Variable convolution UNet tooth image segmentation method and system fusing attention mechanism

Abstract

The invention discloses a variable convolution UNet tooth image segmentation method and system fusing an attention mechanism, and belongs to the technical field of medical image processing and computer-aided diagnosis. A tooth image data set is obtained and preprocessed; using the preprocessed data set to train a DSC-UNet segmentation model; the DSC-UNet segmentation model introduces a multi-modal feature fusion module and a deformable convolution module into a UNet framework; and segmenting an input tooth image by using the trained DSC-UNet segmentation model, and outputting a pixel-level segmentation result of the tooth structure. By constructing a multi-level feature interaction mechanism and a deformable convolution module, the capturing capability of the network on tooth edge details and the segmentation robustness of special-shaped teeth are remarkably improved, more reliable technical support is provided for an intelligent dental diagnosis and treatment system, and the method is particularly suitable for high-precision segmentation of complex tooth structures in oral panoramic X-ray films and CBCT images.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and computer-aided diagnosis technology, specifically relating to a variable convolutional UNet tooth image segmentation method and system that integrates attention mechanisms. Background Technology

[0002] With the development of digital oral diagnosis and treatment technology, automatic tooth structure segmentation technology based on medical images has become a core requirement of modern dental diagnosis and treatment systems. Accurate tooth segmentation in images such as panoramic X-rays and cone-beam computed tomography (CBCT) is of great value for clinical decisions such as orthodontic planning, implant placement, and periodontal disease diagnosis. However, due to technical challenges in dental imaging, such as low inter-tissue contrast, large variations in tooth morphology, and blurred edges of pathological structures, traditional image processing methods (such as thresholding and region growing) struggle to achieve reliable automated segmentation.

[0003] In recent years, semantic segmentation technology based on deep learning has provided new solutions for medical image analysis. The encoder-decoder architecture represented by UNet (Ronneberger et al., 2015) has made breakthroughs in medical scenarios such as cell microscopy image segmentation due to its symmetrical structure and skip connection design. However, the original UNet still has significant limitations when processing dental images: (1) the rigid geometric constraints of conventional convolution operations are difficult to adapt to complex morphological features such as root curvature and adjacent tooth overlap; (2) the single-channel attention mechanism is not capable of selecting features of multi-scale dental structures, which can easily lead to missegmentation of key anatomical structures such as the enamel-dentin junction; (3) the problem of deep feature loss is particularly prominent in low-contrast CBCT images, affecting the segmentation accuracy of thin-walled tissues such as periodontal ligament.

[0004] To address these issues, researchers have proposed several improvement schemes: Zhang et al. (2018) introduced residual connections in UNet to enhance gradient propagation, but the improvement in modeling ability for deformed structures was limited; Gu et al. (2019) proposed Attention UNet to enhance the attention of the target region through a gating mechanism, but its single-channel attention could not simultaneously optimize the feature correlation between spatial and channel dimensions; Chen et al. (2021) used deformable convolution to improve the encoder, which improved the accuracy of crown segmentation, but failed to effectively solve the problem of deep root feature loss. Existing technologies indicate that how to achieve accurate modeling and context awareness of multi-scale deformed structures in dental images remains a key technical bottleneck restricting clinical application. Summary of the Invention

[0005] Based on the aforementioned technical deficiencies, this invention proposes a variable convolutional UNet tooth image segmentation method and system that integrates an attention mechanism. By constructing a multi-level feature interaction mechanism and a deformable convolutional module, the network's ability to capture tooth edge details and the robustness of segmenting irregular teeth are significantly improved, providing more reliable technical support for intelligent dental diagnosis and treatment systems.

[0006] The technical solution adopted in this invention is as follows:

[0007] In a first aspect, this invention discloses a variable convolutional UNet tooth image segmentation method incorporating an attention mechanism, comprising:

[0008] Acquire and preprocess dental image datasets;

[0009] Train a DSC-UNet segmentation model using a preprocessed dataset;

[0010] The DSC-UNet segmentation model introduces a multimodal feature fusion module and a deformable convolution module into the UNet architecture. The multimodal feature fusion module contains a spatial-channel collaborative attention mechanism located between the encoder and decoder. It is used to fuse different-dimensional encoded features obtained from the encoding results of two adjacent layers based on the spatial-channel collaborative attention mechanism to generate a feature of the same size as the low-dimensional encoded feature. This feature is then concatenated and fused with the decoded feature of the same size generated after upsampling in the corresponding decoding layer according to the channel dimension. The high-dimensional encoded feature guides the generation of attention weights for the low-dimensional encoded feature, realizing cross-scale feature interaction. The deformable convolution module replaces the double convolutional layer of the decoder in the UNet architecture and enhances the modeling ability of deformable structures by dynamically predicting the sampling point offset.

