Hippocampus segmentation method and system based on coronal T1-MRI

Abstract

The present invention provides a hippocampus segmentation method and system based on coronal T1-MRI, including: obtaining AD coronal T1MRI images and corresponding hippocampus labels, and obtaining a data set after preprocessing; constructing a feature extraction module to obtain a feature map; constructing an uncertainty attention module to calculate uncertainty factors; building a segmentation network NU-Mamba; constructing a loss function to balance dynamic balance through uncertainty factors; NU-Mamba is trained and tested on the data set and outputs the final model, and the T1MRI image is segmented to obtain a segmentation result. The present invention solves the problem that the existing model optimization can only achieve suboptimal results by constructing different loss functions to balance the importance of foreground and background. Compared with other hippocampus segmentation algorithms, the segmentation accuracy is improved by more than 10%; it has the characteristics of automation and intelligence, and provides help for doctors' clinical diagnosis and formulation of patient treatment plans through simpler operations and faster result presentation, and has better clinical practicality, high immediacy and popularity.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing, and in particular, relates to a hippocampus segmentation method and system based on coronal T1-MRI. Background Art

[0002] Alzheimer's disease (AD) is a progressive neurodegenerative disease characterized by severe impairment of cognitive and intellectual functions. Currently, there are approximately 90 million AD patients worldwide, and this number continues to grow. The hippocampus is a pair of medial subcortical structures located near the temporal horn of the lateral ventricle, responsible for the formation and storage of memory. In the development of AD, the volume and morphology of the hippocampus are the first to be affected. Studies have shown that early diagnosis is important for slowing or even reversing the course of AD. Hippocampal atrophy is considered a quantifiable biomarker that can provide an important basis for early detection and disease progression of AD. Coronal T1-weighted magnetic resonance imaging (T1 MRI) can generate high-resolution brain images to help clinicians comprehensively evaluate and monitor changes in brain structure.

[0003] With the rapid development of computer image processing technology, more and more researchers have begun to explore hippocampal segmentation methods based on deep learning, including convolutional neural networks (CNN) and Transformer architectures. However, most existing AD studies are limited to the application of existing models, and no special model design is performed for the characteristics of the hippocampus, which leads to low accuracy of hippocampal boundary segmentation and limits its clinical application. In addition, CNN and Transformer have limitations in modeling global feature relationships.

[0004] The patent document "Automatic Hippocampus Segmentation Method and System Based on V-Net and Combined Loss Function" (CN115311453A) discloses the use of V-Net and combined loss function to mine hippocampus image features and improve the efficiency and accuracy of automatic hippocampus segmentation. However, it also does not pay attention to the importance of hippocampus features in balancing foreground and background, and the segmentation accuracy is still insufficient.

[0005] Therefore, there is an urgent need to develop a hippocampus segmentation method based on coronal T1-MRI to improve the segmentation accuracy and thereby assist doctors in early diagnosis and clinical intervention of AD. Summary of the invention

[0006] In view of the defects in the prior art, the object of the present invention is to provide a hippocampus segmentation method and system based on coronal T1-MRI.

[0007] A hippocampus segmentation method based on coronal T1-MRI provided by the present invention comprises:

[0008] Data acquisition steps: obtain AD coronal T1 MRI images and corresponding hippocampal labels, and obtain the data set after preprocessing;

[0009] Feature extraction step: construct a feature extraction module to extract multi-scale features from T1 MRI images and obtain feature maps;

[0010] Uncertainty attention module construction steps: Construct uncertainty attention module to process feature graph and calculate uncertainty factor;

[0011] Segmentation network construction steps: Based on the feature extraction module and the uncertainty attention module, build the segmentation network NU-Mamba;

[0012] Loss function construction steps: construct loss function balance dynamic balance through uncertainty factors;

[0013] Model training and segmentation steps: NU-Mamba is trained and tested on the dataset and outputs the final model, which segments the T1 MRI images to obtain segmentation results.

[0014] Preferably, in the data acquisition step, the T1 MRI image and the corresponding hippocampus label are respectively denoted as I m 、mask Hip In the data acquisition step, the T1 MRI image and the corresponding hippocampus label are respectively denoted as I m 、mask Hip ,include:

[0015] Step S1.1: Screen the acquired coronal T1 MRI images, exclude low-quality images, delete the patient's metadata, and resample the screened T1 MRI images to obtain I T1W1 , resize each image to 224×224×16;

[0016] Step S1.2: Manually annotate the coronal T1 MRI image to obtain the hippocampus label mask Hip And build AD database D all , the ratio of 8:2 is divided into training set D train With the test set D test .

[0017] The meta information includes name, gender, age and personal information that can identify the patient.

[0018] The low-quality image is an image with excessive noise, blur and incomplete image.

[0019] The feature extraction module includes a MB module and a RSU module.

[0020] The uncertainty attention module uses the extracted segmentation edge information and the mean and variance of the weight map to dynamically adjust the model's attention on the foreground and background areas.

[0021] The segmentation network NU-Mamba includes an encoder, a decoder and a deep supervision mechanism.

[0022] The encoder includes four MBs and a downsampling operation.

[0023] The decoder includes four RSUs and an upsampling operation.

[0024] The loss function L total Including Dice loss function, cross entropy loss function and uncertainty factor μ, balancing the dynamic balance between the hippocampal tissue area and the background area.

[0025] In the model training segmentation step, NU-Mamba train Upper pair I T1W1 Train until L total Tends to stabilize and save the final model parameters, in D test NU-Mamba was tested on the Dice Not less than 0.9, output the model and segment the T1 MRI image to obtain the segmentation result.

[0026] Preferably, the feature extraction step comprises:

[0027] MB module build step: define the input as During the coding phase, After 5-1 VSS layers and 4-1 patch fusion layers, the encoded feature map is obtained In the decoding stage, After 5-1 VSS layers and 4-1 patch expansion layers, the decoded feature map is obtained

[0028] In each VSS layer, the feature map Divided into feature extension branch F path1 and semantic compression branch F path2 Parallel processing;

[0029] For F path1 , the VSS layer will The number of channels is expanded to λ×C to generate enhanced feature representation.

[0030] F path1=LayerNorm(SSM(SiLU(DWConv(F 1 ))));

[0031] For F path2 ,

[0032] Fusion and generation of the final output feature map F out =Linear(F path1 ⊕F path2 );

[0033] Among them, W, H, C, and l represent the width, height, number of channels, and depth of the module respectively;

[0034] Represents a three-dimensional vector;

[0035] represents the encoding stage characteristics of the lth MB module;

[0036] represents the decoding stage characteristics of the lth MB module;

[0037] represents the input characteristics of the lth VSS module;

[0038] Linear() represents a linear transformation operation;

[0039] LayerNorm() represents the layer normalization operation;

[0040] SSM() represents the local feature enhancement operation;

[0041] SiLU() represents the activation function operation;

[0042] DWConv() represents the depth convolution operation;

[0043] The patch fusion layer is a downsampling operation;

[0044] The VSS layer captures spatial and channel information and combines semantic features to learn the feature representation of the input feature map;

[0045] The patch expansion layer is an upsampling operation.

