An oral image segmentation method, system, electronic device and storage medium

Abstract

This application discloses a method, system, electronic device, and storage medium for oral cavity image segmentation. The method includes preprocessing oral cavity three-dimensional volume data to obtain three-dimensional image blocks and corresponding mask blocks; obtaining image features based on the three-dimensional image blocks; obtaining cue features based on the mask blocks; and using the image features and cue features as input to a noise-resistant segmentation decoding network. The noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel. The first mask decoder receives the image features and cue features and outputs a first-level filtering mask; the second mask decoder receives the image features, cue features, and the first-level filtering mask and outputs a second-level filtering mask; and the third mask decoder receives the image features, cue features, and the second-level filtering mask and outputs the segmentation result. This application improves the geometric continuity and boundary positioning accuracy of the segmentation result and enhances the segmentation integrity of slender structures such as the mandibular nerve canal.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, system, electronic device and storage medium for oral image segmentation. Background Technology

[0002] In the field of cone-beam computed tomography (CBCT) image segmentation, achieving accurate three-dimensional segmentation of anatomical structures such as teeth, jawbone, and mandibular canal is a core prerequisite for clinical diagnosis and treatment scenarios.

[0003] In related technologies, when the SAM (Segment Anything Model) segmentation method is applied to oral CBCT, oral CBCT images have characteristics such as artifact interference and naturally blurred anatomical boundaries. This inevitably leads to subjective bias and inconsistency in the boundary regions of manually labeled mask blocks, forming boundary-level label noise. Existing methods lack explicit modeling and robust training mechanisms for this type of noise, and are prone to overfitting incorrect boundary signals, resulting in problems such as fine structural breaks and boundary positioning deviations in the segmentation results, making it difficult to meet the stringent clinical requirements for segmentation accuracy. Summary of the Invention

[0004] To address the shortcomings mentioned above, this application provides a method, system, electronic device, and storage medium for oral image segmentation.

[0005] Firstly, to achieve the above objectives, this application provides a method for segmenting oral images, comprising: The three-dimensional volume data of the oral cavity is preprocessed, and then block-level sampling is performed to obtain three-dimensional image blocks and corresponding mask blocks. Image features are obtained from 3D image patches; corresponding cue features, including the center point coordinates of the 3D image patches and the binary features of the mask, are obtained from mask blocks; the image features and cue features are used as input to the noise-resistant segmentation and decoding network; The noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel. The first mask decoder is constrained based on a preset first loss function, the second mask decoder is constrained based on a preset second loss function, and the third mask decoder is constrained based on a preset third loss function. The first mask decoder receives image features and cue features, and outputs a first-level filtering mask. The second mask decoder receives image features, cue features, and the first-level filtering mask, and outputs a second-level filtering mask. The third mask decoder receives image features, cue features, and the second-level filtering mask, and outputs the segmentation result of the oral cavity three-dimensional volume data.

[0006] Furthermore, the calculation method for the first loss function includes: A three-dimensional distance transformation is performed on the mask block to obtain voxel distance features, and the mask block is divided into core area and non-core area based on the voxel distance features; The robust loss function is used to calculate the loss value of the core region, and the weighted combination loss function of cross-entropy loss and Dice loss is used to calculate the loss value of the non-core region. The loss result of the first loss function is obtained based on the loss value of the core region and the loss value of the non-core region.

[0007] Furthermore, the calculation method for the third loss function includes: The third mask decoder encodes the input image features and prompt features to obtain a mask feature map; Based on the mask feature map, the voxel feature vectors of the core area, the voxel feature vectors of the non-core area, and the boundary voxel feature vectors are obtained respectively, and the feature center of each type of voxel feature vector is determined. The weight coefficients of each class of voxel feature vectors are determined based on the Euclidean distance from each voxel feature vector to each feature center in the mask feature map. Based on the basic segmentation loss and weight coefficients of each voxel, the loss result of the third loss function is obtained.

