A method and system for brain image target structure segmentation based on progressive channel compression

Abstract

This invention discloses a method and system for target structure segmentation in brain images based on progressive channel compression. The method includes the following steps: acquiring and preprocessing a three-dimensional medical image volume block of the subject's brain; determining the channel sequence of the encoder-decoder network according to the progressive channel compression efficiency criterion; setting channel-adaptive attention gating modules at the skip connections of each stage of the encoder-decoder network; training the encoder-decoder network to obtain a trained encoder-decoder network; inputting the preprocessed three-dimensional medical image volume block into the trained encoder-decoder network, and finally outputting a three-dimensional segmentation probability map; and thresholding the three-dimensional segmentation probability map to obtain a three-dimensional segmentation mask. This invention has advantages such as high parameter efficiency, strong generalization ability, high boundary localization accuracy, and excellent segmentation performance for small target structures.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method and system for segmenting target structures in brain images based on progressive channel compression. Background Technology

[0002] Hippocampal shrinkage is the earliest and most sensitive imaging biomarker of Alzheimer's disease (AD). Magnetic resonance imaging (MRI) is the preferred method for measuring hippocampal volume. However, hippocampal boundaries are blurred and morphology is complex. Manual annotation is time-consuming and has poor repeatability. Therefore, the development of automated segmentation algorithms is of great significance.

[0003] Existing deep learning segmentation methods suffer from the following main technical problems:

[0004] (1) Existing standard 3D encoder-decoder segmentation networks (such as 3D U-Net) use a uniform doubling channel configuration [64,128,256,512,1024], which has a large number of parameters. In small sample scenarios, the overfitting gap is significant and it cannot generalize stably.

[0005] (2) Although existing automated medical image segmentation frameworks (such as nnU-Net) can adaptively adjust network hyperparameters, their channel configuration is based on heuristic rules constrained by hardware resources, and they have not established a formal constraint relationship with the training sample size, nor have they performed non-uniform optimization for the differences in semantic expression at each coding level in small sample scenarios.

[0006] (3) Existing lightweight segmentation methods (such as methods based on depth-separable convolution or channel pruning) reduce the number of parameters by compressing the number of channels in a uniform and proportional manner. The growth ratio of each coding level channel is completely equal. There is a lack of a systematic method to incorporate the training sample size into the channel design constraints. Shallow features are redundant while deep semantic expression is limited.

[0007] (4) Existing attention gating mechanisms (such as Attention U-Net) use a fixed projection dimension. The attention gating was not designed in coordination with the channel configuration. The deep and shallow attention gating were allocated the same proportion of projection capacity, resulting in limited expression of deep attention or redundancy of shallow attention.

[0008] (5) The boundary-aware scale parameters in existing loss function designs (such as Dice loss, Tversky loss, boundary loss, etc.) are independent of the network channel design. The hyperparameter settings rely on manual experience search and lack a principled design method coupled with the network structure.

[0009] The root cause of the above problems is that no formal constraint relationship has been established between channel configuration design and training sample size, which makes it impossible to optimize channel configuration, attention mechanism and loss function in a small sample scenario of medical imaging. Summary of the Invention

[0010] To address the technical problems existing in the prior art, this invention provides a brain image target structure segmentation method and system based on progressive channel compression, which features high parameter efficiency, strong generalization ability, high boundary localization accuracy, and excellent small target structure segmentation performance.

[0011] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:

[0012] A brain image target structure segmentation method based on progressive channel compression includes the following steps:

[0013] Three-dimensional medical image volume blocks of the subject's cranium were acquired and preprocessed to obtain preprocessed three-dimensional medical image volume blocks. ;

[0014] The channel sequence of the encoder-decoder network is determined according to the progressive channel compression efficiency criterion. ; Set up channel adaptive attention gating modules at the jump connections of each level of the encoder-decoder network; based on the channel sequence The first in Level channel compression ratio Number of input channels Adaptive determination of the first Level of attention projection dimension Then the encoder-decoder network is trained to obtain a trained encoder-decoder network.