[0046] RSU module construction steps: define the input as During the coding phase, After 5-1 feature extraction layers using convolution operations with a convolution kernel size of 3×3, batch normalization operations and activation function operations and 4-1 downsampling layers using maximum pooling operations, the encoded feature map is obtained.

[0047] In the decoding stage, After 5-1 feature extraction layers and 4-1 upsampling layers, the nearest neighbor interpolation function is used to obtain the decoding feature map

[0048] The uncertainty attention module construction steps include:

[0049] The uncertainty attention module UAM constructs a segmentation edge pixel weight map through the feature map. For σ(P m ) to perform derivative calculations, Then perform the combined transformation to obtain the uncertainty factor μ,

[0050] Among them, σ() represents the Sigmoid activation function;

[0051] P m , P v Respectively represent the mean and variance of the weight graph;

[0052] σ′(P m ) represents σ(P m )’s derivative;

[0053] ε represents a very small constant.

[0054] Preferably, the segmentation network building step includes:

[0055] Step S4.1: In the encoder, I m As input, the size is 224×224, and the convolution operation with the convolution kernel size of 1×1 is performed to obtain the feature map After a MB 4 And the downsampling operation is

[0056] right The MB and downsampling operations are performed three times to obtain the sizes of 56×56×128, 28×28×256, and 14×14×512 respectively.

[0057] Among them, MB 4 Indicates a MB operation with a depth of 4;

[0058] CBD() represents a convolution operation with a convolution kernel of 3×3, a batch normalization operation, and an activation function operation.

[0059] Step S4.2: In the decoder, After one RSU Through the Concat operation and Fusion generation Input into UAM to get the uncertainty factor μ 4 ,right Perform upsampling operation to obtain feature map

[0060] Step S4.3: Get through an RSU Through the Concat operation and Fusion generation Send to UAM to get μ 3 ,right Perform upsampling to obtain

[0061] Step S4.4: Repeat step S4.3 twice to obtain μ 2 and μ 1 and The sizes are 112×112×64 and 224×224×32 respectively.

[0062] Step S4.5: Upsampling the feature map The size is unified to 224×224×32 and the feature map is obtained

[0063] Step S4.6: Use the Concat operation to The fusion is sent to the 3×3 convolution layer to obtain the feature map F fuse , sent to UAM to get μ fuse .

[0064] Preferably, the loss function construction step includes:

[0065] Step S5.1: Construct Dice loss function,

[0066] Among them, N represents the number of pixel values ​​of the model segmentation result;

[0067] p i represents the probability of pixel i in the segmentation result, p i ≥0.5 means that the pixel belongs to the hippocampus region, p i <0.5 means the pixel belongs to the background area;

[0068] q i Indicates pixel i in mask Hip Category of;

[0069] ε represents a very small constant.

[0070] Step S5.2: Construct the cross entropy loss function,

[0071] Where (r,c) represents the pixel coordinates;

[0072] H and W represent the length and width of the segmentation result, respectively;

[0073] P G(r,c) , P F(r,c) Represent the segmentation result and mask respectively. Hip The value of the corresponding pixel.

[0074] Step S5.3: Loss function L total For UAM module, L Dice , L BCE The proposed modules are weighted.

[0075]

[0076] in, Respectively represent the feature maps after the size is unified mask Hip segmentation loss;

[0077] Represents the feature map F of the fusion output respectively fuse 、mask Hip segmentation loss;

[0078] μ l represents the lth uncertainty factor;

[0079] μ fuse represents the uncertainty factor of the fusion output;

[0080] L represents the maximum depth.

[0081] The model training segmentation step includes:

[0082] Step S6.1: The training set D obtained by NU-Mamba after T1 MRI image preprocessing train For training, use the Adam optimizer with a learning rate of 10 -4 , and save the model parameters after 500 iterations.

[0083] Step S6.2: Preprocess the T1 MRI images to obtain the test set D test Load the model parameters to predict the segmentation result Seg , according to Result Seg and maskHip Calculate the Dice coefficient.

[0084] Step S6.3: Determine L Dice , if L Dice If it is lower than 0.9, repeat steps S6.1 and S6.2;

[0085] If L Dice If it is not less than 0.9, the model at this time is output and the T1 MRI image is segmented to obtain the segmentation result.

[0086] A hippocampus segmentation system based on coronal T1-MRI provided by the present invention includes: a data acquisition module and a segmentation network NU-Mamba.

[0087] The segmentation network NU-Mamba consists of a feature extraction module and an uncertainty attention module.

[0088] The data acquisition module obtains AD coronal T1 MRI images and corresponding hippocampal labels, and obtains the data set after preprocessing;

[0089] The feature extraction module extracts multi-scale features from the T1 MRI image to obtain a feature map;

[0090] The uncertainty attention module processes the feature map and calculates the uncertainty factor;

[0091] The loss function is constructed through uncertainty factors to balance the dynamic equilibrium;

[0092] NU-Mamba is trained and tested on the dataset and outputs the final model to segment T1 MRI images to obtain segmentation results.

[0093] Preferably, in the data acquisition module, the T1 MRI image and the corresponding hippocampus label are respectively denoted as I m 、mask Hip , to I m Data screening was performed to exclude low-quality images, delete patient metadata, and resample the screened T1 MRI images to 224×224×16 to obtain I T1W1 , manually annotate the coronal T1 MRI image to obtain the hippocampus label mask Gip And build AD database D all , the ratio of 8:2 is divided into training set D train With the test set D test .

[0094] The meta information includes name, gender, age and personal information that can identify the patient.

[0095] The low-quality image is an image with excessive noise, blur and incomplete image.

[0096] The feature extraction module includes a MB module and a RSU module.

[0097] The uncertainty attention module uses the extracted segmentation edge information and the mean and variance of the weight map to dynamically adjust the model's attention on the foreground and background areas.

[0098] The segmentation network NU-Mamba includes an encoder, a decoder and a deep supervision mechanism.

[0099] The encoder includes four MBs and a downsampling operation.

[0100] The decoder includes four RSUs and an upsampling operation.

[0101] The loss function L total Including Dice loss function, cross entropy loss function and uncertainty factor μ, balancing the dynamic balance between the hippocampal tissue area and the background area.

