Lumbar intervertebral disc segmentation method, device and equipment based on improved U-Net network
By improving the U-Net network, Mamba's dual attention fusion module and context enhancement module are used to solve the problems of insufficient segmentation accuracy and high computational complexity in lumbar disc segmentation, and efficient lumbar disc segmentation is achieved.
Patent Information
- Application Number
- CN202510454369.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-11
- Publication Date
- 2025-08-01
AI Technical Summary
The prior art has problems of insufficient segmentation accuracy and high computational complexity in lumbar disc segmentation, especially the semantic gap caused by jump connections based on UNet models and the high computational complexity of the Transformer model.
Using an improved U-Net network, a feature map is passed between the encoder and the decoder through a Mamba-based dual attention fusion module, and a context enhancement module is set up at the bottom of the encoder to enhance the underlying feature extraction capability, and a combination of Dice loss and binary cross entropy loss functions are trained.
It improves the segmentation accuracy of the lumbar disc and reduces the computational complexity, achieving a more efficient medical image segmentation effect.
Smart Images
Figure CN120411501A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of lumbar intervertebral disc segmentation, and specifically to a method, device and equipment for lumbar intervertebral disc segmentation based on an improved U-Net network. Background Art
[0002] Lumbar disc herniation (LDH), as a chronic disease, brings a heavy economic burden to patients. Its typical symptoms are low back pain, and at the same time, a series of neurological symptoms such as paresthesia, anesthesia, and muscle weakness may occur. The pathological manifestation of LDH is that the nucleus pulposus tissue in the center of the intervertebral disc breaks through the normal boundary, and then protrusion, extrusion or even free state occurs, compressing the surrounding nerve roots. During the clinical diagnosis process, it mainly relies on MRI or CT images. Radiologists evaluate the condition by observing the shape and position of the lumbar intervertebral disc, so as to formulate corresponding treatment plans. Radiologists need to carefully view each image to judge the pathological conditions of the lumbar intervertebral disc. This process is both time-consuming and laborious, and the diagnostic results are easily affected by observation errors. Some doctors with insufficient experience may misjudge, and then wrongly judge the type of lumbar protrusion, resulting in a decline in the treatment effect. Therefore, achieving rapid and accurate detection and delineation of the lumbar intervertebral disc is of great significance for improving the prognosis of patients and optimizing the treatment plan.
[0003] In past studies, some scholars have been committed to designing computer vision-based automatic auxiliary diagnosis methods for lumbar protrusion. For example, the UNet model is used for lumbar intervertebral disc segmentation. However, there are defects in the skip connections of this model, resulting in a semantic gap between the encoder and the decoder, which in turn affects the segmentation accuracy. In order to narrow the semantic gap, UNet++ introduced dense skip connections, enhancing the ability to capture detailed information, but at the cost of a substantial increase in the number of parameters. UDTransNet, which emerged later, improved the skip connections based on Transformer, but this also greatly increased the computational complexity and had poor performance in lumbar intervertebral disc segmentation.
[0004] In view of the above problems existing in the lumbar intervertebral disc segmentation task, the present invention proposes a method for lumbar intervertebral disc segmentation based on an improved U-Net network. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method for lumbar intervertebral disc segmentation based on an improved U-Net network, which can achieve accurate segmentation of the lumbar intervertebral disc and has low computational complexity.
[0006] To solve the above technical problems, the technical solution of the present invention is: A method for lumbar intervertebral disc segmentation based on an improved U-Net network, comprising:
[0007] Obtain a dataset, where the dataset includes multiple labeled lumbar disc images;
[0008] Build an improved U-Net network, where feature maps are transmitted between the encoder and decoder of the improved U-Net network through a dual attention fusion module based on Mamba; a context enhancement module is set at the bottom layer of the encoder to enhance the bottom layer feature extraction ability;
[0009] Train the improved U-Net network using the dataset to obtain a trained improved U-Net network;
[0010] Use the trained improved U-Net network to segment the lumbar disc image to be segmented.
[0011] Furthermore, the obtaining of the dataset specifically includes:
[0012] Collect the MRI images of multiple patients with lumbar disc herniation;
[0013] Use the pre-trained nnU-Net network to perform preliminary segmentation on specific lumbar segments in the MRI images. After correction and annotation, slice them to obtain the dataset.
