] Step S1.5: Perform feature annotation on the image obtained in step S1.4.

[0148] Step S2: Automatic segmentation of the brainstem is performed using a 3D U-net network model. The 3D U-net network model includes an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers. The loss function is the standard Dice score.

[0149] Step S2.1: Construct a 4-layer 3D U-net network model;

[0150] Step S2.2: Use the dataset processed in step S1.5 to train the 3D U-net network model and obtain the loss function and segmentation results;

[0151] Step S2.3: Adjust and optimize the network model parameters based on the loss function and segmentation results;

[0152] Step S2.4: Generate and save the optimized model after parameter tuning;

[0153] Step S2.5: Save the model segmentation results to obtain the brainstem image Stem. origin .

[0154] Step S3: Acquire mirror images of the brainstem and a brainstem atlas containing the midbrain, pons, and medulla oblongata regions. stem Regions were divided into brainstem and mirror brainstem images;

[0155] Step S3.1: Process brainstem images. origin Perform a mirroring operation to obtain the mirror brainstem. mirror ;

[0156] Step S3.2: Based on the brainstem atlas Atlas stem to brainstem images Stem origin The mapping divides the brainstem regions;

[0157] Step S3.3: Based on the brainstem atlas Atlas stem To the mirror brainstem Stem mirrorThe brainstem region is divided by mapping. Step S4: Construct a brainstem infarction lesion detection and segmentation network model; The brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale.

[0158] Step S4.1: Construct a brainstem infarction lesion detection and segmentation network model;

[0159] Step S4.2: The brainstem image Stem obtained in step S3.4... origin Mirror brainstem Stem mirror and the obtained brainstem atlas Atlas stem Input model;

[0160] Step S4.3: The three input datasets are encoded by the three encoders in the model to obtain feature maps;

[0161] Step S4.4: The brainstem image Stem obtained in step S4.2... origin The feature map is input into the decoder for decoding;

[0162] Step S4.5: Input the feature maps obtained in step S4.3 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas;

[0163] Step S4.6: In the difference calculation module, the two input feature maps M1 and M2 are first subtracted element-wise to obtain the difference feature map S. Then, pyramid convolution is performed at multiple scales. Assume that the difference feature map at the i-th scale is S. i The convolution result is obtained.

[0164]

[0165]

[0166] Where W1 represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction, w i This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution;

[0167] Step S4.7: Input the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, use average pooling to extract dense features F1(X) in its c-th channel. c ), and extract dense features F2(X) using the maximized output. c ):

[0168]

[0169]

[0170] Where H, W, Z, and C represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively, and h, w, z, and c represent the coordinates of a specific input during the calculation.

[0171] Step S4.8: First, combine the two dense features F1(X) c ) and F2(X c The activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and

[0172]

[0173] in, Represents the convolution operation. The parameters represent the dimensionality reduction fully connected layer, and μ represents the ReLU function. ω represents the parameters of the upgraded fully connected layer, and ω represents the sigmoid function.

[0174] Step S4.9: Element-wise summation of the two densely activated features yields the summed feature S. sum ;

[0175]

[0176] Where W2 represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation;

[0177] Step S4.10: Utilize the summation characteristic S sum The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features S. mix ;

[0178] S mix =S sum ⊙X

[0179] Where ⊙ represents channel-level multiplication operation;

[0180] Step S4.11: Use a skip connection to connect S mix Connecting the features of the corresponding resolution in the decoder obtained in step S4.4 yields the connection feature S. skip The connection features were obtained through a total of three connections. and

[0181] Step S4.12: For Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution

[0182]

[0183]

[0184] Among them, W 1′ and W 2′ They are respectively the corresponding The parameters of the deconvolution layer, Represents the convolution operation;

[0185] Step S4.13: For and Element-wise addition is performed using the ReLU activation function, and a multi-resolution fusion coefficient map S is generated using a 1×1×1 convolutional layer and the sigmoid function. map ;

[0186]

[0187] in, This represents element-wise addition, where μ represents the ReLU function. ω represents the convolution operation, W3 represents the parameters of a 1×1×1 convolutional layer, and ω represents the sigmoid function.

[0188] Step S4.14: Analyze the fusion coefficient map S map and the last level connection features Element-wise multiplication is used to obtain the segmentation feature S. seg ;

[0189]

[0190] in, Represents element-wise multiplication;

[0191] Step S4.15: Based on segmentation feature S seg The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

[0192] Step S5: Perform a predictive score based on the segmented lesions.