[0102] The NU-Mamba in D train Upper pair I T1W1 Train until L total Tends to stabilize and save the final model parameters, in D test NU-Mamba was tested on the Dice Not less than 0.9, output the model and segment the T1 MRI image to obtain the segmentation result.

[0103] Preferably, the feature extraction module comprises:

[0104] MB module: Define the input as During the coding phase, After 5-1 VSS layers and 4-1 patch fusion layers, the encoded feature map is obtained In the decoding stage, After 5-1 VSS layers and 4-1 patch expansion layers, the decoded feature map is obtained

[0105] In each VSS layer, the feature map Divided into feature extension branch F path1 and semantic compression branch F path2 Parallel processing;

[0106] For F path1 , the VSS layer will The number of channels is expanded to λ×C to generate enhanced feature representation.

[0107] F path1 =LayerNorm(SSM(SiLU(DWConv(F 1 ))));

[0108] For F path2 ,

[0109] Fusion and generation of the final output feature map F out =Linear)F path1 ⊕F path2 );

[0110] Among them, W, H, C, and l represent the width, height, number of channels, and depth of the module respectively;

[0111] Represents a three-dimensional vector;

[0112] Represents the encoding stage characteristics of the lth MB module;

[0113] represents the decoding stage characteristics of the lth MB module;

[0114] represents the input characteristics of the lth VSS module;

[0115] Linear() represents a linear transformation operation;

[0116] LayerNorm() represents the layer normalization operation;

[0117] SSM() represents the local feature enhancement operation;

[0118] SiLU() represents the activation function operation;

[0119] DWConv() represents the depth convolution operation;

[0120] The patch fusion layer is a downsampling operation;

[0121] The VSS layer captures spatial and channel information and combines semantic features to learn the feature representation of the input feature map;

[0122] The patch expansion layer is an upsampling operation.

[0123] RSU module: Define the input as During the coding phase, After 5-1 feature extraction layers using convolution operations with a convolution kernel size of 3×3, batch normalization operations and activation function operations and 4-1 downsampling layers using maximum pooling operations, the encoded feature map is obtained.

[0124] In the decoding stage, After 5-1 feature extraction layers and 4-1 upsampling layers, the nearest neighbor interpolation function is used to obtain the decoding feature map

[0125] The uncertainty attention module constructs a segmentation edge pixel weight map through the feature map. For σ(P m ) to perform derivative calculations, Then perform the combined transformation to obtain the uncertainty factor μ,

[0126] Among them, σ() represents the Sigmoid activation function;

[0127] P m , P v Respectively represent the mean and variance of the weight graph;

[0128] σ′(P m ) represents σ(P m )’s derivative;

[0129] ε represents a very small constant.

[0130] Preferably, the segmented network NU-Mamba comprises:

[0131] Module M4.1: In the encoder, I m As input, the size is 224×224, and the convolution operation with the convolution kernel size of 1×1 is performed to obtain the feature map After a MB 4 And the downsampling operation is

[0132] right The MB and downsampling operations are performed three times to obtain the sizes of 56×56×128, 28×28×256, and 14×14×512 respectively.

[0133] Among them, MB 4 Indicates a MB operation with a depth of 4;

[0134] CBD() represents a convolution operation with a convolution kernel of 3×3, a batch normalization operation, and an activation function operation.

[0135] Module M4.2: In the decoder, After one RSU Through the Concat operation and Fusion generation Input into UAM to get the uncertainty factor μ 4 ,right Perform upsampling operation to obtain feature map

[0136] Module M4.3: Obtained through an RSU Through the Concat operation and Fusion generation Send to UAM to get μ 3 ,right Perform upsampling to obtain

[0137]

[0138] Module M4.4: Yes Repeat triggering module M4.3 twice to obtain μ 2 and μ 1 and The sizes are 112×112×64 and 224×224×32 respectively.

[0139] Module M4.5: Upsampling operation is used to convert feature maps The size is unified to 224×224×32 and the feature map is obtained

[0140] Module M4.6: Concat operation The fusion is sent to the 3×3 convolution layer to obtain the feature map F fuse , sent to UAM to get μ fuse .

[0141] Preferably, the constructing the loss function comprises:

[0142] Construct the Dice loss function,

[0143] Among them, N represents the number of pixel values ​​of the model segmentation result;

[0144] p i represents the probability of pixel i in the segmentation result, p i ≥0.5 means that the pixel belongs to the hippocampus region, p i <0.5 means the pixel belongs to the background area;

[0145] q iIndicates pixel i in mask Hip Category of;

[0146] ε represents a very small constant.

[0147] Construct the cross entropy loss function,

[0148] Where (r,c) represents the pixel coordinates;

[0149] H and W represent the length and width of the segmentation result, respectively;

[0150] P G(r,c) , P F(r,c) Represent the segmentation result and mask respectively. Hip The value of the corresponding pixel.

[0151] Loss function L total For UAM module, L Dice , L BCE The proposed modules are weighted.

[0152] in, Respectively represent the feature maps after the size is unified mask Hip segmentation loss;

[0153] Represents the feature map F of the fusion output respectively fuse 、mask Hip segmentation loss;

[0154] μ l represents the lth uncertainty factor;

[0155] μ fuse represents the uncertainty factor of the fusion output;

[0156] L represents the maximum depth.

[0157] The training test includes:

[0158] The training set D obtained by NU-Mamba after T1 MRI image preprocessing train For training, use the Adam optimizer with a learning rate of 10 -4 , save the model parameters after 500 iterations;

[0159] The test set D is obtained after preprocessing the T1 MRI images. test Load the model parameters to predict the segmentation result Seg , according to Result Seg and maskHip Calculate the Dice coefficient;

[0160] Judge L Dice , if L Dice If it is lower than 0.9, repeat the training and predict the calculation of L Dice , if L Dice If it is not less than 0.9, the model at this time is output and the T1 MRI image is segmented to obtain the segmentation result.

[0161] Compared with the prior art, the present invention has the following beneficial effects:

[0162] 1. The present invention solves the problem that the existing model optimization can only achieve suboptimal results by constructing different loss functions to balance the importance of foreground and background. Compared with other hippocampus segmentation algorithms, the segmentation accuracy is improved by more than 10%.

[0163] 2. The present invention is automated and intelligent. Doctors only need to input image data to obtain segmentation results, which has better clinical practicality.

[0164] 3. The present invention provides assistance to doctors in clinical diagnosis and formulation of patient treatment plans through simpler operation and faster result presentation, and has high immediacy and popularity. BRIEF DESCRIPTION OF THE DRAWINGS

[0165] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0166] Figure 1 Schematic diagram of the hippocampus segmentation method based on coronal T1-MRI. DETAILED DESCRIPTION

[0167] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0168] A hippocampus segmentation method based on coronal T1-MRI provided by the present invention comprises:

[0169] Data acquisition step S1: Acquire AD coronal T1 MRI image data and the corresponding hippocampal label, denoted as I m and mask Hip To ensure data consistency and standardization, m Preprocessing is performed and the image is resampled to 224×224×16 to obtain IT1W1 . Build AD database D all , divided into training set (D train ) and the test set (D test ), where D train For model training, D test Used for model performance evaluation and verification.

