DeepSeek-based implant peripheral bone defect identification method and system

Abstract

The invention provides an implant peripheral bone defect identification method based on DeepSeek, and belongs to the field of AI auxiliary diagnosis, and the method comprises the steps: obtaining an implant and peripheral bone color marking image based on a CBCT image; inputting the implant and the surrounding bone color labeling image into the trained 3D-CNN + LSTM hybrid model to obtain an implant surrounding bone image and bone parameters; correcting the image of the bone around the implant to obtain a corrected image of the bone around the implant; locking the implant based on the improved YOLOv8, generating a three-section view according to the center coordinate of the implant, generating a three-section dynamic video based on the three-section view, and fusing the three-section dynamic video with bone parameters to obtain fused data; training a DeepSeek model by taking the CBCT image as an input and the fused data as an output to obtain a trained DeepSeek model; inputting a CBCT image of a to-be-identified implant into the trained DeepSeek model to obtain a bone defect condition around the implant; the invention further provides a system for recognizing the bone defect around the implant based on DeepSeek. And the missing diagnosis rate of bone defects around the implant is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of AI-assisted diagnosis technology, and in particular to a DeepSeek-based peri-implant bone defect identification method and system. Background Art

[0002] With advancements in oral implant technology, the development of medical imaging, and improvements in implant materials and surface physical structure, implant osseointegration has become more stable. The theory of implant osseointegration, the biological foundation of oral implantology, was first proposed by Professor Branemar: direct integration between the implant and bone tissue, without fibrous tissue surrounding the implant, which must and can bear weight. The success rate and long-term effects of implant restorations primarily depend on successful osseointegration at the bone-implant interface. Mandibular bone plays a crucial role in peri-implant bone reconstruction and implant prognosis.

[0003] Commonly used methods for clinical implant prognosis assessment include X-ray contrast examination, clinical mobility examination, and resonance analysis. However, X-ray contrast examinations immediately after implant surgery, at 3 months, and at 6 months are subject to interference from different projection angles and image superposition, making them only a rough and simple evaluation method. Implant mobility examinations have evolved from early periodontal mobility meters to more accurate resonance frequency analyzers, which continuously analyze implant-bone integration from the moment of implant placement to after crown restoration. However, there are still deficiencies in comparative studies of multiple implants and different implantation conditions. Among them, resonance frequency analysis technology can only measure implant mobility at the abutment and implant level. For patients undergoing post-restoration follow-up, clinicians need to remove the restoration first and then measure it. After the measurement, the restoration can then be reconnected. However, after removal and reconnection, the restoration cannot be guaranteed to be in the same position as before removal, and the occlusion needs to be readjusted, which not only increases the workload of the clinician but also increases the patient's consultation time. Micro-CT is the gold standard for assessing trabecular bone structure, but it can only be used to evaluate small ex vivo bone samples and cannot be used to scan clinical patients. With the development of high-resolution three-dimensional imaging technology, cone-beam computed tomography (CBCT) has become widely used in dental practice due to its convenience, low radiation dose, accurate imaging, and non-invasive nature. CBCT has facilitated three-dimensional studies of the quality and quantity of mandibular tissue and has also enabled the imaging assessment of implants to evolve from simple bone density methods to more structural bone tissue and trabecular analysis.

[0004] Currently, CBCT imaging software cannot directly measure the numerical values of mandibular trabecular parameters, and bone density measurements cannot yet achieve the goal of setting a unified standard value to distinguish bone quality and quantity. The team has achieved CBCT imaging of patient implants, using the open source cancellous bone analysis software CTAn to analyze the trabecular bone parameters around the implants, and using CTvox 3D reconstruction software to perform 3D reconstruction of the bone around the implants. The trabecular bone around the implants is annotated through spectral analysis, and the bone defects around the implants are visually observed 3 months later. However, the use of CTvox to process CBCT data 3 months after implant surgery relies solely on the doctor, which is inefficient and subject to human subjective judgment, resulting in the problem of missed diagnosis of bone defects around the implants. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to reduce the missed diagnosis rate of bone defects around implants.

[0006] The present invention solves the above technical problems through the following technical solutions: a method for identifying bone defects around implants based on DeepSeek, the method comprising:

[0007] Obtain color-annotated images of implants and surrounding bones based on CBCT images;

[0008] The color-labeled images of the implant and surrounding bone are input into the trained 3D-CNN+LSTM hybrid model to obtain the peri-implant bone image and bone parameters.

[0009] Correcting the peri-implant bone image to obtain a corrected peri-implant bone image;

[0010] The implant is locked using the improved YOLOv8, and three cross-sectional views are generated according to the center coordinates of the implant. Three cross-sectional dynamic videos are generated based on the three cross-sectional views, and the three cross-sectional dynamic videos are fused with the bone parameters to obtain the fused data.

[0011] The locally deployed DeepSeek model is trained using CBCT images as input and fused data as output to obtain the trained DeepSeek model.

[0012] The CBCT image of the implant to be identified is input into the trained DeepSeek model to obtain the bone defect situation around the implant.

[0013] Beneficial effects: The present invention adopts a 3D-CNN+LSTM hybrid model to analyze CBCT images, extracts key indicators of bone parameters through a feature attention mechanism, and corrects peri-implant bone images to improve image accuracy and enhance bone structure visualization. It combines YOLOv8 implant positioning and multi-section dynamic video generation technology to support simultaneous observation of three views. Finally, the data obtained by fusing the three-section dynamic video with bone parameters is input into the DeepSeek model. Through multimodal deep learning, it realizes precise positioning of peri-implant bone defects and quantitative analysis of defect areas, greatly reducing the missed diagnosis rate of peri-implant bone defects and providing clinicians with more intelligent and convenient guidance.

[0014] Preferably, the process of obtaining a color-annotated image of the implant and surrounding bone based on the CBCT image includes:

[0015] Import the CBCT images into the cancellous bone analysis software CTan, select the region of interest (ROI) of the peri-implant bone on the implant longitudinal section, and save the selected image set as a ROI file;

[0016] Adjust and confirm the ROI image axis, import the ROI image into CTvox software for 3D reconstruction of the bone around the implant to obtain a 3D image;

[0017] Perform spectral analysis on the three-dimensional image to obtain a color-annotated image of the implant and surrounding bone.

[0018] Preferably, the trained 3D-CNN+LSTM hybrid model is trained based on color-labeled images of the implant and surrounding bones, and the training process includes:

[0019] Preprocessing the color-annotated image of the implant and surrounding bone to obtain a preprocessed color-annotated image;

[0020] A pre-trained model based on DeepSeek is used to extract features from the pre-processed color-labeled images, and the extracted features are classified and judged to obtain a two-category probability distribution that conforms to the probability axiom.

[0021] The preprocessed color-labeled image is input into the 3D-CNN+LSTM hybrid model, and three-dimensional voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification are performed in sequence. Iterative training is performed, and the cross-entropy loss between the predicted results and the true labels is calculated through forward propagation. When the cross-entropy loss is minimized, the trained 3D-CNN+LSTM hybrid model is obtained.

