CN-121998999-A - CBCT tooth image segmentation method and device and electronic equipment

CN121998999ACN 121998999 ACN121998999 ACN 121998999ACN-121998999-A

Abstract

The application provides a CBCT tooth image segmentation method, a device and electronic equipment, wherein the CBCT tooth image to be segmented is acquired and is input into a tooth image segmentation model to obtain a tooth segmentation result, the tooth image segmentation model is obtained by training a pre-constructed segmentation model through a tooth sample image, the pre-constructed segmentation model comprises an encoder provided with a double-scale cooperative attention module, a layered decoder provided with a jump fusion door module and a multi-scale fusion head, after the pre-constructed segmentation model is trained through the tooth sample image, the CBCT tooth image is segmented through the tooth image segmentation model to obtain the tooth image segmentation model, adverse effects on CBCT tooth image segmentation caused by problems such as metal artifacts and boundary blurring are reduced, the segmentation accuracy of the CBCT tooth image is improved, the dependence on labeling data is reduced, and the generalization capability of the tooth image segmentation model is improved.

Inventors

ZHANG KANG
QIU GANG
SUN JIE
WANG JINGJING
Che Quanjiang
WANG HUAN
PENG SIYU
GUO LING

Assignees

昌吉学院

Dates

Publication Date: 20260508
Application Date: 20260128

Claims (10)