[0193] Step S5.1: Based on the brainstem lesion mask, perform a step on the brainstem image. origin The region of interest (ROI) is obtained from the brainstem image and then placed within the brainstem image.origin By outlining the contours, a brainstem image containing the lesion region is obtained. focus ;

[0194] Step S5.2: Based on the brainstem atlas Atlas stem Stem images of the brainstem containing lesions focus The brain was divided and located in the midbrain, pons, and medulla oblongata regions.

[0195] Step S5.3: Calculate the volume of lesions in each region of the midbrain, pons, and medulla oblongata using voxels;

[0196] Step S5.4: Calculate the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel;

[0197] Step S5.5: Score the lesions in each region separately. Add 2 points if the volume of the lesion is greater than 50% of the volume of the same site on one side, add 1 point if it is less than 50% of the volume of the same site on one side, and add no points if there are no lesions.

[0198] Example 2:

[0199] Example 2 is a preferred embodiment of Example 1, and is used to illustrate the present invention in more detail.

[0200] The present invention also provides an automatic lesion segmentation and scoring system for ischemic stroke in the brainstem, the system comprising the following modules:

[0201] Module M1: Acquire plain CT images of the brain, acquire standard brain atlases, and perform registration and preprocessing on the brain CT images;

[0202] Module M1.1: Acquires the deformation field from plain brain CT images to standard brain atlases;

[0203] Module M1.2: Registering brain CT images using a deformation field;

[0204] Module M1.3: Performs skull removal on the image obtained from module M1.2;

[0205] Module M1.4: Normalizes the grayscale values ​​of the image obtained from module M1.3 to 0-255;

[0206] Module M1.5: Performs feature annotation on the image obtained from module M1.4.

[0207] Module M2: Automatic brainstem segmentation is performed using a 3D U-net network model. The 3D U-net network model includes an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers. The loss function is the standard Dice score.

[0208] Module M3: Acquires mirror images of the brainstem and brainstem atlases containing the midbrain, pons, and medulla oblongata regions. stem Regions were divided into brainstem and mirror brainstem images;

[0209] Module M4: Constructs a brainstem infarction lesion detection and segmentation network model; The brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale.

[0210] Module M5: Predicts and scores based on the segmented lesions.

[0211] Module M2 includes the following modules:

[0212] Module M2.1: Constructs a 4-layer 3D U-net network model;

[0213] Module M2.2: Uses the dataset processed by Module M1.5 to train the 3D U-net network model, and obtains the loss function and segmentation results;

[0214] Module M2.3: Adjusts and optimizes network model parameters based on loss function and segmentation results;

[0215] Module M2.4: Generates and saves the optimized model after parameter tuning;

[0216] Module M2.5: Saves the brainstem image Stem from the model segmentation results. origin .

[0217] Preferably, module M3 includes the following modules:

[0218] Module M3.1: For brainstem images... origin Perform a mirroring operation to obtain the mirror brainstem. mirror ;

[0219] Module M3.2: Based on brainstem atlas Atlas stem to brainstem images Stem origin The mapping divides the brainstem regions;

[0220] Module M3.3: Based on brainstem atlas Atlas stem To the mirror brainstem Stem mirror The brainstem region is mapped and divided; module M4 includes the following modules:

[0221] Module M4.1: Constructing a network model for detecting and segmenting brainstem infarcts;

[0222] Module M4.2: Module M3.4 generates brainstem images. origin Mirror brainstem Stem mirror and the obtained brainstem atlas Atlas stem Input model;

[0223] Module M4.3: The three input datasets are encoded into feature maps by the three encoders in the model, respectively;

[0224] Module M4.4: Module M4.2 generates brainstem images. origin The feature map is input into the decoder for decoding;

[0225] Module M4.5: Input the feature maps obtained from module M4.3 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas;

[0226] Module M4.6: In the difference calculation module, the two input feature maps M1 and M2 are first subtracted element-wise to obtain the difference feature map S. Then, pyramid convolution is performed at multiple scales. Assume that the difference feature map at the i-th scale is S. i The convolution result is obtained.

[0227]

[0228]

[0229] Where W1 represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction, w i This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution;

[0230] Module M4.7: Inputs the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, average pooling is used in its c-th channel to extract dense features F1(X). c ), and extract dense features F2(X) using the maximized output. c ):

[0231]

[0232]

[0233] Where H, W, Z, and C represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively, and h, w, z, and c represent the coordinates of a specific input during the calculation.

