Image segmentation method and system for tooth root tip image

Abstract

The invention relates to the technical field of tooth root apex image processing, and particularly discloses an image segmentation method and system for a tooth root apex image, and the method comprises the steps: 1, enabling the tooth root apex image to pass through a convolution initialization module, and obtaining an initial feature map; step 2, enabling the initial feature map to pass through an encoder module and a decoder module based on a Unet structure to obtain decoder features; wavelet convolution is introduced into the encoder module; and step 3, enabling the decoder features to pass through a segmentation head to obtain a segmented image. According to the method, wavelet convolution is introduced into medical image segmentation, the best effect is achieved on the root tip image segmentation task, and the multi-scale problem can be effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of tooth root apex image processing, and in particular to an image segmentation method and system for tooth root apex images. Background Art

[0002] Radiographic evaluation plays a key role in dentistry, enabling dentists to diagnose conditions, develop treatment plans, and evaluate treatment outcomes through a variety of imaging techniques. The main dental radiographic imaging modalities fall into three categories: panoramic radiography (PAN), cone-beam computed tomography (CBCT), and periapical radiography (PR).

[0003] Most existing methods focus on modalities such as oral panoramic X-rays or CBCT, and there are few intelligent segmentation methods for endodontic periapical radiographs. For example, Chinese invention patent application publication number CN118365964A discloses a method for identifying tooth position and periapical periodontitis in oral curved surface slices based on deep learning. The method uses an oral curved surface slice dataset to train a deep learning contour segmentation model; the oral curved surface slice dataset is input into the trained deep learning contour segmentation model, outputting tooth position and periapical periodontitis contour data, and converting the contour data features into sequence data to obtain a sequence dataset; the sequence dataset is used to train a deep learning sequence classification model; and the trained deep learning contour segmentation model and the trained deep learning sequence classification model are used to identify actual tooth position and periapical periodontitis. Another example is Chinese invention patent application publication number CN118135217A, which discloses an oral image recognition method based on image segmentation. The image is divided, fixed, and calculated by an image processing unit, and the randomness and diversity of image processing are increased by arbitrarily selecting starting points and diffraction.

[0004] PR is an intraoral imaging method in which sensors are strategically placed inside the patient's mouth to capture images, facilitating detailed examination of one or two teeth, including the pulp, root apex, and periodontal condition. While PAN can provide a comprehensive view of the dental arch, it lacks the local details provided by PR. In contrast, CBCT, as a 3D imaging modality, can provide detailed visualization of the teeth and surrounding bone structures. However, its high cost and increased radiation exposure make it less popular in endodontics. Therefore, it is necessary to use deep learning (DL) technology to perform intelligent analysis of PR. However, due to the significant multi-scale characteristics of root apex images, a method combining global and local features is required for their segmentation, and the existing methods have poor segmentation results. Summary of the Invention

[0005] The present invention aims to solve the above-mentioned problems. To this end, the present invention provides an image segmentation method and system for tooth apex images, which introduces wavelet convolution into medical image segmentation. Two levels of multi-scale wavelet convolution layers and convolution layers are used in the encoder. At the same time, a fusion channel attention mechanism is introduced in the skip connection part to enrich the multi-scale features of the decoder. This method achieves the best results in the task of tooth apex image segmentation and can effectively solve the multi-scale problem.

[0006] The present invention provides an image segmentation method for tooth root apex images, which adopts the following technical solutions: Step 1: Pass the root apex image through the convolution initialization module to obtain the initial feature map; Step 2: The initial feature map is passed through the encoder module and decoder module based on the Unet structure to obtain the decoder features; wavelet convolution is introduced in the encoder module; Step 3: The decoder features pass through the segmentation head to obtain the segmented image.

[0007] Furthermore, the encoder module includes multiple encoder sub-modules, and the encoder sub-modules include a multi-scale wavelet convolutional network module and a maximum pooling layer.

[0008] Furthermore, the working process of the multi-scale wavelet convolutional network module is as follows: After the input features are normalized through the layers, they pass through the first convolutional layer and the first wavelet convolutional layer respectively. After the two features are added, they pass through the second convolutional layer and the second wavelet convolutional layer respectively. After the two features are added, they pass through the convolutional layer, the Leaky Relu layer and the convolutional layer in sequence to obtain the output features.

