A rodent classification method and device based on deep learning
Through deep learning-based methods, CT scans and feature extraction of rodent teeth are solved, and the problems of inconspicuous and subjective characteristics in traditional classification methods are achieved, and efficient and accurate classification of rodents is achieved.
Patent Information
- Application Number
- CN202510226681.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-27
- Publication Date
- 2025-05-13
- Estimated Expiration
- 2045-02-27
AI Technical Summary
The existing rodent classification methods are based on traditional geometric morphological measurements, making it difficult to accurately classify small teeth, and there are problems of subjectivity and query heavy problems.
A deep learning-based method is used to CT scan the teeth of rodents to generate a three-dimensional model, and the overall features are extracted, such as the ridge between the tooth tip, the diameter of the incisor pulp cavity and the shape of the molar transverse section, and the shape of the molar transverse section for classification.
It realizes efficient, objective and non-destructive classification of rodents, overcomes the problems of inconspicuous characteristics and subjectivity in traditional methods, and improves the accuracy and efficiency of classification.
Smart Images
Figure CN119723222B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of rodent classification, and specifically to a rodent classification method and device based on deep learning. Background Art
[0002] Small mammals generally refer to mammals that are small in size and live in various environments. According to their size and lifestyle, small mammals can be divided into many categories, the most common of which include: bats, lagomorphs, carnivorous small mammals, some primates, rodents, etc. Among them, rodents, such as mice, voles, squirrels, etc., are the most numerous type of mammals on earth. Rodents play a vital role in the ecosystem, involving many aspects such as food chain, ecological balance, species propagation and environmental change. The classification of rodents not only provides us with profound insights into biology, ecology and evolution, but also has a wide range of applications in species protection, agricultural management, disease prevention and control, etc. Through the classification and study of rodents, we can better understand their diversity and ecological functions, so as to effectively protect the ecological environment and human interests.
[0003] In the classification of rodents, although classification based on factors such as skull morphology and animal genes has also made great progress, teeth, as the easiest samples to preserve, still play an important role in classification. In previous classification work based on teeth, researchers used the size and shape of teeth and traditional geometric morphological measurement methods to determine their species. However, the size of rodent teeth is often small, resulting in unclear characteristics, and the external morphology is affected by factors such as the preservation environment. There are certain omissions. For example, the tooth tip is the most important consideration in the existing classification, but the tooth tip is also the structure that is most vulnerable to damage from the external environment. In addition, the classification work is accompanied by an increase in reference materials, the query work is heavy, and it is easy to be dominated by experts, and the classification results are somewhat subjective.
[0004] In summary, how to more accurately classify animals, especially rodents, based on dental information is a technical challenge that researchers need to solve. Summary of the invention
[0005] In order to overcome the shortcomings of existing rodent classification methods, the present invention provides a rodent classification method and device based on deep learning, which can classify rodents efficiently, objectively and non-destructively.
[0006] To achieve the above objectives, the technical contents of the present invention include the following contents.
[0007] A rodent classification method based on deep learning, the method comprising:
[0008] Performing CT scanning on rodent teeth and generating a three-dimensional model of the rodent teeth; wherein the rodent teeth include: incisors, first molars, second molars and third molars;
[0009] Extracting overall features of the rodent teeth based on the three-dimensional model of the rodent teeth; wherein the overall features include: ridges between tooth tips, the diameter of the incisor pulp cavity, and the cross-sectional shape of the molars;
[0010] Classification is performed based on the overall characteristics of the teeth of the rodent to obtain a classification result of the teeth of the rodent.
[0011] Furthermore, the CT scanning of the rodent teeth and the generation of the three-dimensional model of the rodent teeth include:
[0012] Obtain CT images of rodent teeth;
[0013] Performing semantic segmentation on the CT image of the rodent's teeth to obtain a semantic segmentation result;
[0014] The semantic segmentation result is subjected to isosurface extraction and surface volume rendering to generate a three-dimensional model of the rodent teeth of the CT image.
[0015] Further, CT images of rodent teeth are obtained, including:
[0016] Constructing a rodent tooth micro-CT scanning sample stage; wherein the rodent tooth micro-CT scanning sample stage comprises a cylindrical support mechanism and a tray I, a tray M1, a tray M2, and a tray M3 fixed to the cylindrical support mechanism from top to bottom;
[0017] The rodent teeth are placed on the rodent tooth micro-CT scanning sample stage; wherein, the incisors of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the I tray respectively, the first molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M1 tray respectively, the second molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M2 tray respectively, and the third molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M3 tray respectively;
[0018] A full-scale fast CT scanner was used to perform four-segment scans from the bottom of the M3 tray upwards, and after each segment of scanning was completed, the scanning center automatically dropped to obtain four-segment CT images;
[0019] Perform beam hardening correction, automatic geometric correction and optimized filter processing on each CT image;
[0020] According to the overlapping areas in each CT image segment, each processed CT image segment is spliced to obtain the original CT image of the rodent's teeth;
[0021] The grayscale value of the original CT image is obtained according to the average grayscale value of each segment of the CT image, and the original CT image is normalized based on the grayscale value of the original CT image to obtain a CT image of the rodent's teeth.