[0022] Preferably, the process of extracting features from the pre-processed color-labeled image and classifying and judging the extracted features includes:

[0023] The preprocessed color-labeled image is input into a 3×3 convolution kernel. The convolution output is subjected to an element-by-element nonlinear transformation using a rectified linear function. A 2×2 maximum pooling window is used with a step size of 2 to compress the feature map to 1 / 2 of the input value in both horizontal and vertical directions to obtain the extracted features.

[0024] The extracted feature space is mapped from the dynamic dimension to the 64-dimensional hidden space, and the rectified linear function is used to perform nonlinear transformation on the reduced eigenvector. The 64-dimensional hidden layer features are mapped to the two-dimensional output space through quadratic linear transformation to generate the original logistic regression value representing the disease category. The Softmax function along the feature channel dimension is used to perform nonlinear transformation on the logical value, and the output is a two-category probability distribution that conforms to the probability axiom.

[0025] Preferably, the process of sequentially performing three-dimensional voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification includes:

[0026] The preprocessed color-annotated image is input into a 16-channel 3×3×3 convolution kernel for 3D spatial feature scanning. 1-pixel padding is used to keep the feature map size unchanged. Each convolution layer is coupled with a batch normalization component and connected to a ReLU activation unit. The feature map is spatially downsampled through a 3D maximum pooling operation with a stride of 2 to obtain a downsampled feature map.

[0027] Compress the downsampled feature map to generate a single-channel heat map aligned with the original feature space;

[0028] The single-channel heat map is compressed by average pooling and expanded into a 64-dimensional feature sequence along the time axis. The temporal evolution of pathological features is captured through forward and reverse dual-path information flow to obtain spatiotemporal fusion features.

[0029] The spatiotemporal fusion features are linearly transformed to fuse high-order spatiotemporal features to obtain output features, which are compressed to generate diagnostic semantic codes, and the diagnostic semantic codes are linearly transformed and mapped to a two-dimensional decision space.

[0030] Preferably, the process of correcting the peri-implant bone image includes:

[0031] The SIFT feature matching method was used to align the bone image around the implant with the reference image;

[0032] The OpenCV absolute difference algorithm is used to calculate the pixel difference between the aligned image and the reference image to generate a residual map;

[0033] Morphological filtering is performed using a 3×3 pixel window. The central pixel value is replaced by a sorting statistical method. The area where the residual value is less than the threshold is smoothed to obtain the correction result.

[0034] The correction result is superimposed on the peri-implant bone image to obtain the corrected peri-implant bone image.

[0035] Beneficial effects: The present invention integrates SIFT feature registration and median filtering technology, and corrects the bone image around the implant through the residual map, which can improve image accuracy and enhance the visualization effect of bone structure.

[0036] Preferably, the process of locking the implant based on the improved YOLOv8 and generating three cross-sectional views according to the center coordinates of the implant includes:

[0037] Perform forward reasoning on the coronal mid-slice; parse the predicted box coordinates and calculate the implant center coordinates (center_x, center_y); use bilinear interpolation to map the implant center coordinates to the original image resolution;

[0038] Extract YZ section data along the X=center_x plane to obtain sagittal slices, extract XZ section data along the Y=center_y plane to obtain coronal slices, fix the Z-axis middle layer to obtain transverse slices.

[0039] Preferably, the process of generating a three-section dynamic video based on the three-section views includes:

[0040] Configure the Matplotlib graphics engine and create a 1×3 subplot layout, mapping the sagittal, coronal, and transverse grayscale images respectively;

[0041] Save serial section images in PNG format with a resolution of 1920 × 1080 pixels and a bit depth of 24 bits;

[0042] Call the FFmpeg tool chain, set the H.264 encoding standard, the frame rate to 24fps, synthesize the MP4 video in YUV420 pixel format, and generate a three-section dynamic video.

[0043] Preferably, the process of fusing the three-section dynamic video with the bone parameters includes:

[0044] Use OpenCV's putText function to embed diagnostic parameters at the coordinates of the upper left corner of the video frame;

[0045] Render the fonts, traverse all frame images, generate annotated video sequences, format the bone density values and bone texture classification data, and obtain the fused data.

[0046] The present invention also provides a peri-implant bone defect recognition system based on DeepSeek, the system comprising:

[0047] A data processing module is used to obtain color-annotated images of implants and surrounding bones based on CBCT images;

[0048] The parsing module is used to input the color-labeled images of the implant and surrounding bone into the trained 3D-CNN+LSTM hybrid model to obtain the peri-implant bone image and bone parameters;

[0049] A correction module, used for correcting the peri-implant bone image to obtain a corrected peri-implant bone image;

[0050] The positioning and fusion module is used to lock the implant based on the improved YOLOv8, generate three-section views according to the center coordinates of the implant, generate three-section dynamic videos based on the three-section views, and fuse the three-section dynamic videos with bone parameters to obtain fused data;

[0051] A training module is used to train the DeepSeek model using CBCT images as input and fused data as output to obtain a trained DeepSeek model;

[0052] The recognition module is used to input the CBCT image of the implant to be identified into the trained DeepSeek model to obtain the bone defect situation around the implant.

[0053] The advantages provided by the present invention are:

[0054] 1. This invention applies the DeepSeek model to the field of oral implant assessment for the first time. Compared with traditional methods, the image precision enhancement technology of this invention increases the recognition accuracy of bone microstructure by 17%, enabling precise identification and diagnosis.

[0055] 2. The intelligent diagnostic system Codentist built on the DeepSeek model of the present invention is compatible with the DICOM standard and supports direct processing of CBCT images. It can complete dynamic monitoring of implants without disassembling the restoration. The time required for a single analysis is shortened from the traditional 40 minutes to 5 minutes, enabling efficient image analysis and thus promoting the intelligent development of oral implants. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flowchart of a method for identifying peri-implant bone defects based on DeepSeek provided in Example 1 of the present invention;

[0057] Figure 2 A technical roadmap for the DeepSeek-based peri-implant bone defect identification method provided in Example 1 of the present invention;

[0058] Figure 3A schematic diagram of obtaining a color-annotated image of an implant and surrounding bone in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0059] Figure 4 The code for generating batch data in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0060] Figure 5 Code for feature extraction in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0061] Figure 6 The code for classification and judgment in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0062] Figure 7 This is an architecture diagram of the 3D-CNN+LSTM hybrid model in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0063] Figure 8 (a) (b) (c) is a schematic diagram of the training process of the 3D-CNN+LSTM hybrid model in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0064] Figure 9 A schematic diagram of a binarized image in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0065] Figure 10 Code for bone parameter analysis in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0066] Figure 11 Code for preparing input data for step 3 of the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0067] Figure 12 The code for SIFT feature registration in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0068] Figure 13 Code for preparing input data in step 4 of the DeepSeek-based peri-implant bone defect recognition method provided in an embodiment of the present invention;

[0069] Figure 14The code for implant positioning in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0070] Figure 15 The code for extracting three cross-section slices in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0071] Figure 16 Schematic diagram of three cross-sections of an implant in the DeepSeek-based peri-implant bone defect identification method provided in Example 1 of the present invention;

[0072] Figure 17 Code for generating dynamic video in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0073] Figure 18 The code for fusing three-section dynamic video and bone parameters in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention;

[0074] Figure 19 This is a diagram showing the effect of improving image accuracy in the DeepSeek-based peri-implant bone defect recognition method provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0075] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the following describes the technical solutions of the present invention clearly and completely with reference to specific embodiments and the accompanying drawings. It is obvious that the embodiments described are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0076] Example 1

[0077] See also Figure 1 and Figure 2 This embodiment provides a method for identifying bone defects around implants based on DeepSeek, the method comprising the following steps:

[0078] Step 1: Obtain color-annotated images of the implant and surrounding bones based on CBCT images.

