A method and system for fine-grained segmentation of three-dimensional cbct skull images

Abstract

The application discloses a kind of three-dimensional CBCT skull image fine granularity segmentation method and system.The present application includes the following steps: first, construct first training dataset;Then, in combination with first training dataset, after training secondary segmentation neural network model, obtain trained secondary segmentation neural network model;Then, using trained secondary segmentation neural network model to the three-dimensional CBCT skull image to be processed is segmented, and the segmentation result of each structure of skull is obtained;Finally, using the improved region growing method is carried out the segmentation of the bone cortex and bone marrow of upper and lower jaw bones, and the bone cortex and bone marrow segmentation result corresponding to upper and lower jaw bones is obtained, so as to complete the fine granularity segmentation of three-dimensional CBCT skull image.The method proposed in the present application can separate the fine granularity region of oral and maxillofacial region, including complete skull, upper and lower jaw bones, teeth, bone cortex, bone marrow, nerve canal, maxillary sinus, incisor canal and other structures.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical image processing, and specifically relates to a fine-grained segmentation method and system for three-dimensional CBCT skull images. BACKGROUND

[0002] In the field of oral medicine, three-dimensional imaging technology, especially cone beam computed tomography (CBCT), has become an important diagnostic tool. CBCT, with its high resolution and low radiation dose, is widely used in oral implantation, orthodontic treatment, maxillofacial surgery and other clinical fields. Through analysis of CBCT images, doctors can clearly observe the detailed structure of the jaw bone, teeth and surrounding tissues, thereby providing accurate basis for disease diagnosis, treatment planning and surgical operation.

[0003] However, CBCT image data is large and complex in detail, and doctors still face challenges such as image interpretation difficulty and insufficient segmentation accuracy when analyzing anatomical structures. Existing manual segmentation methods usually rely on the experience of doctors, which is time-consuming and easily affected by human factors, resulting in low segmentation accuracy and increasing the risk of diagnosis and treatment decisions; automatic segmentation technology can achieve partial region segmentation, but it is difficult to perform fine automatic separation for complex anatomical structures (such as cortical bone, cancellous bone, nerve canal, maxillary sinus, etc.).

[0004] Chinese invention patent CN118736218A proposes a method for segmenting the mandible and condylar process in the mandible; Chinese invention patents CN119152212A, CN118823766A, etc. disclose a method for segmenting the mandibular nerve canal in CBCT images. These methods partially solve the segmentation problem of certain fine-grained structures in the upper and lower jawbone structure, but overall, more complex structures and larger spatial range of fine-grained segmentation need to be achieved. SUMMARY

[0005] To overcome the shortcomings of the prior art, the present application proposes a fine-grained segmentation method and system for three-dimensional CBCT skull images, which can separate the fine-grained regions of the oral and maxillofacial region, including the complete skull, mandible, teeth, cortical bone, cancellous bone, nerve canal, maxillary sinus, incisor canal, etc.

[0006] To achieve this purpose, the present application provides the following technical solutions:

[0007] I. A fine-grained segmentation method for three-dimensional CBCT skull images

[0008] Step 1: Construct a first training data set using the obtained original three-dimensional CBCT skull image;

[0009] Step 2: after training the secondary segmentation neural network model by combining the first training data set, a trained secondary segmentation neural network model is obtained;

[0010] Step 3: after segmenting the three-dimensional CBCT skull image to be processed by using the trained secondary segmentation neural network model, the segmentation results of the tooth, skull, mandible, nerve canal, maxillary sinus, and incisor canal structures are obtained.

[0011] Step 4: the segmentation results of the skull and mandible are extracted; then, based on the segmentation results of the skull and mandible, the improved region growing method is used to segment the cortical bone and cancellous bone of the upper and lower jaws, and the segmentation results of the cortical bone and cancellous bone corresponding to the upper and lower jaws are obtained, thereby completing the fine-grained segmentation of the three-dimensional CBCT skull image.

