Oral cavity region segmentation method and system based on head and neck CT (Computed Tomography) image
The head and neck CT images are segmented through multi-scale convolution attention block U-Net model, which solves the problems of low segmentation accuracy and insufficient anti-artifact performance in the traditional method, and achieves higher segmentation accuracy and accuracy. Head and neck radiation therapy provides reliable technical support.
Patent Information
- Application Number
- CN202411199185.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-08-29
AI Technical Summary
The traditional CT image segmentation method has low accuracy in oral area segmentation in head and neck radiation therapy, and is susceptible to noise and artifacts. It lacks multi-scale information, which leads to unstable and inaccurate segmentation results, affecting the accuracy and treatment of radiation therapy plans. Evaluation of effectiveness.
The multi-scale convolution attention block U-Net model is adopted, combined with the U-Net structure and the multi-scale convolution attention mechanism, and the head and neck CT images are preprocessed and segmented, and the image feature information is enriched using multi-scale convolution layers and hollow convolution. , improve segmentation accuracy and anti-artifact performance.
It effectively improves the accuracy and accuracy of oral area segmentation, enhances the anti-artifact performance and segmentation performance of the model, and provides more accurate technical support for head and neck radiation therapy.
Smart Images

Figure CN119027408A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of medical image processing, and in particular to an oral region segmentation method and system based on head and neck CT images. Background Art
[0002] Head and neck radiotherapy is a common cancer treatment method, in which CT (Computed Tomography) images play a key role in the diagnosis and treatment process. CT images provide detailed information about the patient's anatomical structure and therefore play a vital role in radiotherapy planning, dose distribution, and treatment effect evaluation. Traditional CT image segmentation methods are mainly based on basic image processing techniques such as threshold segmentation, region growing, and edge detection.
[0003] The problems of traditional CT image segmentation methods include:
[0004] (1) Low segmentation accuracy: Traditional methods such as threshold and region growing have low segmentation accuracy for the oral region in head and neck radiotherapy CT images, which limits the accuracy of radiotherapy planning and the evaluation of treatment effects.
[0005] (2) Sensitive to noise and artifact structures: Traditional methods are sensitive to noise and complex structures in images and are easily affected by external interference, resulting in unstable segmentation results, which in turn affects the formulation and implementation of radiotherapy plans. In addition, most current technologies focus on processing artifact-free images and outlining the oral area based on this, but relatively little consideration is given to artifact interference. Oral artifacts are one of the common problems in head and neck radiotherapy CT images, which may come from a variety of factors, such as metal prostheses, intraoral implants, and masticatory muscle movements. These artifacts may not only distort the morphology of the oral area, but also affect the accuracy of oral segmentation, leading to inaccurate radiotherapy plans and limited evaluation of treatment effects.
[0006] (3) Lack of multi-scale information: Traditional methods lack effective use of multi-scale information in images and are unable to fully capture the details and structural information inside the oral area, resulting in a lack of completeness and accuracy in the segmentation results, which in turn affects the accuracy of radiotherapy planning and the evaluation of treatment effects. Summary of the invention
[0007] In view of the defects in the prior art, the present invention provides an oral region segmentation method and system based on head and neck CT images, which effectively solves the technical problem of low oral segmentation accuracy in the prior art.
[0008] In a first aspect, the present invention provides an oral region segmentation method based on a head and neck CT image, comprising:
[0009] Obtain original head and neck radiotherapy CT images;
[0010] Preprocessing the CT image to obtain a preprocessed image;
[0011] Construct a multi-scale convolutional attention block U-Net model, train and adjust the parameters of the multi-scale convolutional attention block U-Net model, and obtain the trained optimal multi-scale convolutional attention block U-Net model;
[0012] Use the trained optimal multi-scale convolutional attention block U-Net model to verify the model performance on the validation set and obtain the oral cavity segmentation model;
[0013] The preprocessed image is input into the oral cavity segmentation model to obtain an oral cavity region segmentation result.