[0234] Module M4.8: First, combine two dense features F1(X) c ) and F2(X c The activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and

[0235]

[0236] in, Represents the convolution operation. The parameters represent the dimensionality reduction fully connected layer, and μ represents the ReLU function. ω represents the parameters of the upgraded fully connected layer, and ω represents the sigmoid function.

[0237] Module M4.9: Element-wise addition of two densely activated features yields the summed feature S. sum ;

[0238]

[0239] Where W2 represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation;

[0240] Module M4.10: Utilizing the summation feature S sum The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features S. mix ;

[0241] S mix =S sum ⊙X

[0242] Where ⊙ represents channel-level multiplication operation;

[0243] Module M4.11: Using skip connections to connect S mix The connection feature S is obtained by connecting the corresponding resolution features in the decoder obtained from module M4.4. skip The connection features were obtained through a total of three connections. and

[0244] Module M4.12: for Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution

[0245]

[0246]

[0247] Among them, W 1′ and W 2′ They are respectively the corresponding The parameters of the deconvolution layer, Represents the convolution operation;

[0248] Module M4.13: For and Element-wise addition is performed using the ReLU activation function, and a multi-resolution fusion coefficient map S is generated using a 1×1×1 convolutional layer and the sigmoid function. map ;

[0249]

[0250] in, This represents element-wise addition, where μ represents the ReLU function. ω represents the convolution operation, W3 represents the parameters of a 1×1×1 convolutional layer, and ω represents the sigmoid function.

[0251] Module M4.14: For the fusion coefficient map S map and the last level connection features Element-wise multiplication is used to obtain the segmentation feature S. seg ;

[0252]

[0253] in, Represents element-wise multiplication;

[0254] Module M4.15: Based on segmentation feature S seg The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

[0255] Preferably, module M5 includes the following modules:

[0256] Module M5.1: Based on brainstem lesion masks in brainstem images origin The region of interest (ROI) is obtained from the brainstem image and then placed within the brainstem image. origin By outlining the contours, a brainstem image containing the lesion region is obtained. focus ;

[0257] Module M5.2: Based on the brainstem atlas Atlas stem Stem images of the brainstem containing lesions focus The brain was divided and located in the midbrain, pons, and medulla oblongata regions.

[0258] Module M5.3: Calculates the volume of lesions in the midbrain, pons, and medulla oblongata by voxel;

[0259] Module M5.4: Calculates the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel;

[0260] Module M5.5: Scores are given for lesions in each region. 2 points are added when the volume of the lesion is greater than 50% of the volume of the same region on one side, 1 point is added when it is less than 50% of the volume of the same region on one side, and no points are added when there are no lesions.

[0261] Example 3:

[0262] Example 3 is a preferred example of Example 1, and is used to illustrate the present invention in more detail.

[0263] This invention provides a method for detecting lesions and scoring brainstem ischemic stroke based on deep learning, in order to overcome the shortcomings of existing methods.

[0264] The technical solution adopted in this invention is:

[0265] A deep learning-based method for lesion detection and brainstem scoring in ischemic stroke of the brainstem, the method comprising the following steps:

[0266] Preprocessing of brain CT images:

[0267] Obtain plain CT images of the brain, obtain standard brain atlases, obtain the deformation field from plain CT images of the brain to standard brain atlases, use the deformation field to perform affine registration on the brain CT images, remove skull bones from the registered images, normalize the image grayscale to 0-255, and perform feature annotation on the processed images.

[0268] Automatic segmentation of the brainstem:

[0269] A 3D U-net automatic brainstem segmentation model was constructed. Preprocessed brain CT images were input for training, yielding a loss function and segmentation results. Based on the training results, the network model parameters were adjusted and optimized. The optimized model was generated and saved, and the brainstem image Stem was obtained by saving the model's segmentation results. origin .

[0270] 3D U-net Automatic Brainstem Segmentation Model: This model employs a 4-layer 3D U-net network for brainstem segmentation, comprising an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers and a corresponding ReLU (Rectified Linear Unit). The encoder also includes three downsampling operations. The decoder also contains three sub-modules, each containing two convolutional layers and a corresponding ReLU. Skip connections are used to connect the upsampled results from the decoder to the outputs of sub-modules in the encoder with the same resolution, serving as the input to the next sub-module in the decoder. The model's input data is a 132×132×116 voxel image with three channels, and the output of the final layer is a 44×44×28 voxel image in the x, y, and z directions. By training the model, the brainstem region in the output brain CT image can be segmented to obtain a brainstem image.