[0011] The trained DSC-UNet segmentation model is used to segment the input tooth image, and the output tooth structure pixel-level segmentation results are obtained.

[0012] Furthermore, the preprocessing includes resolution standardization using an adaptive scaling strategy of bilinear or nearest neighbor interpolation, pixel value normalization, and channel dimension reconstruction.

[0013] Furthermore, in the decoder path, the cascaded position of the multimodal feature fusion module is located before the deformable convolution module. Except for the deepest layer of the UNet architecture, each layer is configured with a multimodal feature fusion module.

[0014] The calculation process of the multimodal feature fusion module includes:

[0015] The low-dimensional and high-dimensional encoded features of the input are aligned by 1x1 convolutional projection, bilinear interpolation is used to solve the resolution difference, and basic fusion features are constructed by element-wise addition and ReLU activation.

[0016] The basic fusion features are used to compute spatial attention and channel attention in parallel.

[0017] The two attention results are dynamically weighted and fused using learnable parameters, and the optimized fused feature is output after Sigmoid gating and residual connection.

[0018] Furthermore, in the decoder path, the input high-dimensional decoding features are first upsampled to obtain low-dimensional decoding features, which are then concatenated with the fusion features of the same size according to the channel dimension before the deformable convolution module is executed to complete one layer of decoding. After the last layer of decoding is completed, a 1×1 convolution is used to obtain the pixel-level segmentation result of the tooth structure with 1 channel.

[0019] Furthermore, the size of the basic fusion feature is the same as the size of the low-dimensional encoded feature.

[0020] Furthermore, the fusion feature optimized by Sigmoid gating and residual connection output includes:

[0021] The weighted sum of the two attention results is calculated using Sigmoid gating to obtain the first attention weight map;

[0022] Feature enhancement is performed by multiplying the first attention weight map with the low-dimensional encoded features channel by channel. The feature enhancement result is then concatenated with the low-dimensional encoded feature residuals to generate the optimized fused features.

[0023] Furthermore, the spatial-channel collaborative attention mechanism achieves collaborative optimization of spatial domain and channel domain attention through a dual-branch parallel structure.

[0024] Furthermore, during the spatial attention calculation, the max pooling feature and average pooling feature of the basic fusion features are calculated respectively. The two pooling features are aggregated using a 7x7 convolution. The aggregated features are then processed by Sigmoid gating to generate a second attention weight map as the spatial attention result.

[0025] Furthermore, during the channel attention calculation, the spatial information of each channel of the basic fusion feature is compressed into a scalar through global average pooling and max pooling, respectively, thereby obtaining the channel-level global statistical features. Subsequently, the dependencies between channels are learned through fully connected layers, a set of channel weight vectors are generated and added together, and then a third attention weight map is generated through Sigmoid gating calculation. This map is multiplied with the basic fusion feature channel by channel, thereby dynamically adjusting the response intensity of each channel and generating the channel attention result.

[0026] Furthermore, a composite loss function is used to train the DSC-UNet segmentation model, including cross-entropy loss and Dice loss.

[0027] Secondly, this invention discloses a variable convolutional UNet tooth image segmentation system with an attention fusion mechanism, used to implement the aforementioned variable convolutional UNet tooth image segmentation method with an attention fusion mechanism.

[0028] The beneficial effects of this invention are:

[0029] This invention proposes an improved scheme for dental image segmentation that integrates spatial channel collaborative attention and deformable convolution. Through a multimodal feature fusion module, it dynamically guides low-level features with high-level semantic features, significantly improving segmentation accuracy and clinical applicability. Its core advantages are:

[0030] High-level semantic features guide the generation of low-level feature weights, and a dual-branch parallel structure enables cross-scale feature interaction. Specifically, high-dimensional encoded features are aligned with low-dimensional features after 1×1 convolutional projection and bilinear interpolation to generate basic fused features. Subsequently, spatial attention and channel attention are computed in parallel, and the two types of attention results are dynamically weighted and fused using learnable parameters. This high-level semantic-driven attention mechanism enables the model to adaptively focus on key structures. For example, in the root curvature region, channel attention is strengthened to enhance the response of low-contrast features and avoid fractures caused by the loss of deep features; in the crown margin region, spatial attention is emphasized to sharpen the boundary gradient and accurately separate adjacent tooth adhesions; and at the enamel-dentin junction, the two types of attention are balanced to improve sub-pixel-level segmentation accuracy.