[0014] In a second aspect, an embodiment of the present invention provides an oral region segmentation system based on head and neck CT images, comprising: an image acquisition module, a preprocessing module, a model building module, a model verification module and an image segmentation module;
[0015] The image acquisition module is used to acquire original head and neck radiotherapy CT images;
[0016] The preprocessing module is used to preprocess the CT image to obtain a preprocessed image;
[0017] The model building module is used to build a multi-scale convolutional attention block U-Net model, train and adjust the parameters of the multi-scale convolutional attention block U-Net model, and obtain a trained optimal multi-scale convolutional attention block U-Net model;
[0018] The model verification module is used to use the trained optimal multi-scale convolutional attention block U-Net model to verify the model performance on the verification set to obtain the oral cavity segmentation model;
[0019] The image segmentation module inputs the preprocessed image into the oral cavity segmentation model to obtain an oral cavity region segmentation result.
[0020] Beneficial effects of the present invention:
[0021] The embodiment of the present invention provides a method and system for oral region segmentation based on head and neck CT images, and constructs a multi-scale convolutional attention block U-Net model. Combining the U-Net structure and the multi-scale convolutional attention mechanism, it can segment the oral cavity containing artifacts and the oral cavity without artifacts, effectively solving the gradient vanishing and explosion problems existing in the traditional U-Net model, and improving the convergence speed and training stability of the model. At the same time, multi-scale convolutional layers and dilated convolutions are used to enrich image feature information, improve segmentation precision and accuracy, and greatly improve the anti-artifact performance and segmentation performance of the model, providing reliable technical support for the precision treatment of head and neck radiotherapy. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.
[0023] Figure 1 A flow chart of an oral region segmentation method based on head and neck CT images provided by the first embodiment of the present invention is shown;
[0024] Figure 2 A model diagram of a multi-scale convolutional attention block U-Net model in the first embodiment of the present invention is shown;
[0025] Figure 3 A schematic diagram of the structure of three sub-models in the multi-scale convolutional attention block U-Net model in the first embodiment of the present invention is shown;
[0026] Figure 4 A schematic diagram of the structure of the attention branch in the multi-scale convolutional attention block U-Net model in the first embodiment of the present invention is shown;
[0027] Figure 5 A flowchart of a multi-scale convolutional attention block U-Net model in the first embodiment of the present invention is shown;
[0028] Figure 6 A structural block diagram of an oral region segmentation system based on head and neck CT images provided in another embodiment of the present invention is shown. DETAILED DESCRIPTION
[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0030] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.
[0031] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0032] It should be further understood that the term "and / or" used in the present specification and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0033] As used in this specification and the appended claims, the term "if" may be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrases "if it is determined" or "if [described condition or event] is detected" may be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.
[0034] It should be noted that, unless otherwise specified, the technical terms or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.
[0035] See also Figure 1-5 The first embodiment of the present invention provides an oral region segmentation method based on a head and neck CT image, comprising:
[0036] Obtain original head and neck radiotherapy CT images;
[0037] Preprocessing the CT image to obtain a preprocessed image;
[0038] Construct a multi-scale convolutional attention block U-Net model, train and adjust the parameters of the multi-scale convolutional attention block U-Net model, and obtain the trained optimal multi-scale convolutional attention block U-Net model;
[0039] Use the trained optimal multi-scale convolutional attention block U-Net model to verify the model performance on the validation set and obtain the oral cavity segmentation model;
[0040] The preprocessed image is input into the oral cavity segmentation model to obtain an oral cavity region segmentation result.
[0041] Specifically, the original data of head and neck radiotherapy CT images with outlined oral cavity were divided into training set (52 cases), validation set (8 cases) and test set (12 cases). At the same time, data augmentation was used to expand the training set to avoid model overfitting and improve model stability. The data augmentation methods included rotation, flipping, scaling and cropping techniques.