[0012] The step 2 is specifically:

[0013] The secondary segmentation neural network model includes a coarse segmentation neural network model and a fine segmentation neural network model; first, the coarse segmentation neural network model is trained by using the first training data set, and a trained coarse segmentation neural network model is obtained; then, the original three-dimensional CBCT skull image in the first training data set is input into the trained coarse segmentation neural network model, and the model outputs the corresponding image coarse segmentation result; after adding all image coarse segmentation results to the first training data set, a second training data set is obtained; then, after performing image resolution enhancement processing on each image sample in the second training data set, a third training data set is obtained; then, the fine segmentation neural network model is trained by using the third training data set, and a trained fine segmentation neural network model is obtained, thereby obtaining the trained secondary segmentation neural network model.

[0014] The image resolution enhancement processing on each image sample in the second training data set includes:

[0015] First, the size of the original three-dimensional CBCT skull image, the image coarse segmentation result, and the corresponding label image in each image sample is modified at the same time, so that the image resolution is uniformly changed to r millimeters per pixel; then, each three-dimensional CBCT skull image after modifying the size and the corresponding image coarse segmentation result are spliced to form a CBCT and coarse segmentation fusion image; finally, each CBCT and coarse segmentation fusion image is cut into a plurality of d2xd2xd2 local fusion images.

[0016] In the step 3, the three-dimensional CBCT skull image to be processed is input into the trained coarse segmentation neural network model, and the model outputs a coarse segmentation result of the image; after the three-dimensional CBCT skull image to be processed and the coarse segmentation result of the image are subjected to enhancement processing of image resolution, the image is input into the trained fine segmentation neural network model, and the model outputs a segmentation result of each structure of teeth, skull, mandible, mandibular nerve canal, maxillary sinus and incisor canal.

[0017] In the step 4, the segmentation of the cortical bone and spongy bone of the maxilla includes the following steps:

[0018] S1: determining a maxilla region according to the segmentation result of the skull; finding all voxels located at the outer edge of the maxilla and in contact with the background in the maxilla region, taking the found voxels as initial seed points, and forming an initial growth region from the initial seed points;

[0019] S2: judging and updating the candidate voxels of the edge of the latest growth region according to a preset update rule round by round until a preset condition is met, and obtaining a final growth region, wherein all voxels in the final growth region together constitute a cortical bone region of the maxilla and are recorded as a cortical bone segmentation result;

[0020] S3: subtracting the cortical bone region from the maxilla region to obtain a spongy bone region and recording the spongy bone region as a maxilla spongy bone segmentation result.

[0021] The preset update rule includes a threshold control rule based on gray scale features, a spatial position constraint rule based on adjacent relationships, a morphological optimization rule based on edge smoothness, a count limit rule and a structure exclusivity rule.

[0022] II. A fine-grained segmentation system for three-dimensional CBCT skull images

[0023] A first training data set construction unit is configured to construct a first training data set by using the obtained original three-dimensional CBCT skull image.

[0024] An image coarse segmentation unit is configured to store a coarse segmentation neural network model and train the coarse segmentation neural network model by using the first training data set.

[0025] A third training data set construction unit is configured to generate a second training data set by using the trained coarse segmentation neural network model, and further construct a third training data set.

[0026] An image fine segmentation unit is configured to store a fine segmentation neural network model and train the fine segmentation neural network model by using the third training data set.

[0027] The mandibular fine-grained segmentation unit is used to segment the cortical bone and cancellous bone of the maxillary and mandibular bones based on the segmentation results of the skull and mandible using an improved region growing method to obtain the corresponding cortical bone and cancellous bone segmentation results of the maxillary and mandibular bones.

[0028] 3. A computer device

[0029] The device includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, the steps of the method for fine-grained segmentation of a three-dimensional CBCT skull image are implemented.

[0030] 4. A Computer-Readable Storage Medium

[0031] The medium stores a computer program, which, when executed by a processor, implements the steps of the method for fine-grained segmentation of a three-dimensional CBCT skull image.

[0032] 5. A computer program product

[0033] The product includes a computer program / instruction, which, when executed by a processor, implements the steps of the method for fine-grained segmentation of a three-dimensional CBCT skull image.

[0034] The beneficial effects of the present invention are:

[0035] The present invention performs coarse and fine segmentation on three-dimensional CBCT skull images through a two-level segmentation neural network model, and can realize automatic segmentation of the skull, mandible, teeth, neural canal, maxillary sinus, and incisive canal.