[0015] preprocessed 3D medical image volume blocks The input is fed into the trained encoder-decoder network to obtain encoder features. With decoder gate signal , in the first Level of attention projection dimension Calculate the attention coefficient within the spanned adaptive projection space And with attention coefficient Encoder features Perform element-wise weighting to obtain the weighted skip connectivity features. Then feature fusion and progressive upsampling are performed to finally output a three-dimensional segmentation probability map;

[0016] Thresholding is performed on the 3D segmentation probability map to obtain the 3D segmentation mask. .

[0017] Preferably, the first Level of attention projection dimension The calculation formula is:

[0018]

[0019] in, This is a preset lower bound for positive integers.

[0020] Preferably, after the channel sequence is determined Then, the channel non-uniformity index is further calculated. To quantify the non-uniformity of channel sequences; when The channel sequence is then determined to satisfy the non-uniform growth condition.

[0021] Among them, the channel non-uniformity index The calculation formula is:

[0022]

[0023] in These represent the maximum and minimum values ​​of the overall growth ratio. The growth rate is the arithmetic mean. For the first The growth rate of the first level, This represents the pre-bottleneck suppression ratio.

[0024] Preferably, attention coefficient The calculation process is as follows:

[0025] Encoder features With decoder gate signal Projected onto dimension through linear transformation To obtain the intermediate feature vector :

[0026]

[0027] in: , The weight matrix is ​​a linear transformation matrix. For bias terms;

[0028] Will Applying the ReLU activation function Then convolution and Sigmoid activation function Obtain attention coefficient :

[0029]

[0030] in, For the first Level-learnable weight matrix; This is a bias term.

[0031] Preferably, a boundary adaptive composite loss function is used. The encoder-decoder network is trained, where the boundary adaptive composite loss function is used. The calculation formula is:

[0032]

[0033] in These are the weighting coefficients; For volume overlap rate Similarity loss function; This is the weighted cross-entropy loss function.

[0034] Preferably, The calculation formula is:

[0035]

[0036] in For voxel indexing; For category; Category weights; For real labels, To predict probabilities; voxels The boundary-aware weight at a given location is calculated using the following formula:

[0037]

[0038] in voxels The Euclidean distance transformation value to the nearest target boundary; This refers to the boundary sensing intensity coefficient. The radius of the boundary perception.

[0039] Preferably, the boundary sensing radius Pre-bottleneck suppression ratio of encoder-decoder network With volume blocks Side length Confirmed, the specific calculation formula is as follows:

[0040] .

[0041] Preferably, a structured clinical quantitative report is generated based on a three-dimensional segmentation mask of the target structure, specifically as follows:

[0042] Volume normalization: Calculating the voxel volume of the target structure Then, based on the total intracranial volume of the subject Normalization is performed to obtain the standardized volume. ;

[0043] Asymmetric quantization: Calculate the normalized volume on the left side separately. With the normalized volume on the right Generate asymmetric index As an imaging quantitative parameter reflecting the relative difference in the volume of the target structure on both sides;

[0044] Intergroup difference assessment: When data from multiple groups of subjects are available, the standardized volume of each group is assessed. Perform a statistical significance test.

[0045] Preferably, the preprocessing of the three-dimensional medical image volume block of the brain includes sequentially performing skull stripping, affine space registration to a standard brain template, z-score intensity normalization, and region of interest (ROI) cropping.

[0046] The present invention also discloses a brain image target structure segmentation system based on progressive channel compression, comprising an interconnected memory and a processor, wherein the memory stores a computer program, and the computer program executes the steps of the method described above when run by the processor.

[0047] Compared with the prior art, the advantages of the present invention are as follows:

[0048] This invention's method is applicable to brain MRI, small-sample medical image segmentation, and 3D target structure segmentation tasks. It employs the Progressive Channel Compression Efficiency (APCC) design criterion to determine a non-uniformly increasing channel sequence. Channel Adaptive Attention Gating (CAAG) modules are introduced at each skip connection level, allowing the attention projection dimension at each level to adaptively adjust with the channel compression ratio. Training is performed using a boundary adaptive composite loss function, outputting a 3D segmentation mask of the target structure and a structured clinical quantitative report. In small-sample medical image segmentation scenarios, this invention significantly reduces the number of parameters compared to the standard 3D U-Net channel configuration, effectively controlling overfitting gaps. It boasts advantages such as high parameter efficiency, strong generalization ability, high boundary localization accuracy, and excellent segmentation performance for small target structures.