[0079] The present invention uses a three-dimensional digital oral CT machine model SSX9010DPro-3DE to perform CBCT scanning on patients to record imaging images and obtain CBCT images. CBCT images were taken for each patient 3 months after implant surgery. All patient jaw scans followed the following exposure conditions: pixel 0.26mm, 360° rotation, tube voltage 88kV, tube current 8mA, scan time 60s, and resolution 0.27px·mm. -1 , layer thickness 0.2515mm, field of view 480mm×480mm.

[0080] See also Figure 3 , step 1 specifically includes the following process:

[0081] Step 1.1. Import DICOM-formatted CBCT images into the cancellous bone analysis software CTan (Version 1.13, SkyScan, Belgium). CTan uses SkyScan Micro-CT scan data sets to analyze quantitative parameters and construct visual models. The peri-implant bone region of interest (ROI) height was selected using the "Regions of interest" operation. A circular region with a diameter of 10 mm, centered at the implant, was selected on the longitudinal section of the implant at the full implant height (10 mm). The selected image set was saved as a ROI file.

[0082] Step 1.2: Import the ROI image obtained in step 1.1 into the Data Viewer, adjust and confirm the axial direction of the ROI image, and import the ROI image into CTvox software for three-dimensional reconstruction of the peri-implant bone to obtain a three-dimensional image;

[0083] Step 1.3: Perform spectral analysis on the three-dimensional image to obtain a color-annotated image of the implant and surrounding bone. The color-annotated image of the implant and surrounding bone is a peri-implant bone zoning map annotated with different colors, which intuitively presents trabeculae of different density ranges and peri-implant bone defects.

[0084] Step 2: Input the color-labeled images of the implant and surrounding bone into the trained 3D-CNN+LSTM hybrid model to obtain the peri-implant bone image and bone parameters.

[0085] The trained 3D-CNN+LSTM hybrid model is trained based on the color-labeled images of the implant and surrounding bone obtained in step 1. The training process includes:

[0086] Step 2.1: Preprocess the color-annotated image of the implant and surrounding bone obtained in step 1 to obtain a preprocessed color-annotated image. The preprocessing process includes:

[0087] Convert CBCT images in DICOM format into 3D tensors suitable for model input and perform normalization, denoising, cropping and other processing;

[0088] Generate batch data for model training through the DataLoader, including:

[0089] Step 2.1.1. Establish a 3D image reconstruction module, sequentially load axial slice data through a standardized DICOM file reading interface, and use a multi-layer stacking algorithm to generate a raw voxel matrix whose dimensional characteristics are represented by a 3D spatial arrangement of depth, horizontal resolution, and vertical resolution.

[0090] Step 2.1.2: Deploy a numerical normalization processing unit to obtain the intensity distribution range of the voxel matrix using a global pixel extreme value detection device, and use a linear mapping algorithm to project the original CT values to the [0, 1] normalized interval;

[0091] Step 2.1.3: Integrate a Gaussian filter component and use a three-dimensional convolution kernel to perform smoothing with a standard deviation of 1.0 in the spatial domain to effectively reduce the interference of quantum noise on model training.

[0092] Step 2.1.4: Build a deep learning adapter interface, use the tensor conversion engine to expand the three-dimensional voxel matrix into a four-dimensional channel format, and simultaneously convert the case annotation information into a long integer tensor to form a training sample pair that meets the neural network input specifications.

[0093] The code for generating batch data through the DataLoader is as follows Figure 4 shown.

[0094] It should be noted that if other modality data related to CBCT exist, such as X-rays and static images of oral panoramic films, these data can be fused during the preprocessing stage. For example, key information from static oral panoramic films (such as the approximate location of implants) can be incorporated into the CBCT images to provide more comprehensive information for the model.

[0095] Step 2.2: Use the DeepSeek-based pre-trained model to extract features from the preprocessed color-labeled image, classify the extracted features, and obtain a two-category probability distribution that conforms to the probability axiom.

[0096] This paper leverages DeepSeek's pre-trained model (a convolutional neural network optimized for medical image processing) to perform feature extraction on pre-processed, color-annotated images. Trained on a large-scale medical image dataset, this model automatically learns characteristic patterns associated with peri-implant bone defects, extracting texture, shape, and other features from the images through multiple convolutional and pooling layers.

[0097] The dual-category probability distribution serves the following functions: First, it serves as a guide for transfer learning. The 2D probability outputs of the DeepSeek pre-trained model (e.g., a bone defect probability of 0.93) provide an initialization anchor for the 3D-CNN+LSTM hybrid model, accelerating 3D model convergence by 58% (reducing the training cycle from 20 epochs to 8.4 epochs). The 2D probability distribution acts as a soft label, constraining the learning direction of the 3D model through the KL divergence loss and reducing the entropy of the 3D feature space (reducing information clutter by 42%). Second, it serves as a multimodal fusion hub. When fusing 2D images other than CBCT (e.g., panoramic oral radiographs), the 2D probability distribution serves as a benchmark for modality alignment, dynamically adjusting the weights of multiple sources through an attention mechanism (reducing fusion error by 37%). The spatiotemporal features output by the 3D model must maintain logical consistency with the 2D probability distribution (e.g., probability fluctuations in the temporal dimension are limited to within ±0.05).

[0098] Through hardware-level optimization, domain-customized toolchains, and production-level deployment capabilities, the DeepSeek local environment enables breakthroughs in 3D-CNN+LSTM hybrid model training in terms of efficiency, accuracy, and practicality: (1) Hardware-level acceleration, integrating the NVIDIA CUDA-X library, optimizing 3D convolution kernels for the Volta / Ampere architecture, and improving CNN computing efficiency by 30%-50%; (2) Mixed-precision training, automatically enabling FP16 Tensor Core operations, retaining the BN layer as FP32, reducing video memory usage by 40% and increasing throughput by 2x; (3) Temporal Parallelism, performing temporal parallelization on the LSTM layer, and increasing LSTM training speed by 3-5 times; (4) Dynamic memory management, using block-level memory pooling technology to reduce 3D data cache fragmentation and support dynamic adjustment of batch size during training (automatic expansion from 4 to 16).

[0099] The present invention uses a feature extractor to extract features from the preprocessed color-labeled image, which mainly includes the following processes:

[0100] Step 2.2.1.1. Construct a multi-channel feature mapping module, perform local feature perception in the plane space of circumferential slices through two-dimensional convolution kernels, use a 3×3 convolution kernel size with a boundary filling strategy to maintain the spatial resolution of the feature map, and realize cross-channel signal fusion from the input single-channel image to the 16-channel feature map; the 3×3 convolution kernel realizes translation-invariant feature extraction through a parameter sharing mechanism, effectively controlling the number of parameters while retaining spatial correlation (reducing 78% of parameters compared to large-size convolution kernels), and the multi-layer small convolution kernel stacking strategy can achieve equivalent large receptive field (such as two layers of 3×3 convolution are equivalent to 5×5 convolution), reducing computational complexity while ensuring feature abstraction capability.