[0022] Furthermore, semantic segmentation is performed on the CT image of the rodent's teeth to obtain a semantic segmentation result, including:
[0023] Based on the CT image of the rodent's teeth, the grayscale mean of the rodent's teeth is obtained. , the grayscale mean of the surrounding noise And the overall gray value of the CT image ;
[0024] By calculating the optimal threshold The between-class variance , get the optimal threshold ;in, is the probability of a rodent tooth, is the probability of noise;
[0025] Applying the best threshold Performing binary image segmentation on the CT image of the rodent's teeth to obtain a binary CT image of the rodent's teeth;
[0026] After removing the external small particles and internal small gaps in the binarized rodent tooth CT image, the adjacent pixels with similar attributes are merged by selecting the seed points of the rodent teeth. The watershed segmentation algorithm is combined with the manual circle selection method to group pixels with similar properties into semantic regions, and the semantic segmentation results of rodent teeth are obtained.
[0027] Further, the semantic segmentation result is subjected to isosurface extraction and surface volume rendering to generate a three-dimensional model of rodent teeth of the CT image, including:
[0028] Generate the original 3D tooth model based on the semantic segmentation results;
[0029] Move each vertex in the original three-dimensional tooth model toward the average position of its neighboring vertices to obtain a three-dimensional tooth model after noise removal, and extract the incisor part of the three-dimensional tooth model after noise removal;
[0030] The three-dimensional tooth model after noise removal is optimized to obtain an optimized three-dimensional tooth model, and the molar part in the optimized three-dimensional tooth model is extracted; wherein the optimization process includes:
[0031] The denoised 3D tooth model is smoothed in positive and negative directions using Laplacian, and local smoothing weights are calculated based on the principal curvature to retain high curvature features.
[0032] Use QEM simplification to reduce the number of triangles while maintaining key geometric features;
[0033] Remove isolated fragments through topology optimization and repair of self-intersections and abnormal meshes;
[0034] The extracted incisor parts and molar parts are combined to obtain a three-dimensional model of rodent teeth in the CT image.
[0035] Furthermore, based on the three-dimensional model of the rodent teeth, the overall features of the rodent teeth are extracted, including:
[0036] Preprocessing the three-dimensional model of rodent teeth; wherein the preprocessing includes: posture normalization, mean curvature map calculation and multi-view rendering;
[0037] The VGG16 network model trained in the ImageNet dataset was used as a migration network, and the overall features of the rodent teeth were extracted based on the migration network.
[0038] Furthermore, the overall features of the rodent teeth are extracted based on the migration network, including:
[0039] Outputting a low-level feature map of the three-dimensional model of the rodent tooth using the front part of the migration network, wherein the low-level features in the low-level feature map include: an outer contour of the tooth pulp cavity and an outer contour line of the tooth tip;
[0040] The low-level features are recombined using the middle part of the migration network to obtain an intermediate feature map; wherein the intermediate features in the intermediate feature map include: the tooth tip structure of rodent teeth and the cavity connectivity of the dental pulp cavity;
[0041] The intermediate features are recombined using the latter part of the migration network to obtain a high-level feature map; wherein the high-level features in the high-level feature map include: ridges between tooth tips, diameter of the pulp cavity of the incisors, and cross-sectional shapes of the molars.
[0042] A rodent classification device based on deep learning, the device comprising:
[0043] A model generation module, used for performing CT scanning on rodent teeth and generating a three-dimensional model of rodent teeth; wherein the rodent teeth include: incisors, first molars, second molars and third molars;
[0044] A feature extraction module, for extracting overall features of the rodent teeth based on the three-dimensional model of the rodent teeth; wherein the overall features include: ridges between tooth tips, the diameter of the incisor pulp cavity, and the cross-sectional shape of the molars;
[0045] The species classification module is used to classify the teeth of the rodent according to the overall characteristics of the teeth to obtain the classification results of the teeth of the rodent.
[0046] An electronic device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the deep learning-based rodent classification method described in any one of the above items is implemented.
[0047] A computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement any of the above-mentioned deep learning-based rodent classification methods.
[0048] A computer program product, characterized in that when the computer program product is run on a computer device, the computer device executes any of the deep learning-based rodent classification methods described above.