[0049] The APCC criterion of this invention establishes the training sample size for the first time. Formal constraints on channel sequence design elevate channel configuration from empirical design to a verifiable systematic approach; shallow sparsity conditions establish formal constraints on the number of shallow channels and the size of training samples, suppressing overfitting in small sample scenarios; the CNUI index provides a quantifiable evaluation criterion for the non-uniformity of channel configuration.

[0050] The CAAG module of this invention serves as a structured extension of APCC, with a channel compression ratio The system automatically derives the attention projection dimensions at each level using a single input, enabling a verifiable linkage between attention capacity and channel semantic depth. The deep attention projection capacity is significantly improved compared to fixed-dimensional AG, and the overall AG parameter overhead accounts for a very low proportion of the total number of parameters.

[0051] The boundary adaptive composite loss function of the present invention From network structure parameters ( )according to It was determined that the loss function hyperparameters and network structure parameters were automatically adapted across stages, eliminating the dependence on manual search, and maintaining consistency between the data augmentation scale and the loss function boundary awareness scale. Attached Figure Description

[0052] Figure 1 This is a flowchart of a brain image target structure segmentation method based on progressive channel compression, according to an embodiment of the present invention.

[0053] Figure 2 This is a diagram of the encoder-decoder network architecture according to an embodiment of the present invention.

[0054] Figure 3 This is an architecture diagram of the channel adaptive attention gating module according to an embodiment of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0056] like Figure 1 As shown, the brain image target structure segmentation method based on progressive channel compression provided in this embodiment of the invention includes the following steps:

[0057] S1. Obtain the three-dimensional medical image volume block of the subject's brain and preprocess it to obtain the preprocessed three-dimensional medical image volume block. ;

[0058] Specifically, three-dimensional magnetic resonance imaging (MRI) images of the subjects' brains were acquired, and the images underwent preprocessing including skull removal, affine space registration to a standard brain template, z-score intensity normalization, and region of interest (ROI) cropping. Skull removal involved using brain extraction tools to remove the skull and non-brain tissue, obtaining the brain parenchyma. Affine space registration to a standard brain template involved using affine transformation to register the skull-removed images to a standard brain template space to eliminate inter-individual anatomical differences. Z-score intensity normalization involved normalizing the registered images to ensure a mean intensity of 0 and a standard deviation of 1 for all brain voxels. ROI cropping involved extracting a three-dimensional volume block containing either the left or right hippocampus from the standard space, centered on the target structure (hippocampus). During cropping, the proportion of target structure voxels in the total ROI voxels was ensured to be at least 5% to maintain a balance between positive and negative samples. The preprocessed three-dimensional volume blocks served as input to the subsequent encoder-decoder network.

[0059] Specifically, the channel sequence of the encoder-decoder network is determined according to the Progressive Channel Compression Efficiency Criterion (APCC criterion). The APCC criterion requires the channel sequence to be... ( For the first The number of channels in a level encoder, (The number of channels in the bottleneck layer) must simultaneously meet the following four conditions:

[0060] Monotonically increasing property: ;in For encoding levels; For the first Number of output channels of the level encoder;

[0061] Non-uniform growth: Definition of the first Growth rate The growth ratio sequence must have at least one pair of adjacent levels that satisfy the condition. That is, the growth rate is not proportional throughout the entire process;

[0062] Shallow sparsity: ,in This represents the number of channels in the first-stage encoder. Γ represents the number of training samples in the specific implementation, and Γ is a positive integer allocation coefficient. In this embodiment, Γ=2. Give an upper bound on the number of shallow channels;

[0063] Pre-bottleneck inhibition: Pre-bottleneck inhibition ratio ,in To preset the minimum suppression ratio, Must be no less than ; The constraint range is taken ;

[0064] One specific implementation of the channel sequence is It meets the APCC criteria for monotonically increasing and non-uniform growth (growth ratio sequence). (at least one strictly decreasing), shallow sparsity ( (satisfying shallow sparsity constraints) and pre-bottleneck inhibition ( (Satisfies the constraint [0.65, 0.85]).