[0271] Obtain mirrored brainstem images and register brainstem atlases:

[0272] For the segmented brainstem image Stem origin Perform mirroring operations to obtain the mirror brainstem. mirror Atlas brainstem mapping stem According to the Atlas brainstem map stem to brainstem images Stem origin and mirror brainstem Stem mirror The mappings are used to divide the brainstem regions.

[0273] a) Mirror brainstem: The mirror brainstem is also a three-dimensional image, which is a mirror image of the brainstem image.

[0274] b) Brainstem Atlas: An Atlas template was created to annotate the midbrain, pons, and medulla oblongata. stem By registering the template, the corresponding midbrain, pons, and medulla oblongata in each image can be obtained.

[0275] Lesion segmentation based on brainstem mirror images and atlases:

[0276] Constructing a lesion detection and segmentation network model, brainstem image Stem origin Mirror brainstem Stem mirror and brainstem atlas Atlas stem As input, the three input datasets are encoded by the three encoders in the model to obtain feature maps, which are then used to transform the brainstem images. originThe feature maps are input into the decoder for decoding, and then the feature maps are input into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas, to calculate the multi-resolution fusion segmentation feature S. seg Based on S seg The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

[0277] a) Lesion detection and segmentation network model: The model contains three encoders, one decoder, and six difference calculation modules. Each encoder consists of three convolutional layers, with each encoder corresponding to one input. Each decoder consists of three deconvolutional layers, and each decoder only decodes brainstem image inputs.

[0278] b) Differential Calculation Module: Located between each convolutional layer of the encoder, the input of each module is the convolution result of the brainstem image and mirror brainstem of that convolutional layer in the encoder, or the convolution result of the brainstem image and brainstem atlas. In the differential calculation module, the two input feature maps M1 and M2 are first subtracted element-wise to obtain the differential feature map S. Then, pyramid convolution is performed at multiple scales. Assuming the differential feature map at the i-th scale is S... i The convolution result is obtained. Input it into the feature fusion module.

[0279] c) Feature Fusion Module: The pyramid convolution results at each scale are input into the feature fusion module. Assuming the input at the nth scale is X, average pooling is used in its c-th channel to extract dense features F1(X). c And extract dense features F2(X) using Max-out. c ), which combines two dense features F1(X) c ) and F2(X c The activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and The summation feature S is obtained by element-wise addition of two densely activated features. sum Using S sum The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features S. mix .

[0280] d) Segmentation features S seg Using skip connections to fuse differential features S mix The connection feature S is obtained by concatenating the features at the corresponding resolution in the decoder. skip A total of three connections were obtained sequentially. and right Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution right and Element-wise addition is performed using the ReLU activation function, and a multi-resolution fusion coefficient map S is generated using a 1×1×1 convolutional layer and the sigmoid function. map For the fusion coefficient map S map and the last level connection features Element-wise multiplication is used to obtain the segmentation feature S. seg .

[0281] Lesion prediction scoring model:

[0282] Brainstem lesion masks were used as input for predictive scoring, and quantitative scores were performed on three brainstem regions: the midbrain, pons, and medulla oblongata. Regions of interest (ROIs) were located in the brainstem images using the brainstem lesion mask, and their outlines were marked to obtain a brainstem image containing the lesion region. focus Based on the brainstem atlas Atlas stem Stem images of the brainstem containing lesions focus The brain is divided into regions, including the midbrain, pons, and medulla oblongata. The volume of lesions in each region and the overall volume of the region are calculated using voxels. The severity of the lesions is then predicted and scored according to a scoring method.

[0283] Scoring method: Scoring is based on the proportion of the lesion volume in the midbrain, pons, and medulla oblongata to the volume of the unilateral area. For each lesion, 2 points are added to the total score when its voxel volume is greater than 50% of the volume of the unilateral area; 1 point is added to the total score when it is less than 50% of the volume of the unilateral area; no points are added when there is no lesion.