[0042] The divided data sets were preprocessed, mainly HU value conversion, window width and window level adjustment, adaptive histogram equalization and image normalization. HU value conversion was automatically performed by the Simple ITK software package. Window width and window level are used to adjust the range of gray levels in CT images. The image contrast of different tissues can be more obvious through adjustment, thereby highlighting the structures and tissues of interest. Adaptive histogram equalization can adjust the local features of the image according to the gray level, thereby effectively enhancing the contrast and details of the image and further highlighting the oral cavity. Image normalization normalizes the pixel values to make the model more stable during training. The image range is normalized to [0,1].
[0043] Hu value conversion process:
[0044] 1. Read original head and neck radiotherapy CT image data.
[0045] 2. Use sitk.GetArrayFromImage to convert the raw data into Hounsfield units (HU) and convert it to Numpy format.
[0046] Window width and window level adjustment process:
[0047] 1. Select the Window Width and Window Center parameters.
[0048] 2. Adjust the image display range according to the selected window width and window position.
[0049] The window width and window level adjustment is achieved through the following formula:
[0050]
[0051] In the formula, window_width represents the window width, window_center represents the window position, and pixel represents pixel.
[0052] Adaptive histogram equalization process:
[0053] 1. Input the image obtained after HU value conversion and window width and window level adjustment for format judgment;
[0054] 2. Use cv2.createCLAHE algorithm to perform histogram equalization on the image;
[0055] 3. Input the equalization result image.
[0056] The original image is processed by the above preprocessing method to further enhance the contrast between the region of interest and the surrounding tissues, so as to reduce the noise in the image and highlight the features of the oral region.
[0057] Tensorflow and Keras were used to construct the model. The constructed multi-scale convolutional attention block U-Net model includes 9 convolutional layers, 9 normalization layers and 6 attention branch modules. The convolutional layers use convolution kernels of different sizes (such as 1×1, 3×3, 5×5 full convolution) and 3×3 dilated convolution for feature extraction at different scales. The residual connection part consists of a full convolution layer with a convolution kernel size of 1×1 and a normalization layer. The front end is connected to the original input and added to the subsequent feature map after processing to retain key information. The model performs three different convolutions on the original input and divides the three different convolutions into three feature extraction channels, named the first feature extraction sub-model, the second feature extraction sub-model and the third feature extraction sub-model. The first feature extraction sub-model and the second feature extraction sub-model are used to extract the basic features of the oral cavity to obtain the first feature map and the second feature map respectively. The third feature extraction sub-model is used to extract the detailed features of the oral cavity to obtain the third feature map. The first feature map, the second feature map and the third feature map are added and merged to obtain the relevant features of the entire oral structure.
[0058] Among them, the first feature extraction sub-model includes 2 convolutional layers to extract basic oral features. The convolution kernel of the convolutional layer is 3×3. A batch normalization (BN) layer and a ReLU activation layer are added after each convolutional layer. An attention branch module is added after each convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the first feature map. After parallel convolution operations, the model can extract multi-scale features of oral structures such as teeth and gums, such as the edges of teeth, the position of the oral cavity, and the microstructure of periodontal tissues. These multi-scale features are input into the subsequent attention branch module for further processing. The attention branch module receives the multi-scale features from the parallel convolution branch as input.
[0059] The second feature extraction sub-model includes two convolutional layers to extract basic oral features. The convolution kernel of the convolutional layer is 5×5. Each convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the second feature map.
[0060] The third feature extraction sub-model includes two layers of dilated convolutional layers to extract oral detail features. The convolution kernel of the convolutional layer is 3×3. Each layer of dilated convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each layer of dilated convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the third feature map.
[0061] The main part of the U-Net model is the multi-scale convolutional attention block, and its core ideas are divided into three parts:
[0062] (1) A normalization layer is added after each convolution, and residual connections are introduced at each resolution level for feature extraction. This can effectively solve the gradient explosion problem that occurs during training in traditional U-Net models due to simple convolution feature extraction. In addition, as the depth of the model increases, simple convolution stacking may also cause gradient vanishing or exploding during training. Residual connections can effectively solve or alleviate the information loss and gradient vanishing or exploding problems of deep learning models during training.