[0036] The present invention proposes an improved region growing method, which can realize the automatic separation of cortical bone and cancellous bone in maxillary and mandibular bones.

[0037] Therefore, the present invention can segment the skull, mandible, teeth, cortical bone, cancellous bone, neural canal, maxillary sinus, and incisive canal, achieving fine-grained segmentation of CBCT skull images and possessing promising applications. For example, in oral implants, this helps select the optimal implant position and angle; in orthodontic treatment, detailed bone structure segmentation can help doctors design appropriate tooth movement pathways and accurately assess bone development and correction outcomes; and in maxillofacial surgery, precise anatomical annotations can provide more accurate references for surgical procedures, reducing surgical risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Flow chart of the method of the present invention.

[0039] Figure 2 Schematic diagram of the network structure of the SwinUNETR model based on Swin Transformer.

[0040] Figure 3 Schematic diagram of the selection of initial seed points in the region growing algorithm proposed in this invention.

[0041] Figure 4 This is the final cortical bone selection result diagram in the region growing algorithm proposed in this invention.

[0042] Figure 5 This is the segmentation result of the upper and lower jaw parts in the embodiment of the present invention.

[0043] Figure 6 This is the segmentation result of the mandibular ramus and condyle in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] This specification is further described below in conjunction with the accompanying drawings and implementation methods. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this specification.

[0045] like Figure 1 As shown, the fine-grained segmentation method of a three-dimensional CBCT skull image proposed in the present invention includes the following steps:

[0046] Step 1: Construct a first training dataset using the acquired original three-dimensional CBCT skull images; the first training dataset includes several original three-dimensional CBCT skull images and corresponding label images.

[0047] Three-dimensional CBCT skull images were collected from multiple oral centers. Labeling tools such as ITK-SNAP were then used to manually annotate the teeth, skull, mandible, mandibular canal, maxillary sinus, and incisor canal. The annotated images were then converted into labeled images. In the labeled images, the voxel value of the background label was 0, the tooth label was 1, the skull label was 2, the mandible label was 3, the mandibular canal (bilateral) label was 4, the maxillary sinus (bilateral) label was 5, and the incisor canal label was 6. Because the maxilla is already included in the skull, only the mandible needs to be annotated. The original three-dimensional CBCT skull images of different sizes were scaled and compressed to the same length, width, and height dimensions, with dimension d1 set to 128. The image resolution was 1.2 mm per voxel width. The input CBCT skull image has three dimensions: length, width, and height. For each dimension, the following judgment was made: if the number of voxels in that dimension is d, then the image is labeled. v , with a resolution of r v mm per voxel, then the length in that dimension is d v ×r v mm. If dv ×r v >128×1.2, then compression is performed on this dimension, and the resolution of this dimension becomes d after compression. v ×r v / 128 mm per voxel, the number of voxels is compressed to 128. If d v ×r v ≤128×1.2, then fill in this dimension. The specific filling method is: first, the resolution of this dimension is changed to about 1.2 mm, and the number of voxels is changed to [d v ×r v / 1.2], and then fill this dimension with voxels with a value of 0 to make the length of this dimension 128 voxels. In this way, CBCT images of large, medium, and small fields of view can have approximately the same resolution (no less than 1.2 mm per voxel) and the same size (128×128×128) after transformation.

[0048] Step 2: After training the secondary segmentation neural network model based on the first training data set, a trained secondary segmentation neural network model is obtained;

[0049] Step 2 is as follows:

[0050] The two-level segmentation neural network model includes a coarse segmentation neural network model and a fine segmentation neural network model; the role of the coarse segmentation neural network model is to achieve preliminary segmentation of the skull structure on a lower-resolution image, and the role of the fine segmentation neural network model is to accurately segment the skull structure on a higher-resolution image. Both the coarse segmentation neural network model and the fine segmentation neural network model use the SwinUNETR model based on Swin Transformer. The network structure diagram is shown in the figure below. Figure 2 As shown in the figure, the SwinUNETR model consists of two main modules: an encoder and a decoder. The encoder is based on the Swin Transformer structure. It extracts features from the input image through a layered window attention mechanism while preserving the spatial locality and global context information of the image. The decoder uses a layer-by-layer upsampling method to fuse multi-scale features step by step to restore high-resolution segmentation results. In addition, multi-scale features are directly transferred between the encoder and decoder through skip connections. Except for the first layer of the encoder and the last layer of the decoder, the remaining parameters of the encoder and decoder are consistent between the coarse segmentation neural network model and the fine segmentation neural network model.

