1. A detecting apparatus for automatically detecting six-instar dental caries, comprising:

the shooting device is used for shooting the images of the oral cavity areas of the children;

a network cloud server configured to perform the steps of:

inputting the children oral cavity area image into a depth residual error network for feature extraction so as to obtain a feature map of the children oral cavity area image;

performing feature enhancement processing based on a feature pyramid network on the feature map to obtain an enhanced feature map;

selecting corresponding anchor point frames according to different scales of the enhanced feature map to generate six-age tooth candidate frames;

inputting the six-age tooth candidate frame into a regional regression subnetwork to obtain the coordinate offset of the six-age tooth candidate frame to a real target, and obtaining the coordinate data of the six-age tooth region according to the coordinate offset;

inputting the six-age tooth candidate box into a regional classification sub-network to obtain a prediction category of six-age teeth;

and combining and outputting the coordinate data of the six-age tooth area and the prediction category of the six-age tooth to provide a diagnostic image.

2. The detection device according to claim 1, further comprising:

and the network cloud database is used for storing the children oral cavity area images.

3. The detection device according to claim 1, further comprising:

a display for displaying the diagnostic image.

4. The detecting device according to claim 1, wherein the six-aged-tooth candidate box is input into a local regression subnetwork to obtain the coordinate offset of the six-aged-tooth candidate box to a real target, and the specific steps include:

connecting a fully connected neural network as a regional regression sub-network on each layer of the feature pyramid network, inputting the enhanced feature graph into the fully connected neural network for convolution operation, and activating the feature graph after the convolution operation by adopting a Relu function.

5. The detection apparatus of claim 4, wherein the regional classification subnetwork is configured as a fully-connected neural network in parallel with the regional regression subnetwork, the enhanced feature map is input to the parallel fully-connected neural network for convolution operation, and the feature map after convolution operation is activated by using a Softmax function.

6. The detection apparatus according to claim 4, wherein the network cloud server is further configured to perform the steps of:

and respectively training the region regression sub-network and the region classification sub-network by taking a cross entropy function with weight as a loss function.

7. The detection apparatus of claim 1, wherein the deep residual network is ResNet 50.

Background

In China, dental caries has become one of common oral diseases which seriously affect physical and mental health of children, has the characteristics of high morbidity and wide distribution, and the attack targets of the dental caries are mainly concentrated on six-instar teeth at the fossa sulcus part. The six-instar teeth are used as the first permanent molars which sprout at the earliest position of the fossa sulcus of the early children and play a key role in maintaining healthy permanent dentition. However, the early stage surface layer of the dental caries-resistant dental caries-resistant dental caries-resistant dental caries-resistant dental. The dental caries lesion of six ages of children can directly influence the chewing function, delay treatment and even influence the growth and development of children. Six-age dental caries teeth are screened in time and early treatment and intervention are carried out, so that the children can be helped to prevent the aggravation of the six-age dental caries teeth and the subsequent occurrence of other oral diseases, and the social public health economic burden can be reduced while the investment cost is low. Therefore, the screening of the early six-age dental caries has important significance for preventing and treating the dental caries of children.

The traditional caries diagnosis method mainly comprises visual examination, probing, X-ray imaging, fluorescence imaging detection and the like. The visual examination and the probing rely on the vision of the dentist and the professional technique to diagnose the enamel outside the teeth, and the X-ray or fluorescence imaging is a relatively objective and sensitive caries detection means, but has requirements on equipment, places and the like. The shortage of stomatologists and the uneven distribution of medical resources also aggravate the resistance of children to receive oral health services, so that the oral health work in China faces severe examination. Screening is one of important contents of basic oral hygiene services, the traditional diagnosis method for six-year-old dental caries is mainly characterized in that a clinician carries out comprehensive judgment according to symptoms, probe inspection, imaging inspection and the like, the method is seriously dependent on the personal experience of the clinician, and the diagnosis difference among different doctors causes different treatment schemes of oral diseases. The traditional caries detection method is only relied on to comprehensively screen the health problems of the six-age teeth of the children, which is undoubtedly extremely difficult and has huge manpower and material resource consumption. Therefore, in order to enhance the oral health care service capability, the automatic screening technology for the six-year-old dental caries is realized efficiently and accurately by utilizing the artificial intelligence technologies such as deep learning and the like, and has important scientific research value and social benefit.