[0284] like Figure 1 The diagram shown is a flowchart of the method of the present invention, which includes the following steps:

[0285] Step 1: Preprocess the brain CT images;

[0286] Obtain plain CT images of the brain and obtain standard brain atlases;

[0287] Step 101: Obtain the deformation field from plain brain CT images to standard brain atlas;

[0288] Step 102: Register brain CT images using a deformation field;

[0289] Step 103: Remove the skull from the image obtained in step 102;

[0290] Step 104: Normalize the grayscale values ​​of the image obtained in Step 103 to 0-255;

[0291] Step 105: Perform feature annotation on the image obtained in step 104;

[0292] Step 2: Automatic segmentation of the brainstem;

[0293] The brainstem was segmented using a 3D U-net network model, which consists of an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers, and the loss function is the standard Dice score.

[0294] Step 201: Construct a 4-layer 3D U-net network model;

[0295] Step 202: Use the dataset processed in step 105 to train the 3D U-net network model and obtain the loss function and segmentation results;

[0296] Step 203: Adjust and optimize the network model parameters based on the loss function and segmentation results;

[0297] Step 204: Generate and save the optimized model after parameter tuning;

[0298] Step 205: Save the model segmentation results to obtain the brainstem image Stem. origin ;

[0299] Step 3: Obtain mirror brainstem images and register brainstem atlases;

[0300] Atlas, a brainstem atlas containing the midbrain, pons, and medulla oblongata regions, was obtained. stem ;

[0301] Step 301: Process brainstem images. origin Perform a mirroring operation to obtain the mirror brainstem. mirror ;

[0302] Step 302: Based on the brainstem atlas Atlas stem to brainstem images Stem origin The mapping divides the brainstem regions;

[0303] Step 303: Based on the brainstem atlas Atlas stem To the mirror brainstem Stem mirror Step 4: Construct a brainstem infarction lesion detection and segmentation network model;

[0304] like Figure 2The diagram shows the structure of the brainstem infarction lesion detection and segmentation model in this invention. The model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale.

[0305] Step 401: Construct a brainstem infarction lesion detection and segmentation network model;

[0306] Step 402: The brainstem image Stem obtained in step 304... origin Mirror brainstem Stem mirror and the obtained brainstem atlas Atlas stem Input model;

[0307] Step 402: The three input datasets are encoded by the three encoders in the model to obtain feature maps;

[0308] Step 403: Convert the brainstem image Stem obtained in step 402 into a digital format. origin The feature map is input into the decoder for decoding;

[0309] Step 404: Input the feature maps obtained in step 402 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas;

[0310] Step 405: In the difference calculation module, the two input feature maps M1 and M2 are first subtracted element-wise to obtain the difference feature map S. Then, pyramid convolution is performed at multiple scales. Let S be the difference feature map at the i-th scale. i The convolution result is obtained.

[0311]

[0312]

[0313] Where W1 represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction, w i This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution;

[0314] Step 406: Input the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, use average pooling in its c-th channel to extract dense features F1(X). c ),

[0315] Dense features F2(X) are extracted using Max-out. c );

[0316]

[0317]

[0318] Where H, W, Z, and C represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively, and h, w, z, and c represent the coordinates of a specific input during the calculation.

[0319] Step 407: First, combine the two dense features F1(X) c ) and F2(X c The activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and

[0320]

[0321] in, Represents the convolution operation. The parameters represent the dimensionality reduction fully connected layer, and μ represents the ReLU function. ω represents the parameters of the upgraded fully connected layer, and ω represents the sigmoid function.

[0322] Step 408: Element-wise summation of the two densely activated features yields the summed feature S. sum ;

[0323]

[0324] Where W2 represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation;

[0325] Step 409: Utilize the summation characteristic S sum The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features S. mix ;

[0326] S mix =S sum ⊙X

[0327] Where ⊙ represents channel-level multiplication operation;

[0328] Step 410: Use a jump connection to connect S mixConnecting the features at the corresponding resolution in the decoder obtained in step 403 yields the connection feature S. skip The connection features were obtained through a total of three connections. and

[0329] Step 411: For Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution

[0330]

[0331]

[0332] Among them, W 1′ and W 2′ They are respectively the corresponding The parameters of the deconvolution layer, Represents the convolution operation;

[0333] Step 412: [Regarding...] and Element-wise addition is performed using the ReLU activation function, and a multi-resolution fusion coefficient map S is generated using a 1×1×1 convolutional layer and the sigmoid function. map ;

[0334]

[0335] in, This represents element-wise addition, where μ represents the ReLU function. ω represents the convolution operation, W3 represents the parameters of a 1×1×1 convolutional layer, and ω represents the sigmoid function.