[0051] First, a coarse segmentation neural network model is trained using the first training dataset to obtain a trained coarse segmentation neural network model. During training, the cross-entropy loss function is used to compare the difference between the probability map and the true labels. Based on the results of this loss function, the coarse segmentation neural network model parameters are adjusted using a gradient descent method. The original 3D CBCT skull images from the first training dataset are then input into the trained coarse segmentation neural network model, which outputs the corresponding coarse segmentation result, a class probability map of 7× d1× d1× d1. Each value in the class probability map has a range of R. The class probability map refers to the probability distribution of each voxel belonging to each label class in a 128×128×128 voxel space. After adding all coarse segmentation results to the first training dataset, a second training dataset is obtained. Next, each image sample in the second training dataset (i.e., the original 3D CBCT skull image, the coarse segmentation result, and the corresponding label image) is enhanced to obtain a third training dataset. The fine segmentation neural network model is then trained using the third training dataset to obtain a trained fine segmentation neural network model, thereby obtaining a trained secondary segmentation neural network model. During training, the cross-entropy loss function is used to compare the difference between the class probability map and the true label. The formula and calculation method are the same as those for the coarse segmentation neural network model. Based on the results of this loss function, the parameters of the fine segmentation neural network model are adjusted using gradient descent.

[0052] The formula of the cross entropy loss function L is as follows:

[0053]

[0054] Where x is the position of a pixel, ranging from (0,0,0) to (128,128,128), which covers all pixel positions of the image; C is the number of channels, which is equal to 7; x i,x is the image at the i-th channel and position (x i ,x j ,x k )’s predicted value; x c,x is the image at the cth channel and position (x i ,x j ,x k )’s predicted value; c,x In the category image after converting the manually annotated class label map into a one-hot encoding, the image at the cth channel and position (x i ,x j ,x k) is the true value of the gradient. Specifically, the AdamW gradient descent optimizer is used to update the exponentially weighted moving average of the first and second moments of the gradient. A weight decay term is introduced with each parameter update. This directly applies regularization constraints to the model parameters to control model complexity and prevent overfitting. The initial learning rate is set to 0.001, and an exponentially decaying learning rate control formula is used, whereby the learning rate is updated to 0.5 times the current learning rate every 20 training epochs.

[0055] The image resolution enhancement processing of each image sample in the second training data set (i.e., the original three-dimensional CBCT skull image, the image coarse segmentation result, and the corresponding label image constitute one image sample) includes:

[0056] First, the original 3D CBCT skull image, image coarse segmentation result, and corresponding label image in each image sample are modified at the same time so that the image resolution is uniformly changed to r mm per pixel, where r is 0.3. Then, each modified 3D CBCT skull image and the corresponding image coarse segmentation result are spliced ​​to form an 8-channel CBCT and coarse segmentation fusion image with a resolution of r mm per pixel, and the size meets (8, W L ,H L ,D L ), W L ,H L ,D L are the number of voxels in the width, height, and length directions of the image after the resolution is modified. L = round down (W × r W / r), W is the number of voxels in the width direction of the original image, r W H is the resolution of the original image in the width direction; L = round down (H × r H / r), H is the number of voxels in the original image in the height direction, r H is the resolution of the original image in the high direction; D L = round down (D × r D / r), D is the number of voxels in the original image in the long direction, r Dis the resolution of the original image in the longitudinal direction. Rounding down ( ) indicates that the values ​​in the brackets are rounded down. Finally, due to the large size of the 0.3 mm per pixel CBCT and coarse segmentation fusion image and hardware limitations, each CBCT and coarse segmentation fusion image can be segmented into several local fusion images of size d2 × d2 × d2, with d2 set to 196. The third training dataset consists of all these local fusion images and labeled images of corresponding sizes. In practice, the CBCT and coarse segmentation fusion images are segmented using a sliding window approach. A small sliding window of size 196 × 196 × 196 is constructed. Starting from the left, bottom, and bottom sides, the sliding window is moved to extract small regions of size (8, 196, 196, 196) from the corresponding locations of the CBCT and coarse segmentation fusion image. The fine segmentation neural network is then used to predict the probability of the seven categories corresponding to each small region until the sliding window covers the entire image space. Each sliding step is 64 voxels in length, width, and height. For example, if the current sliding window ranges from (0, 0, 0) to (195, 195, 195), the next position the sliding window moves to should be from (64, 0, 0) to (259, 195, 195). All predictions are aggregated to obtain the final 7-category segmentation result. Because the sliding window step is smaller than the window length, there is some overlap. In areas where multiple sliding windows overlap, the segmentation result should be the average of the predictions of each sliding window for that area.

