Lumbar plexus nerve root segmentation method, medium and device based on magnetic resonance diffusion tensor imaging

Abstract

The present invention discloses a lumbar plexus nerve root segmentation method, medium and device based on magnetic resonance diffusion tensor imaging. First, lumbar plexus nerve DTI imaging DICOM data is obtained and stored in NRRD format; then, the region of interest is extracted from the DTI data, and its region of interest annotation data is converted into a grayscale image in PNG format for storage; finally, the DTI data and its region of interest annotation data are input into a deep learning segmentation model for training, and the lumbar plexus nerve root is segmented and predicted by the trained segmentation model. The present invention utilizes a deep learning segmentation model to learn richer image information, adopts a multi-head self-attention mechanism based on windows and moving windows to enhance the model's global feature learning ability at different resolution scales, and constructs a SwinDW fusion module by combining a convolution module and a wavelet feature extraction module to better fuse global and low-frequency features; the present invention also uses a multi-scale feature adaptive fusion module to fuse the feature map of the current encoding module with the feature map of the encoding module at different resolution scales, so as to learn the feature information of the encoding modules at different scales, alleviate the spatial information loss caused by encoder downsampling, and thus better improve the segmentation effect of the lumbar plexus nerve root.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image post-processing applications, and in particular relates to a lumbar sacral plexus nerve root segmentation method, medium and equipment based on magnetic resonance diffusion tensor imaging. Background Art

[0002] The lumbar plexus is a complex neural network composed of motor nerves and sensory nerves. It is an important channel for controlling and transmitting sensations to the pelvis and lower limbs. Lumbar plexus nerve root compression is the most common peripheral neuropathy in clinical practice, and the incidence rate gradually increases with age. The clinical symptoms mainly include motor weakness, sensory loss, decreased muscle strength, pain, intermittent claudication and other functional disorders in the corresponding nerve innervation area.

[0003] For lumbar plexus nerve root compression, the first step in preoperative diagnosis is to locate the affected nerve segments and determine the responsible nerve roots. Magnetic resonance imaging (MRI), as a non-invasive imaging method, has been widely used in clinical diagnosis and evaluation of lumbar spine-related diseases. Routine MRI lumbar spine examination can reveal morphological changes that lead to nerve root compression, such as spinal stenosis, intervertebral disc herniation, and thickening of the ligamentum flavum. However, indirectly determining nerve root compression through morphological changes and the relative relationship between nerve roots, diagnosing nerve root compression injuries and determining the responsible nerve roots, has a certain false positive rate and false negative rate, and cannot accurately locate the responsible nerve roots.

[0004] With the development and progress of functional magnetic resonance imaging, diffusion tensor imaging (DTI) technology can be used to establish diffusion tracing images in combination with dedicated post-processing software for DTI data, thus realizing diffusion tensor tractography (DTT). Compared with conventional magnetic resonance imaging, DTT can reconstruct the nerve fiber bundles of the lumbar and sacral nerve roots, directly provide anatomical information of the nerve root course and subtle changes in nerve structure, and can be overlapped and fused with conventional magnetic resonance images of the lumbar and sacral region to more intuitively display the site of the lesion. For damaged and degenerated nerve roots, it can also provide a rough location of the site of nerve lesions, providing more information for the formulation of surgical plans for clinical surgical treatment and pain nerve blocks.

[0005] At present, clinicians use special processing software to outline nerve root target areas and reconstruct nerve fiber bundles, which not only consumes a lot of time, but also places higher requirements on the experience and level of doctors. Therefore, those skilled in the art are in urgent need of establishing a lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging to establish an efficient and accurate imaging diagnosis and clinical decision-making system. Summary of the invention

[0006] In view of the deficiencies in the prior art, the present invention provides a lumbar sacral plexus nerve root segmentation method, medium and device based on magnetic resonance diffusion tensor imaging. By using a deep learning-based segmentation model, the relevant position and shape of the lumbar sacral plexus nerve roots can be intuitively displayed on DTI, providing technical support for clinicians to carry out diagnostic work.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging is characterized by comprising:

[0009] Collect DTI imaging data of patients;

[0010] Extract regions of interest from DTI imaging data, and process DTI imaging data and region of interest annotation data;

[0011] The processed DTI imaging data and region of interest annotation data are randomly divided into training sets and test sets, and then the divided training sets and test sets are input into the deep learning segmentation model for training;

[0012] Use the trained deep learning segmentation model to predict the segmentation of lumbar plexus nerve roots.