[0336] Step 413: Analyze the fusion coefficient map S map and the last level connection features Element-wise multiplication is used to obtain the segmentation feature S. seg ;

[0337]

[0338] in, Represents element-wise multiplication;

[0339] Step 414: Based on segmentation feature S seg The softmax function is used to segment brainstem images to obtain brainstem lesion masks;

[0340] Step 5: Predictive scoring based on the segmented lesions;

[0341] Step 501: Based on the brainstem lesion mask, perform a brainstem lesion masking on the brainstem image. origin The region of interest (ROI) is obtained from the brainstem image and then placed within the brainstem image. origin By outlining the contours, a brainstem image containing the lesion region is obtained. focus ;

[0342] Step 502: Based on the brainstem atlas Atlas stem Stem images of the brainstem containing lesions focus The brain was divided and located in the midbrain, pons, and medulla oblongata regions.

[0343] Step 503: Calculate the volume of lesions in each region of the midbrain, pons, and medulla oblongata using voxels;

[0344] Step 504: Calculate the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel;

[0345] Step 505: Score the lesions in each region. Add 2 points if the volume of the lesion is greater than 50% of the volume of the same region on one side, add 1 point if it is less than 50% of the volume of the same region on one side, and add no points if there are no lesions.

[0346] Those skilled in the art can understand this embodiment as a more specific description of Embodiment 1 and Embodiment 2.

[0347] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0348] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for automatic lesion segmentation and scoring of brainstem ischemic stroke, characterized in that, The method includes the following steps: Step S1: Obtain plain CT images of the brain, obtain standard brain atlases, and perform registration and preprocessing on the brain CT images; Step S2: Automatic segmentation of the brainstem is performed using a 3D U-net network model. The 3D U-net network model includes an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers. The loss function is the standard Dice score. Step S3: Obtain a mirror image of the brainstem and a brainstem atlas containing the midbrain, pons, and medulla oblongata regions. Regions were divided into brainstem and mirror brainstem images; Step S4: Construct a brainstem infarction lesion detection and segmentation network model; The brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale. Step S5: Calculate a predictive score based on the segmented lesions; Step S5 includes the following steps: Step S5.1: Based on the brainstem lesion mask on the brainstem image The region of interest (ROI) is obtained from the brainstem image, and then the ROI is placed in the brainstem image. By outlining the contours, brainstem images containing lesion regions can be obtained. ; Step S5.2: Based on the brainstem atlas Brainstem images containing lesion regions The brain was divided and located in the midbrain, pons, and medulla oblongata regions. Step S5.3: Calculate the volume of lesions in each region of the midbrain, pons, and medulla oblongata using voxels; Step S5.4: Calculate the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel; Step S5.5: Score the lesions in each region. Add 2 points if the volume of the lesion is greater than 50% of the volume of the same region on one side, add 1 point if it is less than 50% of the volume of the same region on one side, and add no points if there are no lesions.

2. The method for automatic lesion segmentation and scoring of ischemic stroke in the brainstem according to claim 1, characterized in that, Step S1 includes the following steps: Step S1.1: Obtain the deformation field from the plain brain CT image to the standard brain atlas; Step S1.2: Register brain CT images using a deformation field; Step S1.3: Remove the skull from the image obtained in step S1.2; Step S1.4: Normalize the grayscale values ​​of the image obtained in step S1.3 to 0-255; Step S1.5: Perform feature annotation on the image obtained in step S1.

4.

3. The method for automatic lesion segmentation and scoring of ischemic stroke in the brainstem according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Construct a 4-layer 3D U-net network model; Step S2.2: Use the dataset processed in step S1.5 to train the 3D U-net network model and obtain the loss function and segmentation results; Step S2.3: Adjust and optimize the network model parameters based on the loss function and segmentation results; Step S2.4: Generate and save the optimized model after parameter tuning; Step S2.5: Save the brainstem image by saving the model segmentation results. .

4. The method for automatic lesion segmentation and scoring of ischemic stroke in the brainstem according to claim 1, characterized in that, Step S3 includes the following steps: Step S3.1: Process brainstem images Perform a mirror operation to obtain a mirror brainstem. ; Step S3.2: Based on the brainstem atlas brainstem images The mapping divides the brainstem regions; Step S3.3: Based on the brainstem atlas To the mirror brainstem The mapping divides the brainstem regions.