[0057] Step 3: After segmenting the 3D CBCT skull image using the trained two-level segmentation neural network model, the segmentation results of the teeth, skull, mandible, neural canal, maxillary sinus, and incisor canal structures are obtained. The size meets (7, W L ,H L ,D L );

[0058] In step 3, the 3D CBCT skull image to be processed is preprocessed. This preprocessing process involves performing a z-score normalization transformation on the grayscale values ​​of the 3D CBCT skull image, followed by scaling and deformation to obtain a low-resolution image (IS) with a voxel size of 128×128×128. The 3D CBCT skull image to be processed (i.e., image IS) is input into a trained coarse segmentation neural network model, which outputs a coarse image segmentation result. The 3D CBCT skull image to be processed (i.e., image IS) and the coarse image segmentation result are then subjected to image resolution enhancement processing to obtain several locally fused images. These locally fused images are then sequentially input into a trained fine segmentation neural network model, which outputs segmentation results for the teeth, skull, jaw, mandibular canal, maxillary sinus, and incisive canal.

[0059] Step 4: Extract the skull and mandible segmentation results. Then, based on the skull and mandible segmentation results, use the improved region growing method to segment the cortical bone and cancellous bone of the maxillary and mandibular bones, and obtain the corresponding cortical bone and cancellous bone segmentation results of the maxillary and mandibular bones, thereby completing the fine-grained segmentation of the 3D CBCT skull image, that is, obtaining a fine-grained segmentation result including the complete skull, maxillary and mandibular bones, teeth, cortical bone, cancellous bone, neural canal, maxillary sinus, incisive canal and other structures.

[0060] In step 4, for the skull segmentation result, the segmentation of the maxillary cortical bone and cancellous bone segmentation results includes the following steps:

[0061] S1: In the skull segmentation results, the area with a distance from the maxillary teeth less than a preset threshold is considered the maxillary bone for which cortical bone segmentation is required. Within this maxillary region, all voxels located at the outer edge of the maxillary bone and in contact with the background are found. These voxels are used as initial seed points, and the initial growth region, i.e., the high-density region in contact with the background, is formed from these initial seed points.

[0062] S2: According to the preset update rules, the candidate voxels at the edge of the latest growth area are judged and updated in rounds until the preset conditions are met and the final growth area is obtained. All voxels in the final growth area together constitute the cortical bone area of ​​the maxillary bone and are recorded as the cortical bone segmentation result; specifically, in each iteration round, the boundary front edge of the current growth area is identified, and multi-dimensional generation evaluation is performed on all candidate pixels in it one by one. If a voxel meets at least one growth condition, it is included in the growth area; if it meets the exclusion condition, it is removed from the area. As the iteration proceeds, the growth area gradually expands to form a continuous structure that conforms to the morphological characteristics of the cortical bone. This process continues until the preset maximum number of iterations is reached, or when no new voxels are included or removed from the area in the current round, the growth process terminates. At this point, all voxels in the growth area together constitute the cortical bone area of ​​the mandible. The shape of the cortical bone area is a reasonable shape that includes a part of the mandible boundary with a higher grayscale value, a smoother inner side, and different thicknesses at different locations. After the regional growth process is completed, the cortical bone result obtained is as follows Figure 4 shown.

[0063] S3: The cancellous bone area is obtained by subtracting the cortical bone area from the maxillary area and recorded as the cancellous bone segmentation result.

