CN114549523A - Single-step depth network-based automatic detection method for multiple raw teeth in center of curved surface layer graph - Google Patents
Single-step depth network-based automatic detection method for multiple raw teeth in center of curved surface layer graph Download PDFInfo
- Publication number
- CN114549523A CN114549523A CN202210436787.0A CN202210436787A CN114549523A CN 114549523 A CN114549523 A CN 114549523A CN 202210436787 A CN202210436787 A CN 202210436787A CN 114549523 A CN114549523 A CN 114549523A
- Authority
- CN
- China
- Prior art keywords
- curved surface
- bounding box
- teeth
- surface layer
- layer image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10081—Computed x-ray tomography [CT]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20092—Interactive image processing based on input by user
- G06T2207/20104—Interactive definition of region of interest [ROI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30036—Dental; Teeth
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a single-step depth network-based automatic detection method for multiple fresh teeth in the center of a curved surface layer graph, which is characterized by comprising the following steps of: collecting a new curved surface body layer image; inputting the new curvohedral layer image into a single-step depth network model obtained by training, and obtaining a plurality of candidate bounding box positions and target confidence coefficients containing target classes belonging to a plurality of median teeth; and screening a final boundary frame from the plurality of candidate boundary frames by using a non-maximum inhibition method to obtain the final positioning of the median molar excess raw teeth. The invention can quickly and automatically identify whether the image contains more middle raw teeth and give the position of the middle raw teeth only by scanning the curved surface layer image once, thereby avoiding the influence of experience difference of doctors on the precision and assisting the doctors to quickly and correctly diagnose more middle raw teeth.
Description
Technical Field
The invention relates to a single-step depth network-based automatic detection method for multiple raw teeth in the center of a curved surface layer graph, and belongs to the technical field of computer vision prediction.
Background
The multiple raw teeth refer to teeth or tooth-like tissues which are present in jawbone and are not 20 in deciduous dentition or 32 in permanent dentition, and are common diseases of abnormal development of tooth number. Multiple raw teeth often cause cystic lesions, delayed sprouting of adjacent teeth, ectopic sprouting of adjacent teeth, root absorption of adjacent teeth or torsion of adjacent teeth and other complications. Most of the multiple teeth are buried in the jaw bone, so two-dimensional curvographic body layer images are a common means of diagnosing and locating multiple teeth. Compared with Cone Beam Computed Tomography (CBCT), the method has the advantages of small radiation amount and low cost. Currently, clinicians mainly rely on artificial diagnosis to determine whether there are more middle raw teeth from the curved surface layer images. However, the accuracy of the detection results obtained by manual diagnosis is greatly affected by the difference in experience of clinicians. Therefore, the development of the automatic detection algorithm for the multiple middle raw teeth of the curveyoid layer image has important clinical value.
In view of the excellent performance of deep learning in the field of computer vision, many researchers are beginning to use deep convolutional neural networks for oral intelligent detection diagnosis. For example, Fukuda et al established a CNN-based deep learning model for detecting root fissure using detectenet, and performed five-fold cross validation for protecting test data selection bias and improving reliability. Ekert et al proposed a 7-layer CNN model to detect root tip lesions on dental panoramic films, trained and validated through 10 repeated data rewashing. Inspired by the above work, Kuwada et al created three learning models using AlexNet, VGG-16 and DetectNet for automatic detection of multiple fresh teeth in the center of a curved surface layer image, trained learning of the curved surface layer image of a patient in the upper jaw incisor area, and verification and comparison of the classification effect of 3 deep learning systems on the upper jaw multiple fresh teeth of a patient with full molar teeth. While the above work has performed well in the classification task, there are some deficiencies. An important step of the Kuwada method is to manually crop the image blocks including the possible locations of the lesion area from the test image into the network before classification begins. The size of the cropped image block is also manually set according to experience, but not according to the image content, and errors are often generated in the clinical diagnosis and data annotation processes. In addition, inputting image blocks containing only local information into the training and testing phase may lose critical global information.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides an automatic detection method for multiple median raw teeth of a curved surface layer graph based on a single-step depth network.