5. The method for automatic lesion segmentation and scoring of ischemic stroke in the brainstem according to claim 1, characterized in that, Step S4 includes the following steps: Step S4.1: Construct a brainstem infarction lesion detection and segmentation network model; Step S4.2: Process the brainstem image obtained in step S3.4 Mirror brainstem and the obtained brainstem atlas Input model; Step S4.3: The three input datasets are encoded by the three encoders in the model to obtain feature maps; Step S4.4: Process the brainstem image obtained in step S4.2 The feature map is input into the decoder for decoding; Step S4.5: Input the feature maps obtained in step S4.3 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas; Step S4.6: In the difference calculation module, first process the two input feature maps... and Element-level subtraction yields the difference feature map. Then, pyramid convolution is performed at multiple scales. Assume the difference feature map at the i-th scale is... The convolution result is obtained. ; in, This represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction. This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution; Step S4.7: Input the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, use average pooling to extract dense features in its c-th channel. And extract dense features using the maximized output. : in, These represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively. These represent the coordinates of a specific input during the calculation; Step S4.8: First, combine the two dense features and Activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and ; in, Represents the convolution operation. The parameters representing the dimensionality reduction fully connected layer, Represents the ReLU function. The parameters representing the upgraded fully connected layer, Represents the sigmoid function; Step S4.9: Perform element-wise addition of the two densely activated features to obtain the summed features. ; in, This represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation; Step S4.10: Utilize the summation characteristics The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features. ; in, This represents a channel-level multiplication operation; Step S4.11: Use skip connections to... Connect the features obtained in step S4.4 to the corresponding resolution features in the decoder to obtain the connection features. The connection features were obtained through a total of three connections. , and ; Step S4.12: For , Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution , ; in, and They are respectively the corresponding , The parameters of the deconvolution layer, Represents the convolution operation; Step S4.13: For , and Element-wise addition is performed using the ReLU activation function, and then... Convolutional layers and the sigmoid function generate multi-resolution fusion coefficient maps ; in, Represents element-level addition. Represents the ReLU function. Represents the convolution operation. represent Convolutional layer parameters Represents the sigmoid function; Step S4.14: Analyze the fusion coefficient map and the last level connection features Element-wise multiplication is used to obtain segmentation features. ; in, Represents element-wise multiplication; Step S4.15: Based on segmentation features The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

6. An automatic lesion segmentation and scoring system for ischemic stroke in the brainstem, characterized in that, The system includes the following modules: Module M1: Acquire plain CT images of the brain, acquire standard brain atlases, and perform registration and preprocessing on the brain CT images; Module M2: Automatic brainstem segmentation is performed using a 3D U-net network model. The 3D U-net network model includes an encoder and a decoder. The encoder has three sub-modules, each containing two convolutional layers. The decoder also has three sub-modules, each containing two convolutional layers. The loss function is the standard Dice score. Module M3: Acquire mirror images of the brainstem and brainstem atlases containing the midbrain, pons, and medulla oblongata regions. Regions were divided into brainstem and mirror brainstem images; Module M4: Constructs a brainstem infarction lesion detection and segmentation network model; The brainstem infarction lesion detection and segmentation network model contains three encoders, one decoder and six difference calculation modules. Each encoder consists of three convolutional layers and each decoder consists of three deconvolutional layers. The difference calculation module includes multi-scale pyramid convolution and feature fusion modules at each scale. Module M5: Predicts and scores based on the segmented lesions; Module M5 includes the following modules: Module M5.1: Based on brainstem lesion masks in brainstem images The region of interest (ROI) is obtained from the brainstem image, and then the ROI is placed in the brainstem image. By outlining the contours, brainstem images containing lesion regions can be obtained. ; Module M5.2: Based on brainstem atlas Brainstem images containing lesion regions The brain was divided and located in the midbrain, pons, and medulla oblongata regions. Module M5.3: Calculates the volume of lesions in the midbrain, pons, and medulla oblongata by voxel; Module M5.4: Calculates the overall volume of each region of the midbrain, pons, and medulla oblongata by voxel; Module M5.5: Scores are given for lesions in each region. 2 points are added when the volume of the lesion is greater than 50% of the volume of the same region on one side, 1 point is added when it is less than 50% of the volume of the same region on one side, and no points are added when there are no lesions.