[0049] Compared with the prior art, the present invention has at least the following beneficial effects.
[0050] 1) Non-destructively combine the surface structure of rodent teeth with the internal information of the pulp cavity to provide more complete and accurate identification indicators, overcoming the technical difficulty of being unable to classify due to partial surface structure damage.
[0051] 2) Apply the specially designed rodent tooth sample stand to improve the acquisition efficiency of rodent tooth CT data.
[0052] 3) The innovative segmented stitching scanning method meets the research needs of paleontologists for detailed observation of the three-dimensional microscopic structure of rodent teeth.
[0053] 4) According to the characteristics of teeth, incisors and molars are modeled separately within one technical framework, which not only improves the classification accuracy but also improves the efficiency of modeling. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 It is a flowchart of the rodent teeth classification method provided in an embodiment of the present invention.
[0055] Figure 2 It is a schematic diagram of the identification of the surface structure of rodent teeth provided by an embodiment of the present invention.
[0056] Figure 3 It is a schematic diagram of CT image identification of the pulp cavity of a rodent incisor provided by an embodiment of the present invention.
[0057] Figure 4 It is a schematic diagram of CT image identification of the dental pulp cavity of a rodent molar provided by an embodiment of the present invention.
[0058] Figure 5 It is a schematic diagram of the structure of a rodent tooth micro-CT scanning sample stage provided in an embodiment of the present invention.
[0059] Figure 6 It is a flowchart of a rodent tooth three-dimensional model classification network model creation process based on deep learning provided by an embodiment of the present invention.
[0060] Figure 7 is a block diagram of a rodent classification device provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0061] In order to make the purpose, scheme and advantages of the present invention more clear, the present invention is further described in detail by taking experiments conducted on real data sets as an example. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0062] The rodent classification method of the present invention includes rodent tooth identification, rodent tooth CT data acquisition and high-precision CT reconstruction, rodent tooth CT image semantic segmentation, rodent tooth CT image surface volume rendering based on walking cubes, and rodent tooth three-dimensional model classification based on deep learning.
[0063] Specifically, if Figure 1 As shown, the present invention has the following steps:
[0064] Step 1: Identify the rodent tooth.
[0065] The present invention constructs a training data set by identifying rodent teeth. Figure 2 As shown, rodent teeth were mainly collected from common species of the family Muridae in northern China, including Rattus norvegicus Rattus norvegicus , social rat Niviventer confucianus , Mus musculus Mus musculus , Apodemus agrarius Apodemus agrariusEtc. The crown length of the maxillary first molar M1 of Apodemus agrarius exceeds half of the maxillary dentition length, that is, the length of M1 exceeds the sum of the lengths of M2 and M3, the cusp t12 is moderately developed, weak or missing; the cusp t3 of the maxillary second molar M2 is missing; the maxillary third molar M3 is short. The cusp t1 of the maxillary first molar M1 of the radix raptor is thick, the cusp t3 is underdeveloped, and the cusp t9 is weak; the cusp t3 of the maxillary second molar M2 is missing; the cusp t3 of the maxillary third molar M3 is missing. The cusp t7 of the maxillary first molar M1 of the house mouse is missing, the cusp t4 is not connected to the cusp t8, the cusp t9 is weak and slightly far from the cusp t6; the cusp t1 of the maxillary second molar M2 is thick, and the cusp t3 is missing; the maxillary third molar M3 is a small triangle. The cusp t3 of the maxillary first molar M1 of Rattus norvegicus is weak and tends to fuse with the cusp t2, and the cusps t9 and t8 tend to fuse; the cusp t3 of the maxillary second molar M2 is sometimes present or sometimes missing; the cusp t3 of the maxillary third molar M3 is missing. Figure 3 and Figure 4 As shown in the figure, the pulp cavity of the incisor of Apodemus agrarius is not too large, and its diameter is generally around 0.3-0.4 mm. The pulp cavity of the molar is complex in shape with multiple branches, and the cross section is irregular in shape, with multiple small chambers connected by narrow channels. The overall size of the pulp cavity of the incisor of the social mouse is usually larger than that of the black-lined mouse, and the diameter of the pulp cavity of the incisor is about 0.5-0.6 mm. The branches of the pulp cavity of the molar are more regular than those of the black-lined mouse, with relatively obvious main branches and secondary branches, and the overall structure is relatively clear. The pulp cavity of the incisor of the house mouse is the smallest, and the diameter of the pulp cavity of the incisor is usually between 0.2-0.3 mm. The pulp cavity of the molar has simple branches, and the connecting channels between the chambers are relatively wide from the cross section. The pulp cavity of the incisor of the brown rat is the largest, and the diameter of the pulp cavity of the incisor can reach between 0.9-1 mm. The branches of the pulp cavity of the molar are extremely complex, with multiple chambers of different sizes interconnected to form a highly complex network structure.