[0063] The specific process of the convolutional neural network with normalization layer and residual connection:
[0064] Step 1: Input Layer: Pass the raw input data to the network.
[0065] Step 2: Convolutional Layer: Use Tensflow and Keras's Conv2d package for feature extraction, and the convolution kernel size is 3×3.
[0066] Step 3: Add a normalization (BN) layer after each convolutional layer. The BN layer code is completed by keras.layers.BatchNormalization, and its calculation formula is:
[0067] Calculate the mean:
[0068] Calculate the variance:
[0069] standardization:
[0070] Scale and displacement:
[0071] Among them, μ is the sample mean, σ 2 is the sample variance, γ and β are hyperparameters, xi Represents the i-th sample value in the input data, where i is an integer greater than or equal to 1. yi represents the i-th sample value after scale and shift processing, m represents the number of samples in the current batch, and ε represents a very small positive number used to prevent division by zero during the normalization process.
[0072] Step 4: Residual Connection: Introduce residual connection at each resolution level for feature extraction. The specific implementation is: after convolution, add the original input and the convolution result. This step is to ensure that the network can learn the residual, so as to better retain and propagate information. The calculation formula is:
[0073] z=F(x input )+x input (6)
[0074] In the formula, x input represents the input of the residual block, z represents the output of the residual block, and F(x) represents the mapping function of the residual block, that is, operations such as convolution.
[0075] Step 5: Activation function: Activate each feature map and use the ReLU loss function. The specific formula is:
[0076]
[0077] (2) A multi-scale convolutional layer is used, which combines ordinary convolution and dilated convolution for feature extraction. The multi-scale convolutional layer can enrich image features by adjusting the convolution kernel size without changing the size of the original feature map, and can encode and decode the feature information of interest in the image from a global perspective, thereby improving the segmentation performance. In traditional U-Net models, downsampling or strided convolution is usually used to increase the receptive field, which leads to a decrease in spatial resolution. Dilated convolution can expand the receptive field without losing resolution, thereby improving the accuracy and details of segmentation. In addition, dilated convolution can effectively capture multi-scale contextual information, allowing the model to pay more attention to global features during training. However, the calculation method of dilated convolution may lead to the loss of local information, so dilated convolution is used on one path of the multi-scale convolutional layer, while the remaining paths are trained using traditional convolution to supplement the local information lost during training.
[0078] Specifically, the model performs three different convolutions on the original input data, of which the three convolutions are 3×3 normal convolution, 5×5 normal convolution, and 3×3 dilated convolution. The three different convolutions are divided into three feature extraction pathways, and the first feature extraction sub-model, the second feature extraction sub-model, and the third feature extraction sub-model are used to extract features of different scales. The first and second pathways are used to extract the basic features of the oral cavity, and the steps include a BN layer and two attention mechanisms. The third pathway, the dilated convolution pathway, is used to effectively capture multi-scale contextual information, so that the model pays more attention to the detailed features of the oral cavity during training, thereby improving the segmentation ability of the model. Finally, the extraction results of the three pathways are combined by addition to summarize the relevant features of the entire oral structure.
[0079] The specific steps of the dilated convolution algorithm are as follows:
[0080] 1. Input:
[0081] Feature map size: H×W×C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels of the feature map.
[0082] Convolution kernel size: F×F×C×K, where F is the size of the convolution kernel (usually a square, so F×F), and C is the same as the number of channels in the input feature map. K is the number of convolution kernels, which determines the number of channels in the output feature map.
[0083] Stride: s, s is the stride of the convolution operation, which determines the step size of the convolution kernel moving on the feature map.
[0084] Dilation Rate: d. The dilation rate indicates the number of zeros inserted into the convolution kernel. The dilation rate is used to increase the receptive field of the convolution kernel.
[0085] 2. Padding: Zero padding is performed on the input feature map I so that the size of the feature map after the dilated convolution is the same as the size of the input feature map.