[0008] Furthermore, based on the Euclidean distance from each voxel feature vector in the mask feature map to each feature center, the weight coefficients of each class of voxel feature vectors are determined, including: The boundary uncertainty of each voxel is quantified based on the Euclidean distance from the feature vector of each voxel in the mask feature map to each feature center. Boundary uncertainty is mapped to weight coefficients with values ​​ranging from [0, 1].

[0009] Furthermore, the second loss function is a combination of cross-entropy loss and Dice loss.

[0010] Furthermore, the target anatomical structures corresponding to the segmentation results include at least one of the following: teeth, jawbone, maxillary sinus, and mandibular canal.

[0011] Secondly, this application also provides an oral image segmentation system, comprising: The acquisition module is used to preprocess the three-dimensional volume data of the oral cavity. After preprocessing, block-level sampling is performed to obtain three-dimensional image blocks and corresponding mask blocks. The processing module is used to obtain image features based on 3D image blocks; obtain corresponding prompt features based on mask blocks, including the center point coordinates of the 3D image blocks and the binary features of the mask; and use the image features and prompt features as input to the noise-resistant segmentation decoding network. The noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel. The first mask decoder is constrained based on a preset first loss function, the second mask decoder is constrained based on a preset second loss function, and the third mask decoder is constrained based on a preset third loss function. The execution module is used for the first mask decoder to receive image features and prompt features, and output a first-level filtering mask; the second mask decoder to receive image features, prompt features and the first-level filtering mask, and output a second-level filtering mask; and the third mask decoder to receive image features, prompt features and the second-level filtering mask, and output the segmentation result of the oral cavity three-dimensional volume data.

[0012] Thirdly, this application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0013] Fourthly, this application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application achieves precise suppression of boundary-level label noise by using a three-parallel mask decoder in an anti-noise segmentation decoding network, combined with differentiated loss function constraints, thereby improving the geometric continuity and boundary positioning accuracy of the segmentation results and enhancing the segmentation integrity of slender structures such as the mandibular nerve canal. Attached Figure Description

[0015] Figure 1 This is a flowchart of the oral image segmentation method of this application; Figure 2 This is the image segmentation result of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] Reference Figure 1 This application provides a method for segmenting oral images, including: The three-dimensional volume data of the oral cavity is preprocessed, and then block-level sampling is performed to obtain three-dimensional image blocks and corresponding mask blocks. As one implementation method, three-dimensional volumetric data of the oral cavity obtained from CBCT and corresponding voxel-level manually annotated labels are acquired. The CBCT three-dimensional volumetric data is preprocessed, including resampling, grayscale intensity truncation and normalization, and zero-mean unit variance normalization. After preprocessing, a fixed-size sliding window block sampling strategy is used to perform block-level sampling, obtaining three-dimensional image blocks and corresponding mask blocks. Regions with a voxel value of 1 in the mask block represent the target anatomical structure region, and regions with a voxel value of 0 represent the background region.

[0018] Image features are obtained from 3D image patches; corresponding cue features, including the center point coordinates of the 3D image patches and the binary features of the mask, are obtained from mask blocks; the image features and cue features are used as input to the noise-resistant segmentation and decoding network; As one implementation method, the three-dimensional image block is input into the SAM-Med3D image encoder for forward inference to obtain image features F; the three-dimensional center point coordinates of the target anatomical structure region and the mask binary features are extracted based on the mask block as prompt information, and the prompt information is input into the prompt encoder for feature encoding to obtain prompt features P.

[0019] The noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel. The first mask decoder is constrained based on a preset first loss function, the second mask decoder is constrained based on a preset second loss function, and the third mask decoder is constrained based on a preset third loss function. As one implementation method, the calculation of the first loss function includes: A three-dimensional distance transformation is performed on the mask block to obtain voxel distance features, and the mask block is divided into core area and non-core area based on the voxel distance features; The transformation formula is shown below: ; In the formula, c is the category index (c∈{1,…,C}, C is the number of segmentation categories); M c For the binary prediction mask (M) of category c c (i)=1 indicates that voxel i belongs to category c, otherwise it is 0); i is the three-dimensional voxel index (which can be equivalently represented as coordinates (x,y,z)); Indicates mask M c The set of non-core regions; dist(i, Let be the distance metric function, representing the distance from voxel i to the boundary set. The minimum Euclidean distance; dc (i) represents the interior distance of category c at voxel i; Based on a preset distance threshold Tc, the mask block is divided into a core region and a non-core region, as shown in the following formula: ; In the formula, T c The distance threshold for category c; The core region voxel set of category c; This is the set of non-core voxels for category c; and forms the global core region. and non-core areas : ; In the formula, This is the union operation for sets; Represents the union of all core region voxels of all categories; This represents the union of all non-core region voxels of all categories.