[0066] Step 2: Obtain CT images of rodent teeth.
[0067] The present invention uses a full-scale fast CT scanner to obtain a rodent tooth CT image based on the rodent tooth CT image acquisition standard, thereby reducing the noise and artifact effects of the rodent tooth CT image. The rodent tooth CT image can be a rodent tooth CT image in DICOM format or a rodent tooth CT image in a non-DICOM ordinary image BMP / TIF format.
[0068] In one example, in order to better classify and organize dental data, the present invention designs a rodent tooth micro-CT scanning sample stage based on the characteristics of rodent teeth. Although the traditional sample stage and its scanning method can also realize rodent CT image acquisition and have the characteristics of small amount of calculation, when a large number of rodent tooth CT images need to be collected, the sample stage needs to be frequently replaced, and the scanning efficiency is low, which is not conducive to data classification and organization. In order to achieve fast and efficient CT scanning of rodent teeth, the present invention designs a rodent tooth micro-CT scanning sample stage, such as Figure 5 As shown, it includes: a cylindrical support mechanism with an inner diameter of 2 cm, a wall thickness of 0.2 cm, a total height of 10 cm, a hollow structure depth of 3 cm, and a width of 0.5 cm for the slots on both sides; and four layers of circular trays that can be stacked through the slots, with an inner diameter of 1.5 cm, a wall thickness of 0.2 cm, a hollow structure depth of 0.4 cm, a bottom thickness of 0.1 cm, and a width of 0.5 cm for the protruding parts on both sides. The outer side of the tray is engraved with the words I, M1, M2, and M3, representing the incisor, the first molar, the second molar, and the third molar. The bottom area of the circular tray is divided into four equal parts by a cross line, and the words TL, TR, BL, and BR are engraved, representing the left side of the upper jaw, the right side of the upper jaw, the left side of the mandible, and the right side of the mandible.
[0069] During scanning, the present invention places the incisors, first molars, second molars, and third molars of the same species of rodents into the trays I, M1, M2, and M3 of the rodent tooth micro-CT scanning sample stage designed by the present invention, and the teeth of the left side of the upper jaw, the right side of the upper jaw, the left side of the mandible, and the right side of the mandible are placed into the corresponding partitions TL, TR, BL, and BR of the tray, respectively. For example, the first molar of the left side of the upper jaw of the same species of rodents is placed in the TL partition of the M1 tray, and a full-scale fast CT scanner (GE v|tome|xm300&180) is used to obtain the CT image of the rodent teeth. The supporting structure of the rodent tooth micro-CT scanning sample stage is placed vertically in the CT scanning room, and the segmented splicing scanning method is used to perform segmented scanning along the vertical direction of the supporting structure. The CT scan is performed from the bottom of the M3 tray upward, and four segments are scanned. After each segment scan is completed, the scanning center will automatically drop at equal intervals for subsequent scanning. After completing four segmented scans, the Multi Scan Reconstruction algorithm was used to obtain high-precision CT images of each segment, and the seamless splicing of rodent teeth was achieved based on the overlapping areas in each segment of the CT image. The average grayscale value of each scanned CT image was defined as the grayscale value of the entire rodent tooth CT image, and the spliced rodent tooth CT image was normalized to finally obtain the CT image of the rodent tooth. Among them, the parameters of the full-scale fast CT scanner (GE v|tome|x m300&180) were set as follows: voltage: 120kV; current: 120uA; exposure time: 1000ms; number of superimposed frames: 3 frames; number of projection images: 1500; voxel size: 4.5um; filter; 0.2mm thick copper.
[0070] Step 3: Perform semantic segmentation on the CT image to obtain a semantic segmentation result.
[0071] This embodiment uses the Otsu method to confirm the intra-class variance of the rodent teeth and the noise around the teeth in the whole CT image, and determines the optimal threshold value by traversing the maximum value of the intra-class variance to perform the binarization image threshold segmentation of the whole CT image. The optimal threshold value can divide the grayscale histogram of the whole CT image into two categories (teeth and the noise around them), so that the variance between the two categories is maximized. The grayscale value of the whole CT image of this embodiment is 0-65535 levels, 16-bit grayscale image. In general, when the grayscale distribution histogram of the whole CT image is bimodal, the optimal threshold value is T It should fall at the trough between the two peaks. The optimal threshold calculation formula for the overall CT image is as follows:
[0072]
[0073] in, is the threshold value; When the threshold is The between-class variance at time ; , are the probabilities of the tooth and the noise around it, respectively; and are the mean values of the noise of the tooth and its surroundings, respectively; is the gray value of the overall CT image.