In order to achieve the above object, the present invention provides a method for automatically detecting multiple median raw teeth in a curved surface layer map based on a single-step depth network, comprising:
collecting a new curved surface body layer image;
inputting the new curvohedral layer image into a single-step depth network model obtained by training, and obtaining a plurality of candidate bounding box positions and target confidence coefficients containing target classes belonging to a plurality of median teeth;
and screening a final boundary frame from the plurality of candidate boundary frames by using a non-maximum inhibition method to obtain the final positioning of the median molar excess raw teeth.
Preferably, the training obtains a single-step deep network model, comprising:
(1) expanding a training set, wherein the training set comprises a curved surface layer image, a cone beam CT image, classification labeling information and positioning labeling information;
(2) a single-step deep network model is constructed by iterative training of a training set;
in each iterative training, setting an initial value of a hyper-parameter by adopting automatic learning anchor distribution and an Adam optimizer;
and (3) carrying out hyper-parameter search by using a genetic algorithm, training the single-step deep network model according to the set epoch, and sequencing all hyper-parameter combinations according to a total loss function to obtain the optimal hyper-parameter combination.
Preferably, the single-step deep network model comprises a basic framework end, a detection end and an output end;
the basic skeleton end comprises a focusing layer,S 1 a plurality of convolution layers,S 2 A cross-stage local-1 module, a spatial pyramid pooling module, a cross-stage local-2 module, a focusing layer,S 1 a plurality of convolution layers,S 2 The cross-stage local-1 module, the spatial pyramid pooling module and the cross-stage local-2 module are sequentially connected;
the focusing layer performs connection operation on the input curved surface layer image in channel dimension after 4 slicing operations, and obtains a characteristic diagram with resolution reduced to half of the original resolution;
the cross-stage local-2 module comprises an upper branch and a lower branch, the upper branch is subjected to convolution-batch normalization-LeakyRelu module, a residual error module and a convolutional layer extraction feature map, the lower branch is subjected to convolutional layer extraction feature map, and the feature map of the upper branch and the feature map of the lower branch are subjected to batch normalization module, LeakyRelu activation function module and convolution-batch normalization-LeakyRelu module after connection operation at the channel level;
the detection end comprises a three-layer characteristic pyramid and a three-layer path aggregation network, wherein the three-layer characteristic pyramid is connected with the three-layer path aggregation network in parallel;
and the output end encodes the multi-scale hierarchical feature map, and outputs a plurality of candidate bounding box positions and the target confidence of each candidate bounding box after convolution operation.
Preferentially, the positioning labeling information and the classification labeling information of the training set are obtained by the following steps:
marking a middle multiple-tooth part of the curved surface layer image as an ROI by using a bounding box;
note the bookR i For curved body layer imageiROI, all of which are extracted from the curved surface layer imageR i Is sitting at the center pointSubject matter (x i ，y i ) And width and height: (w i ，h i ) As positioning label information;
and judging whether the median multiple raw teeth exist or not according to the cone beam CT image acquired by the curved surface layer image corresponding to the case, wherein the cone beam CT image is used as the classification marking information of the curved surface layer image.
Preferably, total loss functionL total The expression of (a) is:
L total =L box +L obj +L cls ，
in the formula (I), the compound is shown in the specification,L box to loss of generalized cross-over ratioGIOU，L obj In order to target the loss of the material,L cls to predict object class loss.
Preferably, the first and second liquid crystal films are,L obj andL cls are all binary cross entropy loss functions
；
L box =1-GIOU，
In the formula (I), the compound is shown in the specification,B label in order to be a bounding box for the annotation,B pred for the predicted candidate bounding box, C isB label AndB pre the smallest bounding box of (1);IOUis the ratio of the intersection and union of the predicted candidate bounding box and the labeled bounding box;
in the formula (I), the compound is shown in the specification,i∈[1,N]， Nrepresenting the number of samples of the training set,y i ∈[0,1]representing a sampleiThe standard category label value, the bounding box contains 1 for more raw teeth in the middle and 0 for less raw teeth in the middle;
∈[0,1]representing single-step multi-scale deep learning network model prediction samplesiContains the target confidence for more raw teeth in the middle.