[0013] To optimize the above technical solutions, the specific measures taken also include:

[0014] Furthermore, the acquisition of DTI imaging data of the patient is specifically as follows:

[0015] The patient's DTI image DICOM data was collected and stored in NRRD format using the 3D Slicer tool.

[0016] Furthermore, the processing of the DTI imaging data and the region of interest annotation data is specifically as follows:

[0017] The DTI imaging data and region of interest annotation data were converted into grayscale images and normalized, and then stored in PNG format.

[0018] Furthermore, the deep learning segmentation model includes an image block division module, a wavelet transform module, an encoder, a multi-scale feature adaptive fusion module, a bottleneck module, a decoder and a linear mapping layer;

[0019] The image block division module divides the DTI imaging data into a plurality of non-overlapping image blocks and converts them into continuous sequence features, which are then input into the encoder;

[0020] The wavelet transform module extracts the low-frequency information of the DTI imaging data and converts it into low-frequency features of different scales, which are then input into the encoder;

[0021] The encoder is composed of four groups of continuous encoding modules and one LN layer in sequence. The four groups of encoding modules correspond to four different resolution scales and are used to obtain feature maps of different scales according to the input sequence features and low-frequency features. During the learning process, each group of encoding modules fuses the low-frequency features of the corresponding scale, and inputs the obtained feature map into the multi-scale feature adaptive fusion module for feature information interaction;

[0022] The decoder is composed of four groups of continuous decoding modules and one image block expansion layer in sequence. The four groups of decoding modules correspond to the four groups of encoding modules in scale one by one, and are used to perform upsampling learning according to the input feature map. During the learning process, the feature maps of different scales obtained by the interaction of the multi-scale feature adaptive fusion module are respectively input into the decoding modules of the corresponding scales for fusion. The feature map output by the fourth decoding module is passed through the linear mapping layer to obtain the segmentation result;

[0023] The LN layer of the encoder is connected to the first group of decoding modules through a bottleneck module, and the bottleneck module is composed of two groups of consecutive Swin Transformer modules.

[0024] Furthermore, in the encoder, the first group of encoding modules is sequentially composed of a linear embedding layer and a SwinDW fusion module, and the other three groups of encoding modules are sequentially composed of a image block merging layer and a SwinDW fusion module;

[0025] In the decoder, the first group of decoding modules is composed of a layer of LN layer and a Swin-Conv fusion module in sequence, and the other three groups of decoding modules are composed of a layer of image block expansion layer and a Swin-Conv fusion module in sequence;

[0026] The SwinDW fusion module is composed of a parallel Swin Transformer layer, a downsampling Conv layer, and a wavelet feature channel expansion layer, which perform three-way feature learning and then aggregate them into a feature fusion layer;

[0027] The Swin-Conv fusion module is composed of a Swin Transformer layer and an upsampling Conv layer in sequence.

[0028] Furthermore, in the SwinDW fusion module, the Swin Transformer layer is composed of two groups of continuous sub-layers s1 and s2, wherein s1 is composed of layer normalization LN, window-based multi-head self-attention W-MSA, LN and multi-layer perceptron MLP in sequence, and s2 is composed of LN, moving window-based multi-head self-attention SW-MSA, LN and MLP in sequence, and the feature transmission process between s1 and s2 is as follows:

[0029]

[0030] In the formula, F l-1 is the output feature of the l-1th encoding module, and Respectively represent the output features of W-MSA and SW-MSA in the l-th encoding module, and denote the output features of MLP in s1 and s2, respectively. W-MSA(·), LN(·), MLP(·), and SW-MSA(·) denote the processing of window-based multi-head self-attention, layer normalization, multi-layer perceptron, and multi-head self-attention, respectively.