[0014] Furthermore, the working process of the dual attention fusion module based on Mamba is expressed as:
[0015] G'1, G'2, G'3, G'4 = Split(M(Norm(XCat(G1, G2, G3, G4))))
[0016] S1, S2, S3, S4 = Split([M(Norm([YCat(G'1, G'2, G'3, G'4) T ))] T ) + [G1, G2, G3, G4]
[0017] p i = s i + MLP(Norm(s i ))
[0018] In the formula, G i (i = 1, 2, 3, 4) are features with the same size after Patch Embedding of the feature maps of the corresponding layers of the encoder; XCat(·) represents a feature concatenation operation with a dimension of one, Norm(·) represents a regularization operation, M(·) represents processing using the Mamba module, Split(·) represents an operation of equally dividing the feature map; YCat(·) represents a feature concatenation operation with a dimension of two, [·] T represents a transpose operation; MLP(·) represents a multi-layer perceptron processing operation; si (i = 1, 2, 3, 4) are feature maps connected to the corresponding layers of the decoder.
[0019] Furthermore, the working process of the context enhancement module is expressed as:
[0020] K = f5 + Reshape(M(Norm(Extend(f5)))))
[0021] In the formula, f5 represents the underlying feature, Extend(·) represents the operation of stretching the feature into a sequence, ReShape(·) represents the operation of restoring the feature map to the dimension of the underlying feature, and K represents the output of the context enhancement module.
[0022] Furthermore, during the process of training the improved U-Net network using the dataset, the loss function is:
[0023] L seg = L Dice + L BCE
[0024] In the formula, L seg represents the total segmentation loss, L Dice represents the Dice loss, and L BCE represents the binary cross-entropy loss.
[0025] The present invention also relates to a lumbar intervertebral disc segmentation device based on an improved U-Net network, including:
[0026] An acquisition module for acquiring a dataset, where the dataset includes multiple labeled lumbar intervertebral disc images;
[0027] A construction module for constructing an improved U-Net network, where the feature maps are transmitted between the encoder and the decoder of the improved U-Net network through a dual attention fusion module based on Mamba; a context enhancement module is set at the bottom layer of the encoder to enhance the underlying feature extraction ability;
[0028] A training module for training the improved U-Net network using the dataset to obtain a trained improved U-Net network;
[0029] A segmentation module for segmenting the lumbar intervertebral disc image to be segmented using the trained improved U-Net network.
[0030] The present invention also relates to a device, including:
[0031] A memory for storing a computer program; <l
[0032] A processor for executing the computer program, and when the computer program is executed by the processor, it implements the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network.
[0033] The present invention also relates to a readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network.
[0034] The present invention also relates to a computer program product, including a computer program, and when the computer program is executed by a processor, it implements the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network.
[0035] After adopting the above technical solutions, the present invention has the following beneficial effects:
[0036] 1. The dual attention fusion module based on Mamba innovatively integrates Mamba technology with efficient spatial and channel attention mechanisms, aiming to capture semantic features at different scales. Reasonably integrating this information helps to improve the effect of complex medical image segmentation tasks, thereby improving the segmentation accuracy of lumbar intervertebral discs.
[0037] 2. The context enhancement module can further learn the features extracted by the shallow network, enhance the underlying feature extraction ability, and further improve the segmentation ability. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a flowchart of the lumbar intervertebral disc segmentation method based on the improved U-Net network of the present invention;
[0039] Figure 2 It is a framework diagram of the improved U-Net network of the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0040] In order to make the content of the present invention easier to be clearly understood, the present invention will be further described in detail below according to specific embodiments and in conjunction with the accompanying drawings.
[0041] As Figure 1 and Figure 2 shown, a lumbar intervertebral disc segmentation method based on an improved U-Net network includes:
[0042] Step S1, obtaining a data set, where the data set includes multiple labeled lumbar intervertebral disc images;
[0043] Step S2, building an improved U-Net network, and transmitting feature maps between the encoder and decoder of the improved U-Net network through a dual attention fusion module based on Mamba; a context enhancement module is set at the bottom layer of the encoder to enhance the underlying feature extraction ability;
[0044] Step S3: Train the improved U-Net network using the dataset to obtain a trained improved U-Net network;
[0045] Step S4: Use the trained improved U-Net network to segment the lumbar disc image to be segmented.
[0046] Specifically, in this embodiment, aiming at the semantic gap problem existing in the skip connections of the traditional U-Net network, a dual attention fusion module based on Mamba is designed to effectively fuse the feature maps of different resolutions of the encoder, thereby optimizing the information transmission between the encoder and the decoder, finally narrowing the semantic gap, not only greatly reducing the computational overhead of the model, but also improving the segmentation accuracy, and then comprehensively optimizing the model performance; in addition, this embodiment also designs a context enhancement module to enhance the underlying feature extraction ability and further enhance the segmentation ability of the network.
[0047] In this embodiment, in step S1, obtaining the dataset specifically includes:
[0048] Step S11: Collect the MRI images of multiple patients with lumbar disc herniation;
[0049] Step S12: Use the pre-trained nnU-Net network to perform a preliminary segmentation on specific lumbar segments in the MRI images. After correction and annotation, slice them to obtain the dataset.
[0050] Among them, the source of the MRI images is medical institutions. The pre-trained nnU-Net network performs a preliminary segmentation on the lumbar discs, L1 to L5 vertebral bodies, and spinal canals from L1 / L2 to L5 / S1 segments. Subsequently, the segmentation results are manually corrected and annotated by radiologists, and then sliced to obtain the dataset, which contains 310 images and labels each. The dataset can be divided into a training set, a validation set, and a test set according to a ratio of 8:1:1, and data augmentation is performed on the training set.
[0051] The purpose of slicing is that considering that the hospital's MRI images are three-dimensional and 16×512×512, and the doctor's manual annotation labels are also of the same size. The network is for segmenting 2D medical images, so we take the middle 4 to 6 channels (slices) of the 16 channels. These slices are all 512×512, and the corresponding label slices are also taken out in the same way. Because the lumbar discs in the middle channels are the most complete and clear, and the clinicians in the hospital mainly observe the middle channels.
[0052] In this embodiment, as Figure 2 shown, the working process of the dual attention fusion module based on Mamba is as follows:
[0053] Step A: Perform Xconcat fusion on input feature maps of the same size, input the fused feature maps into the Mamba module to fuse semantic information, and then sequentially split them into feature maps of the same size. Among them, the input feature maps of the same size are obtained by respectively passing the feature maps of each layer of the encoder through PatchEmbedding.
[0054] G'1, G'2, G'3, G'4 = Split(M(Norm(XCat(G1, G2, G3, G4))))
[0055] Step B: Perform Yconcat fusion on the split feature maps, transpose them and input them into the Mamba module to fuse semantic information, then transpose again, split, and introduce residual connections.
[0056] S1, S2, S3, S4 = Split(M(Norm(YCat(G'1, G'2, G'3, G'4) T ))) T ) + [G1, G2, G3, G4]
[0057] Step C: Perform multi-layer perceptron processing operations.
[0058] p i = s i + MLP)Norm(s i ))
[0059] In the formula, G i (i = 1, 2, 3, 4) are the feature maps of the same size after passing the feature maps of the corresponding layers of the encoder through PatchEmbedding; XCat(·) represents the feature concatenation operation with a dimension of one, Norm(·) represents the regularization operation, M(·) represents the processing using the Mamba module, Split(·) represents the operation of equally dividing the feature maps; YCat(·) represents the feature concatenation operation with a dimension of two, [·] T represents the transpose operation; MLP(·) represents the multi-layer perceptron processing operation; s i (i = 1, 2, 3, 4) are the feature maps connected to the corresponding layers of the decoder.
[0060] In this embodiment, as Figure 2 shown, the working process of the context enhancement module is as follows:
[0061] Stretch the bottom layer feature maps into sequences for processing, then perform regularization operations to enhance the stability of the model, then use the Mamba module for processing to further learn the features extracted by the shallow network and enhance the bottom layer feature extraction ability; finally, add residual connections to prevent overfitting.
[0062] The formula is expressed as follows:
[0063] K = f5 + Reshape(M(Norm(Extend(f5))))
[0064] In the formula, f5 represents the underlying feature, Extend(·) represents the operation of stretching the feature into a sequence, ReShape(·) represents restoring the feature map to the dimension of the underlying feature, and K represents the output of the context enhancement module.