[0009] Furthermore, the kernel size of the first convolutional layer and the first wavelet convolutional layer is 3, and the kernel size of the second convolutional layer and the second wavelet convolutional layer is 5.

[0010] Furthermore, when two features are added, a weighted summation method is adopted; the weights corresponding to the output features of the first convolutional layer and the second convolutional layer are the same, and the weights corresponding to the output features of the first wavelet convolution and the second wavelet convolution are the same.

[0011] Furthermore, the encoder module includes four encoder sub-modules connected in sequence.

[0012] Furthermore, a CFA module is provided on each jump connection between the encoder module and the decoder module.

[0013] Furthermore, the working process of the CFA module is: The input feature map , divided into blocks of size s and 2s, and the feature map is obtained and ; Will 、 and After passing through the average pooling layer, the features are averaged along the H and W dimensions. and The average pooled features are randomly grouped according to s and 2s and summed to obtain the feature map 、 and ; Will 、 and Connect them, shuffle them and average them along the channel dimension, then pass them through a point-by-point convolution layer and a sigmoid function to obtain the attention map A; A and Multiply to get the features .

[0014] Furthermore, the categories in the segmented image include tooth, alveolar bone, root canal, root canal filling, dental implant, dental filling, denture crown, apical periodontitis, and orthodontic appliance.

[0015] The present invention also provides an image segmentation system for tooth root apex images, which adopts the following technical solutions: Convolution initialization module, used to perform convolution initialization on the root apex image to obtain the initial feature map; The encoder-decoder module is used to encode and decode the initial feature map to obtain decoder features; wavelet convolution is introduced in the encoder-decoder module; The segmentation head is used to segment the decoder features to obtain a segmented image.

[0016] The above one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: 1. The present invention introduces wavelet convolution and uses a multi-scale wavelet convolution module and a fusion channel attention mechanism to perform pixel-level semantic segmentation of oral apical imaging organs and lesions in endodontics.

[0017] 2. This invention uses two layers of multi-scale wavelet convolutional layers and convolutional layers in the encoder. The wavelet convolutional layer branches are responsible for capturing global features with a large receptive field, while the convolutional layer branches extract local details. The two complement each other. Two independent weights are used for weighting and fusion, allowing the network to adaptively and dynamically adjust the importance of global and local features during the fusion process.

[0018] 3. Experiments show that the present invention surpasses the most advanced medical image segmentation network on the PRAD-10K dataset and can effectively solve the multi-scale challenges inherent in PR image segmentation tasks.

[0019] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0021] Figure 1 It is a flow chart of the method provided by the present invention.

[0022] Figure 2 This is a flow chart of the multi-scale wavelet convolutional network module provided by the present invention.

[0023] Figure 3 It is a flow chart of the CFA module provided by the present invention.

[0024] Figure 4 This is a segmentation effect diagram provided by the present invention. DETAILED DESCRIPTION

[0025] To make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the embodiments described are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0026] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiment of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0027] The following combination Figures 1 to 4The present invention is further described in detail, and an image segmentation method and system for tooth root apex images of the present invention are described as follows: In this embodiment, Figure 1 As shown, a method for image segmentation of root apex images is provided, comprising the following steps: Step 1: Pass the root apex image through the convolution initialization module to obtain the initial feature map.

[0028] The convolution initialization module includes 1X1 convolution layer, 3X3 convolution layer and Rule layer. The root apex image is used as the input image. , input the convolution initialization module, and pass through the 1X1 convolution layer, 3X3 convolution layer and Rule layer in sequence to obtain the initial feature map .

[0029] Step 2: The initial feature map is passed through the encoder module and decoder module based on the Unet structure to obtain the decoder features; wavelet convolution is introduced in the encoder module.

[0030] The encoder module includes multiple encoder sub-modules, which include a multi-scale wavelet convolutional network module (MWCN module) and a maximum pooling layer.

[0031] In this embodiment, the encoder module includes four encoder submodules connected in sequence. Among the four encoder submodules, the number of multi-scale wavelet convolutional network modules is 、 、 and .