[0031] The downsampling Conv layer consists of two groups of sub-layers c1 and c2 with the same structure, including downsampling convolution Conv2d, batch normalization BN and activation function ReLU. The feature transmission process is as follows:

[0032]

[0033] In the formula, and Respectively represent the output features of sub-layer c1 and sub-layer c2 in the l-th encoding module;

[0034] The wavelet feature channel expansion layer is composed of downsampling convolution Conv1d, BN and ReLU in sequence, which is used to expand the channel of the low-frequency features extracted by the wavelet transform module to and The same number of channels, thus obtaining low-frequency features

[0035] The output features of the Swin Transformer layer, the downsampling Conv layer and the wavelet feature channel expansion layer are fused by three-way feature weighting to obtain the output feature F of the lth encoding module l , the specific process is as follows:

[0036]

[0037] Where W s ,W c ,Ww represents a learnable hyperparameter, σ(·) represents the activation function sigmoid, and λ s ,λ c ,λ w Respectively The feature weight of , ⊙ represents the matrix sum.

[0038] Furthermore, the feature transmission process in the multi-scale feature adaptive fusion module is as follows:

[0039] F all =Concat(LN(F1),…,LN(F l ),…,LN(F4)),

[0040] q l =LN(F l )W Q ,k=F all W K ,v=F all W V ,

[0041]

[0042] In the formula, F l represents the output feature of the lth encoding module, F all W represents the output features of all encoding modules that are first layer-normalized LN(·) and then channel-level concatenated Concat(·) to obtain the output features. Q ,W K ,W V represents a learnable hyperparameter, q l Indicates F l The transformation features of F all The transformation features of , AF-MHA(·) represents adaptive fusion multi-head attention, and O l They represent the attention calculation features and final output features of the lth encoding module respectively.

[0043] Furthermore, the processing flow of the adaptive fusion multi-head attention is as follows:

[0044] index l =topk(softmax(mm(q l ,k T ))),

[0045] k l-g ,v l-g =gather(index l ,k,v),

[0046]

[0047] O AF-MHA =Linear(Att o ),

[0048] In the formula, mm(·) represents matrix multiplication, softmax(·) function represents the calculation of probability distribution, and topk(·) function represents the index number of the largest k probability values ​​in the calculation probability distribution. l , k T , They represent k, k respectively. l-g The transposed matrix of l Applied to k and v to extract feature information, conv(·) represents the deep convolutional layer, Att o represents the output feature of the attention operation, Linear(·) represents the linear layer, O AF-MHA represents the output characteristics of AF-MHA(·).

[0049] Accordingly, the present invention also proposes a computer-readable storage medium storing a computer program, characterized in that the computer program enables a computer to execute the lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging as described above.

[0050] Correspondingly, the present invention also proposes an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging as described above is implemented.

[0051] The beneficial effects of the present invention are: the segmentation method of the present invention utilizes a deep learning segmentation model to learn richer image information; a multi-head self-attention mechanism based on windows and moving windows is adopted to enhance the model's ability to learn global features at different resolution scales, and a SwinDW fusion module is constructed by combining a convolution module and a wavelet feature extraction module to better fuse global and low-frequency features; the present invention also uses a multi-scale feature adaptive fusion module to perform attention fusion on the feature map of the current encoding module with the feature map of the encoding module at different resolution scales, so as to learn the feature information of encoding modules at different scales, alleviate the spatial information loss caused by encoder downsampling, thereby better improving the segmentation effect of the lumbar sacral plexus nerve roots, helping clinicians to understand the relative position and organ status of the patient's lumbar sacral plexus nerve roots before surgery, providing technical support for clinicians to carry out diagnostic work, and reducing the difficulty and risk of clinicians' surgery. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The figure is a flow chart of the lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging of the present invention.

[0053] Figure 2 Schematic diagram of the network architecture of the deep learning segmentation model of the present invention.

[0054] Figure 3 Schematic diagram of the structure of the image patch merging layer in the deep learning segmentation model.