Preferentially, before inputting the new curved surface body layer image into the single-step depth network model obtained by training, dividing the collected new curved surface body layer image into an S multiplied by S grid;
screening a final bounding box from a plurality of candidate bounding boxes by using a non-maximum suppression method to obtain the final positioning of the median molar excess, comprising the following steps:
multiplying the target confidence coefficient of the bounding box by the target confidence coefficient of each grid containing the target class belonging to the median toothings to obtain the probability value of the candidate bounding box containing the target class belonging to the median toothings;
according to the fact that the candidate bounding box comprises the probability value and the IOU intersection of the target class belonging to the median prolific dentiform, the redundant candidate bounding box is removed, and the final bounding box is determined;
and if the probability value of the target class belonging to the central excessive tooth in the final bounding box exceeds a set threshold value, judging that the position of the central excessive tooth in the new curvaceous layer image indicated by the final bounding box is qualified.
Preferably, the training set is augmented, including:
and performing brightness modification, random pixel clipping, random saturation modification, image inversion and mosaic enhancement on the curved surface layer image.
An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of any of the above methods when executing the program.
A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method of any of the above.
The invention achieves the following beneficial effects:
the method utilizes the cone beam CT image to evaluate and judge the existence condition of multiple middle fresh teeth, adds additional 3D space information for training the single-step depth network model, and reduces the manual error when the data set is manufactured.
According to the method, the training set is expanded in a data enhancement mode, a more generalized model can be trained by amplifying the size of the training set, and the image-enhanced data set can effectively help a single-step deep network model to reduce negative effects caused by disturbance such as position, illumination relation, saturation and the like when characteristics are extracted and learned;
the method utilizes a genetic algorithm to search the hyperparameters to finally obtain the optimal hyperparameter combination, and hyperparameter evolution enables a single-step deep network model to minimize a given loss function in an independently distributed data to achieve better prediction accuracy;
according to the method, GIOU loss is used as a loss function of a regression bounding box, and when GIOU is used as measurement, L = 1-GIOU and has nonnegativity, uncertainty, symmetry, triangle inequality and scale invariance; GIOU considers the non-overlapping area not considered by IOU and can reflect the overlapping mode of A and B;
the invention inputs the new curvy body layer image into the trained single-step depth network model for testing, the testing stage does not need manual operation, and only needs to scan the whole curvy body layer image once, thus whether the image contains multiple middle raw teeth or not can be rapidly and automatically identified and the position of the image can be given at the same time, and doctors are assisted to rapidly and correctly diagnose the multiple middle raw teeth;
the method adopts a non-maximum value inhibition method to obtain the final positioning of the middle multiple raw teeth, a large number of different boundary frames can be detected at the position of the same target, a plurality of overlapped boundary frames are generated, the non-maximum value inhibition method can inhibit the maximum value, the maximum value of the local space is searched, the optimal solution in the local space is obtained, and the redundant boundary frames are eliminated;
the invention does not need manual operation, and can quickly and automatically identify whether the image contains multiple median raw teeth and give the positions of the multiple median raw teeth by only scanning the curved surface layer image once, thereby avoiding the influence of experience difference of doctors on the precision and assisting the doctors to quickly and correctly diagnose the multiple median raw teeth.
Drawings
FIG. 1 is a flow chart of the method of the present invention;
FIG. 2 is a schematic diagram of a single-step deep network model structure adopted by the method of the present invention;
fig. 3 is a diagram illustrating the result of detecting multiple median raw teeth in a curved surface layer image according to a second embodiment of the present invention.
Detailed Description
The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.
Example one
As shown in fig. 1, the method for automatically detecting multiple raw teeth in the center of a curved surface layer map based on a single-step depth network comprises the following steps:
collecting a new curved surface body layer image;
inputting the new curvohedral layer image into a single-step depth network model obtained by training, and obtaining a plurality of candidate bounding box positions and target confidence coefficients containing target classes belonging to a plurality of median teeth;
and screening a final boundary frame from the plurality of candidate boundary frames by using a non-maximum inhibition method to obtain the final positioning of the median molar excess raw teeth.