[0032] The large receptive field of the wavelet convolution layer is crucial for learning global knowledge in PR images. However, in order to ensure that both global features and local details can be captured, this embodiment designs a multi-scale wavelet convolution network module, such as Figure 2 As shown in the figure, two layers of feature extraction are performed, with two branches in each layer: the convolutional layer branch and the wavelet convolutional layer branch. The wavelet convolutional layer branch is responsible for capturing global features with a large receptive field, while the convolutional layer branch extracts local details. The two complement each other.

[0033] like Figure 2 As shown in Figure 2, the working process of the multi-scale wavelet convolutional network module is as follows: Input features After layer normalization, we get ; The first layer of feature extraction is performed, and two features are obtained by passing through the first convolution layer and the first wavelet convolution layer respectively. After the two features are added, the second layer of feature extraction is performed, and two features are obtained by passing through the second convolution layer and the second wavelet convolution layer respectively. After the two features are added, the feature , Then pass through the 1X1 convolution layer, the Leaky Relu layer and the 1X1 convolution layer in turn to obtain the feature map .

[0034] For the first multi-scale wavelet convolutional network module in the encoder module, its input feature is the initial feature map .

[0035] In this embodiment, the kernel size (K) of the first convolution layer and the first wavelet convolution layer is 3, and the kernel size of the second convolution layer and the second wavelet convolution layer is 5.

[0036] In this embodiment, during the two-layer feature extraction process, a weighted summation approach is used when adding the two features, and two independent global-local feature weighting matrices (GFWMs) are used to weight and fuse the two features. Specifically, the features output by the first and second convolutional layers have the same weights, both using the weighting matrix α; the features output by the first and second wavelet convolution layers have the same weights, both using the weighting matrix β.

[0037] In this embodiment, a CFA module is provided at each skip connection between the encoder and decoder modules. The CFA module's primary function is to weight features from a channel-wise perspective, integrating local features at different levels within the feature layer. This process enhances the decoder's ability to recognize objects of varying sizes by feeding feature layers rich in multi-scale information into the corresponding decoder.

[0038] like Figure 3 As shown in Figure 2, the working process of the CFA module is: The input feature map , divided into blocks of size s and 2s, and the feature map is obtained and ; In this embodiment, s is set to 2; Will After the average pooling layer, the features are averaged along the H and W dimensions to obtain the feature map ;Will After the average pooling layer, The average pooled features are randomly grouped according to s and summed to integrate local features and reduce the dimension to obtain the feature map ;Will After the average pooling layer, The average pooled features are randomly grouped according to 2s and summed to obtain the feature map ; Will 、 and Connect them, shuffle the channels, and take the average along the channel dimension, then pass through the point-by-point convolution layer and sigmoid function to obtain the attention map A; A and Multiply to get the features , completing channel weighting under multi-scale information.

[0039] Assume that the feature maps of the output of the multi-scale wavelet convolutional network module in the four encoder submodules are 、 、 and . 、 、 and It is necessary to input the CFA module separately, obtain the weighted feature map through the CFA mechanism, and input the features output by the CFA module into the decoder module. After passing through the maximum pooling layer, it is input to the decoder module.

[0040] The structure of the decoder module is consistent with that of the UNet. Each decoder in the decoder module consists of a transposed convolution, a 3x3 convolution, and a ReLU activation function. The input features are first upsampled by the transposed convolution, then concatenated with the features of the corresponding dimension along the channel dimension. The concatenated features are then fed into the 3x3 convolution and activation function, and then output to the next decoder, where the above process repeats.

[0041] Step 3: The decoder features pass through the segmentation head, which processes them to generate a segmentation mask and then obtain the segmentation image Y.

[0042] In this embodiment, the segmentation head includes a 3x3 convolutional layer and a 1x1 convolutional layer. The categories in the segmented image include teeth, alveolar bone, root canal, root canal filling, dental implant, tooth filling, denture crown, apical periodontitis, and orthodontic appliance. Figure 4 As shown, four root apex images and their segmented images after processing are displayed.

[0043] This example verifies the quantitative comparison of this method (PRNet) with other currently optimal medical image segmentation methods on the PRAD-10K dataset. Table 1 shows the Dice Similarity Coefficient (DSC) used as the evaluation metric. Tooth, Bone, Pulp, RCF, DC, DF, IM, OD, and AP represent tooth, alveolar bone, root canal, root canal filling, dental implant, dental filling, denture crown, orthodontic appliance, and apical periodontitis, respectively. Avg. represents the average value.

[0044] Table 1 Quantitative comparison results of this method with other current optimal medical image segmentation methods This embodiment also conducted an ablation experiment, and the results are shown in Table 2. In Table 2, √ indicates inclusion and o indicates exclusion.

[0045] Table 2 Ablation experiment results In this embodiment, an image segmentation system for root apex images is also provided, which adopts the following technical solutions: Convolution initialization module, used to perform convolution initialization on the root apex image to obtain the initial feature map; The encoder-decoder module is used to encode and decode the initial feature map to obtain decoder features; wavelet convolution is introduced in the encoder-decoder module; The segmentation head is used to segment the decoder features to obtain a segmented image.

[0046] 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 make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An image segmentation method for tooth root apex images, characterized in that: include: Step 1: Pass the root apex image through the convolution initialization module to obtain the initial feature map; Step 2: Pass the initial feature map through the encoder module and decoder module based on the Unet structure to obtain the decoder features; Wavelet convolution is introduced in the encoder module; Step 3: The decoder features pass through the segmentation head to obtain the segmented image.

2. The image segmentation method for tooth root apex images according to claim 1, characterized in that: The encoder module includes multiple encoder sub-modules, and the encoder sub-modules include a multi-scale wavelet convolutional network module and a maximum pooling layer.

3. The image segmentation method for root apex images according to claim 2, characterized in that: The working process of the multi-scale wavelet convolutional network module is: After the input features are normalized through the layers, they pass through the first convolutional layer and the first wavelet convolutional layer respectively. After the two features are added, they pass through the second convolutional layer and the second wavelet convolutional layer respectively. After the two features are added, they pass through the convolutional layer, the Leaky Relu layer and the convolutional layer in sequence to obtain the output features.

4. The image segmentation method for root apex images according to claim 3, characterized in that: The kernel size of the first convolution layer and the first wavelet convolution layer is 3, and the kernel size of the second convolution layer and the second wavelet convolution layer is 5.

5. The image segmentation method for root apex images according to claim 3, characterized in that: When two features are added, a weighted summation method is adopted; the weights corresponding to the output features of the first convolutional layer and the second convolutional layer are the same, and the weights corresponding to the output features of the first wavelet convolution and the second wavelet convolution are the same.

6. The image segmentation method for tooth root apex images according to any one of claims 2 to 5, characterized in that: The encoder module includes four encoder sub-modules connected in sequence.

7. The image segmentation method for root apex images according to claim 1, characterized in that: A CFA module is set on each jump connection between the encoder module and the decoder module.

8. The image segmentation method for tooth root apex images according to claim 7, characterized in that: The working process of the CFA module is: The input feature map , divided into blocks of size s and 2s, and the feature map is obtained and ; Will 、 and After passing through the average pooling layer, the features are averaged along the H and W dimensions. and The average pooled features are randomly grouped according to s and 2s and summed to obtain the feature map 、 and ; Will 、 and Connect them, shuffle them and take the average along the channel dimension, then pass them through the point-by-point convolution layer and sigmoid function to obtain the attention map A; A and Multiply to get the features .

9. The image segmentation method for root apex images according to claim 1, wherein: The categories in the segmented images include tooth, alveolar bone, root canal, root canal filling, dental implant, dental filling, denture crown, apical periodontitis, and orthodontic appliance.

10. An image segmentation system for tooth root apex images, characterized in that: The method for performing an image segmentation method for a tooth root apex image according to any one of claims 1 to 9 comprises: Convolution initialization module, used to perform convolution initialization on the root apex image to obtain the initial feature map; The encoder-decoder module is used to encode and decode the initial feature map to obtain decoder features; wavelet convolution is introduced in the encoder-decoder module; The segmentation head is used to segment the decoder features to obtain a segmented image.

Images