[0055] Figure 4 Schematic diagram of the structure of the SwinDW fusion module in the deep learning segmentation model.

[0056] Figure 5 Schematic diagram of the structure of adaptive fusion multi-head attention in the multi-scale feature adaptive fusion module.

[0057] Figure 6 It is a schematic diagram of the segmentation result of the present invention. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0059] In one embodiment, if Figure 1 As shown, the present invention proposes a lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging, comprising the following steps:

[0060] (1) DTI imaging data acquisition and storage: DTI imaging data of patients were collected and stored in NRRD format using the 3D Slicer tool;

[0061] (2) Region of interest extraction: Extract the region of interest from the DTI imaging data, then convert the DTI and region of interest annotation data into grayscale images and perform normalization, and then store them in PNG format;

[0062] (3) Deep learning segmentation model training: The DTI and its annotated data obtained in step (2) are randomly divided into training set and test set, and then the data is input into the deep learning segmentation model for training;

[0063] (4) Lumbar plexus nerve root segmentation prediction: Use the segmentation model trained in step (3) to perform lumbar plexus nerve root segmentation prediction.

[0064] like Figure 2As shown, the deep learning segmentation model in step (3) includes an image block partitioning module (PatchPartition), a wavelet transform module (Wavelet Transform), an encoder (Encoder), an image block merging layer (PatchMerging), a SwinDW fusion module (SwinDW Fusion), a linear embedding layer (Linear Embedding), a layer normalization (Layer Normalization, LN), a bottleneck module (Bottleneck), a Swin Transformer module, a multi-scale feature adaptive fusion module (Multi-scale Feature Adaptive Fusion), a decoder (Decoder), a Swin-Conv fusion module, an image block expansion layer (Patch Expanding) and a linear projection layer (Linear Projection).

[0065] The encoder consists of four groups of continuous encoding modules and one LN layer. The first group of encoding modules consists of a linear embedding layer and a SwinDW fusion module, and the other three encoding modules consist of an image block merging layer and a SwinDW fusion module. The decoder consists of four groups of continuous decoding modules and an image block expansion layer. The first group of decoding modules consists of an LN layer and a Swin-Conv fusion module, and the other three decoding modules consist of an image block expansion layer and a Swin-Conv fusion module. The bottleneck module is located between the bottom layer of the encoder and the first layer of the decoder, and consists of two groups of continuous SwinTransformer modules. The encoder and decoder are enhanced and transmitted through a multi-scale feature adaptive fusion module.

[0066] The normalized DTI image is first passed to the image block division module and the wavelet transform module. The image block division module can divide a complete two-dimensional plane image into multiple non-overlapping sub-image blocks according to the image block size and convert it into continuous sequence features, while the wavelet transform module extracts the low-frequency information of the original image and converts it into low-frequency features of different resolution scales. Then, the converted sequence features and low-frequency features are passed to the encoder. The encoder contains four encoding modules with different resolution scales for downsampling feature learning. Each encoding module can obtain feature maps of different scales, and each encoding module will fuse the wavelet transform features of the corresponding scale to learn the low-frequency information of the image at different scales; the last layer of the encoder is connected to the first layer of the decoder through a bottleneck module to learn and transmit the deep feature map of the encoder; then, the feature map of each encoding module in the encoder is exchanged with the feature information through the multi-scale feature adaptive fusion module, and then fused with the feature map of the corresponding scale in the decoder; finally, the output feature of the last decoding module in the decoder is passed to the image block expansion layer and the segmentation result of the network is obtained through the linear mapping layer.

[0067] The input data of the deep learning segmentation model in step (3) is subjected to downsampling learning by using an encoder after the sequence features and low-frequency features with 4 times resolution reduction are extracted through the image block division module and the wavelet transform module; the encoder calculates the corresponding feature map through four sets of encoding modules with different resolution scales. Except for the first encoding module that does not reduce the resolution scale, the other three sets of encoding modules are all subjected to the built-in image block merging layer (such as Figure 3 The resolution scale is reduced to 8, 16, and 32 times, and then three-way feature learning is performed through the SwinDW fusion module consisting of the Swin Transformer layer, the downsampling Conv layer (DownSampling Conv), the wavelet feature channel expansion layer (Wt-Channel Expanding), and the feature fusion layer (FeatureFusion) to enhance the spatial feature information at this scale.