Further, in this embodiment, the training to obtain the single-step deep network model includes:
(1) expanding a training set, wherein the training set comprises a curved surface layer image, a cone beam CT image, classification labeling information and positioning labeling information;
(2) a single-step deep network model is constructed by iterative training of a training set;
in each iterative training, an automatic learning anchor distribution and an Adam optimizer are adopted, and initial hyper-parameters are set, wherein the hyper-parameters comprise an initial learning rate, a training round number, a depth coefficient and a width coefficient;
and (3) carrying out hyper-parameter search by using a genetic algorithm, training the single-step deep network model according to the set epoch, and sequencing all hyper-parameter combinations according to a total loss function to obtain the optimal hyper-parameter combination.
Further, as shown in fig. 2, the single-step depth network model in this embodiment includes a basic skeleton end, a detection end, and an output end;
the basic skeleton end comprises a focusing layer,S 1 a plurality of convolution layers,S 2 A cross-stage local-1 module, a spatial pyramid pooling module, a cross-stage local-2 module, a focusing layer,S 1 a plurality of convolution layers,S 2 The cross-stage local-1 module, the spatial pyramid pooling module and the cross-stage local-2 module are sequentially connected;
the focusing layer performs connection operation on the input curved surface layer image in channel dimension after 4 slicing operations, and obtains a characteristic diagram with resolution reduced to half of the original resolution;
the cross-stage local-2 module comprises an upper branch and a lower branch, the upper branch is subjected to convolution-batch normalization-LeakyRelu module, a residual error module and a convolutional layer extraction feature map, the lower branch is subjected to convolutional layer extraction feature map, and the feature map of the upper branch and the feature map of the lower branch are subjected to batch normalization module, LeakyRelu activation function module and convolution-batch normalization-LeakyRelu module after connection operation at the channel level;
the detection end comprises a three-layer characteristic pyramid and a three-layer path aggregation network, wherein the three-layer characteristic pyramid is connected with the three-layer path aggregation network in parallel;
and the output end encodes the multi-scale hierarchical feature map, and outputs a plurality of candidate bounding box positions and the target confidence of each candidate bounding box after convolution operation.
Further, in this embodiment, the positioning labeling information and the classification labeling information of the training set are obtained through the following steps:
marking a middle multiple-tooth part of the curved surface layer image as an ROI by using a bounding box;
note the bookR i For curved body layer imageiROI, all of which are extracted from the curved surface layer imageR i Coordinates of center point of (1:x i ，y i ) And width and height: (w i ，h i ) As positioning label information;
and judging whether the median multiple fresh teeth exist or not through the acquired cone beam CT image of the case corresponding to the curved surface layer image, wherein the cone beam CT image is used as the classification marking information of the curved surface layer image.
Further, the total loss function in this embodimentL total The expression of (a) is:
L total =L box +L obj +L cls ，
in the formula (I), the compound is shown in the specification,L box to loss of generalized cross-over ratioGIOU，L obj In order to target the loss of the material,L cls to predict object class loss.
Further, in the present embodimentL obj AndL cls are all binary cross entropy loss functions
；
L box =1-GIOU，
In the formula (I), the compound is shown in the specification,B label in order to be a bounding box for the annotation,B pred for the predicted candidate bounding box, C isB label AndB pre the smallest bounding box of (1);IOUis the ratio of the intersection and union of the predicted candidate bounding box and the labeled bounding box;
in the formula (I), the compound is shown in the specification,i∈[1,N]， Nrepresenting training setsThe number of samples is such that,y i ∈[0,1]representing a sampleiThe standard category label value, the bounding box contains 1 for more raw teeth in the middle and 0 for less raw teeth in the middle;
∈[0,1]representing single-step multi-scale deep learning network model prediction samplesiContains the target confidence for more raw teeth in the middle.
The acquired new curved surface body layer image is automatically divided into S multiplied by S grids, and a plurality of candidate bounding boxes are generated in each grid.