[0074] Applying the best threshold The whole CT image is binarized and segmented, and the teeth are selected as the region of interest. Morphological operation methods are used to remove external small particles and internal small gaps in the binarized tooth CT image through corrosion, expansion and other operations, and the main structure in the tooth CT image is retained. The regional growing algorithm is used to select the seed point of the tooth to merge adjacent pixels with similar attributes and combine the manual circle selection method to group pixels with similar properties to form a semantic region. Since rodent teeth, especially molars on the complex occlusal surface of murines, are characterized by close contact, the present invention uses a watershed segmentation method to separate rodent teeth and realize semi-automatic semantic segmentation of rodent tooth CT images.
[0075] Step 4: Perform isosurface extraction and surface volume rendering on the semantic segmentation results to generate a three-dimensional model of rodent teeth in the CT image.
[0076] Incisors of rodents have a simple structure, usually a single block structure, a distinct enamel layer, a clear distinction from dentin, a smooth surface, and sharp edges, which require the retention of precise geometric features. Molars, on the other hand, have a complex structure, with multiple cusps and ridges on the occlusal surface, and a fuzzy boundary between enamel and dentin, requiring more detailed reconstruction. In view of the different characteristics of rodent teeth, the present invention uses a specific method to reconstruct incisors and molars within the same technical framework, and while ensuring the reconstruction efficiency, both incisors and molars can obtain good reconstruction effects.
[0077] Step 4.1: Generate the original tooth 3D model based on the semantic segmentation results.
[0078] The present invention first sets an isovalue surface for segmenting the foreground and background, and then uses the MarchingCubes algorithm to traverse the voxel data in the semantic segmentation result, calculates the vertices for each cube unit and connects them to generate the original three-dimensional tooth model.
[0079] Step 4.2: Move each vertex in the original three-dimensional tooth model toward the average position of its neighboring vertices to obtain a three-dimensional tooth model from which noise is removed, and extract the incisor part of the three-dimensional tooth model from which noise is removed.
[0080] Since the original tooth 3D model may have a jagged surface (high-frequency noise), small floating fragments (topological noise), and a high-density but redundant mesh, the original tooth 3D model needs to be further processed.
[0081] The present invention iteratively updates vertex coordinates to move each vertex toward the average position of neighboring vertices, thereby obtaining vertex New coordinates of , and based on the new coordinates Generate a 3D tooth model with noise removed. The new coordinates , Vertex The original coordinates of Represents a vertex The set of neighbor vertices of Neighbor vertices The original coordinates of is the smoothing factor.
[0082] Due to the above-mentioned incisor characteristics, the accuracy of the reconstructed incisors can meet the requirements of subsequent classification, while excessive processing will not only increase the amount of calculation of the algorithm, but also reduce the classification accuracy due to individual differences of the same type. Therefore, the present invention retains the incisor part of the tooth three-dimensional model after removing the noise.
[0083] Step 4.3: Optimize the tooth three-dimensional model after removing the noise to obtain an optimized tooth three-dimensional model, and extract the molar part in the optimized tooth three-dimensional model.
[0084] The present invention first alternately uses Laplacian smoothing in the positive and negative directions to reduce the shrinkage effect, and then calculates the local smoothing weight based on the principal curvature to retain the high curvature features. Based on QEM simplification, the number of unimportant triangles is reduced while maintaining key geometric features. Isolated fragments are removed through topology optimization and repair of self-intersection and abnormal meshes to enhance the crown edge features of the molars.
[0085] Step 4.4: Combine the extracted incisor parts and molar parts to obtain a three-dimensional model of the rodent teeth of the CT image.
[0086] In addition, in one example, the present invention uses the visualization toolkit VTK to provide a support environment for generating a three-dimensional model of rodent teeth for CT images. The present invention uses the vtkMarchingCubes class in VTK to implement surface volume rendering of rodent teeth CT images. The output rodent tooth three-dimensional model file is usually in the format of STL, OBJ, PLY, etc. The rodent tooth three-dimensional model is composed of voxels containing color and measurement value information. The appearance of the three-dimensional model can be changed arbitrarily (such as color, transparency, etc.), and can be used for research on three-dimensional data visualization. At present, STL files are one of the most popular three-dimensional model format files for sharing. They can be read by three-dimensional data visualization software such as Meshlab or ImageJ, and can also be applied to digital-physical tasks such as 3D printing.
[0087] Step 5: Classify based on the three-dimensional model of rodent teeth to obtain the classification results of rodents.