[0065] To quantify the non-uniformity of channel sequences, a channel non-uniformity index is defined for channel sequences. :

[0066]

[0067] in For the first Growth rate These represent the maximum and minimum values ​​of the overall growth ratio. The growth rate is the arithmetic mean. This represents the pre-bottleneck suppression ratio.

[0068] The index quantifies both the dispersion of the growth ratio sequence and the pre-bottleneck suppression ratio: Corresponding to a uniform and proportional configuration (without non-uniformity); This indicates that the configuration meets the non-uniform growth condition. The larger the value, the higher the degree of non-uniformity; When the value is too large, the expression of deep semantics is limited. The effectiveness should be judged based on the overfit gap Δ not exceeding the preset threshold. The effective range is determined by the combined effect of the training sample size and the proportion of the target structure for the specific segmentation task.

[0069] S2. Determine the channel sequence of the encoder-decoder network according to the progressive channel compression efficiency criterion. ; Set up channel adaptive attention gating modules at the jump connections of each level of the encoder-decoder network; based on the channel sequence The first in Level channel compression ratio Number of input channels Adaptive determination of the first Level of attention projection dimension Then the encoder-decoder network is trained to obtain the trained encoder-decoder network.

[0070] like Figure 2 As shown, the encoder-decoder network is constructed using three-dimensional convolutions and contains four encoding levels (corresponding to...). Figure 2The four encoder blocks are En1, En2, En3, and En4, and the four decoding stages are corresponding to... Figure 2 The system consists of four decoder blocks (Dec1, Dec2, Dec3, and Dec4) and a bottleneck layer; each encoder level contains two 3D convolutional layers, followed by downsampling using 3D max pooling with a stride of 2. Each decoder level contains one upsampling layer and two 3D convolutional layers; skip connections pass encoder features to the corresponding decoder level.

[0071] Channel Adaptive Attention Gating (CAAG) modules are set at each hop connection in the encoder-decoder network. The attention projection dimension of the CAAG module is no longer fixed, and the first... The level attention projection dimension is based on the channel compression ratio of that level. Number of input channels The determination is adaptive and has a pre-defined lower bound of positive integers. .

[0072] Specifically, in the Channel Adaptive Attention Gating (CAAG) module, the first... Attention projection dimension within the adaptive projection space The calculation steps are as follows:

[0073] APCC guidelines determine the channel compression ratio for each coding level. CAAG Automatically derive attention projection dimensions at each level for the input. This establishes a verifiable mathematical link between the expressive capacity of the attention mechanism and the semantic depth of the channel configuration, rather than being fixed as in existing methods. Its independent design. Specifically, the first Level of attention projection dimension The calculation formula is:

[0074]

[0075] in, This is a preset lower bound for positive integers.

[0076] The fundamental difference between CAAG and existing fixed-dimensional AG (attention gating) is that deep AG in CAAG (Attention Gating) ) to obtain sufficient projection expression capacity; shallow AG in CAAG ( Keep it lightweight, the sum of the total parameters of each AG stage satisfy ,in This represents the total number of trainable parameters for the network. This design allows the complexity of the attention mechanism to increase in tandem with the semantic depth of the channel configuration.

[0077] S3. Convert the preprocessed 3D medical image volume blocks As input, these features are fed into the trained encoder-decoder network to obtain the encoder features. With decoder gate signal The CAAG module receives encoder features. With decoder gate signal In the first step determined by step S2 Level of attention projection dimension Calculate the attention coefficient within the spanned projection space And with attention coefficient The encoder features are weighted element-wise to obtain the weighted skip connection features. The data is then sent to the corresponding decoding block for feature fusion and progressive upsampling, ultimately outputting a 3D segmentation probability map. The 3D medical image volume block input to the network is a region of interest (ROI) containing the left or right hippocampus extracted from cranial MRI.