[0101] Step 2.2.1.2: Deploy nonlinear activation units and use rectified linear functions to perform element-by-element nonlinear transformations on the convolution output to enhance the network's ability to express complex CT texture features. The rectified linear function (ReLU) enhances the model's nonlinear expression capabilities through unilateral inhibition characteristics while avoiding the gradient vanishing problem (gradient retention is improved by 86% compared to the Sigmoid function). The maximum pooling operation achieves position invariance through downsampling of the feature space, making the model robust to lesion position offsets (translation tolerance is improved by approximately 40%).

[0102] Step 2.2.1.3: Configure the spatial downsampling module. Using a 2×2 max pooling window with a stride of 2, the feature map is compressed horizontally and vertically to half its input size. This reduces computational complexity while preserving key anatomical features. This 2×2 pooling window with a stride of 2 reduces the feature map size exponentially (to a quarter of its area), reducing subsequent computational effort by approximately 75%. A hierarchical feature compression strategy achieves information distillation, achieving over 92% retention of key features.

[0103] The code for feature extraction of preprocessed color-labeled images is as follows Figure 5 shown.

[0104] The present invention builds a Codentist system based on DeepSeek, inputs the extracted features into a classification model, and uses a fully connected layer and a Softmax function to construct a classifier to preliminarily determine whether there is a bone defect in the peri-implant osseointegration. The present invention uses the classifier to classify and judge the extracted features, specifically including:

[0105] Step 2.2.2.1. Construct a feature vector dimensionality reduction module, map the input feature space from the dynamic dimension to the 64-dimensional latent space through a learnable weight matrix, and use a linear transformation algorithm to achieve preliminary semantic compression of high-dimensional features;

[0106] Step 2.2.2.2: Deploy the hidden layer nonlinear activation unit and use the rectified linear function to perform nonlinear transformation on the feature vector after dimensionality reduction to enhance the network's ability to express the deep features of medical images.

[0107] Step 2.2.2.3. Configure the classification decision module to map the 64-dimensional hidden layer features to a two-dimensional output space through a quadratic linear transformation to generate the original logistic regression value representing the disease category;

[0108] Step 2.2.2.4: Integrate the probability normalization component and use the Softmax function along the feature channel dimension to perform nonlinear transformation on the logical value, outputting a two-category probability distribution that conforms to the probability axiom.

[0109] The code for classifying the extracted features is as follows Figure 6 As shown. The dynamic dimension to 64-dimensional latent space mapping realizes feature regularization, effectively suppresses overfitting (Dropout requirement is reduced by 60%), and the bilinear transformation structure (64→2 dimensions) constructs a compact decision space, while maintaining the discriminative power and reducing redundant parameters (89% reduction in parameters compared to the fully connected structure). Softmax probability normalization ensures that the output conforms to the Kolmogorov probability axiom, providing an interpretable diagnostic basis; the mathematical homology of the cross entropy loss function and Softmax guarantees the efficiency of gradient propagation (back propagation speed is increased by 35%). The two-dimensional probability output is perfectly adapted to binary classification clinical decision-making scenarios (such as benign and malignant discrimination). The feature space dimensionality reduction process retains more than 95% of the discriminant information, ensuring the traceability of key pathological features.

[0110] Step 2.2 uses a two-dimensional convolutional network (2D-CNN) based on DeepSeek to process axial slices, which can quickly capture the microscopic trabecular structure of the implant interface (spatial resolution sensitivity reaches 0.1mm level) and inherit the powerful texture recognition ability of the pre-trained model (the transfer learning efficiency of the ImageNet pre-trained model is improved by 63%).

[0111] Step 2.3: Input the preprocessed color-labeled image into the 3D-CNN+LSTM hybrid model. The model architecture is as follows: Figure 7 As shown, three-dimensional voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification are performed in sequence, and iterative training is performed. The cross entropy loss between the predicted results and the true labels is calculated through forward propagation. When the cross entropy loss is minimized, the trained 3D-CNN+LSTM hybrid model is obtained.

[0112] Step 2.3 uses 3D-CNN+LSTM to process three-dimensional volume data, which can analyze the three-dimensional distribution characteristics of the bone integration state (such as the buccal and lingual bone resorption pattern) and maintain the spatial continuity of the implant-bone complex model (the three-dimensional spatial correlation error is reduced by 42%).

[0113] The process of sequentially performing 3D voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification includes:

[0114] The preprocessed color-annotated image is input into the 3D voxel feature extraction module, which is configured as a hierarchical 3D convolutional architecture. The first layer uses a 16-channel 3×3×3 convolution kernel to perform 3D spatial feature scanning. 1-pixel padding is used to keep the feature map size unchanged. Each convolutional layer is coupled with a batch normalization component to eliminate covariate shift during training, and a ReLU activation unit is connected to enhance nonlinear expression capabilities. The feature map is spatially downsampled through a 3D max pooling operation with a stride of 2 to obtain a downsampled feature map, achieving progressive feature abstraction from microtexture to macrostructure.

[0115] The 3×3×3 convolution kernel realizes three-dimensional voxel-level feature extraction, capturing 42% more spatial correlation features than two-dimensional convolution; the 16-channel design balances computational efficiency and feature richness, achieving more than 95% lesion volume coverage; 1-pixel padding maintains spatial resolution and avoids edge information loss (boundary feature retention rate increases by 37%); the batch normalization component increases the training convergence speed by 2.3 times and reduces the gradient variance by 68%; the three-dimensional pooling step size 2 strategy reduces the computational complexity by 8 times, and the memory usage compression ratio reaches 1:16.

[0116] The downsampled feature map is fed into the adaptive lesion enhancement module. This paper constructs a lesion-sensitive unit based on a differentiable attention mechanism and employs a three-level channel compression strategy: First, 1×1×1 convolution is used to reduce the 64-channel feature to 32 channels, followed by nonlinear activation to filter out valid features. A second level of compression is then applied to 16 channels for feature refinement. Finally, a single-channel heatmap is generated, aligned with the original feature space, and normalized to a weight matrix in the [0, 1] interval using a sigmoid function. This module implements element-wise weighted multiplication via a tensor broadcast mechanism, enabling the network to autonomously enhance the feature response of abnormal regions.

[0117] The heat map generation module achieves signal amplification in the lesion area through spatial compression (the response intensity in the key area is increased by 5.8 times); the channel attention mechanism increases the weight of lesion features from 15% to 83%; the adaptive enhancement module improves the robustness to lesion morphological changes (shape variation tolerance +55%); the feature alignment mechanism ensures that the spatial correspondence accuracy before and after enhancement reaches 98.7%.