[0088] The rodent tooth 3D model to be classified is input into the trained rodent tooth 3D model classification network model based on deep learning to obtain the rodent 3D model classification result, realize the rodent 3D model identification, and output the rodent 3D model identification result. The process mainly includes: 3D model data preprocessing, feature extraction, feature fusion and modeling classification.
[0089] The 3D model data preprocessing includes posture normalization, mean curvature map calculation and multi-view rendering functions. The present invention adopts the PCA method to perform posture normalization on the 3D model of rodent teeth, applies the mean curvature map to enhance the local geometric features of the 3D model of rodent teeth, such as the connectivity of the dental pulp cavity chamber and the ridges between the tooth tips, and uses the multi-view rendering tool to collect multi-views of the 3D model of rodent teeth.
[0090] The feature extraction of the three-dimensional model is mainly achieved by transfer learning technology, using the VGG16 network model trained in the ImageNet dataset as the transfer network. The low-level feature map output by the front part of the transfer network, such as the outer contour of the pulp cavity and the outer contour line of the tooth tip of the rodent tooth. The intermediate feature map in the middle part of the transfer network is a recombination of low-level features. The intermediate feature is close to the cavity connectivity of the tooth tip structure and the pulp cavity of the rodent tooth. The high-level feature map output by the back part of the transfer network is a recombination of the intermediate features. The high-level features can highly summarize the overall structure of rodent teeth, including the ridges between the tooth tips and the branches of the pulp cavity. The feature fusion part is mainly completed by the long short-term memory network in the recurrent neural network.
[0091] The support vector machine multi-classifier is selected for the classification of the three-dimensional model of rodent teeth. A series of hyperplanes are calculated by the support vector machine multi-classifier to distinguish the predicted class of the three-dimensional model of rodent teeth from other classes. The hyperparameters of each hyperplane are represented by w j , the set of hyperparameter vectors is represented as W={w1,w2,…,w k}, k represents the type of rodent teeth in the experimental data set. The SVM classifier needs to calculate k hyperparameters to complete the classification task of the three-dimensional model of rodent teeth. The quadratic optimization method is used to solve the unknown hyperparameter vector set W of the SVM multi-class classifier.
[0092] During training, the main shape features of rodent teeth are combined with the corresponding genus and species names to mark and classify the three-dimensional models of rodent teeth to obtain the genus and species name labels. The three-dimensional models of rodent teeth and the corresponding genus name labels are input as experimental data sets to train, verify and test the network model. The original weight file of the pre-trained improved MVCNN network model is loaded, and the classification results of the three-dimensional models of rodent teeth are compared with the real labels, and the three-dimensional models of rodent teeth classification network model based on deep learning are trained through continuous regression, the accuracy of the classification of the three-dimensional models of rodent teeth is adjusted, and the parameter weights of the optimized improved MVCNN network model are saved.
[0093] like Figure 6As shown, the process of creating a network model for classification of rodent teeth three-dimensional models based on deep learning provided by an embodiment of the present invention. The three-dimensional models of rodent teeth in the training set, validation set and test set of the experimental data set are input into the data preprocessing module, and the posture normalization of the three-dimensional models of rodent teeth is ensured by translation, scaling and rotation operations. The ridges between the tooth tips and the branches of the pulp cavity of the local microstructure of the three-dimensional model of rodent teeth are enhanced by the average curvature, and multi-view views of the three-dimensional model of rodent teeth are collected using an application developed in combination with OpenGL and OpenCV. The multi-view views of the three-dimensional model of rodent teeth are input into the improved MVCNN network model training network. First, the VGG16FT network model for feature extraction is trained, the number of iterations is set to 100, the optimizer is the Adam function, the batch size is set to 4, the initial learning rate is set to 1e-4, and the loss function of the VGG16FT network model for feature extraction is the cross entropy function. Then, the LSTM network model for feature fusion was trained, the number of iterations was set to 100, the optimizer was the Adam function, the batch size was set to 36, the initial learning rate was set to 1e-3, and the loss function of the LSTM network model for feature fusion was the cross entropy function. Finally, the SVM multi-classifier was trained, the number of iterations was set to 3500, and the regularization parameter was set to 0.1 to obtain the optimal deep learning-based rodent tooth three-dimensional model classification network model.
[0094] In summary, in view of the fact that rodents are numerous and numerous, the patent of this invention designs a sample stage specifically for microscopic scanning of rodent teeth; in view of the fact that rodent classification is extremely time-consuming and requires certain professional knowledge to complete, the patent of this invention designs a rodent classification method based on deep learning. This solves the problem that the frequent sample replacement during micro-CT scanning of rodent teeth is inefficient and not conducive to data classification and sorting, and at the same time, the deep learning method is used to make rodent classification more objective and efficient.