[0078] like Figure 3 As shown, in the Channel Adaptive Attention Gating (CAAG) module, the first... Level voxel position Attention coefficient at the location The calculation process is as follows:

[0079] Encoder features With decoder gate signal Projected onto dimension through linear transformation To obtain the intermediate feature vector :

[0080]

[0081] in: , The weight matrix is ​​a linear transformation matrix. For bias terms;

[0082] Will Applying the ReLU activation function Then it goes through a 1×1×1 convolution and a Sigmoid activation function. Obtain attention coefficient :

[0083]

[0084] in, For the first Level-learnable weight matrix; This is the bias term; after expansion, we get:

[0085]

[0086] Attention coefficient Element-wise multiplication by encoder features The weighted skip connection features are obtained. :

[0087]

[0088] in, This is element-wise multiplication;

[0089] Weighted skip connection features The data is sent to the corresponding decoding block and spliced ​​and fused with the upsampled decoding features to obtain a three-dimensional segmentation probability map.

[0090] In this embodiment, a boundary adaptive composite loss function is used. Train the encoder-decoder network using a boundary-adaptive composite loss function. The calculation formula is:

[0091]

[0092] in These are the weighting coefficients; This is the Dice similarity loss function based on volume overlap rate; The weighted cross-entropy loss function is calculated using the following formula:

[0093]

[0094] in For voxel indexing; Category (foreground / background); Category weights (foreground class weights are greater than background class weights for the target structure); For real labels, To predict probabilities; voxels The boundary-aware weight at a given location is calculated using the following formula:

[0095]

[0096] in voxels The Euclidean distance transformation value to the nearest target boundary. During the training phase, the Euclidean distance transformation graph based on the gold standard annotation of the target structure is pre-calculated; it is not required during the inference phase. ; This refers to the boundary sensing intensity coefficient. For the boundary sensing radius, by Pre-bottleneck suppression ratio of encoder-decoder network With the input volume block side length It is confirmed that the loss function hyperparameters can be automatically adapted to the network structure parameters across stages without manual searching.

[0097] in A larger value indicates that the pre-bottleneck stage retains more semantic channel capacity. As it increases, the loss function focuses on a broader boundary neighborhood; Normalize the input block size to the order of side length, so that It is consistent with the voxel space scale.

[0098] The boundary sensing radius of this invention By deriving network structure parameters, the loss function hyperparameters can be automatically adapted to the network structure design without manual searching. Specifically, the boundary-aware radius... Through network structure parameters With input block size The loss function hyperparameters are exported so that they can be automatically adapted to the network structure design without manual search; the linkage design realizes cross-stage collaboration between the network design stage (channel configuration) and the training stage (loss function).

[0099] During the training phase, one or more reversible 3D data augmentation transformations are applied to the 3D volume block, including random flipping, random rotation, random scaling, elastic deformation, and brightness / contrast perturbation. The deformation scale parameter of the elastic deformation is specified. With boundary sensing radius By taking the same value, the scale of data augmentation and the scale of loss function boundary awareness are kept consistent.

[0100] During the inference phase, a test-time enhancement (TTA) strategy is employed for the same input volume. Apply A series of reversible geometric transformations (including original orientation, left and right flip, front and back flip, vertical flip, and combinations thereof) are used to input the transformed images into a trained network. After obtaining the prediction result, it is mapped back to the original space through the corresponding inverse transformation, and finally... The average of the predicted probability maps is taken as the final segmentation probability map. :

[0101]

[0102] in For a well-trained network, For the first A type of reversible enhancement transform, The corresponding inverse operation is given.

[0103] S4. Threshold the final segmentation probability map to obtain the three-dimensional segmentation mask of the target structure. .

[0104] A structured clinical quantitative report is generated based on a 3D segmentation mask of the target structure; the report generation steps include:

[0105] Volume normalization: Calculate the voxel volume of the target structure (such as the left hippocampus). Then, based on the total intracranial volume of the subject Normalization is performed to obtain the standardized volume. ;in It is a quantitative measurement in the intermediate stage of imaging and does not directly constitute a judgment on the health status or disease status of the subject. Its clinical significance must be comprehensively evaluated by professional physicians in combination with the individual information of the subject and standardized reference databases.