[0031] The collaborative design of deformable convolution modules further enhances deformation adaptability. By dynamically predicting the offset vector of convolution kernel sampling points, the feature extraction process can conform to the irregular geometry of teeth, especially solving the modeling challenges of root curvature and thin-walled periodontal ligament structures. The combination of these two approaches forms a multi-level optimization system, ultimately achieving leading performance in core metrics such as Dice, IoU, Precision, and Accuracy on the MICCAI dataset, providing a highly reliable anatomical analysis basis for clinical orthodontic planning and implant placement. Attached Figure Description

[0032] Figure 1 This is a flowchart of the variable convolutional UNet tooth image segmentation method with attention mechanism in this invention;

[0033] Figure 2 This is a schematic diagram of the network structure of the DSC-UNet model in this invention;

[0034] Figure 3 This is a schematic diagram of the dual convolutional structure used in the encoding and decoding parts of the DSC-UNet model in this invention;

[0035] Figure 4 This is a schematic diagram of the structure of the multimodal feature fusion module in this invention. Detailed Implementation

[0036] The present invention will be further described and illustrated below with reference to specific embodiments. The embodiments described are merely examples of the content of this disclosure and do not limit the scope of the invention. The technical features of each embodiment in the present invention can be combined accordingly, provided that there is no mutual conflict.

[0037] The accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0038] The flowchart shown in the attached diagram is merely an illustrative example and does not necessarily include all steps. For example, some steps may be broken down, while others may be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0039] like Figure 1 As shown, this invention proposes a variable convolutional UNet tooth image segmentation method with an attention mechanism, which is suitable for high-precision segmentation of complex tooth structures in panoramic dental X-rays and CBCT images. The method mainly includes the following steps:

[0040] S1: Acquire and preprocess dental image datasets;

[0041] In this embodiment, the MICCAI-Tooth-Segmentation dataset is used. This dataset is the one used in the competition focusing on tooth segmentation at the International Conference on Medical Image Computation and Computer-Assisted Intervention (MICCAI). It contains 1800 24-bit original images and labels. The label refers to the precise annotation of which category each pixel in the original image should belong to. It is an image with the exact same size as the original image, but the value of each pixel no longer represents color information; instead, it is a category identifier. Here, teeth are used as the target region for segmentation. Pixels in the background region are labeled as 0, and each tooth in the image is precisely labeled. All pixels belonging to the tooth region are labeled with a specific integer value. All teeth can be grouped into one category: labeled as 1; or different teeth can be distinguished: the first molar is labeled as 1, the second molar as 2, the incisor as 3, and so on. The labeled image is called the mask image.

[0042] The dataset is preprocessed. This embodiment employs an adaptive scaling strategy using bilinear or nearest-neighbor interpolation to simultaneously normalize the resolution of the original image and the mask image, and performs pixel value normalization and channel dimension reconstruction. A specially designed tensor transformation mechanism ensures data continuity through memory copy optimization, supporting end-to-end deep learning model training.

[0043] S2, Construct a DSC-UNet segmentation model for tooth image segmentation;

[0044] This embodiment uses U-Net as the base model. U-Net is a symmetric encoder-decoder convolutional neural network specifically designed for image segmentation. Its encoder gradually extracts abstract features through convolution and downsampling operations, while the decoder recovers spatial information through upsampling. Furthermore, it fuses shallow high-resolution features with deep semantic features across layers via skip connections, effectively preserving local detail accuracy while significantly enhancing global context awareness. This unique architecture not only alleviates the information loss problem during downsampling but also supports end-to-end training—the input original image can directly output pixel-level segmentation results without complex post-processing.

[0045] like Figure 2 As shown, the DSC-UNet segmentation model constructed in this invention introduces a multimodal feature fusion module and a deformable convolution module into the UNet architecture. The multimodal feature fusion module contains a spatial-channel collaborative attention mechanism, located between the encoder and decoder. It is used to fuse the different-dimensional encoded features obtained from the encoding results of two adjacent layers based on the spatial-channel collaborative attention mechanism to generate a feature of the same size as the low-dimensional encoded feature. This feature is then concatenated and fused with the decoded feature of the same size generated after upsampling in the corresponding decoding layer according to the channel dimension. The high-dimensional encoded feature guides the generation of attention weights for the low-dimensional encoded feature, thereby achieving cross-scale feature interaction. The deformable convolution module replaces the double convolutional layer of the decoder in the UNet architecture and enhances the modeling ability of deformable structures by dynamically predicting the sampling point offset.

[0046] Figure 3This is a schematic diagram of the dual convolutional structure used in the encoding and decoding parts of the DSC-UNet segmentation model of this invention. (a) is applied to the encoder part, where in each encoding layer, the dual convolutional structure shown in (a) is used for processing before downsampling; (b) is applied to the decoder part, where in each decoding layer, the dual convolutional structure shown in (b) is used for processing before upsampling. Here, the dual convolutional structure shown in (b) uses deformable convolution instead of traditional convolution. Deformable convolution introduces learnable spatial offset parameters into traditional convolution operations. It generates convolutional layers using at least one offset to operate on the input feature map, dynamically predicting the position offset vector of each convolutional kernel sampling point. It uses bilinear interpolation to sample features from the offset non-grid points and generates an output feature map through weighted aggregation. This invention enhances the modeling ability for target geometric deformation, occlusion, and complex spatial distributions by adaptively adjusting the sampling position of the convolutional kernel.

[0047] In one specific embodiment of the present invention, in the decoder path, the cascaded position of the multimodal feature fusion module is located before the deformable convolution module, and each layer except the deepest layer of the UNet architecture is configured with a multimodal feature fusion module;

[0048] like Figure 4 As shown, the calculation process of the multimodal feature fusion module includes:

[0049] The input low-dimensional encoded features and high-dimensional encoded features are aligned by 1x1 convolutional projection, and bilinear interpolation is used to solve the resolution difference. The basic fusion features are constructed by element-wise addition and ReLU activation. The size of the basic fusion features is the same as the size of the low-dimensional encoded features.

[0050] The basic fusion features are used to perform parallel computation of spatial attention and channel attention.

[0051] Two attention results are dynamically weighted and fused using a learnable scaling parameter α. The optimized fused feature is then output after Sigmoid gating and residual connection. Specifically, the weighted result of the two attention results is calculated using Sigmoid gating to obtain the first attention weight map. The first attention weight map is then multiplied channel-by-channel with the low-dimensional encoded feature to perform feature enhancement. The feature enhancement result is then connected with the residual of the low-dimensional encoded feature to generate the optimized fused feature.

[0052] by Figure 4Taking the illustrated dimension as an example, the multimodal feature fusion module (SCB module) performs cross-modal projection alignment on the input low-dimensional feature x and high-dimensional feature g through 1x1 convolution. After using bilinear interpolation to solve the feature map resolution difference, it constructs the feature fusion basis through element-wise addition and ReLU activation. In the attention generation stage, the module computes spatial attention (aggregating max / average pooled features through 7x7 convolution kernels) and channel attention (through channel compression-activation mechanism) in parallel. It introduces a learnable scaling parameter α and a balance coefficient β to achieve dynamic weighted fusion of the two attentions: Combined_att = α*(β*spatial_att + (1-β)*channel_att), where spatial_att is the spatial attention result, channel_att is the channel attention result, and Combined_att is the fusion result. Finally, the optimized features are output through Sigmoid gating and residual connections. This design significantly enhances the model's ability to integrate cross-scale features, demonstrating improved capture of subtle structural features in tasks such as medical image segmentation and multimodal object detection, and is particularly suitable for scenarios requiring simultaneous modeling of local details and global semantics. Here, learnable weight parameters allow the network to adaptively adjust the contribution ratio of spatial and channel attention, offering greater flexibility in feature representation compared to traditional sequential attention structures. Assuming... Figure 2 The original input image size is 572×572×1. After one round of double convolution, the feature map becomes 568×568×64. This is the low-dimensional feature map input x of the SCB module. The high-dimensional feature map input y is the feature map after downsampling from x and undergoing double convolution again. The image size should be 280×280×128. After x and y enter the SCB module, the dimensions are aligned and spatial-channel collaborative attention mechanism is performed as described above. Finally, the output size is still the same as x, i.e., 568×568×64.

[0053] In the decoder path, the input high-dimensional decoding features are first upsampled to obtain low-dimensional decoding features, which are then concatenated with the fusion features of the same size according to the channel dimension before the deformable convolution module is executed to complete one layer of decoding. After the last layer of decoding is completed, a 1×1 convolution is used to obtain the pixel-level segmentation result of the tooth structure with 1 channel.

[0054] exist Figure 4 The multimodal feature fusion module introduces a spatial-channel collaborative attention mechanism, which achieves collaborative optimization of spatial and channel domain attention through a dual-branch parallel structure.

[0055] In one specific embodiment of the present invention, during the spatial attention calculation, the max-pooling feature and average-pooling feature of the basic fusion features are calculated separately. A 7x7 convolution is used to aggregate the two pooling features. The aggregated features are then processed by a Sigmoid gate to generate a second attention weight map as the spatial attention result. Spatial attention is a dynamic feature selection mechanism in deep learning. By generating a weight map with the same spatial dimension as the input feature map, it guides the model to focus on key regions of the image, thereby improving the recognition accuracy of fine structures in complex backgrounds.

[0056] During the channel attention calculation, the spatial information of each channel of the basic fusion feature is compressed into a scalar through global average pooling and max pooling, respectively, thereby obtaining the channel-level global statistical features. Then, the dependencies between channels are learned through fully connected layers, and a set of channel weight vectors are generated and added together. Then, a third attention weight map is generated by Sigmoid gating, which is multiplied with the basic fusion feature channel by channel, thereby dynamically adjusting the response intensity of each channel and generating the channel attention result.

[0057] This invention innovatively integrates deformable convolution and a multimodal feature fusion module based on spatial-channel attention mechanism in the UNet architecture. In the decoder path, a cascaded SCB module realizes spatial attention interaction of cross-scale features. High-level semantic features (such as x5) guide the generation of attention weights for low-level features (such as x4), enhancing spatial focusing of the target region. At the same time, the conventional double convolution of the decoder is replaced with a deformable convolution module. The feature sampling position is adaptively adjusted through the dynamic deformation characteristics of the deformable convolution kernel. The two work together to improve the boundary sensitivity of deformable targets and the feature completion capability in occluded scenes. Finally, while maintaining the lightweight backbone network, it significantly enhances the segmentation accuracy and robustness of irregular lesion regions in tooth image segmentation.

[0058] S3, Training the DSC-UNet segmentation model

[0059] In this embodiment, a dynamically adjusted learning rate mechanism is used during training, with the initial learning rate set to 1×10. -5 The data collected in S1 is divided into training, testing, and validation sets. A ReduceLROnPlateau strategy is implemented based on the validation set splitting accuracy metric. When the metric stabilizes, the learning rate decays over a patience period of 5 epochs. The Adam optimization algorithm is used as the optimizer, configured with an L2 weight regularization term and a weight decay coefficient of 1×10⁻⁶. -8 The gradient clipping threshold is set to 5.0; the loss function is a combination of cross-entropy loss and Dice loss, with Softmax normalization used for multi-class segmentation tasks and Sigmoid activation used for single-class tasks.

[0060] Cross-entropy loss measures the difference between the classification probability distribution of each pixel and the true label. Essentially, it optimizes the classification error pixel by pixel to ensure pixel-level classification accuracy. Dice loss directly optimizes the spatial overlap between the segmented region and the true label. It is derived from the Dice coefficient, a medical image evaluation metric, to ensure the integrity of the overall structure.

[0061] S4 uses the trained DSC-UNet segmentation model to segment the input tooth image and outputs pixel-level segmentation results of the tooth structure.

[0062] In this step, the trained DSC-UNet segmentation model can be used to perform tooth image segmentation. After preprocessing the tooth image, it can be directly input into the DSC-UNet segmentation model to generate a mask image of the same resolution, which represents the pixel-level segmentation result of the tooth structure.