[0170] Specifically, step S1 includes:

[0171] Step S1.1: Data screening of the acquired coronal T1 MRI images was performed to exclude low-quality images (such as those with excessive noise, blur, incomplete images, etc.) to ensure that all images met the predetermined quality standards. The patient's personal information, including name, gender, age, and other metadata that may identify the patient, was deleted through desensitization. The screened T1 MRI images were resized using image processing technology to obtain I T1W1 , resize each image to 224×224×16.

[0172] Step S1.2: Two clinicians manually annotated the coronal T1 MRI images to obtain the hippocampus label mask Hip And build AD database D all .

[0173] Feature extraction step S2: Construct a feature extraction module to extract multi-scale features from T1 MRI images. The feature extraction module includes the Mamba module, namely the MB module and the RSU module (RSU). The MB module further refines feature extraction through global modeling and enhances the ability to capture global information. The RSU module uses a multi-level convolutional structure to enhance feature representation capabilities.

[0174] Specifically, step S2 includes:

[0175] Step S2.1: Construct the MB module, including encoding and decoding.

[0176] Define the input of MB as Where W, H, C, and l represent the width, height, number of channels, and depth of the module respectively. In the encoding stage, After (5-l) VSS layers and (4-l) patch fusion layers, the feature map is obtained The patch fusion layer acts as a downsampling operation to gradually compress the spatial resolution of the feature map, reducing the size of the feature map by half. The VSS layer captures spatial and channel information and combines semantic features to learn the feature representation of the input feature map. After (5-1) VSS layers and (4-1) patch expansion layers, we get The patch expansion layer acts as an upsampling operation to double the size of the feature map. In each VSS layer, the feature map It is divided into two branches for parallel processing, namely the feature expansion branch (F path1 ) and semantic compression branch (F path2 ). This dual-branch structure is designed to process the spatial information and semantic information of features simultaneously, enhancing the expressiveness of features. path1 In the VSS layer, The number of channels is expanded to λ×C, and the formula is defined as follows.

[0177]

[0178] Among them, λ represents a manually defined hyperparameter, which is set to 4 here. 1 A series of operations including depth convolution (DWConv), activation function (ReLU), SSM module (for local feature enhancement) and layer normalization (LayerNorm) are taken to generate enhanced feature representation. The formula is defined as follows.

[0179] F path1 =LayerNorm(SSM(SiLU(DWConv(F 1 ))))

[0180] In the second branch F path2 In, yes Perform linear transformation (Linear) and SiLU activation operations, the formula is defined as follows:

[0181]

[0182] Finally, the VSS layer fuses the feature maps generated by the two branches through matrix addition operation, and generates the final output feature map through linear transformation (Linear) The formula is defined as follows:

[0183] F out =Linear(F path1 ⊕F path2 )

[0184] in, Represents a three-dimensional vector;

[0185] Represents the encoding stage characteristics of the lth MB module;

[0186] represents the decoding stage characteristics of the lth MB module;

[0187] Represents the input characteristics of the lth VSS module.

[0188] Step S2.2: Construct the RSU module. The RSU is a U-shaped symmetrical structure consisting of two parts: encoding and decoding. The input of the RSU is defined as Where W, H, C and l represent length, width, number of channels and depth respectively. In the encoding stage, After (5-1) feature extraction layers and (4-1) downsampling layers, the feature map is obtained. The feature extraction layer includes a convolution operation with a kernel size of 3×3, a batch normalization operation, and a ReLU activation operation. The downsampling layer is a maximum pooling operation. In the decoding stage, the feature map After (5-1) feature extraction layers and (4-1) upsampling layers, the feature map is obtained. The upsampling layer uses the nearest neighbor interpolation function.

[0189] Uncertainty attention module construction step S3: To enhance segmentation accuracy, an uncertainty attention module (UAM) is constructed to process the feature map in each decoding process and calculate the uncertainty factor μ. This module uses the extracted segmentation edge information and the mean and variance of the weight map to dynamically adjust the model's attention to the foreground and background areas to improve the robustness of segmentation.

[0190] Specifically, step S3 includes:

[0191] Step S3.1: Construct the UAM module. UAM constructs the segmentation edge pixel weight map through the feature map, and then converts the weight map into the uncertainty factor μ. The formula is defined as follows:

[0192] μ=σ(P m )(1-σ(P m ))P v

[0193]

[0194] For σ(P m ) is used for derivative calculation, and the formula is defined as follows:

[0195]

[0196] The calculation formula of μ is obtained by combining the formulas, and the formula is defined as follows:

[0197]

[0198] Among them, σ() represents the Sigmoid activation function;

[0199] P m , P vThey represent the mean and variance of the weight graph respectively.

[0200] σ′(P m ) represents σ(P m ) is the derivative of .

[0201] ε represents a very small constant.

[0202] Segmentation network construction step S4: Based on the constructed MB, RSU and UAM modules, the segmentation network NU-Mamba is constructed. The network includes an encoder, a decoder and a deep supervision mechanism to process the hippocampus segmentation task of coronal T1 MRI images.

[0203] The proposed MB, RSU and UAM modules are combined to perform hippocampus segmentation on the coronal T1 MRI of Alzheimer's disease. By designing different loss functions to balance the importance of foreground and background, the problem that the existing model optimization can only achieve suboptimal results is solved. Compared with other hippocampus segmentation algorithms, the segmentation accuracy is improved by more than 10%.

[0204] Specifically, the modules proposed in step S2 and step S3 are constructed to obtain NU-Mamba. The network consists of an encoder, a decoder, and a deep supervision mechanism. The encoder includes four MBs and downsampling operations, and the decoder includes four RSUs and upsampling operations.

[0205] In more preferred embodiments, the step S4 specifically includes:

[0206] Step S4.1: In the encoder, I m As the input of the whole model. m The size of I is 224×224. m Perform convolution operation to obtain feature map The convolution kernel size is 1×1. Next, After a MB 4 And the downsampling operation is

[0207] Among them, MB 4 Indicates a depth of 4. The formula is defined as follows:

[0208]

[0209] Among them, CBD represents the convolution operation with a convolution kernel of 3×3, the batch normalization operation, and the ReLU activation operation.

[0210] right Perform MB and downsampling operations three times to obtain The sizes are 56×56×128, 28×28×256, and 14×14×512 respectively.