[0106] Asymmetric quantization: Calculate the normalized volume on the left side separately. With the normalized volume on the right Generate asymmetric index As an imaging quantization parameter output reflecting the relative difference in the volume of the target structure on both sides; HAI The value reflects the relative difference in volume values ​​between the two sides. This parameter is an intermediate measurement result and does not make a direct judgment on the nature of any disease or pathological state. It must be comprehensively analyzed by clinicians in conjunction with the subject's complete clinical data and other assessment methods.

[0107] Intergroup difference assessment: When data from multiple groups of subjects are available, the standardized volume of each group is assessed. Perform between-group statistical significance tests and output the statistical test results (such as effect size and other statistical reference data) to the report to provide imaging measurement data reference for subsequent scientific research or clinical auxiliary analysis. The above statistical test results are intermediate reference data and should not be used directly as the basis for disease diagnosis or cognitive status assessment of individual subjects. Relevant diagnostic conclusions must be made by qualified clinicians based on complete clinical information.

[0108] The parameters in the structured clinical quantitative report output by this invention are all intermediate quantitative measurement data from imaging, providing only objective measurement reference information based on image analysis. The parameters themselves cannot directly lead to a disease diagnosis conclusion. A qualified clinician must make a comprehensive judgment based on the subject's complete clinical data. This invention does not perform or output any disease diagnosis results.

[0109] This invention's method is applicable to brain MRI, small-sample medical image segmentation, and 3D target structure segmentation tasks. It employs the Progressive Channel Compression Efficiency (APCC) design criterion to determine a non-uniformly increasing channel sequence. Channel Adaptive Attention Gating (CAAG) modules are introduced at each skip connection level, allowing the attention projection dimension at each level to adaptively adjust with the channel compression ratio. Training is performed using a boundary adaptive composite loss function, outputting a 3D segmentation mask of the target structure and a structured clinical quantitative report. In small-sample medical image segmentation scenarios, this invention significantly reduces the number of parameters compared to the standard 3D U-Net channel configuration, effectively controlling overfitting gaps. It boasts advantages such as high parameter efficiency, strong generalization ability, high boundary localization accuracy, and excellent segmentation performance for small target structures.

[0110] The APCC criterion of this invention establishes the training sample size for the first time. Formal constraints on channel sequence design elevate channel configuration from empirical design to a verifiable systematic approach; shallow sparsity conditions establish formal constraints on the number of shallow channels and the size of training samples, suppressing overfitting in small sample scenarios; the CNUI index provides a quantifiable evaluation criterion for the non-uniformity of channel configuration.

[0111] The CAAG module of this invention serves as a structured extension of APCC, with a channel compression ratio The system automatically derives the attention projection dimensions at each level using a single input, enabling a verifiable linkage between attention capacity and channel semantic depth. The deep attention projection capacity is significantly improved compared to fixed-dimensional AG, and the overall AG parameter overhead accounts for a very low proportion of the total number of parameters.

[0112] The boundary adaptive composite loss function of the present invention From network structure parameters ( )according to It was determined that the loss function hyperparameters and network structure parameters were automatically adapted across stages, eliminating the dependence on manual search, and maintaining consistency between the data augmentation scale and the loss function boundary awareness scale.

[0113] This invention also provides a brain image target structure segmentation system based on progressive channel compression, including an interconnected memory and a processor. The memory stores a computer program, which, when run by the processor, executes the steps of the method described above. The segmentation system of this invention corresponds to the segmentation method described above and also possesses the advantages described above.