[0068] like Figure 4 As shown in the figure, the Swin Transformer layer is composed of two sets of consecutive sub-layers s1 and s2, where s1 contains layer normalization (LN), window-based multi-head self-attention (W-MSA), LN and multi-layer perceptron (MLP), and s2 contains LN, moving window-based multi-head self-attention (SW-MSA), LN and MLP. The feature transmission process between s1 and s2 is shown as follows:

[0069]

[0070] In the formula, F l-1is the feature output of the l-1th encoding module, and They represent the feature outputs of W-MSA and SW-MSA in the l-th encoding module, respectively. and They represent the feature outputs of MLP in s1 and s2 respectively; within s1 and s2, the output features of every two layers are fused with the previous feature maps through residual connections to enhance the global spatial information learning capability within the module.

[0071] The downsampling Conv layer is composed of two groups of sub-layers c1 and c2 with the same structure, including downsampling convolution (Conv2d), batch normalization (BN) and activation function (ReLU). Its feature transmission process is shown as follows:

[0072]

[0073] In the formula, and They represent the feature outputs of sub-layer c1 and sub-layer c2 in the lth encoding module respectively.

[0074] The wavelet feature channel expansion layer is composed of a set of Conv1d, BN and ReLU, which is used to expand the low-frequency feature channels extracted by the wavelet transform module to sw and c sw The same number of channels, thus obtaining the low-frequency features under this dimension

[0075] Afterwards, the feature outputs of the Swin Transformer layer, the downsampling Conv layer, and the wavelet feature channel expansion layer are weightedly fused to obtain the output feature map F of the encoding module. l The specific process is as follows:

[0076]

[0077] Where W s ,W c ,W w represents a learnable hyperparameter, σ(·) represents the activation function sigmoid, and λ s ,λ c ,λ w Respectively The feature weights of ⊙ represent the matrix sum; the bottleneck module after the encoder uses two consecutive Swin Transformer fusion modules with 32 times reduced resolution to learn deep features.

[0078] The decoding module of the decoder in step (3) has the same number of modules and feature map dimensions as the encoding module of the encoder. Except for the first group of decoding modules consisting of a layer of LN layer and a Swin-Conv fusion module, the other three decoding modules are all composed of a layer of image block expansion layer and a Swin-Conv fusion module, where the Swin-Conv fusion module consists of a SwinTransformer layer and an up sampling Conv layer. Conv), and the same number of layers as the SwinTransformer layer and downsampling Conv layer of the encoding module; when upsampling the decoder, the input feature map of each decoding module must reach the same dimension as the corresponding encoding module feature map; for the first decoding module, since the bottleneck module does not change the resolution scale of the feature map, only one LN layer is used for feature transmission; for the other three groups of decoding modules, their resolution scales are enlarged to 16 times, 8 times and 4 times respectively through the built-in image block expansion module to make them reach the same dimension as the corresponding encoding module feature map; under the feature map of the same dimension, the output feature map of each encoding module needs to learn the feature information of encoding modules of different scales through the multi-scale feature adaptive fusion module before merging with the feature map of the corresponding decoding module, so as to alleviate the spatial information loss caused by downsampling.

[0079] The feature transmission process in the multi-scale feature adaptive fusion module is as follows:

[0080] F all =Concat(LN(F1),…,LN(F l ),…,LN(F4)),

[0081] q l =LN(F l )W Q ,k=F all W K ,v=F all W V ,

[0082]

[0083] In the formula, F l represents the output feature of the lth encoding module, F all It represents the output features of all encoding modules that are first layer-normalized and then channel-level concatenated to obtain the output features. Q ,W K ,W V represents a learnable hyperparameter, q l Indicates F l The transformation features of F allThe transformation features of , AF-MHA(·) represents adaptive fusion multi-head attention, and O l They represent the attention calculation features and final output features of the lth encoding module respectively.