[0211] Step S4.2: In the decoder, After one RSU Secondly, through the Concat operation and Fusion generation Will Input into UAM to get the uncertainty factor μ 4 Next, Perform upsampling operation to obtain feature map

[0212] Step S4.3: After one RSU Through the Concat operation and Fusion generation Next, Send to UAM to get μ 3 .right Perform upsampling to obtain

[0213] Step S4.4: Repeat step S4.3 twice to obtain μ 2 and μ 1 and feature map The sizes are 112×112×64 and 224×224×32 respectively.

[0214] Step S4.5: Upsampling the feature map The size is unified to 224×224×32 and the feature map is obtained

[0215] Step S4.6: Use the Concat operation to The fusion is sent to the 3×3 convolution layer to obtain the feature map F fuse . fuse Send to UAM to get μ fuse .

[0216] Loss function construction step S5: Construct loss function L total , including Dice loss function, cross entropy loss function and uncertainty factor μ. total Balance the dynamic balance of hippocampal tissue areas and background areas.

[0217] Specifically, step S5 includes:

[0218] Step S5.1: Construct the Dice loss function, the formula is defined as follows:

[0219]

[0220] Among them, N represents the number of pixel values ​​of the model segmentation result;

[0221] p i represents the probability of pixel i in the segmentation result, p i ≥0.5 means that the pixel belongs to the hippocampus region, p i <0.5 means the pixel belongs to the background area;

[0222] q i Indicates pixel i in mask Hip Category of;

[0223] ε is a very small constant used to prevent the denominator from being zero.

[0224] Step S5.2: Construct the cross entropy loss function, the formula is defined as follows:

[0225]

[0226] Where (r,c) represents the pixel coordinates;

[0227] H and W represent the length and width of the segmentation result, respectively;

[0228] P G(r,c) , P F(r,c) Represent the segmentation result and mask respectively. Hip The value of the corresponding pixel.

[0229] Step S5.3: L total For UAM module, L Dice , L BCE The proposed modules are weighted and the formula is defined as follows:

[0230]

[0231] in, Respectively represent the feature maps after the size is unified mask Hip segmentation loss;

[0232] Represents the feature map F of the fusion output respectively fuse 、mask Hip segmentation loss;

[0233] μ l represents the lth uncertainty factor;

[0234] μfuse represents the uncertainty factor of the fusion output;

[0235] L represents the maximum depth.

[0236] Model training segmentation step S6: NU-Mamba in D train Upper pair I T1W1 Train until L total The final model parameters are stabilized and saved. test NU-Mamba was tested on the . The calculated Dice coefficient was no less than 0.9.

[0237] Specifically, step S6 includes:

[0238] Step S6.1: Install NU-Mamba in D train Training is performed. The Adam optimizer is used and the learning rate is 10 -4 After 500 iterations, the network becomes stable and the model parameters are saved.

[0239] Step S6.2: Load the test set into the trained model parameters for prediction to obtain the segmentation result Result Seg According to Result Seg and mask Hip Calculate the Dice coefficient.

[0240] Step S6.3: Determine the Dice coefficient, i.e. L Dice , if the Dice coefficient is lower than 0.9, repeat steps S6.1 and S6.2;

[0241] If the Dice coefficient is not less than 0.9, the model is output and the image data is segmented.

[0242] The hippocampus segmentation method proposed in the present invention is automated and intelligent. Doctors only need to input image data to obtain segmentation results. Simpler operations and faster result presentation provide greater assistance to doctors in clinical diagnosis and formulation of patient treatment plans, and have better clinical practicality.

[0243] The present invention also provides a hippocampus segmentation system based on coronal T1-MRI. The hippocampus segmentation system based on coronal T1-MRI can be realized by executing the process steps of the hippocampus segmentation method based on coronal T1-MRI, that is, a person skilled in the art can understand the hippocampus segmentation method based on coronal T1-MRI as a preferred implementation of the hippocampus segmentation system based on coronal T1-MRI.

[0244] A hippocampus segmentation system based on coronal T1-MRI provided by the present invention includes: a data acquisition module and a segmentation network NU-Mamba.

[0245] The segmentation network NU-Mamba consists of a feature extraction module and an uncertainty attention module.

[0246] The data acquisition module obtains AD coronal T1 MRI images and corresponding hippocampal labels, and obtains the data set after preprocessing;

[0247] The feature extraction module extracts multi-scale features from the T1 MRI image to obtain a feature map;

[0248] The uncertainty attention module processes the feature map and calculates the uncertainty factor;

[0249] The loss function is constructed through uncertainty factors to balance the dynamic equilibrium;

[0250] NU-Mamba is trained and tested on the dataset and outputs the final model to segment T1 MRI images to obtain segmentation results.

[0251] In more preferred embodiments, in the data acquisition module, the T1 MRI image and the corresponding hippocampus label are respectively denoted as I m 、mask Hip , to I m Data screening was performed to exclude low-quality images, delete patient metadata, and resample the screened T1 MRI images to 224×224×16 to obtain I T1W1 , manually annotate the coronal T1 MRI image to obtain the hippocampus label mask Hip And build AD database D all , the ratio of 8:2 is divided into training set D train With the test set D test .

[0252] The meta information includes name, gender, age and personal information that can identify the patient.

[0253] The low-quality image is an image with excessive noise, blur and incomplete image.

[0254] The feature extraction module includes a MB module and a RSU module.

[0255] The uncertainty attention module uses the extracted segmentation edge information and the mean and variance of the weight map to dynamically adjust the model's attention on the foreground and background areas.

[0256] The segmentation network NU-Mamba includes an encoder, a decoder and a deep supervision mechanism.

[0257] The encoder includes four MBs and a downsampling operation.

[0258] The decoder includes four RSUs and an upsampling operation.

[0259] The loss function L total Including Dice loss function, cross entropy loss function and uncertainty factor μ, balancing the dynamic balance between the hippocampal tissue area and the background area.

[0260] The NU-Mamba in D train Upper pair I T1W1 Train until L total Tends to stabilize and save the final model parameters, in D test NU-Mamba was tested on the Dice Not less than 0.9, output the model and segment the T1 MRI image to obtain the segmentation result.

[0261] In more preferred embodiments, the feature extraction module includes:

[0262] MB module: Define the input as During the coding phase, After 5-1 VSS layers and 4-1 patch fusion layers, the encoded feature map is obtained In the decoding stage, After 5-1 VSS layers and 4-1 patch expansion layers, the decoded feature map is obtained

[0263] In each VSS layer, the feature map Divided into feature extension branch F psth1 and semantic compression branch F path2 Parallel processing;

[0264] For F path1 , the VSS layer will The number of channels is expanded to λ×C to generate enhanced feature representation.