[0114] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0115] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for target structure segmentation in brain images based on progressive channel compression, characterized in that, Including the following steps: Three-dimensional medical image volume blocks of the subject's cranium were acquired and preprocessed to obtain preprocessed three-dimensional medical image volume blocks. ; The channel sequence of the encoder-decoder network is determined according to the progressive channel compression efficiency criterion. ; Set up channel adaptive attention gating modules at the jump connections of each level of the encoder-decoder network; based on the channel sequence The first in Level channel compression ratio Number of input channels Adaptive determination of the first Level of attention projection dimension Then the encoder-decoder network is trained to obtain a trained encoder-decoder network. The encoder-decoder network is constructed using 3D convolutions and consists of four encoding levels, four decoding levels, and a bottleneck layer. Each encoding level contains two 3D convolutional layers, followed by downsampling using 3D max pooling with a stride of 2. Each decoding level contains one upsampling layer and two 3D convolutional layers. Skip connections pass encoder features to the corresponding decoding level. preprocessed 3D medical image volume blocks The input is fed into the trained encoder-decoder network to obtain encoder features. With decoder gate signal , in the first Level of attention projection dimension Calculate the attention coefficient within the spanned adaptive projection space And with attention coefficient Encoder features Perform element-wise weighting to obtain the weighted skip connectivity features. Then feature fusion and progressive upsampling are performed to finally output a three-dimensional segmentation probability map; Thresholding is performed on the 3D segmentation probability map to obtain the 3D segmentation mask. ; No. Level of attention projection dimension The calculation formula is: in, The lower bound is a preset positive integer; Attention coefficient The calculation process is as follows: Encoder features With decoder gate signal Projected onto dimension through linear transformation To obtain the intermediate feature vector : in: , The weight matrix is ​​a linear transformation matrix. For bias terms; Will Applying the ReLU activation function Then convolution and Sigmoid activation function Obtain attention coefficient : in, For the first Level-learnable weight matrix; For bias terms; Adaptive composite loss function at the boundary The encoder-decoder network is trained, where the boundary adaptive composite loss function is used. The calculation formula is: in These are the weighting coefficients; For volume overlap rate Similarity loss function; This is the weighted cross-entropy loss function.

2. The brain image target structure segmentation method based on progressive channel compression according to claim 1, characterized in that, After determining the channel sequence Then, the channel non-uniformity index is further calculated. To quantify the non-uniformity of channel sequences; when The channel sequence is then determined to satisfy the condition of non-uniform growth. Among them, the channel non-uniformity index The calculation formula is: in These represent the maximum and minimum values ​​of the overall growth ratio. The growth rate is the arithmetic mean. For the first The growth rate of the first level, This represents the pre-bottleneck suppression ratio.

3. The brain image target structure segmentation method based on progressive channel compression according to claim 1, characterized in that, The calculation formula is: in For voxel indexing; For category; Category weights; For real labels, To predict probabilities; voxels The boundary-aware weight at a given location is calculated using the following formula: in voxels The Euclidean distance transformation value to the nearest target boundary; This refers to the boundary sensing intensity coefficient. The radius of the boundary perception.

4. The brain image target structure segmentation method based on progressive channel compression according to claim 3, characterized in that, Boundary sensing radius Pre-bottleneck suppression ratio of encoder-decoder network With volume blocks Side length Confirmed, the specific calculation formula is as follows: 。 5. The brain image target structure segmentation method based on progressive channel compression according to claim 1, 2, or 3, characterized in that, Based on the 3D segmentation mask of the target structure, a structured clinical quantitative report is generated, specifically as follows: Volume normalization: Calculating the voxel volume of the target structure Then, based on the total intracranial volume of the subject Normalization is performed to obtain the standardized volume. ; Asymmetric quantization: Calculate the normalized volume on the left side separately. With the normalized volume on the right Generate asymmetric index As an imaging quantitative parameter reflecting the relative difference in the volume of the target structure on both sides; Intergroup difference assessment: When data from multiple groups of subjects are available, the standardized volume of each group is assessed. Perform a statistical significance test.

6. The brain image target structure segmentation method based on progressive channel compression according to claim 1, 2, or 3, characterized in that, The preprocessing of the three-dimensional medical image volume blocks of the brain includes sequentially performing skull stripping, affine space registration to a standard brain template, z-score intensity normalization, and region of interest (ROI) cropping.

7. A brain image target structure segmentation system based on progressive channel compression, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-6.

Images