[0020] For the core area, a robust loss function based on the Student-t distribution is used to calculate the loss value. For the non-core area, a weighted combination loss function of cross-entropy loss and Dice loss is used to calculate the loss value, using weighting coefficients. Balancing the loss contributions from the core and non-core regions, the total loss formula for the first loss function is: ; In the formula, The total training loss for the LPM branch; For robust supervision loss at voxel i; For example, it is the basic segmentation loss term at voxel i, such as a weighted combination of cross-entropy and Dice loss; The robust loss function is used to calculate the loss value of the core region, and the weighted combination loss function of cross-entropy loss and Dice loss is used to calculate the loss value of the non-core region. The loss result of the first loss function is obtained based on the loss value of the core region and the loss value of the non-core region.

[0021] The second loss function is a combination of cross-entropy loss and Dice loss.

[0022] The calculation methods for the third loss function include: The third mask decoder encodes the input image features and prompt features to obtain a mask feature map; Based on the mask feature map, the voxel feature vectors of the core area, the voxel feature vectors of the non-core area, and the boundary voxel feature vectors are obtained respectively, and the feature center of each type of voxel feature vector is determined. The prototype centers of the three types of features are recursively updated using exponential moving average, and the update formula is as follows: ; In the formula, From At voxel location Extracted feature vectors; For the class prototype center vector, Indicates the core area prototype, Indicates non-core area prototype, Represents the boundary prototype; ; For use in calculating the first The collection of voxels of the class prototype; For set The number of elements; Based on the Euclidean distance from each voxel feature vector in the mask feature map to each feature center, the formula is as follows: ; Boundary uncertainty based on the Euclidean distance quantization voxel i The quantification formula is: ; In the formula: scalars represent absolute values; voxels Measures of uncertainty; To prevent numerical stability constants with a denominator of zero; The boundary uncertainty is mitigated by a weight mapping network consisting of several 1×1×1 convolutional layers. Mapped to voxel dynamic weights with values ​​in the range [0,1]. Weight The formula is shown below: ; In the formula: This is a weighted mapping network.

[0023] Based on the basic segmentation loss and weight coefficients for each voxel, the loss result of the third loss function is obtained, and the formula for the third loss function is shown below: ; In the formula: The training loss for the MRM branch; The gating value for the secondary filter mask output by the second mask decoder ( Voxel representation (If it is selected as a trusted supervised region by the secondary filter, otherwise it is 0). voxels The segmentation loss at the location, For the third mask decoder in voxels The predicted probability at that location. voxels Supervision label at the location.

[0024] The first mask decoder receives image features and cue features, and outputs a first-level filtering mask. The second mask decoder receives image features, cue features, and the first-level filtering mask, and outputs a second-level filtering mask. The third mask decoder receives image features, cue features, and the second-level filtering mask, and outputs the segmentation result of the oral cavity three-dimensional volume data.

[0025] As one implementation, the outputs of the first mask decoder, the second mask decoder, and the third mask decoder are shown below: ; In the formula: P represents image features; P represents cue features. , , These represent the mask prediction headers of the first mask decoder, the second mask decoder, and the third mask decoder, respectively. For the segmentation prediction probability maps output by the first mask decoder, second mask decoder, and third mask decoder, their positions at voxel locations are... The prediction at the location is denoted as ; Based on the segmentation probability map output by the first mask decoder, a voxel-level confidence map is constructed. Voxels with confidence scores greater than a preset first threshold are marked as clean supervised voxels, and a first-level filtering mask is generated. Based on the segmentation probability map output by the second mask decoder, a voxel-level confidence map is constructed. Voxels with confidence scores greater than a preset second threshold are marked as clean supervised voxels, and a second-level filtering mask is generated. Based on the segmentation probability map output by the third mask decoder, the target anatomical structure corresponding to the segmentation result includes at least one of the following: teeth, jawbone, maxillary sinus, and mandibular canal.