Disclosure of Invention

The invention aims at the problems and provides a detection device for automatically detecting six-age dental caries, which can carry out non-invasive and non-contact high-resolution visual imaging on an oral cavity region by utilizing daily mobile equipment and combine a new-generation artificial intelligence technology such as deep learning and the like to carry out intelligent detection and caries classification on the six-age dental caries at a pit and fissure part. Because deep learning has powerful characteristic learning, the detection device can effectively improve the screening efficiency and precision of the six-year-old dental caries of children, can replace the previous diagnosis and screening means, and gets rid of the dependence on oral doctors, professional imaging equipment and the like.

The invention provides a detection device for automatically detecting six-age dental caries, which comprises:

the shooting device is used for shooting the images of the oral cavity areas of the children;

a network cloud server configured to perform the steps of:

inputting the children oral cavity area image into a depth residual error network for feature extraction so as to obtain a feature map of the children oral cavity area image;

performing feature enhancement processing based on a feature pyramid network on the feature map to obtain an enhanced feature map;

selecting corresponding anchor point frames according to different scales of the enhanced feature map to generate six-age tooth candidate frames;

inputting the six-age tooth candidate frame into a regional regression subnetwork to obtain the coordinate offset of the six-age tooth candidate frame to a real target, and obtaining the coordinate data of the six-age tooth region according to the coordinate offset;

inputting the six-age tooth candidate box into a regional classification sub-network to obtain a prediction category of six-age teeth;

and combining and outputting the coordinate data of the six-age tooth area and the prediction category of the six-age tooth to provide a diagnostic image.

Further, the detection device further comprises: and the network cloud database is used for storing the children oral cavity area images.

Further, the detection device further comprises: a display for displaying the diagnostic image.

Further, the method for acquiring the coordinate offset of the six-aged tooth candidate box to the real target by inputting the six-aged tooth candidate box into a regional regression subnetwork comprises the following specific steps of:

connecting a fully connected neural network as a regional regression sub-network on each layer of the feature pyramid network, inputting the enhanced feature graph into the fully connected neural network for convolution operation, and activating the feature graph after the convolution operation by adopting a Relu function.

Further, the region classification sub-network is configured to be a fully-connected neural network in parallel with the region regression sub-network, the enhanced feature map is input to the parallel fully-connected neural network for convolution operation, and a Softmax function is adopted to activate the feature map after the convolution operation.

Further, the network cloud server is further configured to perform the following steps:

and respectively training the region regression sub-network and the region classification sub-network by taking a cross entropy function with weight as a loss function.

Further, the depth residual network is ResNet 50.

The invention provides a detection device for automatically detecting six-instar dental caries, which has the following beneficial effects:

1. different from the traditional X-ray or fluorescence imaging detection, the oral visual imaging mode adopted by the detection device of the invention does not depend on any professional medical equipment and operation, and can realize the complete non-damage, non-contact and high-efficiency imaging of the oral area only by daily mobile equipment. The imaging mode with zero cost has autonomy and universality, and can provide good precondition guarantee for effective popularization and application of the detection device.

2. The automatic detection screening algorithm for the six-age dental caries based on deep learning is adopted by the detection device, and the basic idea is to extract the characteristics of the six-age dental caries through a deep convolutional neural network so as to ensure that the detection algorithm can accurately detect the six-age dental region in the internal imaging of the oral cavity shot by mobile equipment, and simultaneously, the detection accuracy of the six-age dental caries under different angles and light rays is further ensured to be improved by combining the characteristics of different network layers.