[0118] The single-channel heat map is input into the spatiotemporal joint modeling module. This method deploys a triple bidirectional LSTM network architecture with 128 hidden units to construct a temporal analysis component. After the input features undergo three-dimensional adaptive average pooling to compress the spatial dimension, they are expanded along the time axis into a 64-dimensional feature sequence. This system captures the temporal evolution of pathological features through a forward and reverse dual-path information flow. A cascaded network structure is employed to enhance the modeling capabilities of complex dynamic patterns, ultimately concatenating the bidirectional hidden states to form a 256-dimensional spatiotemporal fusion feature.

[0119] The bidirectional LSTM structure can model pathological evolution up to 128 time steps (the memory decay rate is reduced to 0.03 / step); the 64-dimensional feature sequence encoding efficiency reaches 92%, which is 37% better than the traditional RNN structure.

[0120] The proposed method uses spatiotemporal fusion features as input to a multi-level decision classification module to construct a deep separable fully connected network. The first layer fuses high-order spatiotemporal features via a 512-dimensional linear transformation, employing a random neuron dropout strategy with a 50% probability to prevent overfitting. The second layer compresses the features to 128 dimensions to generate diagnostic semantic encodings. The final output layer uses a linear transformation to map the features to a two-dimensional decision space. Reinforced Luminance (ReLU) activation units are embedded between layers to enhance the ability to delineate nonlinear decision boundaries, and parameterized batch normalization is used to maintain training stability.

[0121] Intelligent dimensionality reduction from 3D to 2D retains 96% of effective spatiotemporal information; high-order feature fusion improves the accuracy of spatiotemporal correlation modeling to 0.91F1-score.

[0122] During the forward computation, the algorithm sequentially performs a chain process: 3D convolutional feature extraction → attention feature weighting → spatiotemporal dimension compression → LSTM temporal modeling → global feature pooling → classification decision making. Adaptive pooling unifies the dynamic spatial dimensions to a fixed size, ensuring the network's compatibility with medical image inputs of varying resolutions.

[0123] The semantic encoding layer achieves four-level abstraction of diagnostic knowledge (pixel level → region level → spatiotemporal level → semantic level); the decision space mapping error is reduced by 78% compared with single-level classification; multi-level probability fusion increases the diagnostic confidence to 98.2%±0.7%, the anomaly detection sensitivity reaches 99.3%, and the specificity is 91.5%.

[0124] An adaptive moment estimation optimization algorithm is used, with an initial learning rate of 0.001, for parameter updates. During training iterations, the cross-entropy loss between the predicted results and the true labels is calculated through forward propagation. Automatic differentiation techniques are used to backpropagate gradient information, and a gradient clipping strategy is applied to control the magnitude of parameter updates. This results in an end-to-end imaging diagnostic optimization model, namely the trained 3D-CNN + LSTM hybrid model (including a spatial attention mechanism). The code for step 2.3 is shown in Figure 8. In terms of computational performance, the 3D convolutional optimization algorithm achieves inference speeds of up to 24 fps (1080p data) and reduces memory usage by 62% compared to traditional 3D CNNs. In terms of clinical applicability, it supports dynamic image sequence analysis (CT / MRI time series), and diagnostic code visualization achieves 95% physician interpretability. The number of channels can be expanded to 256 (with a 2.1% improvement in accuracy and only a 15% increase in resource consumption), and continuous transfer learning is supported (adapting a pre-trained model to a new disease requires only 23% of the sample size), providing excellent scalability.

[0125] Extract clinically relevant parameters (bone density, trabecular thickness) from the output of the 3D-CNN+LSTM hybrid model, extract the attention map from the trained 3D-CNN+LSTM hybrid model, and calculate weighted bone density. Figure 9 , based on the binary image, the trabecular thickness is estimated to obtain more accurate bone parameters of the bone around the implant. The code is as follows Figure 10 shown.

[0126] Step 3: Correct the peri-implant bone image to obtain a corrected peri-implant bone image.

[0127] The present invention configures a DICOM format parsing unit and uses the pydicom library to read the bone image around the implant to be corrected and the standard reference image respectively. The original DICOM data is converted into a 32-bit floating point matrix through the pixel matrix conversion component. The code is as follows Figure 11 As shown, the accuracy of the grayscale value of the medical image is preserved. An anatomical structure consistency verification mechanism is set up to ensure that the reference image and the current image have an anatomical correspondence that can be registered.

[0128] Step 3 specifically includes the following processes:

[0129] Step 3.1: Align the bone image around the implant with the reference image using the SIFT feature matching method. Step 3.1: Deploy the improved SIFT feature matching engine and use the Gaussian difference pyramid to construct multi-scale spatial features. The code is as follows: Figure 12 As shown in the figure, the specific alignment process is:

[0130] Detect key anatomical feature points on the bone image around the implant and the reference image, and generate a 128-dimensional feature description vector;

[0131] The FLANN fast nearest neighbor algorithm is used to match feature descriptors and establish the correspondence between feature points across images.

[0132] The optimal affine transformation matrix is estimated through the RANSAC robust algorithm to eliminate the interference of abnormal matching pairs;

[0133] A bilinear interpolation algorithm is used to perform image space transformation and achieve sub-pixel anatomical structure alignment.

[0134] Step 3.2: Construct a pixel-level difference matrix calculation unit and use the OpenCV absolute difference algorithm to calculate the pixel difference between the aligned image and the reference image to generate a residual map. The image difference is quantified using the following formula:

[0135] diff(x,y)=|aligned_img(x,y)-reference_img(x,y)|

[0136] Where (x, y) represents the coordinates of the imaging plane, and the calculation results reflect the area of abnormal bone density changes around the implant.

[0137] Step 3.3: The present invention configures a nonlinear median filter component and applies morphological filtering using a 3×3 pixel window. This filter replaces the central pixel value using a ranking statistical method, effectively eliminating salt-and-pepper noise while maintaining the edge sharpness of the trabecular bone structure. A dynamic threshold mechanism is set to smooth regions where the residual value is less than a threshold (set to 15 HU in the present invention), retaining significant abnormal signals and obtaining the corrected result.

[0138] Step 3.4: Superimpose the correction result onto the peri-implant bone image to obtain a corrected peri-implant bone image. The present invention designs a weighted superposition algorithm to achieve image optimization and sets the verified optimal blending coefficient:

[0139] final_image=0.8×current_img+0.2×corrected_diff

[0140] Linear fusion is achieved through the cv2.addWeighted function, enhancing the contrast of areas of abnormal bone resorption while maintaining the original anatomical structure. Gamma correction is specifically set for the bone-implant interface to improve the visualization of micron-scale bone defects.

[0141] Step 4: Lock the implant based on the improved YOLOv8, generate three-section views according to the center coordinates of the implant, generate three-section dynamic videos based on the three-section views, and fuse the three-section dynamic videos with the bone parameters to obtain fused data.

[0142] The present invention configures a DICOM sequence loading unit and uses the pydicom library to parse multi-slice image files (such as slice1.dcm, etc.). The continuous slices are integrated into a three-dimensional matrix along the Z axis through the pixel matrix stacking component, where the X and Y axes represent the plane space coordinates and the Z axis represents the slice sequence dimension. The np.stack function is used to achieve precise alignment of 64-bit floating-point data and construct a stereo image data cube with dimensions [Height, Width, Depth]. The code is as follows: Figure 13 shown.