[0026] In the implementation of this application, all verification experiments were completed on the NVIDIA computing platform. Using an oral clinical dataset, the performance differences between this method and several existing variants such as SAM3D, 3DSAM-Adapter, SAM-Med3D, and M-SAM, as well as the fully parameter-tuned SAM baseline, were compared under a strictly unified training strategy. The resolution of all input images was uniformly resampled to 128×128×128. Finally, the Dice coefficient was used as the core evaluation index, and the test was conducted on a clean dataset after screening.

[0027] The Dice coefficient is used to measure the degree of overlap between the predicted core region P and the true core region G, and is defined as: ; A larger value indicates a greater spatial overlap between the prediction and the ground truth. In multi-class segmentation scenarios, the average value of the Dice calculated for each class structure is taken as the final average Dice to reflect the overall segmentation performance and boundary consistency.

[0028] Table 1 shows a comparison of the segmentation Dice coefficients between this application and existing methods: Table 1 As shown in Table 1, SAM3D's overall performance on the five anatomical structures is relatively limited, especially in the mandibular nerve canal, a long and slender structure whose boundaries are susceptible to artifacts. This indicates that direct segmentation lacks sufficient robustness to boundary-level noise in dental CBCT. The 3DSAM-Adapter, which incorporates a lightweight adaptation strategy, brings slight improvements across all categories, but its overall gain is limited, suggesting that relying solely on shallow adaptation is insufficient to compensate for domain offset and complex boundary uncertainties in dental CBCT. In contrast, SAM-Med3D, pre-trained on a large scale using medical datasets, shows a significant performance leap, with overall improvements across various foreground structures, validating the crucial role of pre-trained features in stabilizing segmentation boundaries. Furthermore, M-SAM continues to improve overall results across most categories. In comparison, the model proposed in this invention achieves optimal results on all five structures. Notably, the mandibular nerve canal, as the most challenging structure, is most sensitive to noise, yet this invention still achieves further improvements over the strongest baseline in this category. This demonstrates that the proposed method can more effectively suppress boundary noise and enhance geometric continuity modeling, thereby achieving more stable and consistent performance gains on difficult structures.

[0029] Meanwhile, under the same dataset containing noise labels, the repair results of each model for the noisy regions are as follows: Figure 2As shown, it can be observed that the comparative methods are all affected by noise labels to varying degrees during the fine-tuning process, resulting in significant pixel-level misclassification in the prediction. Specifically, when the mandibular label contains noise (incorrectly labeling some tooth root regions as mandible), the prediction results of SAM3D and 3DSAM-Adapter are highly consistent with the original noise label, and they almost fail to suppress the erroneous supervision signal, indicating insufficient robustness to boundary-level noise. Although SAM-Med3D has stronger medical pre-training representation capabilities, it exhibits more significant misclassification problems under this noise supervision, manifested in the incorrect prediction of more tooth root regions as mandible, indicating that relying solely on pre-training cannot avoid the cumulative amplification effect of noise labels during the fine-tuning stage. M-SAM alleviates the misclassification intensity to some extent, reducing the misclassification range of the noise region, but it still fails to fundamentally eliminate noise interference, and the residual error categories remain obvious. In contrast, this application can effectively identify and suppress erroneous gradients caused by noise supervision, significantly reducing misclassification of the mandible and tooth root regions. This allows tooth root regions that were originally incorrectly labeled as mandible to be corrected in the prediction, with noise significantly reduced or even virtually eliminated. Experimental results show that this application possesses stronger pixel-level error correction capabilities and more stable structural boundary recovery effects under noisy labeling conditions, thus demonstrating superior robustness and practical value in automatic annotation tasks.