3. The detection device can effectively carry out intelligent detection and classification on the six-age dental caries of the children based on deep learning, and improves the diagnosis precision and efficiency of the caries of the children. The detection device provided by the invention does not depend on any professional manpower and material resources in the oral cavity, is beneficial to further popularization and application, and improves the primary hygiene service coverage rate of the oral cavity of the primary children.

Drawings

FIG. 1 is a schematic block diagram of a detection apparatus for automatically detecting six-year old dental caries in accordance with an embodiment of the present invention;

FIG. 2 is a schematic diagram of a six-year-old dental caries detection process performed on an image of a child's oral area in an embodiment of the present invention.

Detailed Description

In order to further describe the technical scheme of the present invention in detail, the present embodiment is implemented on the premise of the technical scheme of the present invention, and detailed implementation modes and specific steps are given.

Referring to fig. 1, the detecting apparatus for automatically detecting six-instar dental caries includes: the shooting device 01 is used for shooting the images of the oral cavity areas of the children, and under the condition that the light source is enough, the images of the oral cavity areas of the children are collected only by shooting the oral cavity areas of the children through daily mobile equipment, and do not depend on any professional oral medical imaging equipment; the network cloud database 02 is used for storing the children oral cavity area images, the children oral cavity image data used in the embodiment are from professional hospital institutions, professional doctors carry out accurate labeling on six-age tooth areas in the oral cavity images, and corresponding medical labels (1: decayed teeth; 0: health) are given for each six-age tooth; a display 04 for displaying the diagnostic image; the network cloud server 03 is used for processing the photographed images of the oral cavity area of the child for tooth detection of six-year-old dental caries, and is configured to execute the following steps:

inputting the children oral cavity area image into a depth residual error network for feature extraction so as to obtain a feature map of the children oral cavity area image;

performing feature enhancement processing based on a feature pyramid network on the feature map to obtain an enhanced feature map;

selecting corresponding anchor point frames according to different scales of the enhanced feature map to generate six-age tooth candidate frames;

inputting the six-age tooth candidate frame into a regional regression subnetwork to obtain the coordinate offset of the six-age tooth candidate frame to a real target, and obtaining the coordinate data of the six-age tooth region according to the coordinate offset;

inputting the six-age tooth candidate box into a regional classification sub-network to obtain a prediction category of six-age teeth;

and combining and outputting the coordinate data of the six-age tooth area and the prediction category of the six-age tooth to provide a diagnostic image.

Inputting the six-age tooth candidate box into a regional regression subnetwork to obtain the coordinate offset of the six-age tooth candidate box to a real target, wherein the specific steps comprise:

connecting a fully connected neural network as a regional regression sub-network on each layer of the feature pyramid network, inputting the enhanced feature graph into the fully connected neural network for convolution operation, and activating the feature graph after the convolution operation by adopting a Relu function.

The regional classification sub-network is configured to be a fully-connected neural network in parallel with the regional regression sub-network, the enhanced feature map is input to the parallel fully-connected neural network for convolution operation, and a Softmax function is adopted to activate the feature map after the convolution operation.

The network cloud server is further configured to perform the steps of:

and respectively training the region regression sub-network and the region classification sub-network by taking a cross entropy function with weight as a loss function.

The depth residual network is ResNet 50.

Referring to fig. 2, after the photographed image of the oral cavity area of the child is processed, a six-aged tooth detection part is performed, the six-aged tooth at the pit and fissure part is usually in a deeper area in the oral cavity, the image proportion of the target is small, and the traditional target detection method has certain defects in the target area, and cannot meet the precision requirement of clinical diagnosis. Therefore, a feature extraction method based on a depth residual error network and a feature pyramid is provided for children oral cavity area images, oral cavity image features of different scales are fused through a feature extraction module, features with semantic information and spatial information are obtained, and the accuracy of caries detection is improved. Specifically, after the features are extracted by using the depth residual error network, feature enhancement is carried out on the feature pyramid network. The feature maps of the lower layer in the depth residual error network have rich image information, the feature maps of the upper layer have rich semantic information, each layer of feature maps in the depth residual error network correspond to target information with different scales, and the detection effect of small-scale targets, namely six-year old teeth, can be effectively improved by combining the feature maps of different layers.

Meanwhile, aiming at the oral cavity image characteristics obtained by the characteristic extraction method based on the depth residual error network and the characteristic pyramid, a full-connection network-based six-age dental caries detection and classification algorithm is provided, and two parallel full-connection networks are used as a region regression sub-network and a region classification sub-network. In order to realize multi-scale target detection, anchor points of different scales are used on feature maps of different scales, and generated candidate frames cover the whole picture, so that missing detection is avoided to the greatest extent. In the proposed six-year-old dental caries detection and classification algorithm, the regional regression subnetwork is responsible for outputting coordinate offsets of the six-year-old dental candidate box to the real target. And each layer of the characteristic pyramid network is connected with a fully-connected network, the characteristic diagram is input into the fully-connected network, and accurate six-age tooth areas are generated in the candidate six-age tooth areas. Meanwhile, the region classification sub-network is responsible for outputting the category of the six-age tooth candidate box, inputting the feature map into another parallel fully-connected network, judging whether the candidate six-age tooth region belongs to the category of healthy teeth or decayed teeth, and outputting a corresponding prediction result and a confidence score.

Finally, the display 04 outputs all six-aged teeth and their corresponding categories in the children's mouth image.

Referring to fig. 2, an example in the specific embodiment specifically includes:

s100, collecting oral vision imaging;

s200, inputting the image into a deep residual error network for feature extraction, and firstly, adopting the deep residual error network ResNet50 as a main network for feature extraction to effectively obtain the distinguishing features of the image.

S300, performing feature enhancement based on the feature pyramid, and taking the output of the last three layers of convolutional neural networks in the ResNet50 backbone network as a feature map { C₃,C₄,C₅And (4) performing convolution on each feature map by 1 x 1, and changing the number of channels of the feature maps to obtain the feature maps with the same size, wherein the size of each feature map layer is the same as the size of the feature map layer which is subjected to downsampling with the step length of 2. Generally, the features of the lower layer have higher resolution, so 2 times of nearest neighbor upsampling is selected to be performed on each feature map of the higher layer after 1 × 1 convolution, the feature map with the same size as the feature map of the upper layer is obtained, and the corresponding elements of the feature map of the upper layer after 1 × 1 convolution are added, and the like. Finally, eliminating the influence caused by nearest neighbor upsampling by performing convolution of 3 x 3, and obtaining the feature map { P) enhanced under the feature pyramid₃,P₄,P₅}. I.e. the last layer C₅Performing convolution with 1 × 1 and then convolution with 3 × 3 to obtain P₅；C₅Upsampling after 1 × 1 convolution, and upper layers after 1 × 1 convolutionCharacteristic diagram C₄Adding corresponding elements, and performing 3 × 3 convolution on the obtained feature map to obtain P₄；P₃Obtained in the same manner.

S400, generating a six-aged tooth candidate frame based on a preset anchor frame, using anchor frames with different scales on feature maps with different scales in order to realize multi-scale target detection, and enabling the generated candidate frame to cover the whole picture to the greatest extent so as to avoid missing detection. During the training process, the anchor box is divided into a target box and a background box. If the intersection ratio of the candidate frame generated by an anchor point and the real image is larger than a preset threshold value or the intersection ratio of the candidate frame generated by an anchor point and the real image is the highest for a certain decayed tooth real image, the candidate frame is regarded as a positive sample. The candidate frames smaller than the preset threshold are regarded as negative samples, because the negative samples have great deviation with the actual image position, only the positive samples participate in the backward propagation of the regional regression in the training process.

And S500, sending the six-age tooth candidate frame into a regional regression sub-network, wherein the regional regression sub-network is responsible for outputting the coordinate offset of the six-age tooth candidate frame to a real target. And connecting a fully-connected network behind each layer of the characteristic pyramid network, and inputting the fully-connected network into the fully-connected network. Carrying out convolution operation by sequentially using 4 convolution kernels of 3 x 3, and activating by adopting a Relu function; and then convolving by using 4 × a convolution kernels of 3 × 3 (a is the number of anchor points in each layer), and obtaining 4 × a coordinate offsets from the candidate frame to the real target. The regional regression subnetwork finally outputs the coordinate data L, L of all the six-age tooth regions_i＝{x₁,y₁,x₂,y₂E.g., (i ═ 0,1, 2..) where i is the six-instar tooth sequence detected by the target detection network, { x.. times)₁,y₁And { x }₂,y₂Coordinates of the upper left corner and the lower right corner of the tooth area of six ages are respectively.

Meanwhile, the candidate box is input into a regional classification sub-network to realize category prediction, and the regional classification sub-network is responsible for outputting the category confidence degree that the six-instar tooth candidate box is judged as healthy teeth or decayed teeth. And connecting a fully-connected network behind each layer of the characteristic pyramid network, and inputting the fully-connected network into the fully-connected network. Carrying out convolution operation by sequentially using 4 convolution kernels of 3 x 3, and activating by adopting a Relu function; then 2 a-3 convolution kernels (a is the number of anchor points per layer) are used for convolution and activation is carried out by using a Softmax function, and confidence scores of 2 a candidate frames which are judged to be healthy teeth and decayed teeth are obtained. If the corresponding confidence score is larger than a preset threshold value, judging the carious sample; and if the value is less than the preset threshold value, judging the sample to be a healthy sample.

In the training process of the sub-network, a cross-entropy function with weights is used as a loss function.

Wherein p is the confidence score corresponding to the candidate corresponding region, l is the corresponding real label, α and (1- α) represent the weight of the decayed tooth and the healthy tooth sample, and p^γAnd (1-p)^γA weight representing the degree of identification of the candidate region. When the candidate region is difficult to identify, its confidence score p is close to 0, (1-p)^γApproximately 1. The influence of negative samples in target detection can be reduced through the weighted cross entropy loss function, and the detection accuracy is further improved.

S600, outputting screening and predicting results of six-age dental caries, combining and outputting region coordinate information of the six-age dental caries and category predicting results, and achieving intelligent screening of the six-age dental caries of the children.

In summary, the detection device for automatically detecting six-year-old dental caries provided by the invention is different from the traditional X-ray or fluorescence imaging detection, the oral visual imaging mode adopted by the detection device does not depend on any professional medical equipment and operation, and the complete non-damage, non-contact and efficient imaging of the oral area can be realized only by daily mobile equipment. The imaging mode with zero cost has autonomy and universality, and can provide good premise guarantee for effective popularization and application of the detection device; the automatic detection screening algorithm for the six-age dental caries based on deep learning is adopted by the detection device, and the basic idea is that the six-age characteristics are extracted through a deep convolutional neural network, so that the detection algorithm can accurately detect the six-age area in the internal imaging of the oral cavity shot by a mobile device, and meanwhile, the detection precision of the six-age dental caries under different angles and light rays is further improved by combining the characteristics of different network layers; the detection device can effectively carry out intelligent detection and classification on the six-age dental caries of the children based on deep learning, and the diagnosis precision and efficiency of the caries of the children are improved. The detection device provided by the invention does not depend on any professional manpower and material resources in the oral cavity, is beneficial to further popularization and application, and improves the primary hygiene service coverage rate of the oral cavity of the primary children.

In this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process or method.

The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.