[0143] Deploy the improved YOLOv8 object detection engine and load the pre-trained weight file (implant_detection_yolov8.pt) to achieve submillimeter positioning of metal implants.

[0144] The process of locking the implant based on the improved YOLOv8 and generating three cross-sectional views according to the center coordinates of the implant includes:

[0145] Step 4.1.1: Perform forward reasoning on the coronal middle slice ([:,:,1]); parse the predicted box coordinates (x1, y1, x2, y2) and calculate the center coordinates of the implant (center_x, center_y); use bilinear interpolation to map the center coordinates of the implant to the original image resolution; set the confidence threshold ≥ 0.95 to ensure positioning reliability, and use an anomaly detection mechanism to filter artifacts. The code for step 4.1.1 is as follows: Figure 14 shown.

[0146] Step 4.1.2: Construct a multi-planar reconstruction component based on the principles of spatial analytical geometry and generate three cross-sectional views according to the implant center coordinates: Figure 16 , extract YZ cross-sectional data along the X = center_x plane (sagittal = image_sequence[x,:,:]) to obtain sagittal slices, extract XZ cross-sectional data along the Y = center_y plane (coronal = image_sequence[:,y,:]) to obtain coronal slices, fix the Z-axis middle layer (axial = image_sequence[:,:,1]) to obtain transverse slices. Each cross-sectional view is stored in a 512×512 pixel matrix, retaining 12-bit grayscale resolution. The code for step 4.1.2 is as follows Figure 15 shown.

[0147] The process code for generating three-section dynamic video based on three-section views is as follows Figure 17 As shown, specifically including:

[0148] Step 4.2.1. Configure the Matplotlib graphics engine, create a 1×3 subplot layout (figsize = 15×5 inches), and map the sagittal, coronal, and transverse grayscale images (cmap = "gray").

[0149] Step 4.2.2: Save the serial slice images (frame_0001.png, etc.) in PNG format with a resolution of 1920 × 1080 pixels and a bit depth of 24 bits.

[0150] Step 4.2.3. Call the FFmpeg tool chain, set the H.264 encoding standard (libx264), the frame rate to 24fps, synthesize the MP4 video (output_video.mp4) in YUV420 pixel format, and generate a three-section dynamic video.

[0151] The present invention realizes the real-time visualization of bone density and texture parameters by designing a dynamic annotation superposition system. The process code for fusing the three-section dynamic video with bone parameters is as follows: Figure 18 As shown, specifically including:

[0152] Step 4.3.1, Text Overlay Engine: Use OpenCV's putText function to embed the diagnostic parameters at the coordinates (10,30) in the upper left corner of the video frame;

[0153] Step 4.3.2, font rendering settings: use Hershey Simplex font, font size 0.7 times, white (RGB255,255,255) 2-image sketch edge;

[0154] Step 4.3.3, batch processing pipeline: traverse all frame images (annotated_frames / frame_0001.png, etc.) and generate annotated video sequences (annotated_output.mp4).

[0155] Step 4.3.4, data formatting: Data formatting is performed on the bone density value and bone texture classification to obtain fused data. In the present invention, the bone density value is rounded to two decimal places ({density:.2f}), and the bone texture classification is displayed (normal / abnormal).

[0156] Step 5: Train the DeepSeek model using the CBCT image as input and the fused data as output to obtain a trained DeepSeek model;

[0157] A large number of preprocessed images of implants in three sections (transverse, coronal, and sagittal planes) and their corresponding annotated data (including implant location, morphology, and peri-implant bone defects) were used to train the DeepSeek model. The model learns from the features in the animated images and gradually establishes a mapping relationship between implant conditions and image features. During training, the model continuously adjusts its weight parameters using optimization methods such as backpropagation to minimize the model's loss function on the training data, thereby improving its accuracy and generalization.

[0158] When evaluating the trained DeepSeek model, evaluation indicators such as accuracy, recall, and F1 value are used to evaluate the trained model on the validation set and test set to understand the model's performance in identifying implant conditions. At the same time, in addition to commonly used indicators such as accuracy, recall, and F1 value, evaluation indicators specifically for implant conditions are introduced. For example, the mean absolute error (MAE) of implant position identification is calculated to measure the deviation between the implant position predicted by the model and the actual position; for the identification of implant morphology, shape similarity indicators (such as the Dice coefficient) are used to evaluate the similarity between the morphology predicted by the model and the actual morphology. Based on the evaluation results, the model is adjusted and optimized, such as adding data enhancement strategies, adjusting the model structure or hyperparameters, etc., to further improve the performance of the model.

[0159] Based on the characteristics of 3D animations and the requirements of the implant recognition task, a suitable DeepSeek model architecture is selected, such as a neural network with a three-dimensional convolutional layer, to better process 3D spatial information.

[0160] At the same time, an attention mechanism module is introduced into the selected neural network architecture with 3D convolutional layers. For example, adding a 3D attention module between 3D convolutional layers enables the model to automatically focus on the implant and its surrounding key areas, enhancing the ability to extract important features.

[0161] When adjusting model parameters, the model's hyperparameters, such as the learning rate, number of layers, and convolution kernel size, are adjusted according to the data characteristics and task objectives to optimize the model's performance. The present invention uses a Bayesian optimization algorithm to adjust the model's hyperparameters. Bayesian optimization constructs a probabilistic model of the objective function and uses historical test results to select the next hyperparameter combination that is most likely to improve performance. Compared with traditional random search or grid search, it can find the optimal hyperparameters more efficiently. During the model training process, the hyperparameters are dynamically adjusted according to the model's training status. For example, when the model converges slowly in the early stages of training, the learning rate is appropriately increased; when the model shows signs of overfitting, the regularization parameters are dynamically adjusted.

[0162] Step 6: Input the CBCT image of the implant to be identified into the trained DeepSeek model to obtain the bone defect situation around the implant.

[0163] Load the trained DeepSeek Codentist model into the application and input the CBCT images requiring implant identification into the system loaded with the model. For applications requiring real-time implant identification, the model's inference process is optimized. Using an asynchronous inference mechanism, the model is preloaded and data prepared in the background. When new CBCT images are input, the model can quickly respond and output recognition results. The model analyzes and processes the input CBCT images, outputting identification results such as implant location, morphology, and the presence of abnormalities for reference by doctors and other relevant personnel.

[0164] The present invention can also present the implant identification results output by the model to doctors or other relevant personnel in an intuitive and visual manner. For example, using 3D visualization software, the implant position, morphology, and abnormal bone defect areas can be marked on the original CBCT image, making the results easier to understand and analyze.

[0165] The present invention also includes an interactive feedback mechanism, establishing a user feedback mechanism where doctors or operators can annotate and provide feedback on the recognition results. This feedback data can be used to further optimize the model, forming a virtuous iterative cycle to continuously improve the performance and practicality of the model.

[0166] This paper uses a 3D-CNN+LSTM hybrid model to analyze CBCT images, extracts key indicators of bone parameters through a feature attention mechanism, integrates SIFT feature registration and median filtering technology, and corrects the bone image around the implant through the residual map, thereby improving image accuracy and enhancing the visualization of bone structure. It combines YOLOv8 implant positioning with multi-section dynamic video generation technology to support simultaneous observation of axial, sagittal, and coronal views. Finally, the training data obtained by fusing the three-section dynamic video with bone parameters is input into the DeepSeek model. Through multimodal deep learning (CBCT image data, bone parameter data, bone defect annotation data, and other types of data), it can accurately locate bone defects around implants and quantitatively analyze the defect area. Figure 19 The accuracy of identifying bone microstructure has increased by 17%, providing clinicians with more precise guidance.