[0095] Please refer to Figure 7 , which shows a block diagram of a rodent classification device provided by an embodiment of the present invention. The device can be a computer device, or can be set in a computer device. Figure 7 As shown, the device includes the following modules: a model generation module, a feature extraction module and a species classification module.
[0096] A model generation module, used for performing CT scanning on rodent teeth and generating a three-dimensional model of rodent teeth; wherein the rodent teeth include: incisors, first molars, second molars and third molars;
[0097] A feature extraction module, for extracting overall features of the rodent teeth based on the three-dimensional model of the rodent teeth; wherein the overall features include: ridges between tooth tips, the diameter of the incisor pulp cavity, and the cross-sectional shape of the molars;
[0098] The species classification module is used to classify the teeth of the rodent according to the overall characteristics of the teeth to obtain the classification results of the teeth of the rodent.
[0099] For the detailed description of the implementation process, beneficial effects, etc. of the device module, please refer to the introduction of the above method embodiment, which will not be elaborated here.
[0100] In an exemplary embodiment, a computer device is also provided, which includes a memory and a processor, wherein a computer program is stored in the memory, and the computer program is loaded and executed by the processor to implement the above-mentioned deep learning-based rodent classification method.
[0101] In an exemplary embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, and when the computer program is executed by a processor, the deep learning-based rodent classification method as described above is implemented.
[0102] In an exemplary embodiment, a computer program product is also provided. When the computer program product is run on a computer device, the computer device executes the rodent classification method based on deep learning as described above.
[0103] The above description is only an embodiment of the present invention and is not intended to limit the present invention. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.
Claims
1. A rodent classification method based on deep learning, characterized in that: The method comprises: Performing CT scanning on rodent teeth and generating a three-dimensional model of rodent teeth; wherein the rodent teeth include: incisors, first molars, second molars and third molars, and performing CT scanning on rodent teeth and generating a three-dimensional model of rodent teeth include: Obtaining a CT image of the teeth of a rodent; wherein obtaining the CT image of the teeth of a rodent comprises: Constructing a rodent tooth micro-CT scanning sample stage; wherein the rodent tooth micro-CT scanning sample stage comprises a cylindrical support mechanism and a tray I, a tray M1, a tray M2, and a tray M3 fixed to the cylindrical support mechanism from top to bottom; The rodent teeth are placed on the rodent tooth micro-CT scanning sample stage; wherein, the incisors of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the I tray respectively, the first molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M1 tray respectively, the second molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M2 tray respectively, and the third molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M3 tray respectively; A full-scale fast CT scanner was used to perform four-segment scans from the bottom of the M3 tray upwards, and after each segment of scanning was completed, the scanning center automatically dropped to obtain four-segment CT images; Perform beam hardening correction, automatic geometric correction and optimized filter processing on each CT image; According to the overlapping areas in each CT image segment, each processed CT image segment is spliced to obtain the original CT image of the rodent's teeth; Obtaining a grayscale value of an original CT image according to an average grayscale value of each segment of the CT image, and normalizing the original CT image based on the grayscale value of the original CT image to obtain a CT image of the rodent's teeth; Performing semantic segmentation on the CT image of the rodent's teeth to obtain a semantic segmentation result; Performing isosurface extraction and surface volume rendering on the semantic segmentation result to generate a three-dimensional model of rodent teeth of the CT image; Extracting overall features of the rodent teeth based on the three-dimensional model of the rodent teeth; wherein the overall features include: ridges between tooth tips, the diameter of the incisor pulp cavity, and the cross-sectional shape of the molars; Classification is performed based on the overall characteristics of the teeth of the rodent to obtain a classification result of the teeth of the rodent.
2. The method according to claim 1, characterized in that Performing semantic segmentation on the CT image of the rodent's teeth to obtain a semantic segmentation result, including: Based on the CT image of the rodent's teeth, the grayscale mean of the rodent's teeth is obtained. , the grayscale mean of the surrounding noise And the overall gray value of the CT image ; By calculating the optimal threshold The between-class variance , get the optimal threshold ;in, is the probability of a rodent tooth, is the probability of noise; Applying the best threshold Performing binary image segmentation on the CT image of the rodent's teeth to obtain a binary CT image of the rodent's teeth; After removing the external small particles and internal small gaps in the binarized rodent tooth CT image, the adjacent pixels with similar attributes are merged by selecting the seed points of the rodent teeth. The watershed segmentation algorithm is combined with the manual circle selection method to group pixels with similar properties into semantic regions, and the semantic segmentation results of rodent teeth are obtained.