[0064] For the mandibular segmentation result, the segmentation of the mandibular cortical bone and cancellous bone segmentation results includes the following steps:

[0065] S1: Find all voxels located on the outer edge of the mandible and in contact with the background in the mandibular segmentation result, use these voxels as initial seed points, and form the initial growth area from these initial seed points, that is, the high-density area in contact with the background, such as Figure 3 As shown;

[0066] S2: According to the preset update rules, the candidate voxels at the edge of the latest growth area are judged and updated in rounds until the preset conditions are met and the final growth area is obtained. All voxels in the final growth area together constitute the cortical bone area of ​​the maxillary bone and are recorded as the cortical bone segmentation result; specifically, in each iteration round, the boundary front edge of the current growth area is identified, and multi-dimensional generation evaluation is performed on all candidate pixels in it one by one. If a voxel meets at least one growth condition, it is included in the growth area; if it meets the exclusion condition, it is removed from the area. As the iteration proceeds, the growth area gradually expands to form a continuous structure that conforms to the morphological characteristics of the cortical bone. This process continues until the preset maximum number of iterations is reached, or when no new voxels are included or removed from the area in the current round, the growth process terminates. At this point, all voxels in the growth area together constitute the cortical bone area of ​​the mandible. The shape of the cortical bone area is a reasonable shape that includes a part of the mandible boundary with a higher grayscale value, a smoother inner side, and different thicknesses at different locations. After the regional growth process is completed, the cortical bone result obtained is as follows Figure 4 shown.

[0067] S3: The cancellous bone area is obtained by subtracting the cortical bone area from the segmentation result of the lower skull and recorded as the cancellous bone segmentation result.

[0068] The preset update rules include threshold control rules based on grayscale features, spatial position constraint rules based on proximity relationships, morphological optimization rules based on edge smoothness, counting restriction rules, and structural exclusivity rules.

[0069] The threshold control rule based on grayscale features is as follows: If the grayscale value of a voxel being evaluated (hereinafter referred to as the current voxel) and the average grayscale of its neighborhood meet the preset growth condition (i.e., the grayscale value of the current voxel and the average grayscale of its neighborhood are both greater than the corresponding grayscale threshold), then the voxel is more likely to be included in the growth region; otherwise, it may not be included. The grayscale threshold for the current voxel and its neighborhood is determined by taking the average grayscale value of the voxels in its neighborhood that are included in the cortical bone region as the threshold. This rule helps improve the contrast sensitivity between cortical bone and cancellous bone, thereby enhancing segmentation accuracy.

[0070] A spatial position constraint based on proximity relationships: If the current voxel is located outside the jaw and closer to the background boundary, it is more likely to belong to the cortical bone layer; if it is far from the background boundary or close to internal cavity structures, it is likely not to belong to the cortical bone region. Therefore, the present invention introduces a spatial constraint mechanism to evaluate the relative distance and direction relationship between the current voxel and known anatomical structures (such as teeth, the inner and outer edges of the jaw, and the background area). This mechanism effectively improves the spatial rationality of the segmentation results.

[0071] Morphological optimization rules based on edge smoothness: To improve the continuity and consistency of the final segmentation contour, the present invention evaluates the impact of adding or removing candidate voxels on the local boundary during each iteration. For example, if an unincluded voxel can smooth the boundary when included in the growth region, it is given priority for inclusion; conversely, if removing an included voxel can make the boundary clearer and more reasonable, it is removed. This dynamic optimization mechanism helps reduce misjudgments caused by noise interference.

[0072] Count limit rule: To prevent regional oscillation caused by repeated state switching of certain voxels, the present invention records the number of state changes for each voxel. When the number of state changes of the current voxel reaches a preset threshold, the state modification of the voxel will be stopped to avoid instability caused by excessive iterations.

[0073] The structure-exclusion rule specifies that the region growing operation is performed only within the jaw region, excluding all non-jaw structures (such as teeth, neural canals, maxillary sinuses, incisive canals, and air background). This strategy ensures the efficiency and accuracy of the region growing process.

[0074] The segmentation results for the teeth, skull, jaw, mandibular canal, maxillary sinus, and incisive canal, as output from the model, are stored as volume data in the Nrrd format. Before storage, they must be integrated. This integration process involves two steps: aligning the coordinate systems and prioritizing overlapping regions.