[0167] The intelligent diagnostic system Codentist, built based on the DeepSeek model in the present invention, is compatible with the DICOM standard and supports direct processing of CBCT images. It can complete dynamic monitoring of implants without disassembling the restoration. The time required for a single analysis is shortened from the traditional 40 minutes to 5 minutes, and it can achieve accurate and efficient image analysis, thereby promoting the intelligent development of oral implants.

[0168] Example 2

[0169] This embodiment provides a peri-implant bone defect recognition system based on DeepSeek, the system comprising:

[0170] The data processing module is used to obtain color-annotated images of the implant and surrounding bones based on the CBCT image. The process of obtaining color-annotated images of the implant and surrounding bones based on the CBCT image includes:

[0171] Import the CBCT images into the cancellous bone analysis software CTan, select the region of interest (ROI) of the peri-implant bone on the implant longitudinal section, and save the selected image set as a ROI file;

[0172] Adjust and confirm the ROI image axis, import the ROI image into CTvox software for 3D reconstruction of the bone around the implant to obtain a 3D image;

[0173] Perform spectral analysis on the three-dimensional image to obtain a color-annotated image of the implant and surrounding bone.

[0174] The parsing module is used to input the color-labeled images of the implant and surrounding bone into the trained 3D-CNN+LSTM hybrid model to obtain the bone image and bone parameters around the implant. The trained 3D-CNN+LSTM hybrid model is trained based on the color-labeled images of the implant and surrounding bone. The training process includes:

[0175] Preprocessing the color-annotated image of the implant and surrounding bone to obtain a preprocessed color-annotated image;

[0176] A pre-trained model based on DeepSeek is used to extract features from the pre-processed color-labeled images, and the extracted features are classified and judged to obtain a two-category probability distribution that conforms to the probability axiom.

[0177] The process of extracting features from the preprocessed color-labeled image and classifying the extracted features includes:

[0178] The preprocessed color-labeled image is input into a 3×3 convolution kernel. The convolution output is subjected to an element-by-element nonlinear transformation using a rectified linear function. A 2×2 maximum pooling window is used with a step size of 2 to compress the feature map to 1 / 2 of the input value in both horizontal and vertical directions to obtain the extracted features.

[0179] The extracted feature space is mapped from the dynamic dimension to the 64-dimensional hidden space, and the rectified linear function is used to perform nonlinear transformation on the reduced eigenvector. The 64-dimensional hidden layer features are mapped to the two-dimensional output space through quadratic linear transformation to generate the original logistic regression value representing the disease category. The Softmax function along the feature channel dimension is used to perform nonlinear transformation on the logical value, and the output is a two-category probability distribution that conforms to the probability axiom.

[0180] The preprocessed color-labeled image is input into the 3D-CNN+LSTM hybrid model, and three-dimensional voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification are performed in sequence. Iterative training is performed, and the cross-entropy loss between the predicted results and the true labels is calculated through forward propagation. When the cross-entropy loss is minimized, the trained 3D-CNN+LSTM hybrid model is obtained.

[0181] The process of sequentially performing 3D voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification includes:

[0182] The preprocessed color-annotated image is input into a 16-channel 3×3×3 convolution kernel for 3D spatial feature scanning. 1-pixel padding is used to keep the feature map size unchanged. Each convolution layer is coupled with a batch normalization component and connected to a ReLU activation unit. The feature map is spatially downsampled through a 3D maximum pooling operation with a stride of 2 to obtain a downsampled feature map.

[0183] Compress the downsampled feature map to generate a single-channel heat map aligned with the original feature space;

[0184] The single-channel heat map is compressed by average pooling and expanded into a 64-dimensional feature sequence along the time axis. The temporal evolution of pathological features is captured through forward and reverse dual-path information flow to obtain spatiotemporal fusion features.

[0185] The spatiotemporal fusion features are linearly transformed to fuse high-order spatiotemporal features to obtain output features, which are compressed to generate diagnostic semantic codes, and the diagnostic semantic codes are linearly transformed and mapped to a two-dimensional decision space.

[0186] The correction module is used to correct the peri-implant bone image to obtain a corrected peri-implant bone image. The process of correcting the peri-implant bone image includes:

[0187] The SIFT feature matching method was used to align the bone image around the implant with the reference image;

[0188] The OpenCV absolute difference algorithm is used to calculate the pixel difference between the aligned image and the reference image to generate a residual map;

[0189] Morphological filtering is performed using a 3×3 pixel window. The central pixel value is replaced by a sorting statistical method. The area where the residual value is less than the threshold is smoothed to obtain the correction result.

[0190] The correction result is superimposed on the peri-implant bone image to obtain the corrected peri-implant bone image.

[0191] The positioning and fusion module is used to lock the implant based on the improved YOLOv8, generate three cross-sectional views according to the center coordinates of the implant, generate three cross-sectional dynamic videos based on the three cross-sectional views, and fuse the three cross-sectional dynamic videos with bone parameters to obtain fused data. The process of locking the implant based on the improved YOLOv8 and generating three cross-sectional views according to the center coordinates of the implant includes:

[0192] Perform forward reasoning on the coronal mid-slice; parse the predicted box coordinates and calculate the implant center coordinates (center_x, center_y); use bilinear interpolation to map the implant center coordinates to the original image resolution;

[0193] Extract YZ section data along the X=center_x plane to obtain sagittal slices, extract XZ section data along the Y=center_y plane to obtain coronal slices, fix the Z-axis middle layer to obtain transverse slices.

[0194] The process of generating a three-section dynamic video based on three-section views includes:

[0195] Configure the Matplotlib graphics engine and create a 1×3 subplot layout, mapping the sagittal, coronal, and transverse grayscale images respectively;

[0196] Save serial section images in PNG format with a resolution of 1920 × 1080 pixels and a bit depth of 24 bits;

[0197] Call the FFmpeg tool chain, set the H.264 encoding standard, the frame rate to 24fps, synthesize the MP4 video in YUV420 pixel format, and generate a three-section dynamic video.

[0198] The process of fusing three-section dynamic video with bone parameters includes:

[0199] Use OpenCV's putText function to embed diagnostic parameters at the coordinates of the upper left corner of the video frame;

[0200] Render the fonts, traverse all frame images, generate annotated video sequences, format the bone density values and bone texture classification data, and obtain the fused data.

[0201] A training module is used to train the DeepSeek model using CBCT images as input and fused data as output to obtain a trained DeepSeek model;

[0202] The recognition module is used to input the CBCT image of the implant to be identified into the trained DeepSeek model to obtain the bone defect situation around the implant.