Further, in this embodiment, the step of screening out a final bounding box from the multiple candidate bounding boxes by using a non-maximum suppression method to obtain a final location of the middle multiple raw teeth includes:
multiplying the target confidence coefficient of the bounding box by the target confidence coefficient of each grid containing the target class belonging to the median toothings to obtain the probability value of the candidate bounding box containing the target class belonging to the median toothings;
according to the fact that the candidate bounding box comprises the probability value and the IOU intersection of the target class belonging to the median prolific dentiform, the redundant candidate bounding box is removed, and the final bounding box is determined;
and if the probability value of the target class belonging to the central excessive tooth in the final bounding box exceeds a set threshold value, judging that the position of the central excessive tooth in the new curvaceous layer image indicated by the final bounding box is qualified.
Further, the expanding the training set in this embodiment includes:
and performing brightness modification, random pixel clipping, random saturation modification, image inversion and mosaic enhancement on the curved surface layer image.
An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the steps of any of the methods described above when executing the program.
A computer-readable storage medium, having stored thereon a computer program which, when being executed by a processor, carries out the steps of any of the methods described above.
The obtained plurality of candidate bounding box locations includes candidate bounding box locations and candidate bounding box sizes.
Example two
In practical application, a curved surface body layer image to be detected specifically comprises the following steps:
(1) the training set was the curved surface layer image, classification labeling information and positioning labeling information of 342 study cases, wherein the curved surface layer image size for the training was 2976 × 1536, and the pixel size was 0.07 × 0.07mm 2 。
(2) Whether or not there are more middle and middle raw teeth is determined from the cone beam CT image corresponding to the curved surface layer image study case, and the determined result is used as classification labeling information corresponding to the curved surface layer image. And (3) labeling the curvy body layer image of the study case, taking the curvy body layer image training set with the central multiple tooth parts as the ROI, and labeling by using a bounding box. Note the bookR i For curved body layer imageiROI, all of which are extracted from the curved surface layer imageR i Coordinates of center point of (1:x i ，y i ) And width and height: (w i ，h i ) As positioning label information; 230 cases with more raw teeth and 112 cases without more raw teeth.
(3) The training set is extended in a data enhancement mode, and the data enhancement mode comprises the following steps: brightness modification, random pixel clipping, random saturation modification, image flipping, mosaic enhancement, and the like.
(4) The single-step deep network model sets the initial values of the hyper-parameters: the depth coefficient is set to be 1.66, the width coefficient is set to be 1.50, the initial learning rate is set to be 0.0032, the fusion probability is set to be 0.243, the number of training rounds is set to be 250, and automatic learning anchor distribution and an Adam optimizer are adopted;
after the initial hyper-parameters are set, hyper-parameter search is carried out by using a genetic algorithm, the number of the termination cycle iterations of the hyper-parameter search is 300, each pair of wheels trains 100 epochs of the single-step deep network model, and all hyper-parameter combinations are sequenced according to a total loss function, so that the optimal hyper-parameter combination is finally obtained.
The total loss function of the single step deep network model is as follows:
total loss functionL total Including bounding box lossL box Target lossL obj And predicting object class lossL cls ：
L total =L box +L obj +L cls ，
In the formula (I), the compound is shown in the specification,L box to loss of generalized cross-over ratioGIOU，L obj AndL cls are all binary cross entropy loss functions
；
L box =1-GIOU，
In the formula (I), the compound is shown in the specification,B label in order to be a bounding box for the annotation,B pred for the predicted candidate bounding box, C isB label AndB pre the smallest bounding box of (1);IOUis the ratio of the intersection and union of the predicted candidate bounding box and the labeled bounding box;
in the formula (I), the compound is shown in the specification,i∈[1,N]， Nrepresenting the number of samples of the training set,y i ∈[0,1]representing a sampleiThe standard category label value, the bounding box contains 1 for more raw teeth in the middle and 0 for less raw teeth in the middle;
∈[0,1]representing predicted samplesiContains the target confidence for more raw teeth in the middle.
(5) And (5) repeating the steps (3) to (4) to serve as a training stage of the single-step deep network model, and obtaining the final model weight.
(6) And (5) testing the acquired new curved surface body layer image by using the single-step deep network model trained in the step (5). Inputting the new curvohedral layer image into the single-step depth network model, and outputting a candidate bounding box position, a candidate bounding box size, a target confidence coefficient of each candidate bounding box, and a target confidence coefficient that each grid contains a target class belonging to the median polyneuropathy.