3. The method according to claim 1, characterized in that The semantic segmentation result is subjected to isosurface extraction and surface volume rendering to generate a three-dimensional model of rodent teeth of the CT image, including: Generate the original 3D tooth model based on the semantic segmentation results; Move each vertex in the original three-dimensional tooth model toward the average position of its neighboring vertices to obtain a three-dimensional tooth model after noise removal, and extract the incisor part of the three-dimensional tooth model after noise removal; The three-dimensional tooth model after noise removal is optimized to obtain an optimized three-dimensional tooth model, and the molar part in the optimized three-dimensional tooth model is extracted; wherein the optimization process includes: The denoised 3D tooth model is smoothed in positive and negative directions using Laplacian, and local smoothing weights are calculated based on the principal curvature to retain high curvature features. Use QEM simplification to reduce the number of triangles while maintaining key geometric features; Remove isolated fragments through topology optimization and repair of self-intersections and abnormal meshes; The extracted incisor parts and molar parts are combined to obtain a three-dimensional model of rodent teeth in the CT image.
4. The method according to claim 1, characterized in that: Based on the three-dimensional model of the rodent teeth, the overall features of the rodent teeth are extracted, including: Preprocessing the three-dimensional model of rodent teeth; wherein the preprocessing includes: posture normalization, mean curvature map calculation and multi-view rendering; The VGG16 network model trained in the ImageNet dataset was used as a migration network, and the overall features of the rodent teeth were extracted based on the migration network.
5. The method according to claim 4, characterized in that The overall features of the rodent teeth are extracted based on the migration network, including: Outputting a low-level feature map of the three-dimensional model of the rodent tooth using the front part of the migration network, wherein the low-level features in the low-level feature map include: an outer contour of the tooth pulp cavity and an outer contour line of the tooth tip; The low-level features are recombined using the middle part of the migration network to obtain an intermediate feature map; wherein the intermediate features in the intermediate feature map include: the tooth tip structure of rodent teeth and the cavity connectivity of the dental pulp cavity; The intermediate features are recombined using the latter part of the migration network to obtain a high-level feature map; wherein the high-level features in the high-level feature map include: ridges between tooth tips, diameter of the pulp cavity of the incisors, and cross-sectional shapes of the molars.
6. A rodent classification device based on deep learning, characterized in that: The device comprises: A model generation module is used to perform CT scanning on rodent teeth and generate a three-dimensional model of rodent teeth; wherein the rodent teeth include: incisors, first molars, second molars and third molars, and the CT scanning of rodent teeth and the generation of a three-dimensional model of rodent teeth include: Obtaining a CT image of the teeth of a rodent; wherein obtaining the CT image of the teeth of a rodent comprises: Constructing a rodent tooth micro-CT scanning sample stage; wherein the rodent tooth micro-CT scanning sample stage comprises a cylindrical support mechanism and a tray I, a tray M1, a tray M2, and a tray M3 fixed to the cylindrical support mechanism from top to bottom; The rodent teeth are placed on the rodent tooth micro-CT scanning sample stage; wherein, the incisors of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the I tray respectively, the first molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M1 tray respectively, the second molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M2 tray respectively, and the third molars of the upper jaw on the left side, the upper jaw on the right side, the mandibular on the left side and the mandibular on the right side are placed in the TL partition, TR partition, BL partition and BR partition of the M3 tray respectively; A full-scale fast CT scanner was used to perform four-segment scans from the bottom of the M3 tray upwards, and after each segment of scanning was completed, the scanning center automatically dropped to obtain four-segment CT images; Perform beam hardening correction, automatic geometric correction and optimized filter processing on each CT image; According to the overlapping areas in each CT image segment, each processed CT image segment is spliced to obtain the original CT image of the rodent's teeth; Obtaining a grayscale value of an original CT image according to an average grayscale value of each segment of the CT image, and normalizing the original CT image based on the grayscale value of the original CT image to obtain a CT image of the rodent's teeth; Performing semantic segmentation on the CT image of the rodent's teeth to obtain a semantic segmentation result; Performing isosurface extraction and surface volume rendering on the semantic segmentation result to generate a three-dimensional model of rodent teeth of the CT image; A feature extraction module, for extracting overall features of the rodent teeth based on the three-dimensional model of the rodent teeth; wherein the overall features include: ridges between tooth tips, the diameter of the incisor pulp cavity, and the cross-sectional shape of the molars; The species classification module is used to classify the teeth of the rodent according to the overall characteristics of the teeth to obtain the classification results of the teeth of the rodent.
7. An electronic device, characterized in that: The electronic device comprises: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the deep learning-based rodent classification method as described in any one of claims 1 to 5 is implemented.
8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the rodent classification method based on deep learning as described in any one of claims 1 to 5.