[0265] F path1 =LayerNorm(SSM(SiLU(DWConv(F 1 ))));

[0266] For F path2 ,

[0267] Fusion and generation of the final output feature map F out =Linear(Fpath1 ⊕F patp2 );

[0268] Among them, W, H, C, and l represent the width, height, number of channels, and depth of the module respectively;

[0269] Represents a three-dimensional vector;

[0270] Represents the encoding stage characteristics of the lth MB module;

[0271] represents the decoding stage characteristics of the lth MB module;

[0272] represents the input characteristics of the lth VSS module;

[0273] Linear() represents a linear transformation operation;

[0274] LayerNorm() represents the layer normalization operation;

[0275] SSM() represents the local feature enhancement operation;

[0276] SiLU() represents the activation function operation;

[0277] DWConv() represents the depth convolution operation;

[0278] The patch fusion layer is a downsampling operation;

[0279] The VSS layer captures spatial and channel information and combines semantic features to learn the feature representation of the input feature map;

[0280] The patch expansion layer is an upsampling operation.

[0281] RSU module: Define the input as During the coding phase, After 5-1 feature extraction layers using convolution operations with a convolution kernel size of 3×3, batch normalization operations and activation function operations and 4-1 downsampling layers using maximum pooling operations, the encoded feature map is obtained.

[0282] In the decoding stage, After 5-1 feature extraction layers and 4-1 upsampling layers, the nearest neighbor interpolation function is used to obtain the decoding feature map

[0283] The uncertainty attention module constructs a segmentation edge pixel weight map through the feature map. For σ(P m ) to perform derivative calculations, Then perform the combined transformation to obtain the uncertainty factor μ,

[0284] Among them, σ() represents the Sigmoid activation function;

[0285] P m , P v Respectively represent the mean and variance of the weight graph;

[0286] σ′(P m ) represents σ(P m )’s derivative;

[0287] ε represents a very small constant.

[0288] In more preferred embodiments, the segmented network NU-Mamba includes:

[0289] Module M4.1: In the encoder, I m As input, the size is 224×224, and the convolution operation with the convolution kernel size of 1×1 is performed to obtain the feature map After a MB 4 And the downsampling operation is

[0290] right The MB and downsampling operations are performed three times to obtain the sizes of 56×56×128, 28×28×256, and 14×14×512 respectively.

[0291] Among them, MB 4 Indicates a MB operation with a depth of 4;

[0292] CBD() represents a convolution operation with a convolution kernel of 3×3, a batch normalization operation, and an activation function operation.

[0293] Module M4.2: In the decoder, After one RSU Through the Concat operation and Fusion generation Input into UAM to get the uncertainty factor μ 4 ,right Perform upsampling operation to obtain feature map

[0294] Module M4.3: Obtained through an RSU Through the Concat operation and Fusion generation Send to UAM to get μ 3 ,right Perform upsampling to obtain

[0295] Module M4.4: Yes Repeat triggering module M4.3 twice to obtain μ 2 and μ 1 and The sizes are 112×112×64 and 224×224×32 respectively.

[0296] Module M4.5: Upsampling operation is used to convert feature maps The size is unified to 224×224×32 and the feature map is obtained

[0297] Module M4.6: Concat operation The fusion is sent to the 3×3 convolution layer to obtain the feature map F fuse , sent to UAM to get μ fuse .

[0298] In more preferred embodiments, the constructing the loss function includes:

[0299] Construct the Dice loss function,

[0300] Among them, N represents the number of pixel values ​​of the model segmentation result;

[0301] p i represents the probability of pixel i in the segmentation result, p i ≥0.5 means that the pixel belongs to the hippocampus region, p i <0.5 means the pixel belongs to the background area;

[0302] q i Indicates pixel i in mask Hip Category of;

[0303] ε represents a very small constant.

[0304] Construct the cross entropy loss function,

[0305] Where (r,c) represents the pixel coordinates;

[0306] H and W represent the length and width of the segmentation result, respectively;

[0307] P G(r,c) , P F(r,c) Represent the segmentation result and mask respectively. Hip The value of the corresponding pixel.

[0308] Loss function L total For UAM module, L Dice , L BCE The proposed modules are weighted,

[0309] in, Represent the feature maps after the size is unified mask Hip segmentation loss;

[0310] Represent the feature maps F of the fusion output respectively fuse 、mask Hip segmentation loss;

[0311] μ l represents the lth uncertainty factor;

[0312] μ fuse represents the uncertainty factor of the fusion output;

[0313] L represents the maximum depth.

[0314] The training test includes:

[0315] The training set D obtained by NU-Mamba after T1 MRI image preprocessing train For training, use the Adam optimizer with a learning rate of 10 -4 , save the model parameters after 500 iterations;

[0316] The test set D is obtained after preprocessing the T1 MRI images. test Load the model parameters to predict the segmentation result Seg , according to Result Seg and mask Hip Calculate the Dice coefficient;

[0317] Judge L Dice , if L Dice If it is lower than 0.9, repeat the training and predict the calculation of L Dice , if L Dice If it is not less than 0.9, the model at this time is output and the T1 MRI image is segmented to obtain the segmentation result.

[0318] Those skilled in the art know that, in addition to realizing the system and its various devices, modules, and units provided by the present invention in a purely computer-readable program code, it is entirely possible to realize the same functions in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered as a hardware component, and the devices, modules, and units included therein for realizing various functions can also be regarded as structures within the hardware component; the devices, modules, and units for realizing various functions can also be regarded as both software modules for realizing the method and structures within the hardware component.

[0319] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A hippocampus segmentation method based on coronal T1-MRI, characterized in that: include: Data acquisition steps: obtain AD coronal T1 MRI images and corresponding hippocampal labels, and obtain the data set after preprocessing; Feature extraction steps: Construct a feature extraction module to extract multi-scale features from T1 MRI images and obtain feature maps; Uncertainty attention module construction steps: Construct uncertainty attention module to process feature graph and calculate uncertainty factor; Segmentation network construction steps: Based on the feature extraction module and the uncertainty attention module, build the segmentation network NU-Mamba; Loss function construction steps: construct loss function balance dynamic balance through uncertainty factors; Model training and segmentation steps: NU-Mamba is trained and tested on the dataset and outputs the final model, which segments the T1 MRI images to obtain segmentation results.