(7) Screening out a final boundary box from the candidate boundary boxes by using a non-maximum value inhibition method, determining the final positioning of the median multiple raw teeth, and multiplying the target confidence coefficient by the target confidence coefficient of each grid containing the target class belonging to the median multiple raw teeth to obtain the probability value of the candidate boundary box containing the target class belonging to the median multiple raw teeth;
removing redundant bounding boxes according to the probability value and IOU intersection of the target class belonging to the median molar excess, and determining the final bounding box;
if the probability value that the final bounding box contains the target class and belongs to the central multiple fresh teeth exceeds the set threshold value, the position of the central multiple fresh teeth in the indicated new curvaceous layer image is judged to be qualified.
As shown in FIG. 3, Yes indicates that there are more middle teeth in the curved surface layer image detected by the method of the present invention, and No indicates that there are more middle teeth in the curved surface layer image detected by the method of the present invention; in fig. 3, the left side is the curved surface layer image to be detected, the right side is the final bounding box and the target confidence coefficient output by the present invention, and the target confidence coefficients of the three curved surface layer images to be detected are 0.90, 0.93 and 0.08 from top to bottom.
The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The above description is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, several modifications and variations can be made without departing from the technical principle of the present invention, and these modifications and variations should also be regarded as the protection scope of the present invention.
Claims (10)
1. The automatic detection method for multiple fresh teeth in the center of the curved surface layer graph based on the single-step depth network is characterized by comprising the following steps of:
collecting a new curved surface body layer image;
inputting the new curvohedral layer image into a single-step depth network model obtained by training, and obtaining a plurality of candidate bounding box positions and target confidence coefficients containing target classes belonging to a plurality of median teeth;
and screening a final boundary frame from the plurality of candidate boundary frames by using a non-maximum inhibition method to obtain the final positioning of the median molar excess raw teeth.
2. The method for automatically detecting multiple fresh teeth in the center of a curved surface body layer graph based on the single-step depth network as claimed in claim 1, wherein the training obtains a single-step depth network model, comprising:
(1) expanding a training set, wherein the training set comprises a curved surface layer image, a cone beam CT image, classification labeling information and positioning labeling information;
(2) a single-step deep network model is constructed by iterative training of a training set;
in each iterative training, setting an initial value of a hyper-parameter by adopting automatic learning anchor distribution and an Adam optimizer;
and (3) carrying out hyper-parameter search by using a genetic algorithm, training the single-step deep network model according to the set epoch, and sequencing all hyper-parameter combinations according to a total loss function to obtain the optimal hyper-parameter combination.
3. The method for automatically detecting median multiple raw teeth of a curved surface layer map based on a single-step depth network as claimed in claim 2, wherein the single-step depth network model comprises a basic skeleton end, a detection end and an output end;
the basic skeleton end comprises a focusing layer,S 1 a plurality of convolution layers,S 2 A cross-stage local-1 module, a spatial pyramid pooling module, a cross-stage local-2 module, a focusing layer,S 1 a plurality of convolution layers,S 2 The cross-stage local-1 module, the spatial pyramid pooling module and the cross-stage local-2 module are sequentially connected;
the focusing layer performs connection operation on the input curved surface layer image in channel dimension after 4 slicing operations, and obtains a characteristic diagram with resolution reduced to half of the original resolution;
the cross-stage local-2 module comprises an upper branch and a lower branch, the upper branch is subjected to convolution-batch normalization-LeakyRelu module, a residual error module and a convolutional layer extraction feature map, the lower branch is subjected to convolutional layer extraction feature map, and the feature map of the upper branch and the feature map of the lower branch are subjected to batch normalization module, LeakyRelu activation function module and convolution-batch normalization-LeakyRelu module after connection operation at the channel level;
the detection end comprises a three-layer characteristic pyramid and a three-layer path aggregation network, wherein the three-layer characteristic pyramid is connected with the three-layer path aggregation network in parallel;
and the output end encodes the multi-scale hierarchical feature map, and outputs a plurality of candidate bounding box positions and the target confidence of each candidate bounding box after convolution operation.