[0063] To verify the effectiveness of this invention, Table 1 compares the performance of this invention (DSCUNet) with four other models (Unet, SwinUNet, TransUNet, and AttentionUNet) on six evaluation metrics. These metrics are used to evaluate the performance of medical image segmentation or similar classification tasks. The metrics include:

[0064] Dice coefficient (Dice): measures the similarity between the predicted segment and the true label (range 0-1, the higher the better, 1 indicates a perfect match).

[0065] IoU (Intersection over Union): Measures the degree of overlap between the predicted region and the actual region (range 0-1, the higher the better).

[0066] Precision: The proportion of actual positive cases among those predicted as positive (range 0-1, the higher the better, a high value indicates fewer false positives).

[0067] Sensitivity (or recall): This is the proportion of positive cases that are correctly predicted (range 0-1, the higher the better, a high value indicates fewer false negatives).

[0068] Specificity: The proportion of negative cases that are correctly predicted (range 0-1, the higher the better, a high value indicates fewer false positives).

[0069] Accuracy: The percentage of correct predictions out of all predictions (range 0-1, the higher the better).

[0070] Table 1 Experimental Results

[0071]

[0072] Analyzing the table from the perspective of overall performance and key metrics, it can be seen that DSCUNet performs best among all models, especially leading in core metrics such as Dice, IoU, Precision, and Accuracy. AttentionUNet performs well in some metrics, but there are obvious trade-offs; Unet, SwinUNet, and TransUNet are relatively weaker. This indicates that the model proposed in this invention is a highly balanced and robust model, suitable for tasks with high overall performance requirements, such as medical image segmentation.

[0073] In the multimodal feature fusion module of this invention, high-level semantic features guide the generation of low-level feature weights, enhancing the ability to preserve features of deep structures such as tooth roots. By segmenting tooth images using this technique, the influence of gingiva and other structures on tooth images can be eliminated, providing assistance for medical diagnosis. This model exhibits high accuracy, good convergence and stability, and has broad application prospects in the field of computer vision.

[0074] This embodiment also provides a variable convolutional UNet tooth image segmentation system that incorporates an attention mechanism, including:

[0075] The dental image data acquisition module is used to acquire dental image datasets and perform preprocessing.

[0076] The DSC-UNet segmentation model module introduces a multimodal feature fusion module and a deformable convolution module into the UNet architecture. The multimodal feature fusion module includes a spatial-channel collaborative attention mechanism located between the encoder and decoder. It is used to fuse different-dimensional encoded features obtained from two adjacent encoding layers based on the spatial-channel collaborative attention mechanism to generate a feature of the same size as the low-dimensional encoded feature. This feature is then concatenated and fused with the decoded feature of the same size generated after upsampling from the corresponding decoding layer according to the channel dimension. The high-dimensional encoded feature guides the generation of attention weights for the low-dimensional encoded feature, achieving cross-scale feature interaction. The deformable convolution module replaces the double convolutional layer of the decoder in the UNet architecture and enhances the modeling ability of deformable structures by dynamically predicting the sampling point offset.

[0077] The model training module is used to train a DSC-UNet segmentation model using a preprocessed dataset.

[0078] The tooth image segmentation module is used to segment the input tooth image using the trained DSC-UNet segmentation model and output the pixel-level segmentation results of the tooth structure.

[0079] For the system embodiments, since they basically correspond to the method embodiments, relevant details can be found in the descriptions of the method embodiments; the implementation methods of the remaining modules will not be repeated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0080] The system embodiments of the present invention can be applied to any device with data processing capabilities, such as a computer or other similar device. The system embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution.

[0081] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. Those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A variable convolutional UNet tooth image segmentation method incorporating an attention mechanism, characterized in that, include: Acquire and preprocess dental image datasets; Train a DSC-UNet segmentation model using a preprocessed dataset; The DSC-UNet segmentation model introduces a multimodal feature fusion module and a deformable convolution module into the UNet architecture. The multimodal feature fusion module contains a spatial-channel collaborative attention mechanism, located between the encoder and decoder. It is used to fuse the different-dimensional encoded features obtained from the encoding results of two adjacent layers based on the spatial-channel collaborative attention mechanism to generate a feature of the same size as the low-dimensional encoded feature. This feature is then concatenated and fused with the decoded feature of the same size generated after upsampling in the corresponding decoding layer according to the channel dimension. The high-dimensional encoded feature guides the generation of attention weights for the low-dimensional encoded feature, thereby achieving cross-scale feature interaction. The deformable convolutional module replaces the dual convolutional layer of the decoder in the UNet architecture, and enhances the ability to model deformable structures by dynamically predicting the offset of sampling points. The trained DSC-UNet segmentation model is used to segment the input tooth image, and the output tooth structure pixel-level segmentation results are obtained.

2. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 1, characterized in that, The preprocessing includes resolution standardization using an adaptive scaling strategy with bilinear or nearest neighbor interpolation, pixel value normalization, and channel dimension reconstruction.

3. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 1, characterized in that, In the decoder path, the cascaded position of the multimodal feature fusion module is located before the deformable convolution module. Except for the deepest layer of the UNet architecture, each layer is configured with a multimodal feature fusion module. The calculation process of the multimodal feature fusion module includes: The low-dimensional and high-dimensional encoded features of the input are aligned by 1x1 convolutional projection, bilinear interpolation is used to solve the resolution difference, and basic fusion features are constructed by element-wise addition and ReLU activation. The basic fusion features are used to compute spatial attention and channel attention in parallel. The two attention results are dynamically weighted and fused using learnable parameters, and the optimized fused feature is output after Sigmoid gating and residual connection.

4. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 3, characterized in that, In the decoder path, the input high-dimensional decoding features are first upsampled to obtain low-dimensional decoding features, which are then concatenated with the fusion features of the same size according to the channel dimension before the deformable convolution module is executed to complete one layer of decoding. After the last layer of decoding is completed, a 1×1 convolution is used to obtain the pixel-level segmentation result of the tooth structure with 1 channel.

5. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 3, characterized in that, The size of the basic fusion feature is the same as the size of the low-dimensional encoded feature.

6. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 3, characterized in that, The fusion features optimized by Sigmoid gating and residual connection output include: The weighted sum of the two attention results is calculated using Sigmoid gating to obtain the first attention weight map; Feature enhancement is performed by multiplying the first attention weight map with the low-dimensional encoded features channel by channel. The feature enhancement result is then concatenated with the low-dimensional encoded feature residuals to generate the optimized fused features.

7. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 1, characterized in that, The spatial-channel collaborative attention mechanism achieves collaborative optimization of spatial and channel domain attention through a dual-branch parallel structure.

8. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 7, characterized in that, During the spatial attention calculation, the max pooling feature and average pooling feature of the basic fusion feature are calculated separately. The two pooling features are aggregated using a 7x7 convolution. The aggregated features are then processed by Sigmoid gating to generate a second attention weight map as the spatial attention result.

9. The variable convolutional UNet tooth image segmentation method with fused attention mechanism according to claim 7, characterized in that, During the channel attention calculation, the spatial information of each channel of the basic fusion feature is compressed into a scalar by global average pooling and max pooling, respectively, thereby obtaining the channel-level global statistical features; Subsequently, the dependencies between channels are learned through the fully connected layer, a set of channel weight vectors are generated and added together, and then a third attention weight map is generated by Sigmoid gating. This map is multiplied with the basic fusion features channel by channel to dynamically adjust the response intensity of each channel and generate the channel attention result.

10. A variable convolutional UNet tooth image segmentation system incorporating an attention mechanism, characterized in that, include: The dental image data acquisition module is used to acquire dental image datasets and perform preprocessing. The DSC-UNet segmentation model module introduces a multimodal feature fusion module and a deformable convolution module into the UNet architecture. The multimodal feature fusion module contains a spatial-channel collaborative attention mechanism located between the encoder and decoder. It is used to fuse the different-dimensional encoded features obtained from the encoding results of two adjacent layers based on the spatial-channel collaborative attention mechanism to generate a feature of the same size as the low-dimensional encoded feature. This feature is then concatenated and fused with the decoded feature of the same size generated after upsampling of the corresponding decoding layer according to the channel dimension. The high-dimensional encoded feature guides the generation of attention weights for the low-dimensional encoded feature, thereby achieving cross-scale feature interaction. The deformable convolutional module replaces the dual convolutional layer of the decoder in the UNet architecture, and enhances the ability to model deformable structures by dynamically predicting the offset of sampling points. The model training module is used to train a DSC-UNet segmentation model using a preprocessed dataset. The tooth image segmentation module is used to segment the input tooth image using the trained DSC-UNet segmentation model and output the pixel-level segmentation results of the tooth structure.

Images