1. A CBCT dental image segmentation method, the method comprising: Acquiring CBCT tooth images to be segmented; The CBCT tooth image is input into a tooth image segmentation model to obtain a tooth segmentation result, the tooth image segmentation model is obtained by training a pre-constructed segmentation model through a tooth sample image, and the pre-constructed segmentation model comprises an encoder provided with a double-scale cooperative attention module, a layered decoder provided with a jump fusion door module and a multi-scale fusion head.
2. The CBCT dental image segmentation method of claim 1, wherein training a pre-constructed segmentation model from a dental sample image to obtain the dental image segmentation model comprises: Inputting the tooth sample image into the encoder provided with the double-scale cooperative attention module to obtain a multi-scale enhanced feature map; inputting the multi-scale enhanced feature map into the layered decoder provided with the jump fusion gate module to obtain a decoding feature map; inputting the decoding feature map into the multi-scale fusion head to obtain a prediction segmentation result; And aiming at minimizing the difference between the prediction segmentation result and the tooth sample image, adjusting parameters in the pre-constructed segmentation model to obtain the tooth image segmentation model.
3. The method for segmenting the CBCT dental image of claim 2, wherein if the encoder with the dual-scale co-attention module comprises a plurality of cascaded dual-scale co-attention blocks, inputting the dental sample image to the encoder with the dual-scale co-attention module, and obtaining the multi-scale enhanced feature map comprises: inputting the tooth sample image into a plurality of cascaded double-scale cooperative attention blocks to obtain an enhanced feature map output by each double-scale cooperative attention block, wherein the plurality of cascaded double-scale cooperative attention blocks comprise a convolution kernel combination module, a convolution kernel module, a similarity-based attention module and a high-efficiency channel attention mechanism module; and combining the enhancement feature graphs output by each double-scale cooperative attention block to obtain the multi-scale enhancement feature graph.
4. A CBCT dental image segmentation method according to claim 3, wherein if the plurality of cascaded two-dimensional co-attention blocks comprises a first two-dimensional co-attention block, the process of inputting the dental sample image into the first two-dimensional co-attention block comprises: The tooth sample image is input into the convolution kernel combination module to obtain a first extracted feature image, wherein the convolution kernel combination module comprises a first convolution kernel combination module and a second convolution kernel combination module, the first convolution kernel combination module comprises a convolution kernel, a batch normalization layer and an activation function layer, and the second convolution kernel combination module comprises a convolution kernel and a batch normalization layer; inputting the tooth sample image data into the convolution kernel module to obtain a texture feature map; Splicing the first extracted feature map and the texture feature map to obtain a basic feature map; inputting the basic feature map into the attention module based on the similarity to obtain a weighted feature map; Inputting the basic feature map into the efficient channel attention mechanism module to obtain a space-channel feature map; and carrying out weighted summation on the weighted feature map and the space-channel feature map to obtain a first scale enhancement feature map.
5. The method according to claim 2, wherein if the layered decoder with the skip fusion gate module includes three layered decoding branches and a third convolution kernel combination module, and the multi-scale fusion feature map includes four scale fusion feature maps, inputting the multi-scale enhancement feature map into the layered decoder with the skip fusion gate module, and obtaining the decoding feature map includes: inputting the fourth scale fusion feature map into a fourth convolution kernel combination module to obtain a second extraction feature map; inputting a third-scale fusion feature map and a first up-sampling feature map into a third hierarchical decoding branch to obtain a third decoding feature map, wherein the first up-sampling feature map is obtained by up-sampling the second extraction feature map; Inputting a second scale fusion feature map and a second up-sampling feature map into a second level decoding branch to obtain a second decoding feature map, wherein the second up-sampling feature map is obtained by up-sampling the third decoding feature map; Inputting the first scale fusion feature map and the third upsampling feature map into a first level decoding branch to obtain a first decoding feature map; And inputting the first decoding feature map, the second decoding feature map and the third decoding feature map into the third convolution kernel combination module to obtain the decoding feature map.
6. The method of claim 5, wherein if the first hierarchical decoding branch includes a first skip fusion gate module, a second dual-scale cooperative attention block, and a first feature stitching module, inputting a first scale fusion feature map and the third upsampled feature map into the first hierarchical decoding branch, and obtaining a first decoding feature map includes: inputting the first scale fusion feature map and the third upsampling feature map into the first jump fusion door module to obtain a first enhanced fusion feature map; Inputting the first enhanced fusion feature map and the third upsampled feature map into the first multi-scale feature aggregation module to obtain a first aggregation feature map; inputting the first aggregation feature map into the second double-scale cooperative attention block to obtain a first enhancement aggregation feature map; And inputting the first enhancement aggregation feature map and the third upsampling feature map into the first feature splicing module to obtain the first decoding feature map.
7. The CBCT dental image segmentation method of claim 6, wherein inputting the first scale fusion feature map and the third upsampled feature map into the first skip fusion gate module to obtain a first enhanced fusion feature map comprises: Performing connection operation on the first scale fusion feature map and the third upsampling feature map to obtain a first spliced feature map; Performing convolution operation on the first spliced feature map to obtain an intermediate feature map; performing global average pooling operation, full connection operation and activation operation on the intermediate feature map to obtain a fused attention weight value; and carrying out weighted fusion on the first scale fusion feature map and the third upsampling feature map through the fusion attention weight value to obtain the first enhanced fusion feature map.
8. The CBCT dental image segmentation method according to any one of claims 2-7, wherein if the difference between the predicted segmentation result and the dental sample image is characterized in terms of a loss function, the expression of the loss function comprises: , Wherein, the The loss function is represented by a function of the loss, Representing a supervised loss function, The consistency-loss function is represented as such, Representing a consistency loss weighting function; the expression of the consistency loss weighting function includes: , Wherein, the Representing a consistency loss weighting function, The maximum value of the weight is indicated, Representing the current training round of the present training, The round parameters representing the incremental process are presented, Representing increasing behavior based on sigmoid function, the function value will follow the current training round Gradually increasing, tramp to 60, indicating a rising period based on the sigmoid function; the expression based on the increasing behavior of the sigmoid function includes: , Wherein, the Represents an increasing behavior based on the sigmoid function, Representing the current training round of the present training, A round parameter representing an incremental process; The expression of the supervised loss function includes: , Wherein, the Representing a supervised loss function, Representing the Dice loss function, Representing a cross entropy loss function; The expression of the Dice loss function includes: , Wherein, the Representing the Dice loss function, Represents the prediction tags in the prediction segmentation result, A class label corresponding to the image representing the tooth sample, Representing the index of the samples in the batch, The index of the category is represented and, The depth index is represented as such, The height index is indicated as such, Representing the index of the width of the web, Representing a constant; the expression of the cross entropy loss function includes: , Wherein, the Representing the cross-entropy loss function, Represents the prediction tags in the prediction segmentation result, A class label corresponding to the image representing the tooth sample, Representing the index of the samples in the batch, The size of the batch is indicated and, The index of the category is represented and, Representing the total number of categories, The depth index is represented as such, The maximum depth is indicated as such, The height index is indicated as such, Indicating the maximum height of the container, Representing the index of the width of the web, The maximum width is indicated as being the maximum width, Representing a constant; The expression of the consistency loss function includes: , Wherein, the The consistency-loss function is represented as such, Representing a mask matrix for indicating a location Whether or not a consistency loss is calculated, Representing the index of the samples in the batch, The depth index is represented as such, The height index is indicated as such, Representing the index of the width of the web, Indicating that teacher model is in position The probability distribution of the foreground over the course of the image, Representing student model in position The probability distribution of the foreground over the course of the image, The constant is represented by a value that is a function of, Representing Kullback-Leibler divergence for measuring teacher model position The foreground probability distribution and the student model are in position Differences between the foreground probability distributions.
9. A CBCT dental image segmentation apparatus, the apparatus comprising: The image acquisition module is used for acquiring CBCT tooth images to be segmented; The image segmentation module is used for inputting the CBCT tooth image into a tooth image segmentation model to obtain a tooth segmentation result, wherein the tooth image segmentation model is obtained by training a pre-constructed segmentation model through a tooth sample image, and the pre-constructed segmentation model comprises an encoder provided with a double-scale cooperative attention module, a layered decoder provided with a jump fusion door module and a multi-scale fusion head.
10. An electronic device, comprising: One or more processors; storage means for storing one or more programs which when executed by the one or more processors cause the electronic device to implement the CBCT dental image segmentation method of any one of claims 1 to 8.