[0086] 3. Calculate the output size: The size calculation formula of the output feature map is:
[0087]
[0088]
[0089] Among them, H in and W in are the height and width of the input feature map, H out and W out are the height and width of the output feature map, respectively, and padding is the amount of zero padding.
[0090] 4. The calculation formula of the dilated convolution is:
[0091]
[0092] Among them, O i,j,k is the value of the output feature map, and m, n, j, and l are all integers greater than or equal to 1.
[0093] The core of the model design lies in its modular structure, which is combined with the U-Net architecture. The overall structure of the model starts with the input image, which is processed by multiple modules and finally outputs a segmentation result of the oral cavity.
[0094] (3) The cSE attention mechanism (channel attention mechanism) is introduced. The specific steps are as follows:
[0095] 1. A cSE attention mechanism is added after each convolution + BN layer to obtain the key features related to the oral structure at each step.
[0096] 2. After two cSE attention mechanisms, a residual connection is used to add and merge the original input data and the key feature map obtained after the attention mechanism to obtain the feature map to ensure the stability of the model.
[0097] The attention branch module receives the multi-scale features of the parallel convolution branch as input, and the specific method of obtaining the relevant key features of the oral structure at each step includes: the attention branch module compresses the input features spatially through the global average pooling layer to extract the global feature representation; extracts the key features of the oral structure through the linear layer and the ReLU activation function; generates a weight vector through another linear layer and the Sigmoid activation function; and uses the weight vector to weight the input features to generate an enhanced feature map. The attention branch module can selectively focus on features, so that the model pays more attention to the main channel features during training, and suppresses the secondary channel features through squeezing, excitation, and scaling, which helps to highlight the key oral features and suppress irrelevant information, thereby improving the accuracy of segmentation.
[0098] An attention branch module is added after each downsampling convolution operation to obtain the weights of different feature channels. The attention branch module is used to obtain the key feature channels for oral segmentation for the model during training. Finally, to prevent the model from overfitting and reduce the number of model parameters, we added a dropout layer with a parameter of 0.1 in each multi-scale attention branch module. We also replaced the maximum pooling with strided convolution because in the traditional U-Net architecture, the maximum pooling is mainly used to compress feature maps and resolution. Although it effectively reduces parameters and resolution, it also leads to the loss of extracted features.
[0099] The training set is used to train and adjust the parameters of the constructed multi-scale convolutional attention block U-Net model to obtain the trained optimal multi-scale convolutional attention block U-Net model. The validation set is used to verify the model performance. The model performance is evaluated by indicators such as the Dyss similarity coefficient to obtain the final oral segmentation model. Finally, the test set is input into the model for testing to obtain the model prediction results and the oral segmentation results. This segmentation result can clearly mark the teeth and oral area.
[0100] The embodiment of the present invention provides an oral region segmentation method based on head and neck CT images, which introduces a new attention and multi-scale fusion model, and constructs a multi-scale convolutional attention block U-Net model. Combining the U-Net structure and the multi-scale convolutional attention mechanism, it can segment the oral cavity containing artifacts and the oral cavity without artifacts, effectively solving the gradient vanishing and explosion problems existing in the traditional U-Net model, and improving the convergence speed and training stability of the model. At the same time, multi-scale convolutional layers and dilated convolutions are used to enrich image feature information, improve segmentation precision and accuracy, and greatly improve the anti-artifact performance and segmentation performance of the model, providing reliable technical support for the precision treatment of head and neck radiotherapy.
[0101] In the first embodiment described above, a method for segmenting the oral region based on head and neck CT images is provided. Correspondingly, the present application also provides a system for segmenting the oral region based on head and neck CT images. Figure 6 , which is a structural block diagram of an oral region segmentation system based on head and neck CT images provided by the second embodiment of the present invention. Since the device embodiment is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The device embodiment described below is only illustrative.
[0102] like Figure 6 As shown, another embodiment of the present invention provides an oral region segmentation system based on head and neck CT images, including: an image acquisition module, a preprocessing module, a model construction module, a model verification module and an image segmentation module; the image acquisition module is used to acquire original head and neck radiotherapy CT images; the preprocessing module is used to preprocess the CT images to obtain preprocessed images; the model construction module is used to construct a multi-scale convolutional attention block U-Net model, train and adjust the parameters of the multi-scale convolutional attention block U-Net model, and obtain a trained optimal multi-scale convolutional attention block U-Net model; the model verification module is used to use the trained optimal multi-scale convolutional attention block U-Net model to perform model performance verification on a verification set to obtain an oral segmentation model; the image segmentation module inputs the preprocessed image into the oral segmentation model to obtain an oral region segmentation result.
[0103] The preprocessing module preprocesses the CT images including: HU value conversion, window width and window level adjustment, adaptive histogram equalization and image normalization.
[0104] The multi-scale convolutional attention block U-Net model includes a first feature extraction sub-model, a second feature extraction sub-model and a third feature extraction sub-model. The first feature extraction sub-model and the second feature extraction sub-model are used to extract the basic features of the oral cavity to obtain a first feature map and a second feature map respectively. The third feature extraction sub-model is used to extract the detailed features of the oral cavity to obtain a third feature map. The first feature map, the second feature map and the third feature map are added and merged to obtain the relevant features of the entire oral structure.
[0105] The first feature extraction sub-model includes two convolutional layers to extract basic oral features. The convolution kernel of the convolutional layer is 3×3. Each convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the first feature map.
[0106] The second feature extraction sub-model includes two convolutional layers to extract basic oral features. The convolution kernel of the convolutional layer is 5×5. Each convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the second feature map.
[0107] The third feature extraction sub-model includes two layers of dilated convolutional layers to extract oral detail features. The convolution kernel of the convolutional layer is 3×3. Each layer of dilated convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each layer of dilated convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the third feature map.
[0108] The specific method of the attention branch module to obtain the relevant key features of the oral structure at each step includes:
[0109] The attention branch module compresses the input features spatially through the global average pooling layer to extract the global feature representation;
[0110] The key features of the oral structure are extracted through the linear layer and the ReLU activation function;
[0111] Generate a weight vector through another linear layer and Sigmoid activation function;
[0112] The input features are weighted using the weight vector to generate an enhanced feature map.
[0113] The embodiment of the present invention provides an oral region segmentation system based on head and neck CT images, which introduces a new attention and multi-scale fusion model, and constructs a multi-scale convolutional attention block U-Net model. Combining the U-Net structure and the multi-scale convolutional attention mechanism, it can segment the oral cavity containing artifacts and the oral cavity without artifacts, effectively solving the gradient vanishing and explosion problems existing in the traditional U-Net model, and improving the convergence speed and training stability of the model. At the same time, multi-scale convolutional layers and dilated convolutions are used to enrich image feature information, improve segmentation precision and accuracy, greatly improve the model's anti-artifact performance and segmentation performance, and provide reliable technical support for the precision treatment of head and neck radiotherapy.
[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. These modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and specification of the present invention.
Claims
1. A method for oral region segmentation based on head and neck CT images. It is characterized in that include: Obtain original head and neck radiotherapy CT images; Preprocessing the CT image to obtain a preprocessed image; Construct a multi-scale convolutional attention block U-Net model, train and adjust the parameters of the multi-scale convolutional attention block U-Net model, and obtain the trained optimal multi-scale convolutional attention block U-Net model; Use the trained optimal multi-scale convolutional attention block U-Net model to verify the model performance on the validation set and obtain the oral cavity segmentation model; The preprocessed image is input into the oral cavity segmentation model to obtain an oral cavity region segmentation result.
2. The method according to claim 1, It is characterized in that The preprocessing of the CT image includes: HU value conversion, window width and window level adjustment, adaptive histogram equalization and image normalization processing.
3. The method according to claim 1, It is characterized in that The multi-scale convolutional attention block U-Net model includes a first feature extraction sub-model, a second feature extraction sub-model and a third feature extraction sub-model. The first feature extraction sub-model and the second feature extraction sub-model are used to extract the basic features of the oral cavity to obtain a first feature map and a second feature map respectively. The third feature extraction sub-model is used to extract the detailed features of the oral cavity to obtain a third feature map. The first feature map, the second feature map and the third feature map are added and merged to obtain the relevant features of the entire oral structure.
4. The method according to claim 3, It is characterized in that The first feature extraction sub-model includes two convolutional layers to extract basic oral features. The convolution kernel of the convolutional layer is 3×3. Each convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the first feature map.
5. The method according to claim 4, It is characterized in that The second feature extraction sub-model includes two convolutional layers to extract basic oral features. The convolution kernel of the convolutional layer is 5×5. Each convolutional layer is followed by a normalization layer and a ReLU activation layer. An attention branch module is added after each convolutional layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the second feature map.
6. The method according to claim 5, It is characterized in that The third feature extraction sub-model includes 2 layers of hole convolution layers to extract oral detail features. The convolution kernel of the convolution layer is 3×3. Each hole convolution layer is followed by a normalization layer and a ReLU activation layer in sequence. An attention branch module is added after each hole convolution layer. The attention branch module obtains the relevant key features of the oral structure at each step. After two attention mechanisms, a residual connection is used to add the original input and the key features obtained after the attention mechanism to obtain the third feature map.
7. The method according to any one of claims 4 to 6, It is characterized in that The specific method for the attention branch module to obtain the relevant key features of the oral structure at each step includes: The attention branch module compresses the input features spatially through the global average pooling layer to extract the global feature representation; The key features of the oral structure are extracted through the linear layer and the ReLU activation function; Generate a weight vector through another linear layer and Sigmoid activation function; The input features are weighted using the weight vector to generate an enhanced feature map.
8. The method according to any one of claims 4 to 6, It is characterized in that The residual connection consists of a full convolution layer with a 1×1 convolution kernel and a normalization layer.
9. An oral region segmentation system based on head and neck CT images. It is characterized in that include: Image acquisition module, preprocessing module, model building module, model verification module and image segmentation module; The image acquisition module is used to acquire original head and neck radiotherapy CT images; The preprocessing module is used to preprocess the CT image to obtain a preprocessed image; The model building module is used to build a multi-scale convolutional attention block U-Net model, train and adjust the parameters of the multi-scale convolutional attention block U-Net model, and obtain a trained optimal multi-scale convolutional attention block U-Net model; The model verification module is used to use the trained optimal multi-scale convolutional attention block U-Net model to verify the model performance on the verification set to obtain the oral cavity segmentation model; The image segmentation module inputs the preprocessed image into the oral cavity segmentation model to obtain an oral cavity region segmentation result.
10. The system of claim 9, It is characterized in that The multi-scale convolutional attention block U-Net model includes a first feature extraction sub-model, a second feature extraction sub-model and a third feature extraction sub-model. The first feature extraction sub-model and the second feature extraction sub-model are used to extract the basic features of the oral cavity to obtain a first feature map and a second feature map respectively. The third feature extraction sub-model is used to extract the detailed features of the oral cavity to obtain a third feature map. The first feature map, the second feature map and the third feature map are added and merged to obtain the relevant features of the entire oral structure.
Citation Information
Patent Citations
Medical image segmentation method based on multiple scales and attention
CN114359292A
Multi-scale decoupling EM attention module and medical image segmentation method
CN116823856A
Three-dimensional image segmentation method and system based on attention mechanism multi-scale convolutional neural network
CN116843893A
Medical image segmentation method based on multi-scale convolution and attention mechanism
CN117036700A
Oral cavity CBCT image tooth and soft tissue segmentation model method based on improved U-Net model
CN117115132A
Cited By
Method for analyzing morphological parameters of coniferous wood tracheid and intercellular space based on computer vision
CN120071348A
Image enhancement-based mouth opening difficulty degree evaluation system for head and neck radiotherapy patient
CN120318226A