[0065] Among them, the formula of the Mamba module is expressed as:
[0066] M out = L3(SSM(σ(conv(L1(M in )))) · σ(L2(M in )))
[0067] Among them, M in is the input of the Mamba module, M out represents the output of the Mamba module, L1(·), L2(·), L3(·) represent three different linear layers, SSM(·) represents the spatial state model, conv(·) represents the convolutional layer, and σ(·) represents the Sigmoid activation function.
[0068] Specifically, the context enhancement module in this embodiment is based on the long sequence processing ability of Mamba and is deployed at the bottom layer of the network, which can capture the long-range dependencies in the underlying features. In this way, the model can better understand the global context information, thereby significantly improving the segmentation performance.
[0069] In this embodiment, during the process of training the improved U-Net network using the dataset, the loss function is:
[0070] L seg = L Dice + L BCE
[0071] In the formula, L seg represents the total segmentation loss, L Dice represents the Dice loss, and L BCE represents the binary cross-entropy loss.
[0072] Specifically, in this embodiment, a loss function that is the sum of the Dice coefficient and binary cross-entropy is used during training. It can not only effectively measure the overlap degree between the prediction result and the true label using the Dice coefficient, focusing on the segmentation accuracy of foreground objects, but also handle the problem of sample imbalance and enhance the model's learning ability for boundary details with the help of binary cross-entropy. The combination of the two improves the prediction accuracy and stability of the model in the regression task. Therefore, the loss function in this embodiment can simultaneously optimize the regional overlap accuracy of the segmentation task and the pixel-level classification accuracy, improve the prediction accuracy and stability of the model in the regression task, and thus improve the overall performance of the model.
[0073] Table 1 is a comparison table of the quantitative results of the improved U-Net network and the traditional model in the above embodiment on three lumbar disc segmentation datasets, namely Private, Spider, and OSF.
[0074] Table 1
[0075]
[0076] (1) Evaluate the segmentation results.
[0077] The segmentation results of the three lumbar disc segmentation datasets are summarized in Table 1, showing the performance of the proposed improved U-Net network (MNSG-Net) on different datasets and comparing it with the widely used Transformer-based method. On the Private dataset, the Dice score of MNSG-Net is 92.79%, the IoU is 86.91%, and the F1 score is 92.93%. MNSG-Net improves by 1.49%, 2.48%, and 1.33% in DSC, IoU, and F1 score respectively compared to the baseline UNet. Compared with Unet++, which also aims to narrow the semantic gap, our network improves by 0.6%, 1.26%, and 0.72% in DSC, IoU, and F1 score respectively.
[0078] It is worth noting that the three Transformer-based networks perform poorly, and even the three evaluation metrics are worse than those of Unet, with a large difference in the results of the first two datasets. We think there are two reasons. First, the Private dataset has a small sample size, and the Transformer model has a large number of parameters and high complexity, making it easy to overfit. Second, because when annotating private data, nnUnet was used to assist radiologists in annotation, so the Unet model showed good segmentation results. On the Spider dataset and the OSF dataset, our model achieved the highest DSC, IoU, and F1 score.
[0079] (2) Evaluate the number of parameters and computational complexity.
[0080] The computational complexity is an inherent defect of the Transformer model. As shown in Table 1, MNSG-Net is superior to the current state-of-the-art Transformer-based segmentation methods in terms of parameter efficiency. Its number of parameters is significantly lower than those of two advanced Transformer methods, but slightly higher than that of Swin-Unet-tiny. Compared with UDTransNet, MNSG-Net has fewer parameters, which benefits from the fact that the Mamba module is more streamlined than the Transformer module. The results show that the attention mechanism based on the state space model (SSM) has better performance than the Transformer. In terms of the GFLOPs metric, MNSG-Net is not the smallest, but still at a relatively low level, where UNet performs best due to its simple architecture. Overall, MNSG-Net achieves better segmentation performance with fewer parameters and computational volume, reaching a good balance between effectiveness and efficiency.
[0081] A lumbar intervertebral disc segmentation device based on an improved U-Net network, comprising:
[0082] An acquisition module for acquiring a dataset, the dataset including multiple labeled lumbar intervertebral disc images;
[0083] A construction module for constructing an improved U-Net network, where feature maps are transmitted between the encoder and decoder of the improved U-Net network through a dual attention fusion module based on Mamba; a context enhancement module is set at the bottom layer of the encoder to enhance the underlying feature extraction ability;
[0084] A training module for training the improved U-Net network using the dataset to obtain a trained improved U-Net network;
[0085] A segmentation module for segmenting the lumbar intervertebral disc image to be segmented using the trained improved U-Net network.