7. The automatic lesion segmentation and scoring system for ischemic stroke in the brainstem according to claim 6, characterized in that, Module M1 includes the following modules: Module M1.1: Acquires the deformation field from plain brain CT images to standard brain atlases; Module M1.2: Registering brain CT images using a deformation field; Module M1.3: Performs skull removal on the image obtained from module M1.2; Module M1.4: Normalizes the grayscale values ​​of the image obtained from module M1.3 to 0-255; Module M1.5: Performs feature annotation on the image obtained from module M1.4; Module M2 includes the following modules: Module M2.1: Constructs a 4-layer 3D U-net network model; Module M2.2: Uses the dataset processed by Module M1.5 to train the 3D U-net network model, and obtains the loss function and segmentation results; Module M2.3: Adjusts and optimizes network model parameters based on loss function and segmentation results; Module M2.4: Generates and saves the optimized model after parameter tuning; Module M2.5: Saves the brainstem image obtained from the model segmentation results. .

8. The automatic lesion segmentation and scoring system for ischemic stroke in the brainstem according to claim 6, characterized in that, Module M3 includes the following modules: Module M3.1: Brainstem Images Perform a mirror operation to obtain a mirror brainstem. ; Module M3.2: Based on brainstem atlas brainstem images The mapping divides the brainstem regions; Module M3.3: Based on brainstem atlas To the mirror brainstem The mapping divides the brainstem regions; Module M4 includes the following modules: Module M4.1: Constructing a network model for detecting and segmenting brainstem infarcts; Module M4.2: Processes the brainstem images obtained from Module M3.

4. Mirror brainstem and the obtained brainstem atlas Input model; Module M4.3: The three input datasets are encoded into feature maps by the three encoders in the model, respectively; Module M4.4: Processes the brainstem images obtained from Module M4.

2. The feature map is input into the decoder for decoding; Module M4.5: Input the feature maps obtained from module M4.3 into the difference calculation module according to the combination of brainstem image and mirror brainstem, and original image and brainstem atlas; Module M4.6: In the difference calculation module, the two input feature maps are first processed. and Element-level subtraction yields the difference feature map. Then, pyramid convolution is performed at multiple scales. Assume the difference feature map at the i-th scale is... The convolution result is obtained. ; in, This represents the parameters of the convolutional layer after the subtraction operation. Represents the convolution operation. Represents element-wise subtraction. This represents the parameters of the convolutional layer at the i-th scale in pyramid convolution; Module M4.7: Inputs the pyramid convolution results at each scale into the feature fusion module. Assuming the input at the nth scale is X, it uses average pooling to extract dense features in the cth channel. And extract dense features using the maximized output. : in, These represent the height, width, slice thickness, and number of channels of the acquired brain CT plain scan image, respectively. These represent the coordinates of a specific input during the calculation; Module M4.8: First, combine two dense features and Activation-dense features are obtained by sequentially passing the layers through a dimensionality-reducing fully connected layer, a ReLU activation function, an updimensional fully connected layer, and a sigmoid activation function. and ; in, Represents the convolution operation. The parameters representing the dimensionality reduction fully connected layer, Represents the ReLU function. The parameters representing the upgraded fully connected layer, Represents the sigmoid function; Module M4.9: Calculates summation features by element-wise addition of two densely activated features. ; in, This represents the parameters of the convolutional layer after the addition operation. Represents element-level addition. Represents the convolution operation; Module M4.10: Utilizing Summation Features The input X at the nth scale of the feature fusion module is fused using channel-level multiplication to obtain the fused difference features. ; in, This represents a channel-level multiplication operation; Module M4.11: Using skip connections to Connecting features is obtained by concatenating the corresponding resolution features in the decoder obtained from module M4.

4. The connection features were obtained through a total of three connections. , and ; Module M4.12: For , Perform a deconvolution operation to obtain the result. Feature results with the same spatial resolution , ; in, and They are respectively the corresponding , The parameters of the deconvolution layer, Represents the convolution operation; Module M4.13: For , and Element-wise addition is performed using the ReLU activation function, and then... Convolutional layers and the sigmoid function generate multi-resolution fusion coefficient maps ; in, Represents element-level addition. Represents the ReLU function. Represents the convolution operation. represent Convolutional layer parameters Represents the sigmoid function; Module M4.14: For fusion coefficient plots and the last level connection features Element-wise multiplication is used to obtain segmentation features. ; in, Represents element-wise multiplication; Module M4.15: Based on segmentation features The softmax function is used to segment brainstem images to obtain brainstem lesion masks.

Images