[0075] The alignment processing of the coordinate system includes extracting the segmentation results of the skull, mandible, teeth, neural canal, maxillary sinus, and incisive canal predicted by the neural network model; extracting the segmentation results of the maxillary cortical bone, maxillary cancellous bone, mandibular cortical bone, and mandibular cancellous bone calculated by the region growing method and applying affine transformation to them, so as to spatially align them with the segmentation results predicted by the neural network model.

[0076] The priority decision of overlapping areas includes establishing a blank volume data with a voxel space size equal to the output result of the neural network model and multiple layers according to the Nrrd format regulations, placing the skull, mandible, teeth, neural canal, maxillary sinus, and incisor canal in the first layer, and placing the maxillary cortical bone, maxillary cancellous bone, mandibular cortical bone, and mandibular cancellous bone in the second layer, and writing the segmentation result information saved in each layer into the metadata. Figure 5 and Figure 6 The fine-grained segmentation and slicing results of the maxillofacial region, mandibular ascending ramus and condylar region in an embodiment of the present invention are shown as examples. Some tissue structures can be seen in the slices, where the yellow part is the maxillary / mandibular cortical bone, the green part is the maxillary / mandibular cancellous bone, the red part is the neural canal, and the blue part is the maxillary sinus.

[0077] Finally, visualization methods including the marching cube algorithm and the light diffraction algorithm are used to achieve visual simulation of each structure. In actual clinical applications, doctors can use the visualization interface provided by the present invention to conduct intuitive manual review and local correction of the automatic segmentation results. The present invention can also feed the corrected results back into the training process as incremental data for fine-tuning and optimization of the secondary segmentation neural network model, thereby continuously improving the robustness and personalized adaptability of the secondary segmentation neural network model.

[0078] The present invention also proposes a fine-grained segmentation system for three-dimensional CBCT skull images, the system comprising:

[0079] a first training data set construction unit, configured to construct a first training data set using the acquired original three-dimensional CBCT skull image;

[0080] An image coarse segmentation unit is used to store a coarse segmentation neural network model and train the coarse segmentation neural network model using the first training data set, that is, to perform coarse segmentation on the three-dimensional CBCT skull image to be processed using the trained coarse segmentation neural network model to obtain a corresponding image coarse segmentation result;

[0081] A third training data set construction unit is used to generate a second training data set using the trained coarse segmentation neural network model, and then construct a third training data set;

[0082] an image fine segmentation unit, configured to store a fine segmentation neural network model and train the fine segmentation neural network model using a third training data set, i.e., performing fine segmentation based on the to-be-processed three-dimensional CBCT skull image and the corresponding image coarse segmentation result using the trained fine segmentation neural network model to obtain a final image segmentation result;

[0083] The mandibular fine-grained segmentation unit is used to segment the cortical bone and cancellous bone of the maxillary and mandibular bones based on the segmentation results of the skull and mandible corresponding to the 3D CBCT skull image to be processed, using an improved region growing method to obtain the corresponding cortical bone and cancellous bone segmentation results of the maxillary and mandibular bones.

Claims

1. A fine-grained segmentation method for three-dimensional CBCT skull images, characterized in that: The following steps are involved: Step 1: Construct the first training dataset using the original 3D CBCT skull images. Step 2: After training the secondary segmentation neural network model based on the first training data set, a trained secondary segmentation neural network model is obtained; Step 3: After using the trained two-level segmentation neural network model to segment the 3D CBCT skull image to be processed, the segmentation results of the teeth, skull, mandible, neural canal, maxillary sinus, and incisive canal structures are obtained; Step 4: Extract the skull and mandible segmentation results. Then, based on the skull and mandible segmentation results, use the improved region growing method to segment the cortical bone and cancellous bone of the maxillary and mandibular bones, obtaining the corresponding cortical bone and cancellous bone segmentation results of the maxillary and mandibular bones, thereby completing the fine-grained segmentation of the 3D CBCT skull image. The segmentation of the maxillary cortical bone and cancellous bone segmentation results includes the following steps: S1: Determine the maxillary region based on the skull segmentation results; find all voxels located on the outer edge of the maxillary region and in contact with the background in the maxillary region, use these voxels as initial seed points, and form the initial growth region from these initial seed points; S2: According to the preset update rules, the candidate voxels at the edge of the latest growth area are judged and updated in rounds until the preset conditions are met, and the final growth area is obtained. All voxels in the final growth area together constitute the maxillary cortical bone area and are recorded as the cortical bone segmentation result; the preset update rules include a threshold control rule based on grayscale features, a spatial position constraint rule based on neighboring relationships, a morphological optimization rule based on edge smoothness, a count restriction rule, and a structural exclusivity rule; The counting limit rule specifically stops modifying the state of each voxel when the number of state changes of the voxel reaches a preset number threshold; The method of judging and updating the latest candidate voxels at the edge of the growing region in rounds according to the preset updating rules includes: In each iteration, the leading edge of the current growth region is identified, and a multi-dimensional generation evaluation is performed on all candidate pixels one by one. The multi-dimensional generation evaluation includes all the preset update rules mentioned above. If a voxel meets the growth condition, it is included in the growth region; if it meets the exclusion condition, it is removed from the region. S3: The cancellous bone area is obtained by subtracting the cortical bone area from the maxillary area and recorded as the maxillary cancellous bone segmentation result.