[0086] A device, comprising:
[0087] A memory for storing a computer program;
[0088] A processor for executing the computer program, and when the computer program is executed by the processor, the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network in the above embodiments are implemented.
[0089] A readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network in the above embodiments are implemented.
[0090] A computer program product includes a computer program which, when executed by a processor, implements the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network in the above embodiments.
[0091] Inspired by the ideal embodiments of the present invention as described above, through the above description, relevant staff can completely make various changes and modifications without departing from the technical idea of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.
Claims
1. A lumbar disc segmentation method based on an improved U-Net network, characterized in that it includes: Obtain a data set, where the data set includes multiple labeled lumbar disc images; Build an improved U-Net network, and transfer feature maps between the encoder and decoder of the improved U-Net network through a dual attention fusion module based on Mamba; set a context enhancement module at the bottom layer of the encoder to enhance the bottom layer feature extraction ability; Use the data set to train the improved U-Net network to obtain a trained improved U-Net network; Use the trained improved U-Net network to segment the lumbar disc image to be segmented.
2. The lumbar disc segmentation method based on the improved U-Net network according to claim 1, characterized in that The obtaining of the data set specifically includes: Collect the MRI images of multiple lumbar disc herniation patients; Use the pre-trained nnU-Net network to perform preliminary segmentation on specific lumbar segments in the MRI images, and after correction and annotation, slice them to obtain the data set.
3. The lumbar disc segmentation method based on the improved U-Net network according to claim 1, characterized in that The working process of the dual attention fusion module based on Mamba is expressed as: G'1,G'2,G'3,G'4=Split(M(Norm(XCat(G1,G2,G3,G4)))) S1,S2,S3,s4 = sSplit([M(Norm([YCat(G'1,G'2,G'3,G'4)] T ))] T ) + [G1,G2,G3,G4] p i = s i + MLP(Norm(s i )) Where G i (i = 1, 2, 3, 4) are features of the same size after PatchEmbedding of the feature maps of the corresponding layers of the encoder; XCat(·) represents a feature concatenation operation with a dimension of one, Norm(·) represents a regularization operation, M(·) represents processing using the Mamba module, and Split(·) represents an operation of equally dividing the feature map; YCat(·) represents a feature concatenation operation with a dimension of two, [·] T represents a transpose operation; MLP(·) represents a multi-layer perceptron processing operation; s i (i = 1, 2, 3, 4) are the feature maps connected to the corresponding layers of the decoder.
4. The lumbar disc segmentation method based on the improved U-Net network according to claim 1, characterized in that The working process of the context enhancement module is expressed as: K=f5+Reshape(M(Norm(Extend(f5)))) In the formula, f5 represents the bottom layer feature, Extend(·) represents the operation of stretching the feature into a sequence, ReShape(·) represents restoring the feature map to the dimension of the bottom layer feature, and K represents the output of the context enhancement module.
5. The lumbar disc segmentation method based on the improved U-Net network according to claim 1, characterized in that During the process of training the improved U-Net network using the data set, the loss function is: L seg = L Dice + L BCE where, L seg represents the total segmentation loss, L Dice represents the Dice loss, and L BCE represents the binary cross-entropy loss.
6. A lumbar disc segmentation device based on an improved U-Net network, characterized in that it includes: An acquisition module for acquiring a data set, where the data set includes multiple labeled lumbar disc images; A building module for building an improved U-Net network, and transferring feature maps between the encoder and decoder of the improved U-Net network through a dual attention fusion module based on Mamba; setting a context enhancement module at the bottom layer of the encoder to enhance the bottom layer feature extraction ability; A training module for using the data set to train the improved U-Net network to obtain a trained improved U-Net network; A segmentation module for using the trained improved U-Net network to segment the lumbar disc image to be segmented.
7. A device, characterized in that it includes: A memory for storing a computer program; A processor for executing the computer program, wherein when the computer program is executed by the processor, the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network according to any one of claims 1-5 are implemented.
8. A readable storage medium, on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network according to any one of claims 1-5 are implemented.
9. A computer program product, comprising a computer program, characterized in that when the computer program is executed by a processor, the steps of the lumbar intervertebral disc segmentation method based on the improved U-Net network according to any one of claims 1-5 are implemented.