2. The hippocampus segmentation method based on coronal T1-MRI according to claim 1, characterized in that: In the data acquisition step, the T1 MRI image and the corresponding hippocampus label are respectively denoted as I m 、mask Hip ,include: Step S1.1: Screen the acquired coronal T1 MRI images, exclude low-quality images, delete the patient's metadata, and resample the screened T1 MRI images to obtain I T1W1 , resize each image to 224×224×16; Step S1.2: Manually annotate the coronal T1 MRI image to obtain the hippocampus label mask Hip And build AD database D all , the ratio of 8:2 is divided into training set D train With the test set D test ; The meta information is name, gender, age and personal information that can identify the patient; The low-quality image is an image with excessive noise, blur and incomplete image; The feature extraction module includes a MB module and an RSU module; The uncertainty attention module uses the extracted segmentation edge information and the mean and variance of the weight map to dynamically adjust the model's attention on the foreground and background areas; The segmentation network NU-Mamba includes an encoder, a decoder and a deep supervision mechanism; The encoder includes four MBs and a downsampling operation; The decoder includes four RSUs and an upsampling operation; The loss function L total Including Dice loss function, cross entropy loss function and uncertainty factor μ, balancing the dynamic balance between the hippocampal tissue area and the background area; In the model training segmentation step, NU-Mamba train Upper pair I T1W1 Train until L total Tends to stabilize and save the final model parameters, in D test NU-Mamba was tested on the Dice Not less than 0.9, output the model and segment the T1MRI image to obtain the segmentation result.

3. The hippocampus segmentation method based on coronal T1-MRI according to claim 1, characterized in that: The feature extraction step comprises: MB module build step: define the input as During the coding phase, After 5-1 VSS layers and 4-1 patch fusion layers, the encoded feature map is obtained In the decoding stage, After 5-1 VSS layers and 4-1 patch expansion layers, the decoded feature map is obtained In each VSS layer, the feature map Divided into feature extension branch F path1 and semantic compression branch F path2 Parallel processing; For F path1 , the VSS layer will The number of channels is expanded to λ×C to generate enhanced feature representation. F path1 =LayerNorm(SSM(SiLU(DWConv(F1)))); For F path2 , Fusion and generation of the final output feature map Among them, W, H, C, and l represent the width, height, number of channels, and depth of the module respectively; Represents a three-dimensional vector; represents the encoding stage characteristics of the lth MB module; represents the decoding stage characteristics of the lth MB module; represents the input characteristics of the lth VSS module; Linear() represents a linear transformation operation; LayerNorm() represents the layer normalization operation; SSM() represents the local feature enhancement operation; SiLU() represents the activation function operation; DWConv() represents the depth convolution operation; The patch fusion layer is a downsampling operation; The VSS layer captures spatial and channel information and combines semantic features to learn the feature representation of the input feature map; The patch expansion layer is an upsampling operation; RSU module construction steps: define the input as During the coding phase, After 5-1 feature extraction layers using convolution operations with a convolution kernel size of 3×3, batch normalization operations and activation function operations and 4-1 downsampling layers using maximum pooling operations, the encoded feature map is obtained. In the decoding stage, After 5-1 feature extraction layers and 4-1 upsampling layers, the nearest neighbor interpolation function is used to obtain the decoding feature map The uncertainty attention module construction steps include: The uncertainty attention module UAM constructs a segmentation edge pixel weight map through the feature map. For σ(P m ) to perform derivative calculations, Then the uncertainty factor μ is obtained by combined transformation. Among them, σ() represents the Sigmoid activation function; P m , P v Respectively represent the mean and variance of the weight graph; σ′(P m ) represents σ(P m )’s derivative; ε represents a very small constant.

4. The hippocampus segmentation method based on coronal T1-MRI according to claim 1, characterized in that: The segmentation network building step includes: Step S4.1: In the encoder, I m As input, the size is 224×224, and the convolution operation with the convolution kernel size of 1×1 is performed to obtain the feature map After an MB4 and downsampling operation, we get right The MB and downsampling operations are performed three times to obtain the sizes of 56×56×128, 28×28×256, and 14×14×512 respectively. Among them, MB4 represents MB operation with a depth of 4; CBD() represents a convolution operation with a convolution kernel of 3×3, a batch normalization operation, and an activation function operation; Step S4.2: In the decoder, After one RSU Through the Concat operation and Fusion generation The uncertainty factor μ4 is sent to UAM. Perform upsampling operation to obtain feature map Step S4.3: Get through an RSU Through the Concat operation and Fusion generation Send it into UAM to get μ3, Perform upsampling to obtain Step S4.4: Repeat step S4.3 twice to obtain μ2 and μ1 and The sizes are 112×112×64, 224×224×32 respectively; Step S4.5: Upsampling the feature map The size is unified to 224×224×32 and the feature map is obtained Step S4.6: Use the Concat operation to The fusion is sent to the 3×3 convolution layer to obtain the feature map F fuse , sent to UAM to get μ fuse .

5. The hippocampus segmentation method based on coronal T1-MRI according to claim 1, characterized in that: The loss function construction step includes: Step S5.1: Construct Dice loss function, Among them, N represents the number of pixel values ​​of the model segmentation result; p i represents the probability of pixel i in the segmentation result, p i ≥0.5 means that the pixel belongs to the hippocampus region, p i <0.5 means the pixel belongs to the background area; q i Indicates pixel i in mask Hip Category of; ε represents a very small constant; Step S5.2: Construct the cross entropy loss function, Where (r,c) represents the pixel coordinates; H and W represent the length and width of the segmentation result, respectively; P G(r,c) , P F(r,c) Represent the segmentation result and mask respectively. Hip The value of the corresponding pixel; Step S5.3: Loss function L total For UAM module, L Dice , L BCE The proposed modules are weighted, in, Represent the feature maps after the size is unified segmentation loss; Represent the feature maps F of the fusion output respectively fuse 、mask Hip segmentation loss; μ l represents the lth uncertainty factor; μ fuse represents the uncertainty factor of the fusion output; L represents the maximum depth; The model training segmentation step includes: Step S6.1: The training set D obtained by NU-Mamba after T1 MRI image preprocessing train For training, use the Adam optimizer with a learning rate of 10 -4 , save the model parameters after 500 iterations; Step S6.2: Preprocess the T1 MRI images to obtain the test set D test Load the model parameters to predict the segmentation result Seg , according to Result Seg and mask Hip Calculate the Dice coefficient; Step S6.3: Determine L Dice , if L Dice If it is lower than 0.9, repeat steps S6.1 and S6.2; If L Dice If it is not less than 0.9, the model at this time is output and the T1 MRI image is segmented to obtain the segmentation result.

6. A hippocampus segmentation system based on coronal T1-MRI, characterized in that: include: Data acquisition module and segmentation network NU-Mamba; The segmentation network NU-Mamba consists of a feature extraction module and an uncertainty attention module; The data acquisition module obtains AD coronal T1 MRI images and corresponding hippocampal labels, and obtains the data set after preprocessing; The feature extraction module extracts multi-scale features from the T1 MRI image to obtain a feature map; The uncertainty attention module processes the feature map and calculates the uncertainty factor; The loss function is constructed through uncertainty factors to balance the dynamic equilibrium; NU-Mamba is trained and tested on the dataset and outputs the final model to segment T1 MRI images to obtain segmentation results.