4. The method for automatically detecting multiple fresh teeth in the center of a curved surface layer graph based on a single-step depth network as claimed in claim 2, wherein the positioning label information and the classification label information of the training set are obtained by the following steps:
marking a middle multiple-tooth part of the curved surface layer image as an ROI by using a bounding box;
note the bookR i For curved body layer imageiROI, all of which are extracted from the curved surface layer imageR i Coordinates of center point of (1:x i ，y i ) And width and height: (w i ，h i ) As positioning label information;
and judging whether more middle raw teeth exist or not through the cone beam CT image acquired by the curved surface layer image corresponding to the case, and using the cone beam CT image as the classification and labeling information of the curved surface layer image.
5. The method of claim 2, wherein the total loss function is a function of total lossL total The expression of (a) is:
L total =L box +L obj +L cls ，
in the formula (I), the compound is shown in the specification,L box to loss of generalized cross-over ratioGIOU，L obj In order to target the loss of the material,L cls to predict object class loss.
6. The method for automatically detecting central multiple raw teeth of a curvography based on a single-step depth network as claimed in claim 5,L obj andL cls are all binary cross entropy loss functions
；
L box =1-GIOU，
In the formula (I), the compound is shown in the specification,B label in order to be a bounding box for the annotation,B pred for the predicted candidate bounding box, C isB label AndB pre the smallest bounding box of (1);IOUis the ratio of the intersection and union of the predicted candidate bounding box and the labeled bounding box;
in the formula (I), the compound is shown in the specification,i∈[1,N]， Nrepresenting the number of samples of the training set,y i ∈[0,1]representing a sampleiThe standard category label value, the bounding box contains 1 for more raw teeth in the middle and 0 for less raw teeth in the middle;
∈[0,1]representing single step multi-scale deep learning netNetwork model prediction samplesiContains the target confidence for more raw teeth in the middle.
7. The method for automatically detecting central multiple raw teeth of a curved surface body layer map based on a single-step depth network as claimed in claim 1, characterized in that before inputting the new curved surface body layer image into the trained single-step depth network model, the collected new curved surface body layer image is divided into S x S grids;
screening a final bounding box from a plurality of candidate bounding boxes by using a non-maximum suppression method to obtain the final positioning of the median molar excess, comprising the following steps:
multiplying the target confidence coefficient of the bounding box by the target confidence coefficient of each grid containing the target class belonging to the median toothings to obtain the probability value of the candidate bounding box containing the target class belonging to the median toothings;
according to the fact that the candidate bounding box comprises the probability value and the IOU intersection of the target class belonging to the median prolific dentiform, the redundant candidate bounding box is removed, and the final bounding box is determined;
and if the probability value of the target class belonging to the central excessive tooth in the final bounding box exceeds a set threshold value, judging that the position of the central excessive tooth in the new curvaceous layer image indicated by the final bounding box is qualified.
8. The method for automatically detecting multiple fresh teeth in the center of a curved surface layer map based on a single-step deep network as claimed in claim 2, wherein the training set is expanded and comprises:
and performing brightness modification, random pixel clipping, random saturation modification, image inversion and mosaic enhancement on the curved surface layer image.
9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the steps of the method of any of claims 1 to 8 are implemented when the processor executes the program.
10. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out the steps of the method of any one of claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210436787.0A CN114549523A (en) | 2022-04-25 | 2022-04-25 | Single-step depth network-based automatic detection method for multiple raw teeth in center of curved surface layer graph |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210436787.0A CN114549523A (en) | 2022-04-25 | 2022-04-25 | Single-step depth network-based automatic detection method for multiple raw teeth in center of curved surface layer graph |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114549523A true CN114549523A (en) | 2022-05-27 |
Family
ID=81667711
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210436787.0A Pending CN114549523A (en) | 2022-04-25 | 2022-04-25 | Single-step depth network-based automatic detection method for multiple raw teeth in center of curved surface layer graph |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114549523A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117036365A (en) * | 2023-10-10 | 2023-11-10 | 南京邮电大学 | Third molar tooth root number identification method based on deep attention network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109816661A (en) * | 2019-03-22 | 2019-05-28 | 电子科技大学 | A kind of tooth CT image partition method based on deep learning |
| CN112151167A (en) * | 2020-05-14 | 2020-12-29 | 余红兵 | Intelligent screening method for six-age dental caries of children based on deep learning |
| US20210150702A1 (en) * | 2017-06-30 | 2021-05-20 | Promaton Holding B.V. | Classification and 3d modelling of 3d dento-maxillofacial structures using deep learning methods |
| CN112884064A (en) * | 2021-03-12 | 2021-06-01 | 迪比（重庆）智能科技研究院有限公司 | Target detection and identification method based on neural network |
| CN114119950A (en) * | 2021-10-25 | 2022-03-01 | 上海交通大学医学院附属第九人民医院 | Artificial intelligence-based oral cavity curved surface fault layer dental image segmentation method |
-
2022
- 2022-04-25 CN CN202210436787.0A patent/CN114549523A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20210150702A1 (en) * | 2017-06-30 | 2021-05-20 | Promaton Holding B.V. | Classification and 3d modelling of 3d dento-maxillofacial structures using deep learning methods |
| CN109816661A (en) * | 2019-03-22 | 2019-05-28 | 电子科技大学 | A kind of tooth CT image partition method based on deep learning |
| CN112151167A (en) * | 2020-05-14 | 2020-12-29 | 余红兵 | Intelligent screening method for six-age dental caries of children based on deep learning |
| CN112884064A (en) * | 2021-03-12 | 2021-06-01 | 迪比（重庆）智能科技研究院有限公司 | Target detection and identification method based on neural network |
| CN114119950A (en) * | 2021-10-25 | 2022-03-01 | 上海交通大学医学院附属第九人民医院 | Artificial intelligence-based oral cavity curved surface fault layer dental image segmentation method |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117036365A (en) * | 2023-10-10 | 2023-11-10 | 南京邮电大学 | Third molar tooth root number identification method based on deep attention network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Yang et al. | Automated dental image analysis by deep learning on small dataset | |
| CN111667456B (en) | Method and device for detecting vascular stenosis in coronary artery X-ray sequence radiography | |
| CN108205806B (en) | Automatic analysis method for three-dimensional craniofacial structure of cone beam CT image | |
| US20080071160A1 (en) | Displaying A Tracheobronchial Tree | |
| Li et al. | DenseX-net: an end-to-end model for lymphoma segmentation in whole-body PET/CT images | |
| CN111553892B (en) | Lung nodule segmentation calculation method, device and system based on deep learning | |
| CN106056596B (en) | Full-automatic three-dimensional liver segmentation method based on local prior information and convex optimization | |
| JP2012512730A (en) | Method for calculating brightness level in a target area of a digital X-ray image for medical use | |
| CN110246580B (en) | Cranial image analysis method and system based on neural network and random forest | |
| US20220180516A1 (en) | Identifying boundaries of lesions within image data | |
| Eun et al. | Oriented tooth localization for periapical dental X-ray images via convolutional neural network | |
| CN102240212B (en) | Method and apparatus for measuring pneumothorax | |
| CN111627014A (en) | Root canal detection and scoring method and system based on deep learning | |
| CN110874860A (en) | Target extraction method of symmetric supervision model based on mixed loss function | |
| CN114037663A (en) | Blood vessel segmentation method, device and computer readable medium | |
| CN114638852A (en) | Jaw bone and soft tissue identification and reconstruction method, device and medium based on CBCT image | |
| CN108280833B (en) | Skeleton extraction method for plant root system bifurcation characteristics | |
| Liu et al. | Automated classification and measurement of fetal ultrasound images with attention feature pyramid network | |
| CN113516639B (en) | Training method and device for oral cavity abnormality detection model based on panoramic X-ray film | |
| CN114549523A (en) | Single-step depth network-based automatic detection method for multiple raw teeth in center of curved surface layer graph | |
| CN111986217B (en) | Image processing method, device and equipment | |
| Dai et al. | A one-stage deep learning method for fully automated mesiodens localization on panoramic radiographs | |
| US20220122261A1 (en) | Probabilistic Segmentation of Volumetric Images | |
| JP7178016B2 (en) | Image processing device and its image processing method | |
| CN112150422A (en) | Modeling method of oral health self-detection model based on multitask learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |