A method for constructing a classification model for evaluating implant stability based on CBCT image data and a device thereof
By constructing a multi-task convolutional neural network model based on CBCT images, the problems of objectivity and professionalism in implant stability assessment were solved, achieving efficient and accurate implant stability assessment and simplifying the clinical assessment process.
Patent Information
- Application Number
- CN202211287122.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-10-20
- Publication Date
- 2026-01-06
- Estimated Expiration
- 2042-10-20
AI Technical Summary
Existing technologies for assessing implant stability suffer from limitations in the objectivity and professionalism of imaging data, traditional methods cannot provide information on the buccal and lingual alveolar bone levels, and rely on the experience and judgment of clinicians, lacking effective automated assessment tools.
A multi-task convolutional neural network model based on CBCT image data was constructed. Implant prosthesis segmentation was performed using the Mobilenetv2-DeeplabV3+ network, and peri-implant bone image extraction and classification were performed using the ResNet-50 network. An evaluation model for implant stability was established, and deep learning and data augmentation techniques were used to improve the model's accuracy and generalization ability.
It achieves efficient and accurate implant stability assessment. The model performs excellently in binary, triadic, and quadriadic classifications with an accuracy rate of over 90%, providing reliable guidance for implant restoration treatment and simplifying the clinical assessment process.
Smart Images
Figure CN115512167B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of bioinformatics and relates to a method, apparatus and application for constructing a classification model for assessing implant stability. Specifically, it relates to a method and apparatus for constructing a multi-task convolutional neural network model for assessing implant stability based on CBCT image data. Background Technology
[0002] Imaging studies are increasingly being used to assess bone quality and volume in the jawbone. Implant stability is a dynamic process following implantation, constantly changing with bone remodeling. Using medical imaging techniques to assess peri-implant bone quality can aid in implant stability assessment. However, traditional periapical and panoramic radiographs have limitations in assessing implant stability, failing to provide information on the buccal and lingual alveolar bone levels, where bone loss precedes mesial and distal bone loss. Cone-beam computed tomography (CBCT) of the oral cavity, assessing the peri-implant osseointegration zone (implant-bone interface), effectively evaluates implant stability.
[0003] In clinical practice, implant stability is often assessed using resonant frequency analysis (RFA). RFA can be measured using an Osstell device and automatically converted into an implant stability coefficient (ISQ). A higher ISQ score indicates greater implant stability and more extensive osseointegration. Generally, the ISQ of implants in clinical practice is between 40 and 80. The success of early loading protocols depends heavily on the clinician's ability and capacity to control initial stability and assess its changes over healing time. However, relying solely on imaging data requires the assistance of a specialist radiologist, making its widespread adoption among clinicians difficult, and its objectivity cannot be assessed.
[0004] With the development of big data and computer technology, artificial intelligence (AI) may be an effective solution. Convolutional Neural Networks (CNNs) are the core of AI, composed of neurons that learn and optimize themselves, with multiple hidden layers in between. Their main applications in the medical field are semantic segmentation, feature extraction, and classification diagnosis. With the development of deep learning, it has also shown excellent performance in disease diagnosis. For example, using features extracted by CNNs as input and support vector machines for classification, it is possible to diagnose images of diabetic retinopathy. Some researchers have also used deep CNN models to detect the tear film breakup zone, using sliding window detection to screen breakup centers and determine tear film breakup time. This method has been used to assess tear film stability and automatically evaluate the condition of dry eye. In oral health, CNNs with masked regions can detect and segment each tooth in panoramic images. Mu-Qing Liu used a U-network to detect and segment the mandibular third molar and mandibular nerve canal at the pixel level, and used a ResNet-34 network to segment the M3-MC relationship.
[0005] Utilizing deep learning to more effectively assess implant stability in CBCT images, thereby better guiding further implant repair treatment, is of great significance. Summary of the Invention
[0006] To address the aforementioned problems, the present invention aims to provide a method and apparatus for constructing a classification model for evaluating implant stability based on CBCT image data.
[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a method for constructing a classification model for evaluating implant stability based on CBCT image data, comprising the following steps:
[0008] (1) CBCT image data collection and preprocessing:
[0009] CBCT data from several patients were collected, the CBCT data were processed, and cross-sectional images of the dentition and implants were reconstructed. The dentures in the implant cross-sectional images were manually marked.
[0010] (2) Training the Mobilenetv2-DeeplabV3+ network:
[0011] The labeled implant cross-sectional image data obtained in step (1) is input into the Mobilenetv2-DeeplabV3+ semantic segmentation neural network for training. After training, an implant denture segmentation model based on CBCT images is obtained.
[0012] (3) Peri-implant bone image extraction:
[0013] The CBCT images from the training set are input into the implant segmentation model to obtain the corresponding implant cross-sectional image set; the images in the image set are processed in combination with the original information of the implant to obtain the corresponding peri-implant bone image set.
[0014] (4) ResNet-50 network training:
[0015] According to the required classification method, the peri-implant bone images obtained in step (3) are classified according to their ISQ score information, and the peri-implant bone image set and its classification information are input into the ResNet-50 network. The network is trained according to the set classification mode to obtain a classification model for evaluating implant stability.
[0016] The Mobilenetv2-DeeplabV3+ network is as follows: the original improved Xception backbone network is modified to Mobilenetv2; after completing the feature extraction of MobilenetV2, two effective feature layers are obtained. The result of compressing the height and width of the input image twice is used as the low-level feature, and the result of compressing the height and width of the input image four times is processed by ASPP as the high-level feature. The ASPP part and the decoder part are consistent with DeeplabV3+.
[0017] Most semantic segmentation networks currently consist of encoding and decoding blocks. The encoding block progressively downsamples the input image, with the resolution of the feature maps decreasing stepwise to capture deeper feature information. The decoding block uses the upsampling method to generate a segmentation result with the same resolution as the original image. DeeplabV3+ uses the same encoding and decoding block construction, employing an improved Xception network, applying depthwise separable convolutions, and incorporating BN and ReLU to enhance feature extraction capabilities. In the encoding block, ASPP strengthens the feature extraction network, performing parallel Atrous Covnvolution on the initially effective feature layers compressed four times. This is used for feature extraction via parallel convolutions at different ratios, followed by concatenation, and finally, feature compression via 1x1 convolutions. In the decoding block, low-level features undergo 1x1 convolutions to reduce the number of channels, while high-level features are upsampled by 4x bilinearly and then concatenated with the corresponding low-level features. Feature concatenation or finally, feature layers are refined using 3x3 convolutions, followed by a final 4x bilinear upsampling. The Mobilenetv2 is composed of an InvertedResidual block as its core component. The InvertedResidual block has a spindle-shaped network structure, which first increases the dimensionality through 1x1 convolution, then through depthwise separable convolution, and finally reduces the dimensionality. At the same time, the input and output of each block are connected through residual edges to stack tensors.
[0018] The implant segmentation model obtained after training in step (2) exhibits good segmentation performance, with an mIoU of 94.4%, mPA of 96.76%, average recall of 96.87%, average precision of 97.33%, and accuracy of 99.76%, laying the foundation for subsequent evaluation of implant stability. Inputting the patient's CBCT image data into the model yields an output image of the segmented implant cross-section (the prosthesis portion is marked with different colors).
[0019] Further processing of the segmented cross-sectional image of the implant prosthesis can yield an image of the bone surrounding the implant. Since the bone surrounding the implant is highly correlated with implant stability, this image can be used to assess implant stability.
[0020] Depending on the desired classification method, a ResNet-50 network is trained to establish the correlation between peri-implant bone images and implant classification, ultimately yielding a corresponding classification model for assessing implant stability based on CBCT image data. Different classification methods can lead to different classification models.
[0021] In a preferred embodiment of the present invention, the Mobilenetv2-DeeplabV3+ network training includes at least one of the following processes:
[0022] a. Expand the dataset using online random data augmentation;
[0023] b. The Focal loss function is introduced to address the severe imbalance in the ratio of positive to negative samples in one-stage object detection.
[0024] c. Load the MobileNetV2 pre-trained weights trained on the ImageNet dataset.
[0025] This invention employs online random data augmentation, performing operations such as translation, rotation, and folding on each batch of data to expand the dataset and prevent model overfitting. Furthermore, the Focal loss function significantly reduces the weight of negative samples during training, thereby addressing the severe imbalance in the positive-to-negative sample ratio in one-stage object detection. Those skilled in the art can formulate and adopt appropriate processing methods based on the type and quantity of selected data samples, and are not limited to the methods mentioned above.
[0026] As a preferred embodiment of the present invention, the training parameter settings of the Mobilenetv2-DeeplabV3+ network in the construction method include: an initial learning rate of 5e-4, an input image resolution of 256x256, the Adam optimizer, the stepLR learning rate decay function, and a gamma of 0.94; a backbone network downsampling factor of 8, and a loss function obtained by combining the Dice loss function and the cross-entropy loss function.
[0027] As a preferred embodiment of the present invention, the processing of the obtained implant cross-sectional image set in step (3) is as follows: flip all implant images implanted in the mandible so that all implant roots are on top and the perforated part and upper structure are on the bottom; then crop the image according to the implant information to limit the outer frame range; further expand outward by 5 pixels according to the denture edge information obtained by semantic segmentation, then hide the corresponding pixels of the implant and retain the remaining pixels.
[0028] By flipping all images of implants placed in the mandible so that all implant roots are facing upwards and the perforated portions and superstructures are facing downwards, errors in machine recognition can be reduced. This invention provides one method for obtaining peri-implant bone images through image processing. Those skilled in the art can adjust this method or use different approaches to obtain peri-implant bone images.
[0029] In a preferred embodiment of the present invention, the construction method includes at least one of the following processes for training the ResNet-50 network:
[0030] a. Use online data random augmentation.
[0031] b. Train the ResNet-50 network by freezing / thawing.
[0032] c. Sample size estimation steps before training;
[0033] d. Load the pre-trained model based on ImageNet.
[0034] During the training process, transfer learning is carried out based on the pre-trained model and frozen training. Those skilled in the art can formulate and adopt appropriate processing methods according to the type and quantity of selected data samples, and are not limited to the content mentioned above.
[0035] In a preferred embodiment of the present invention, the training parameters for the ResNet-50 network in the construction method include: an input image resolution of 64x64, a cosine dynamic learning rate, an initial learning rate of 1e-3, a minimum learning rate of 1e-5, the use of the Adam optimizer, and the use of cross-entropy as the loss function.
[0036] The present invention also claims a device comprising a classification model for assessing implant stability based on CBCT image data, obtained according to the said construction method.
[0037] The device is a computer-readable storage medium storing a computer program that can be executed by a computer processor to perform computer-readable instructions for the method described therein, in order to assess implant stability.
[0038] In a preferred embodiment of the present invention, the classification model includes a binary classification model, a ternary classification model, and a quaternary classification model; the binary classification model uses 65 as the dividing line; the ternary classification model uses 60 and 70 as the dividing lines; and the quaternary classification model uses 50, 60, and 70 as the dividing lines.
[0039] Those skilled in the art can construct corresponding classification models based on the classification methods shown. This invention constructed and tested the classification performance of the models in binary, triadic, and quadruple classification. Although the model performance slightly decreased as the required classification precision increased, the accuracy remained generally above 90%, making it effective for assessing implant stability.
[0040] In a preferred embodiment of the present invention, the device includes at least one classification model.
[0041] The device can store one or more classification models, allowing applications to select the appropriate model for implant stability analysis as needed.
[0042] This invention provides a multi-task cascaded network for assessing implant stability based on CBCT images. First, a trained Mobilenetv2-DeeplabV3+ network segments the implant in the CBCT image. Then, combining oral implantology knowledge and the segmentation results, the VOI (Vascular Object Image) is extracted from the cross-sectional image. Finally, a trained ResNet-50 deep residual network is used to complete the classification, obtaining the implant stability assessment result. Testing and evaluation show that the model of this invention (binary, tri-, and quadruple classification) has high diagnostic performance, short assessment time, and high accuracy (>90%), making it effective for assessing implant stability. This is of great significance for better guiding further implant restoration treatment and provides important reference for the application of artificial intelligence in image diagnosis. Attached Figure Description
[0043] Figure 1 This is a schematic diagram illustrating the operation of implant denture segmentation using the Mobilenetv2-DeeplabV3+ network according to the present invention.
[0044] Figure 2This is a schematic diagram illustrating the use of the ResNet-50 network to evaluate implant stability classification in this invention.
[0045] Figure 3 This is a flowchart of the implant stability classification and assessment based on CBCT images according to the present invention.
[0046] Figure 4 Violin-box plots showing the prediction results of different classification models in this invention.
[0047] Figure 5 This is a confusion matrix representing the classification results of different classification models in this invention.
[0048] Figure 6 This is the analysis result of performance testing of different classification models of the present invention. Detailed Implementation
[0049] To better illustrate the purpose, technical solution, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.
[0050] Example 1: Construction Method of Implant Denture Segmentation Model Based on CBCT
[0051] I. Basic Information
[0052] This study complies with the Declaration of Helsinki and has been approved by the Ethics Committee of the Third Affiliated Hospital of Guangzhou Medical University. All subjects signed informed consent forms before enrollment, and the study also met the requirements of the TRIPOD test form used for predictive model development and validation.
[0053] This study collected CBCT images and concurrent ISQ values from 41 patients who requested implant restoration treatment at the Department of Stomatology, The Third Affiliated Hospital of Guangzhou Medical University from July 2021 to June 2022. The patients included 23 men and 18 women, with a total of 55 implants.
[0054] Inclusion criteria: The narrowest part of the alveolar ridge is wider than 6mm, and the mesiodistal distance of the edentulous space is at least 6mm.
[0055] Exclusion criteria: severe periodontal disease; need for maxillary sinus lift or complex bone augmentation surgery; CBCT images may have severe distortion or scattered artifacts due to orthodontic treatment, metal restoration, etc.; and poor general health, such as severe liver or kidney disease, history of maxillofacial radiotherapy, severe diabetes or hypertension, etc., which are contraindications for implant surgery.
[0056] All implant placements were performed using standard implant surgery procedures, with Straumann implants placed at least 6 weeks after tooth extraction. Dental implant systems use implants with lengths ranging from 8 to 10 mm and diameters from 3.3 to 4.8 mm. CBCT scans are taken immediately after implantation, 3 months post-surgery, and 1 year post-surgery, and ISQ values are measured using Osstell equipment (Osstell AB, Sampgatan, Goteborg, Sweden). ISQ values are measured on the buccal, lingual, and mesial-distal sides to ensure repeatability, and the average values are recorded.
[0057] II. CBCT Image Data Acquisition and Preprocessing
[0058] CBCT scans were performed using NewTom VG (FSV: 90kV, mAs: 12.92, CTDIvoI: 3.05mGy, exposure time: 3.6s), and the data was stored in DICOM. All personally identifiable information (PII) was first removed through an anonymization process. The CBCT data was imported into NewTom software, and the dentition was reconstructed according to the dentition curve. Cross-sectional images perpendicular to the curve were reconstructed with a thickness of 0.3mm, resulting in 20,488 dentition cross-sectional images and 779 implant cross-sectional images. The 779 implant coronal images were manually segmented by a trained radiologist using the open-source annotation software LabelMe 5.0.1.
[0059] III. Construction of the Segmentation Model for Implant Dentures
[0060] (1) Construction of semantic segmentation neural network (Mobilenetv2-DeeplabV3+ network)
[0061] Currently, most semantic segmentation networks consist of encoding blocks and decoding blocks. The encoding blocks progressively downsample the input image, while the resolution of the feature maps is also gradually reduced to capture deeper feature information. The decoding blocks are used to generate segmentation results with the same resolution as the original image by upsampling the small feature maps.
[0062] DeeplabV3+ employs the same encoding and decoding block construction method, utilizing an improved Xception network. It applies depthwise separable convolutions and incorporates Batch Normalization (BN) and ReLU to enhance feature extraction capabilities. In the encoding block, ASPP strengthens the feature extraction network, performing parallel Atrous Covnvolution on the initially effective feature layers compressed four times. These are then used for feature extraction via parallel convolutions at different ratios, followed by concatenation, and finally, feature compression via 1x1 convolutions. In the decoding block, low-level features undergo 1x1 convolutions to reduce the number of channels, while high-level features are upsampled by 4x bilinearly and then concatenated with the corresponding low-level features. Feature concatenation or final feature layer refinement via 3x3 convolutions is followed by a final 4x bilinear upsampling.
[0063] We made the following improvements to the existing DeeplabV3+ network at the network layer:
[0064] At the backbone network level, considering the potential requirements for this research to run in embedded devices, we modified the original improved Xception backbone network to Mobilenetv2. Mobilenetv2 consists of an Inverted Residual block as its core component. The Inverted Residual block network structure is spindle-shaped, first increasing the dimensionality through 1x1 convolutions, then through depthwise separable convolutions, and finally reducing the dimensionality. Simultaneously, the input and output of each block are connected through residual edges, allowing for the stacking of tensors.
[0065] After completing the feature extraction for MobileNetV2, we obtain two effective feature layers: the result of compressing the height and width of the input image twice as low-level features, and the result of compressing the height and width of the input image four times using ASPP as high-level features. The ASPP part and the decoder part are consistent with DeeplabV3+.
[0066] (2) Training of Mobilenetv2-DeeplabV3+ network
[0067] The 779 implant coronal section images, which were manually segmented using the open-source annotation software labelme 5.0.1, were input into the Mobilenetv2-DeeplabV3+ network for training purposes aimed at implant recognition.
[0068] To prevent overfitting and imbalanced positive-negative sample ratios, we employed online random data augmentation, performing translation, rotation, and flipping operations on each batch of data to expand the dataset. We also used Focal loss to address the severe imbalance in the positive-negative sample ratio in one-stage object detection, significantly reducing the weight of negative samples during training. Furthermore, we loaded pre-trained MobileNetV2 weights trained on the ImageNet dataset before training. During training, we froze the backbone of the ImageNet-based pre-trained model, training only the layers outside the backbone.
[0069] The Mobilenetv2-DeeplabV3+ network was trained for a total of 50 epochs. The specific implementation methods and parameters used during training were as follows: batch size of 16, initial learning rate of 5e-4, input image resolution of 256x256, and the Adam optimizer. To accelerate convergence in a normal direction and avoid oscillations near the minimum point, as well as to prevent excessively slow convergence, a stepLR learning rate decay function with gamma of 0.94 was used; the backbone network downsampling factor was 8; and finally, the loss function used was a combination of the Dice loss function and the cross-entropy loss function, with the specific formula as follows:
[0070]
[0071] In the formula, H is the height of the image, W is the width of the image, pi,j is the predicted probability of the foreground at position (i,j), and gi,j is the label at position (i,j), which is 0 or 1.
[0072] The above training method yields a segmentation model for implant-supported prostheses based on CBCT images. By inputting the cross-sectional image of the entire dentition from the CBCT image to be processed into the model, a cross-sectional image of the implant can be obtained, and the implant-supported prosthesis can be segmented (the implant-supported prosthesis portion is displayed using different colors). A schematic diagram of the model's operation is shown below. Figure 1 The green part represents the implant denture portion marked after image model processing.
[0073] Example 2: A method for constructing a classification model for assessing implant stability
[0074] I. Extraction of peri-implant bone images
[0075] The sample images required for modeling (779 implant coronal section images mentioned in Example 1) are input into the implant prosthesis segmentation model constructed in Example 1 to obtain the corresponding implant cross-sectional images and corresponding implant prosthesis segmentation results. All implant images implanted in the mandible are flipped so that all implant roots are on top, and the perforated portion and superstructure are facing down. Combining the information of each implant (including implant length and tooth position), and with the help of the cropping function supported by the deep learning system, the obtained implant cross-sectional images are further processed: the images are cropped according to the implant information, limiting the outer bounding box range. Based on the implant region obtained from semantic segmentation, the segmentation edges are horizontally extended by 5 pixels, retaining the corresponding implant pixels, and the remaining pixels are preserved to obtain the corresponding bone image around the implant.
[0076] II. Sample Size Estimation
[0077] To explore the sample size of the test set, we used the single-arm diagnostic test sample estimation formula in PASS15 software to calculate the sample size. A preliminary experiment was designed. Based on the results, the expected sensitivity was 95%, the specificity 88%, and the required α was 0.05. The sample size was then estimated. According to the calculation formula, the test set should include at least 133 images as the research subjects.
[0078] To reduce overfitting, the data was divided in an 8:2 ratio, and test sets were randomly selected. The final sample size of each test set was as follows: 155 samples each for the binary and quadrilateral classification models, and 150 samples for the trilateral classification model.
[0079] III. Construction of a classification model for assessing implant stability
[0080] To overcome the problems of decreased learning efficiency and ineffective accuracy improvement caused by increasing network depth, we adopted ResNet-50 as the classification neural network. ResNet directly feeds data from one of the first few layers into the input of subsequent layers, skipping many layers. The content of later feature layers is partially contributed by a layer preceding it. ResNet-50 has two basic modules: Conv Block and Identity Block. The Conv Block has different input and output dimensions, so it cannot be concatenated; its function is to change the network's dimensionality. The Identity Block has the same input and output dimensions and can be concatenated to deepen the network. ResNet-50 goes through four stages, each with 3, 4, 6, and 3 blocks respectively. The first block of each stage is the Conv Block, and the rest are Identity Blocks. Depending on the classification mode, the number of neurons in the last fully connected layer is set to the number of classes (see reference). Figure 2 ).
[0081] The 779 peri-implant bone images obtained were used to create duplicates for training using three different classification modes. Each duplicate was randomly divided into a training set and a test set at an 8:2 ratio. The first mode was a binary classification with a predicted ISQ score of 65 as the cutoff (i.e., <65 or ≥65). The second mode was a triadic classification with cutoffs of 60 and 70 (i.e., <60, 60–69, ≥70). The third mode was a quadratic classification with cutoffs of 50, 60, and 70 (<50, 50–59, 60–69, ≥70).
[0082] According to different classification methods, the training set data is classified according to its original ISQ score information (e.g., classified as ISQ<65 or ≥65); and the peri-implant bone images and their classification information data (e.g., classified as ISQ<65 or ≥65) of the training set are input into the ResNet-50 network to train the model with the aim of establishing the correlation between peri-implant bone images and ISQ score classification information, so as to obtain the corresponding classification model.
[0083] During training, a pre-trained model based on ImageNet is also loaded. First, frozen training is performed (the backbone of the ImageNet-based pre-trained model is frozen during training, and only the layers outside the backbone are trained), for a total of 50 epochs, and then unfrozen training is performed up to 1000 generations.
[0084] Online random data augmentation and a freeze / thaw approach were used for training. During training, transfer learning was performed based on the pre-trained model and the frozen training. A total of 1000 epochs were trained. The input image resolution was 64x64. The cosine dynamic learning rate was used, with an initial learning rate of 1e-3 and a minimum learning rate of 1e-5. The Adam optimizer was used, and cross-entropy was used as the loss function.
[0085] Through the above training, three multi-task convolutional neural network models for evaluating implant stability can be obtained: a binary classification model, a three-class classification model, and a four-class classification model.
[0086] Example 3: A method for assessing implant stability based on CBCT images
[0087] This invention provides a method for evaluating implant stability based on CBCT images, as detailed below:
[0088] (1) Inputting the cross-sectional image of the patient's CBCT image of the entire dentition into the implant denture segmentation model of the present invention can obtain the cross-sectional image of the implant denture marked by the implant.
[0089] (2) Based on the clinical information of the implant, the cross-sectional images of the implants of the implant prosthesis are further processed: all implant images implanted in the mandible are flipped so that all implant roots are on top and the perforated part and the superstructure are on the bottom. Then, the images are cropped according to the implant information to limit the outer frame range. Based on the denture edge information obtained by semantic segmentation, the image is further extended outward by 5 pixels. Then, the corresponding pixels of the implant are hidden and the remaining pixels are retained to obtain the bone image around the implant.
[0090] (3) Input the bone image around the implant into the required classification model to obtain the corresponding implant stability assessment results (workflow reference). Figure 3 ).
[0091] Example 4: Model Performance Testing
[0092] 1. Performance testing of implant denture segmentation models
[0093] Test method: The implant prosthesis segmentation model obtained in Example 1 was used to test the test set. 150 CBCT coronal section images were input into the model constructed in Example 1, and the results were compared with the results of manual segmentation. The results are shown in Table 1.
[0094] Table 1 Performance test results of the implant-supported denture segmentation model of the present invention
[0095]
[0096] The model of this invention has good segmentation performance, with an accuracy of 99.76%. Specifically, the mIoU is 94.4%, the mPA is 96.76%, the average recall is 96.87%, and the average precision is 97.33% (see Table 1).
[0097] 2. Performance testing of a multi-task convolutional neural network model for evaluating implant stability.
[0098] Testing Methods: The binary, tri-class, and quadri-class models constructed in Example 2 were tested respectively. The CBCT coronal section images of the test set were input into the entire multi-task cascaded network, and the results were compared and analyzed with the original ISQ of each test sample. The results are as follows:
[0099] In the binary classification test set of 155 images, 149 images were correctly classified, with an accuracy of 96.13%, precision of 96.20%, and recall of 96.06%. In the tri-class classification test set of 150 images, 143 images were correctly predicted, with an accuracy of 95.33%, precision of 95.33%, and recall of 95.37%. In the quadruple classification test set of 155 images, 141 images were correctly predicted, with an accuracy of 92.90%, precision of 93.71%, and recall of 93.01%.
[0100] Further violin-box diagrams were drawn, such as... Figure 4As shown (a is the binary classification result, b is the tri-class classification result, and c is the quadri-class classification result). The horizontal axis shows the predicted value groupings for this replica, and the vertical axis is the actual ISQ measurement value. The box plot depicts the distribution interval of positive predicted values in the replica test set, as well as indicators such as the median and quartiles; the violin plot depicts the distribution and numerical density of positive predicted values. The points at the upper or lower edge of the box plot represent negative predicted values (i.e., sample values in the test set whose predicted values do not match the actual ISQ measurement values). A confusion matrix was further constructed based on the classification results, as shown below. Figure 5 As shown (a is the result of binary classification, b is the result of tri-class classification, and c is the result of quad-class classification).
[0101] In addition, the multi-category variable was converted into a binary category to detect various indicators, and the results are as follows: Figure 6 As shown. By Figure 6 As can be seen, the classification model of the present invention has good accuracy, sensitivity, specificity, PPV (positive predictive value), NPV (negative predictive value), and F1 score, indicating that the model of the present invention can be effectively used for the stability assessment of implant-supported prostheses based on CBCT images.
[0102] 3. Efficiency Assessment
[0103] In addition, we conducted a time-consuming test on the entire diagnostic process. The test platform consisted of an i5 11400 CPU and an NVIDIA RTX 3060 GPU. We randomly selected 20 implants from the entire dataset and performed ISQ prediction on their coronal section images. Using Python's `time` module, we recorded the time taken for each implant to complete all ISQ predictions. This process was repeated ten times, and the results showed that, on average, ISQ prediction for all coronal section images of one implant could be completed in 3.76 seconds.