[0084] like Figure 5 As shown in Figure 2, in the adaptive fusion multi-head attention, it is necessary to first fuse q through matrix multiplication (mm). l and k, and then use the softmax(·) function to calculate its probability distribution, and then use the topk(·) function to calculate the index number index of the top k probability values ​​in the probability distribution l , and then use the gather(·) method to get the index number index l Applied to the initial features k and v to extract their important feature information, and the enhanced feature k l-g ,v l-g With the initial feature q l ,v performs attention operation, and finally obtains the output feature O of AF-MHA(·) through a linear layer. AF-MHA , so that F l Adaptively integrate important information of output features of other encoding modules. The specific process is as follows:

[0085] index l =topk(softmax(mm(q l ,k T ))),

[0086] k l-g ,v l-g =gather(index l ,k,v),

[0087]

[0088] O AF-MHA =Linear(Att o ),

[0089] In the formula, k T , They represent k, k respectively. l-g The transposed matrix of , conv(·) represents the depth convolution layer, Att o represents the output feature of the attention operation; finally, the feature map is transmitted between the decoding modules and the feature map of the last decoding module is used to obtain the segmentation result of the network through the linear mapping layer (such as Figure 6 shown).

[0090] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging as described in the first embodiment.

[0091] In another embodiment, the present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the lumbar sacral plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging as described in Example 1 is implemented.

[0092] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0093] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0094] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging, characterized in that: include: Collect DTI imaging data of patients; Extract regions of interest from DTI imaging data, and process DTI imaging data and region of interest annotation data; The processed DTI imaging data and region of interest annotation data are randomly divided into training sets and test sets, and then the divided training sets and test sets are input into the deep learning segmentation model for training; The deep learning segmentation model includes an image block division module, a wavelet transform module, an encoder, a multi-scale feature adaptive fusion module, a bottleneck module, a decoder and a linear mapping layer; The image block division module divides the DTI imaging data into a plurality of non-overlapping image blocks and converts them into continuous sequence features, which are then input into the encoder; The wavelet transform module extracts the low-frequency information of the DTI imaging data and converts it into low-frequency features of different scales, which are then input into the encoder; The encoder is composed of four groups of continuous encoding modules and one LN layer in sequence. The four groups of encoding modules correspond to four different resolution scales and are used to obtain feature maps of different scales according to the input sequence features and low-frequency features. During the learning process, each group of encoding modules fuses the low-frequency features of the corresponding scale, and inputs the obtained feature map into the multi-scale feature adaptive fusion module for feature information interaction; In the encoder, the first group of encoding modules is composed of a linear embedding layer and a SwinDW fusion module in sequence, and the other three groups of encoding modules are composed of an image block merging layer and a SwinDW fusion module in sequence; the SwinDW fusion module is composed of a parallel Swin Transformer layer, a downsampling Conv layer and a wavelet feature channel expansion layer, which are aggregated to a feature fusion layer after three-way feature learning; in the SwinDW fusion module, the Swin Transformer layer is composed of two groups of continuous sub-layers s1 and s2, wherein s1 is composed of layer normalization LN, window-based multi-head self-attention W-MSA, LN and multi-layer perceptron MLP in sequence, and s2 is composed of LN, moving window-based multi-head self-attention SW-MSA, LN and MLP in sequence, and the feature transmission process between s1 and s2 is as follows: In the formula, F l-1 is the output feature of the l-1th encoding module, and Respectively represent the output features of W-MSA and SW-MSA in the l-th encoding module, and denote the output features of MLP in s1 and s2, respectively. W-MSA(·), LN(·), MLP(·), and SW-MSA(·) denote the processing of window-based multi-head self-attention, layer normalization, multi-layer perceptron, and multi-head self-attention, respectively. The downsampling Conv layer consists of two groups of sub-layers c1 and c2 with the same structure, including downsampling convolution Conv2d, batch normalization BN and activation function ReLU. The feature transmission process is as follows: In the formula, and Respectively represent the output features of sub-layer c1 and sub-layer c2 in the l-th encoding module; The wavelet feature channel expansion layer is composed of downsampling convolution Conv1d, BN and ReLU in sequence, which is used to expand the channel of the low-frequency features extracted by the wavelet transform module to and The same number of channels, thus obtaining low-frequency features The output features of the Swin Transformer layer, the downsampling Conv layer and the wavelet feature channel expansion layer are subjected to three-way feature weighted fusion through the feature fusion layer to obtain the output feature F of the lth encoding module l , the specific process is as follows: Where W s ,W c ,W w represents a learnable hyperparameter, σ(·) represents the activation function sigmoid, and λ s ,λ c ,λ w Respectively The feature weight of , ⊙ represents the matrix sum; The decoder is composed of four groups of continuous decoding modules and a layer of image block expansion layer in sequence. The four groups of decoding modules correspond to the four groups of encoding modules in scale one by one, and are used to perform upsampling learning according to the input feature map. During the learning process, the feature maps of different scales obtained by the interaction of the multi-scale feature adaptive fusion module are respectively input into the decoding modules of the corresponding scales for fusion. The feature map output by the fourth decoding module is obtained by the linear mapping layer to obtain the segmentation result; the feature transmission process in the multi-scale feature adaptive fusion module is as follows: F all =Concat(LN(F1),…,LN(F l ),…,LN(F4)), q l =LN(F l )W Q ,k=F all W K ,v=F all W V , In the formula, F l represents the output feature of the lth encoding module, F all W represents the output features of all encoding modules that are first layer-normalized LN(·) and then channel-level concatenated Concat(·) to obtain the output features. Q ,W K ,W V represents a learnable hyperparameter, q l Indicates F l The transformation features of F all The transformation features of , AF-MHA(·) represents adaptive fusion multi-head attention, and O l Respectively represent the attention calculation features and final output features of the lth encoding module; The processing flow of the adaptive fusion multi-head attention is as follows: index l =topk(softmax(mm(q l ,k T ))), k l-g ,v l-g =gather(index l ,k,v), O AF-MHA =Linear(To o ), In the formula, mm(·) represents matrix multiplication, softmax(·) function represents the calculation of probability distribution, and topk(·) function represents the index number of the largest k probability values ​​in the calculation probability distribution. l , k T , They represent k, k respectively. l-g The transposed matrix of l Applied to k and v to extract feature information, conv(·) represents the depth convolution layer, Att o represents the output feature of the attention operation, Linear(·) represents the linear layer, O AF-MHA represents the output characteristics of AF-MHA(·); In the decoder, the first group of decoding modules is composed of a LN layer and a Swin-Conv fusion module in sequence, and the other three groups of decoding modules are composed of a block expansion layer and a Swin-Conv fusion module in sequence; the Swin-Conv fusion module is composed of a Swin Transformer layer and an upsampling Conv layer in sequence; The LN layer of the encoder is connected to the first set of decoding modules through a bottleneck module, and the bottleneck module is composed of two sets of consecutive Swin Transformer modules; Use the trained deep learning segmentation model to predict the segmentation of lumbar plexus nerve roots.

2. The method for segmenting lumbar plexus nerve roots based on magnetic resonance diffusion tensor imaging according to claim 1, characterized in that: The acquisition of the patient's DTI imaging data is specifically as follows: The patient's DTI image DICOM data was collected and stored in NRRD format using the 3D Slicer tool.

3. The method for segmenting lumbar plexus nerve roots based on magnetic resonance diffusion tensor imaging according to claim 1, characterized in that: The processing of the DTI imaging data and the region of interest annotation data is specifically as follows: The DTI imaging data and region of interest annotation data were converted into grayscale images and normalized, and then stored in PNG format.

4. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the lumbar plexus nerve root segmentation method based on magnetic resonance diffusion tensor imaging as described in any one of claims 1 to 3.

5. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for segmenting lumbar plexus nerve roots based on magnetic resonance diffusion tensor imaging as described in any one of claims 1 to 3 is implemented.

Images