7. The hippocampus segmentation system based on coronal T1-MRI according to claim 6, characterized in that: In the data acquisition module, the T1 MRI image and the corresponding hippocampus label are respectively denoted as I m 、mask Hip , to I m Data screening was performed to exclude low-quality images, delete patient metadata, and resample the screened T1 MRI images to 224×224×16 to obtain I T1W1 , manually annotate the coronal T1 MRI image to obtain the hippocampus label mask Hip And build AD database D all , the ratio of 8:2 is divided into training set D train With the test set D test ; The meta information is name, gender, age and personal information that can identify the patient; The low-quality image is an image with excessive noise, blur and incomplete image; The feature extraction module includes a MB module and an RSU module; The uncertainty attention module uses the extracted segmentation edge information and the mean and variance of the weight map to dynamically adjust the model's attention on the foreground and background areas; The segmentation network NU-Mamba includes an encoder, a decoder and a deep supervision mechanism; The encoder includes four MBs and a downsampling operation; The decoder includes four RSUs and an upsampling operation; The loss function L total Including Dice loss function, cross entropy loss function and uncertainty factor μ, balancing the dynamic balance between the hippocampal tissue area and the background area; The NU-Mamba in D train Upper pair I T1W1 Train until L total Tends to stabilize and save the final model parameters, in D test NU-Mamba was tested on the Dice Not less than 0.9, output the model and segment the T1 MRI image to obtain the segmentation result.

8. The hippocampus segmentation system based on coronal T1-MRI according to claim 6, characterized in that: The feature extraction module comprises: MB module: Define the input as During the coding phase, After 5-1 VSS layers and 4-1 patch fusion layers, the encoded feature map is obtained In the decoding stage, After 5-1 VSS layers and 4-1 patch expansion layers, the decoded feature map is obtained In each VSS layer, the feature map Divided into feature extension branch F path1 and semantic compression branch F path2 Parallel processing; For F path1 , the VSS layer will The number of channels is expanded to λ×C to generate enhanced feature representation. F path1 =LayerNorm(SSM(SiLU(DWConv(F1)))); For F path2 , Fusion and generation of the final output feature map Among them, W, H, C, and l represent the width, height, number of channels, and depth of the module respectively; Represents a three-dimensional vector; represents the encoding stage characteristics of the lth MB module; represents the decoding stage characteristics of the lth MB module; represents the input characteristics of the lth VSS module; Linear() represents a linear transformation operation; LayerNorm() represents the layer normalization operation; SSM() represents the local feature enhancement operation; SiLU() represents the activation function operation; DWConv() represents the depth convolution operation; The patch fusion layer is a downsampling operation; The VSS layer captures spatial and channel information and combines semantic features to learn the feature representation of the input feature map; The patch expansion layer is an upsampling operation; RSU module: Define the input as During the coding phase, After 5-1 feature extraction layers using convolution operations with a convolution kernel size of 3×3, batch normalization operations and activation function operations and 4-1 downsampling layers using maximum pooling operations, the encoded feature map is obtained. In the decoding stage, After 5-1 feature extraction layers and 4-1 upsampling layers, the nearest neighbor interpolation function is used to obtain the decoding feature map The uncertainty attention module constructs a segmentation edge pixel weight map through the feature map. For σ(P m ) to perform derivative calculations, Then perform the combined transformation to obtain the uncertainty factor μ, Among them, σ() represents the Sigmoid activation function; P m , P v Respectively represent the mean and variance of the weight graph; σ′(P m ) represents σ(P m )’s derivative; ε represents a very small constant.

9. The hippocampus segmentation system based on coronal T1-MRI according to claim 6, characterized in that: The segmented network NU-Mamba includes: Module M4.1: In the encoder, I m As input, the size is 224×224, and the convolution operation with the convolution kernel size of 1×1 is performed to obtain the feature map After an MB4 and downsampling operation, we get right The MB and downsampling operations are performed three times to obtain the sizes of 56×56×128, 28×28×256, and 14×14×512 respectively. Among them, MB4 represents MB operation with a depth of 4; CBD() represents a convolution operation with a convolution kernel of 3×3, a batch normalization operation, and an activation function operation; Module M4.2: In the decoder, After one RSU Through the Concat operation and Fusion generation The uncertainty factor μ4 is sent to UAM. Perform upsampling operation to obtain feature map Module M4.3: Obtained through an RSU Through the Concat operation and Fusion generation Send it into UAM to get μ3, Perform upsampling to obtain Module M4.4: Yes Repeat triggering module M4.3 twice to obtain μ2 and μ1 and The sizes are 112×112×64, 224×224×32 respectively; Module M4.5: Upsampling operation is used to convert feature maps The size is unified to 224×224×32 and the feature map is obtained Module M4.6: Concat operation The fusion is sent to the 3×3 convolution layer to obtain the feature map F fuse , sent to UAM to get μ fuse .

10. The hippocampus segmentation system based on coronal T1-MRI according to claim 6, characterized in that: The construction of the loss function includes: Construct the Dice loss function, Among them, N represents the number of pixel values ​​of the model segmentation result; p i represents the probability of pixel i in the segmentation result, p i ≥0.5 means that the pixel belongs to the hippocampus region, p i <0.5 means the pixel belongs to the background area; q i Indicates pixel i in mask Hip Category of; ε represents a very small constant; Construct the cross entropy loss function, Where (r,c) represents the pixel coordinates; H and W represent the length and width of the segmentation result, respectively; P G(r,c) , P F(r,c) Represent the segmentation result and mask respectively. Hip The value of the corresponding pixel; Loss function L total For UAM module, L Dice , L BCE The proposed modules are weighted, in, Represent the feature maps after the size is unified mask Hip segmentation loss; Represent the feature maps F of the fusion output respectively fuse 、mask Hip segmentation loss; μ l represents the lth uncertainty factor; μ fuse represents the uncertainty factor of the fusion output; L represents the maximum depth; The training test includes: The training set D obtained by NU-Mamba after T1 MRI image preprocessing train For training, use the Adam optimizer with a learning rate of 10 -4 , save the model parameters after 500 iterations; The test set D is obtained after preprocessing the T1 MRI images. test Load the model parameters to predict the segmentation result Seg , according to Result Seg and mask Hip Calculate the Dice coefficient; Judge L Dice , if L Dice If it is lower than 0.9, repeat the training and predict the calculation of L Dice , if L Dice If it is not less than 0.9, the model at this time is output and the T1 MRI image is segmented to obtain the segmentation result.

Images