[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the essence and scope of the technical solutions of the present invention.
Claims
1. A method for constructing a classification model for evaluating implant stability based on CBCT image data, characterized in that, The method comprises the following steps: (1) CBCT image data collection and preprocessing: CBCT data of a plurality of patients is collected, the CBCT data is processed, cross-sectional images of the dentition and cross-sectional images of the implants are reconstructed, and the dentures in the cross-sectional images of the implants are manually marked; (2) Mobilenetv2-DeeplabV3+ network training: The marked cross-sectional image data of the implants obtained in step (1) is input into a Mobilenetv2-DeeplabV3+ semantic segmentation neural network for training, and a CBCT image-based implant denture segmentation model is obtained after training; (3) Implant surrounding bone image extraction: The CBCT images of the training set are input into the implant denture segmentation model to obtain a corresponding set of implant cross-sectional images; the images in the image set are processed in combination with the original information of the implant denture to obtain a corresponding set of implant surrounding bone images; (4) ResNet-50 network training: According to the required classification mode, the implant surrounding bone images obtained in step (3) are classified according to their ISQ score information, and the implant surrounding bone image set and its classification information are input into a ResNet-50 network, and the classification model for evaluating the stability of the implant is obtained according to the set classification mode. The Mobilenetv2-DeeplabV3+ network is: the original improved Xception backbone network is modified into Mobilenetv2; after completing the feature extraction of MobilenetV2, two effective feature layers are obtained, the results of compressing the input image height and width twice are used as low-level features, and the results of compressing the input image height and width four times are used as high-level features after ASPP, and the ASPP part and the decoder part are consistent with DeeplabV3+.
2. The construction method of claim 1, wherein, The Mobilenetv2-DeeplabV3+ network training comprises at least one of the following processes: a. Online random data augmentation is used to expand the data set; b. The Focal loss function is introduced to solve the problem of serious imbalance between positive and negative samples in one-stage target detection; c. The MobilenetV2 pre-training weight trained through the Imagenet data set is loaded.
3. The construction method of claim 1, wherein, The Mobilenetv2-DeeplabV3+ network training parameter setting comprises: the initial learning rate is 5e-4, the input image resolution is 256X256, the Adam optimizer is used, the stepLR learning rate decay function is used, the gamma is 0.94; the downsampling rate of the backbone network is 8, and the loss function is a loss function obtained by combining the Dice loss function and the cross-entropy loss function.
4. The construction method of claim 1, wherein, The processing of the implant cross-sectional image set in step (3) is: flipping all the images of the implants implanted in the mandible, so that all the implant roots are on top, and the transgingival part and the upper structure are downward; then cutting the images according to the information of the implants to limit the range of the outer frame; further expanding 5 pixels outward according to the denture edge information obtained by semantic segmentation, then hiding the pixels corresponding to the implants, and retaining the remaining pixels.
5. The construction method of claim 1 wherein, The ResNet-50 network training includes at least one of the following processes: a. Online data random enhancement is adopted; b. ResNet-50 network training is performed by freezing / thawing; c. Sample size estimation step before training; d. Loading a pre-trained model based on Imagenet.
6. The construction method of claim 1 wherein, The training parameter setting of the ResNet-50 network training includes: input image resolution is 64X64, cos dynamic learning rate, initial learning rate is 1e-3, minimum learning rate is 1e-5, Adam optimizer is adopted, and cross-entropy is adopted as the loss function.
7. An apparatus comprising a classification model for evaluating implant stability based on CBCT image data obtained by the construction method according to any one of claims 1-6.
8. The apparatus of claim 7 wherein, The classification model includes a binary classification model, a three-classification model, and a four-classification model; wherein the binary classification model takes 65 as the boundary; the three-classification model takes 60 and 70 as the boundary; and the four-classification model takes 50, 60 and 70 as the boundary.