[0030] To verify the effectiveness of each component module in the proposed framework, ablation experiments were conducted while maintaining the consistency of the remaining training and inference configurations. A single decoder baseline was used as the control starting point. This baseline did not enable meta-adaptive learning and multi-branch selection mechanisms, and relied solely on basic supervised learning to complete the segmentation. The ablation experiment results are shown in Table 2. Table 2 Specifically, the average Dice of the single decoder baseline is 0.9217. This setting makes it more susceptible to boundary bias and artifact-related errors, limiting the overlap between the predicted boundary and the ground truth annotation. To further quantify its marginal contribution, meta-adaptive learning and multi-branch filtering were added: when only meta-adaptive learning was enabled, the average Dice increased to 0.9281, an improvement of 0.0064 compared to the baseline, indicating that adaptive reweighting of uncertain voxels and suppression of misleading gradients effectively mitigates the interference of boundary noise on the optimization direction. When only multi-branch filtering was enabled, the average Dice reached 0.9265, an improvement of 0.0048 compared to the baseline, demonstrating that purifying the supervision signal through hierarchical filtering and consistency constraints on the same encoded features improves the overall credibility of the supervision domain and reduces pixel-level misclassification.

[0031] When both methods are used together, the average Dice improves to 0.9345, an overall improvement of 0.0128 compared to the baseline. Furthermore, compared to using only meta-adaptive learning and only multi-branch selection, the improvements are additionally 0.0064 and 0.0080 respectively, demonstrating a more significant performance gain. This indicates that the two methods have complementary and synergistic effects in noise suppression: multi-branch selection primarily acts on error removal and enhanced supervision consistency at the supervision source level, thereby gradually expanding the high-confidence supervision domain and improving training stability; meta-adaptive learning primarily acts on voxel-level weight allocation and gradient constraints during the optimization process, characterizing boundary uncertainty through dynamic distance weighting and adaptively emphasizing supervision beneficial to the quality of the true boundary through the meta-learning update mechanism, thus reducing misleading updates introduced by noisy labels. Experimental results demonstrate that the proposed method can effectively mitigate the interference of labeling errors on the model optimization process under the condition of boundary-level noise, and stably improve the spatial overlap consistency and boundary accuracy between the prediction results and the true labels.

[0032] Automatic annotation of unannotated oral CBCT data generates pseudo-labels with an average Dice coefficient of over 0.93 compared to manually annotated data. The boundary uncertainty metric effectively identifies low-quality annotated areas, and manual review only requires correction of less than 5% of the areas. Compared to existing automatic annotation methods, the workload of manual review is reduced by more than 80%, and the rework rate is reduced by more than 75%, significantly improving the efficiency of oral CBCT image segmentation and annotation. It has good application value in clinical scenarios such as preoperative planning of oral implant surgery, orthodontic design, and maxillofacial surgical navigation.

[0033] This application also provides an oral imaging segmentation system, including: The acquisition module is used to preprocess the three-dimensional volume data of the oral cavity. After preprocessing, block-level sampling is performed to obtain three-dimensional image blocks and corresponding mask blocks. The processing module is used to obtain image features based on 3D image blocks; obtain corresponding prompt features based on mask blocks, including the center point coordinates of the 3D image blocks and the binary features of the mask; and use the image features and prompt features as input to the noise-resistant segmentation decoding network. The noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel. The first mask decoder is constrained based on a preset first loss function, the second mask decoder is constrained based on a preset second loss function, and the third mask decoder is constrained based on a preset third loss function. The execution module is used for the first mask decoder to receive image features and prompt features, and output a first-level filtering mask; the second mask decoder to receive image features, prompt features and the first-level filtering mask, and output a second-level filtering mask; and the third mask decoder to receive image features, prompt features and the second-level filtering mask, and output the segmentation result of the oral cavity three-dimensional volume data.

[0034] This application also provides an electronic device, including at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the program is executed by the processing unit, the processing unit performs the above-described method.

[0035] This application also provides a storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the above-described method.