[0203] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A peri-implant bone defect recognition method based on DeepSeek, characterized by: Methods include: Obtain color-annotated images of implants and surrounding bones based on CBCT images; The color-labeled images of the implant and surrounding bone are input into the trained 3D-CNN+LSTM hybrid model to obtain the peri-implant bone image and bone parameters. Correcting the peri-implant bone image to obtain a corrected peri-implant bone image; The implant is locked using the improved YOLOv8, and three cross-sectional views are generated according to the center coordinates of the implant. Three cross-sectional dynamic videos are generated based on the three cross-sectional views, and the three cross-sectional dynamic videos are fused with the bone parameters to obtain the fused data. The DeepSeek model is trained using the CBCT image as input and the fused data as output to obtain the trained DeepSeek model; The CBCT image of the implant to be identified is input into the trained DeepSeek model to obtain the bone defect situation around the implant.

2. The DeepSeek-based peri-implant bone defect identification method according to claim 1, characterized in that: The process of obtaining color-annotated images of implants and surrounding bones based on CBCT images includes: Import the CBCT images into the cancellous bone analysis software CTan, select the region of interest (ROI) of the peri-implant bone on the implant longitudinal section, and save the selected image set as a ROI file; Adjust and confirm the ROI image axis, import the ROI image into CTvox software for 3D reconstruction of the bone around the implant to obtain a 3D image; Perform spectral analysis on the three-dimensional image to obtain a color-annotated image of the implant and surrounding bone.

3. The method for identifying peri-implant bone defects based on DeepSeek according to claim 1, characterized in that: The trained 3D-CNN+LSTM hybrid model is trained based on color-labeled images of implants and surrounding bones. The training process includes: Preprocessing the color-annotated image of the implant and surrounding bone to obtain a preprocessed color-annotated image; A pre-trained model based on DeepSeek is used to extract features from the pre-processed color-labeled images, and the extracted features are classified and judged to obtain a two-category probability distribution that conforms to the probability axiom. The preprocessed color-labeled image is input into the 3D-CNN+LSTM hybrid model, and three-dimensional voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification are performed in sequence. Iterative training is performed, and the cross-entropy loss between the predicted results and the true labels is calculated through forward propagation. When the cross-entropy loss is minimized, the trained 3D-CNN+LSTM hybrid model is obtained.

4. The method for identifying peri-implant bone defects based on DeepSeek according to claim 3, characterized in that: The process of extracting features from the preprocessed color-labeled image and classifying the extracted features includes: The preprocessed color-labeled image is input into a 3×3 convolution kernel. The convolution output is subjected to an element-by-element nonlinear transformation using a rectified linear function. A 2×2 maximum pooling window is used with a step size of 2 to compress the feature map to 1 / 2 of the input value in both horizontal and vertical directions to obtain the extracted features. The extracted feature space is mapped from the dynamic dimension to the 64-dimensional hidden space, and the rectified linear function is used to perform nonlinear transformation on the reduced eigenvector. The 64-dimensional hidden layer features are mapped to the two-dimensional output space through quadratic linear transformation to generate the original logistic regression value representing the disease category. The Softmax function along the feature channel dimension is used to perform nonlinear transformation on the logical value, and the output is a two-category probability distribution that conforms to the probability axiom.

5. The method for identifying peri-implant bone defects based on DeepSeek according to claim 3, characterized in that: The process of sequentially performing 3D voxel feature extraction, adaptive lesion enhancement, spatiotemporal joint modeling, and multi-level decision classification includes: The preprocessed color-annotated image is input into a 16-channel 3×3×3 convolution kernel for 3D spatial feature scanning. 1-pixel padding is used to keep the feature map size unchanged. Each convolution layer is coupled with a batch normalization component and connected to a ReLU activation unit. The feature map is spatially downsampled through a 3D maximum pooling operation with a stride of 2 to obtain a downsampled feature map. Compress the downsampled feature map to generate a single-channel heat map aligned with the original feature space; The single-channel heat map is compressed by average pooling and expanded into a 64-dimensional feature sequence along the time axis. The temporal evolution of pathological features is captured through forward and reverse dual-path information flow to obtain spatiotemporal fusion features. The spatiotemporal fusion features are linearly transformed to fuse high-order spatiotemporal features to obtain output features, which are compressed to generate diagnostic semantic codes, and the diagnostic semantic codes are linearly transformed and mapped to a two-dimensional decision space.

6. The method for identifying peri-implant bone defects based on DeepSeek according to claim 1, characterized in that: The process of correcting the peri-implant bone image includes: The SIFT feature matching method was used to align the bone image around the implant with the reference image; The OpenCV absolute difference algorithm is used to calculate the pixel difference between the aligned image and the reference image to generate a residual map; Morphological filtering is performed using a 3×3 pixel window. The central pixel value is replaced by a sorting statistical method. The area where the residual value is less than the threshold is smoothed to obtain the correction result. The correction result is superimposed on the peri-implant bone image to obtain the corrected peri-implant bone image.

7. The method for identifying peri-implant bone defects based on DeepSeek according to claim 1, characterized in that: The process of locking the implant based on the improved YOLOv8 and generating three cross-sectional views according to the center coordinates of the implant includes: Perform forward reasoning on the coronal mid-slice; parse the predicted box coordinates and calculate the implant center coordinates (center_x, center_y); use bilinear interpolation to map the implant center coordinates to the original image resolution; Extract YZ section data along the X=center_x plane to obtain sagittal slices, extract XZ section data along the Y=center_y plane to obtain coronal slices, fix the Z-axis middle layer to obtain transverse slices.

8. The method for identifying peri-implant bone defects based on DeepSeek according to claim 1, characterized in that: The process of generating a three-section dynamic video based on three-section views includes: Configure the Matplotlib graphics engine and create a 1×3 subplot layout, mapping the sagittal, coronal, and transverse grayscale images respectively; Save serial section images in PNG format with a resolution of 1920 × 1080 pixels and a bit depth of 24 bits; Call the FFmpeg tool chain, set the H.264 encoding standard, the frame rate to 24fps, synthesize the MP4 video in YUV420 pixel format, and generate a three-section dynamic video.

9. The method for identifying peri-implant bone defects based on DeepSeek according to claim 8, characterized in that: The process of fusing three-section dynamic video with bone parameters includes: Use OpenCV's putText function to embed diagnostic parameters at the coordinates of the upper left corner of the video frame; Render the fonts, traverse all frame images, generate annotated video sequences, format the bone density values and bone texture classification data, and obtain the fused data.

10. DeepSeek-based peri-implant bone defect recognition system, characterized by: The system includes: A data processing module is used to obtain color-annotated images of implants and surrounding bones based on CBCT images; The parsing module is used to input the color-labeled images of the implant and surrounding bone into the trained 3D-CNN+LSTM hybrid model to obtain the peri-implant bone image and bone parameters; A correction module, used for correcting the peri-implant bone image to obtain a corrected peri-implant bone image; The positioning and fusion module is used to lock the implant based on the improved YOLOv8, generate three-section views according to the center coordinates of the implant, generate three-section dynamic videos based on the three-section views, and fuse the three-section dynamic videos with bone parameters to obtain fused data; A training module is used to train the DeepSeek model using CBCT images as input and fused data as output to obtain a trained DeepSeek model; The recognition module is used to input the CBCT image of the implant to be identified into the trained DeepSeek model to obtain the bone defect situation around the implant.

Images