2. The fine-grained segmentation method for three-dimensional CBCT skull images according to claim 1, characterized in that: The step 2 is specifically as follows: The two-level segmentation neural network model includes a coarse segmentation neural network model and a fine segmentation neural network model. First, the coarse segmentation neural network model is trained using a first training data set to obtain a trained coarse segmentation neural network model. Then, the original three-dimensional CBCT skull images in the first training data set are input into the trained coarse segmentation neural network model, and the model outputs corresponding image coarse segmentation results. After all the image coarse segmentation results are added to the first training data set, a second training data set is obtained. Then, each image sample in the second training data set is enhanced in image resolution to obtain a third training data set. Then, the fine segmentation neural network model is trained using the third training data set to obtain a trained fine segmentation neural network model, thereby obtaining a trained secondary segmentation neural network model.

3. The fine-grained segmentation method for three-dimensional CBCT skull images according to claim 2, characterized in that: The step of performing image resolution enhancement processing on each image sample in the second training data set includes: First, the original 3D CBCT skull image, image coarse segmentation result, and corresponding label image in each image sample are resized simultaneously so that the image resolution becomes uniformly r mm per pixel. Then, each resized 3D CBCT skull image and the corresponding image coarse segmentation result are spliced ​​to form a CBCT and coarse segmentation fusion image. Finally, each CBCT and coarse segmentation fusion image is cut into several d2×d2×d2 local fusion images.

4. The fine-grained segmentation method for three-dimensional CBCT skull images according to claim 2, characterized in that: In step 3, the three-dimensional CBCT skull image to be processed is first input into a trained coarse segmentation neural network model, and the model outputs a coarse image segmentation result; the three-dimensional CBCT skull image to be processed and the coarse image segmentation result are enhanced in image resolution and then input into a trained fine segmentation neural network model, and the model outputs segmentation results of the teeth, skull, mandible, mandibular nerve canal, maxillary sinus, and incisive canal structures.

5. A fine-grained segmentation system for three-dimensional CBCT skull images for implementing the method of claim 1, characterized in that: include: a first training data set construction unit, configured to construct a first training data set using the acquired original three-dimensional CBCT skull image; An image coarse segmentation unit, configured to store a coarse segmentation neural network model and train the coarse segmentation neural network model using a first training data set; A third training data set construction unit is used to generate a second training data set using the trained coarse segmentation neural network model, and then construct a third training data set; An image fine segmentation unit, configured to store the fine segmentation neural network model and train the fine segmentation neural network model using a third training data set; The mandibular fine-grained segmentation unit is used to segment the cortical bone and cancellous bone of the maxillary and mandibular bones based on the segmentation results of the skull and mandible using an improved region growing method to obtain the corresponding cortical bone and cancellous bone segmentation results of the maxillary and mandibular bones.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the processor implements the steps of the fine-grained segmentation method for three-dimensional CBCT skull images according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the fine-grained segmentation method for three-dimensional CBCT skull images according to any one of claims 1 to 4 are implemented.

8. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the fine-grained segmentation method for three-dimensional CBCT skull images as claimed in any one of claims 1 to 4 are implemented.

Images