Description

CBCT tooth image segmentation method and device and electronic equipment Technical Field The present application relates to the field of image processing technologies, and in particular, to a CBCT dental image segmentation method, device, and electronic apparatus. Background Cone beam computed tomography (Cone Beam Computed Tomography, CBCT) is an imaging technique that has been widely used in the oral field in recent years, which can provide three-dimensional images of the oromaxillofacial region, providing an important basis for the diagnosis and treatment of dental diseases. However, in practical applications, due to the low contrast between the teeth and surrounding soft tissues, and the metal streak artifacts and anatomical complexity (such as crowding of tooth root or thin cortical bone), the problem of blurred boundary and poor segmentation accuracy of the tooth segmentation method (e.g. 3D U-Net architecture) in the related art is often caused. In addition, the labor-intensive nature of manual labeling by dental professionals makes the labeling dataset limited, which not only exacerbates the data scarcity, but also limits the generalization capability of the fully supervised deep learning model, so that the method is difficult to adapt to complex morphological structures, changeable gray features and similarity with surrounding tissues of teeth, and results in inaccurate segmentation results, which affect subsequent tooth diagnosis and treatment scheme formulation. Therefore, an improvement of the tooth segmentation method in the related art is needed to solve the problems of metal artifacts, boundary blurring and the like, so as to realize high-precision segmentation of the CBCT tooth image. Disclosure of Invention In view of the above drawbacks of the prior art, the present application provides a CBCT dental image segmentation method, apparatus and electronic device, so as to solve the above technical problems. According to one aspect of the embodiment of the application, a CBCT tooth image segmentation method is provided, and comprises the steps of obtaining a CBCT tooth image to be segmented, inputting the CBCT tooth image into a tooth image segmentation model to obtain a tooth segmentation result, training a pre-built segmentation model through a tooth sample image by the tooth image segmentation model, wherein the pre-built segmentation model comprises an encoder provided with a double-scale cooperative attention module, a layered decoder provided with a jumping fusion door module and a multi-scale fusion head. According to one aspect of the embodiment of the application, a CBCT tooth image segmentation device is provided, and comprises an image acquisition module, an image segmentation module, a tooth image segmentation module and a multi-scale fusion head, wherein the image acquisition module is used for acquiring a CBCT tooth image to be segmented, the image segmentation module is used for inputting the CBCT tooth image into a tooth image segmentation model to obtain a tooth segmentation result, the tooth image segmentation model is obtained by training a pre-built segmentation model through a tooth sample image, and the pre-built segmentation model comprises an encoder provided with a double-scale cooperative attention module, a layered decoder provided with a jumping fusion door module and the multi-scale fusion head. According to one aspect of an embodiment of the present application, there is provided an electronic device including one or more processors, and storage means for storing one or more programs which, when executed by the one or more processors, cause the electronic device to implement a CBCT dental image segmentation method as described above. The method has the beneficial effects that the CBCT tooth image to be segmented is acquired and input into the tooth image segmentation model to obtain a tooth segmentation result, the tooth image segmentation model is obtained by training a pre-built segmentation model through a tooth sample image, the pre-built segmentation model comprises an encoder provided with a double-scale cooperative attention module, a layered decoder provided with a jump fusion door module and a multi-scale fusion head, the pre-built segmentation model is obtained through the encoder provided with the double-scale cooperative attention module, the layered decoder provided with the jump fusion door module and the multi-scale fusion head, a cooperative self-adaptive fusion network architecture (Collaborative Adaptive Fusion Network, CAF-Net) is formed after the pre-built segmentation model is trained through the tooth sample image, the pre-built segmentation model is obtained, the CBCT tooth image is segmented through the tooth image segmentation model, adverse effects on the CBCT tooth image segmentation caused by problems such as metal artifacts and boundary blurring are reduced, the accuracy of the CBCT tooth image segmentation is improved, the dependence o