[0036] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for segmenting oral cavity images, characterized in that, include: The three-dimensional volume data of the oral cavity is preprocessed, and then block-level sampling is performed to obtain three-dimensional image blocks and corresponding mask blocks. Image features are obtained from the three-dimensional image blocks; corresponding cue features, including the center point coordinates of the three-dimensional image blocks and the binary features of the mask, are obtained from the mask blocks; the image features and the cue features are used as inputs to the noise-resistant segmentation and decoding network; The noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel. The first mask decoder is constrained based on a preset first loss function, the second mask decoder is constrained based on a preset second loss function, and the third mask decoder is constrained based on a preset third loss function. The first mask decoder receives the image features and the cue features, and outputs a first-level filtering mask. The second mask decoder receives the image features, the cue features, and the first-level filtering mask, and outputs a second-level filtering mask. The third mask decoder receives the image features, the cue features, and the second-level filtering mask, and outputs the segmentation result of the oral cavity three-dimensional volume data.

2. The oral cavity image segmentation method according to claim 1, characterized in that, The calculation method of the first loss function includes: A three-dimensional distance transformation is performed on the mask block to obtain voxel distance features, and the mask block is divided into a core region and a non-core region based on the voxel distance features; The loss value of the core region is calculated using a robust loss function, and the loss value of the non-core region is calculated using a weighted combination loss function of cross-entropy loss and Dice loss. The loss result of the first loss function is obtained based on the loss value of the core region and the loss value of the non-core region.

3. The oral cavity image segmentation method according to claim 2, characterized in that, The calculation method of the third loss function includes: The third mask decoder encodes the input image features and the prompt features to obtain a mask feature map; Based on the mask feature map, the voxel feature vectors of the core area, the voxel feature vectors of the non-core area, and the boundary voxel feature vectors are obtained respectively, and the feature center of each type of voxel feature vector is determined. The weight coefficients of each type of voxel feature vector are determined based on the Euclidean distance from each voxel feature vector to each feature center in the mask feature map. The loss result of the third loss function is obtained based on the basic segmentation loss of each voxel and the weight coefficients.

4. The oral cavity image segmentation method according to claim 3, characterized in that, The step of determining the weight coefficient of each type of voxel feature vector based on the Euclidean distance from each voxel feature vector to each feature center in the mask feature map includes: The boundary uncertainty of each voxel is quantified based on the Euclidean distance from the feature vector of each voxel in the mask feature map to each feature center. The boundary uncertainty is mapped to a weighting coefficient with values ​​in the range [0, 1].

5. The oral cavity image segmentation method according to claim 1, characterized in that, The second loss function is a combination of cross-entropy loss and Dice loss.

6. The oral cavity image segmentation method according to claim 1, characterized in that, The target anatomical structure corresponding to the segmentation result includes at least one of the following: teeth, jawbone, maxillary sinus, and mandibular canal.

7. A dental image segmentation system, characterized in that, include: The acquisition module is used to preprocess the three-dimensional volume data of the oral cavity. After preprocessing, block-level sampling is performed to obtain three-dimensional image blocks and corresponding mask blocks. The processing module is configured to obtain image features based on the three-dimensional image block; obtain corresponding prompt features including the center point coordinates of the three-dimensional image block and the binary features of the mask block based on the mask block; and use the image features and the prompt features as input to a noise-resistant segmentation decoding network; the noise-resistant segmentation decoding network includes a first mask decoder, a second mask decoder, and a third mask decoder arranged in parallel, constraining the first mask decoder based on a preset first loss function, constraining the second mask decoder based on a preset second loss function, and constraining the third mask decoder based on a preset third loss function; The execution module is configured to: receive the image features and the prompt features from the first mask decoder and output a first-level filtering mask; receive the image features, the prompt features and the first-level filtering mask from the second mask decoder and output a second-level filtering mask from the second mask decoder; and receive the image features, the prompt features and the second-level filtering mask from the third mask decoder and output the segmentation result of the oral cavity three-dimensional volume data.

8. An electronic device, characterized in that, The method includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the method described in any one of claims 1 to 6.

9. A storage medium, characterized in that, It stores a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the method described in any one of claims 1 to 6.

Images