Method and system for detecting periimplant mucosa red and swollen area based on deep learning

Abstract

The invention relates to the technical field of oral cavity digital image processing, and provides an implant perimucosa red and swollen area detection method based on deep learning, and the method comprises the steps: S1, collecting three-dimensional model data of an implant site of a patient, and exporting a standardized visual angle rendering screenshot with a visual enhancement effect by using a matched software rendering function; s2, constructing a deep convolutional neural network based on a YOLOv8 architecture, training a model by adopting a transfer learning strategy, and realizing automatic extraction of implant perimucosa red and swollen focus features; s3, through a feature fusion module in the deep convolutional neural network, automatically retrieving suspected red and swollen sites on the feature maps with different resolutions, performing coordinate correction on the candidate region, and generating an accurate detection frame; and S4, automatically executing batch prediction on the test set based on the trained model, and outputting a detection frame and a quantitative index. And automatic identification and spatial positioning of the red and swollen mucosa area around the implant are realized, so that a visual basis is provided for clinical precise probing and diagnosis and remote early warning.

Description

Technical Field

[0001] This invention relates to the technical field of digital image processing in the oral cavity, and in particular to a method and system for detecting peri-implant mucosal redness and swelling areas based on deep learning. Background Technology

[0002] Peri-implant mucosal redness and swelling is one of the earliest and most crucial clinical manifestations of peri-implant disease. Its occurrence is related to various factors such as plaque buildup, occlusal trauma, and host immune response, directly reflecting the health status of the soft and hard tissues surrounding the implant. If not identified and intervened in a timely and accurate manner, it may gradually develop into peri-implantitis, leading to alveolar bone resorption, implant loosening, and ultimately jeopardizing the long-term survival and function of the implant. With the rapid development of digital dental technology, intraoral scanning systems and 3D model reconstruction technology have been widely applied in implant restoration and long-term maintenance procedures. Digital dental models, with their advantages of being intuitive, repeatable, and easy to store, are gradually replacing traditional plaster models and becoming an important carrier for dentists to assess the condition of the peri-implant mucosa. Rendered screenshots, as a visual presentation of digital models, after optimization through lighting and shadow rendering algorithms, provide a clearer visual basis for observing the mucosal condition, promoting digital implant maintenance as the mainstream clinical approach.

[0003] While the application of digital models in current digital implant maintenance processes has significantly improved the convenience of diagnosis and treatment, the assessment of peri-implant mucosal redness and swelling still falls short of the limitations of traditional methods. Physicians must visually assess the mucosal condition based on rendered images generated by software accompanying an intraoral scanning system. This process relies entirely on the physician's subjective judgment of color changes and morphological features in the images, and the results are manually recorded. There is no standardized automated recording process for the standardized storage of assessment data, no effective automated positioning methods to accurately pinpoint the site of redness and swelling, and no quantitative feedback mechanism to objectively characterize key indicators such as the extent and severity of redness and swelling. Consequently, the entire assessment process remains at the level of qualitative observation and manual recording, failing to meet the clinical demands for precision, standardization, and efficiency in diagnosis and treatment.

[0004] Specifically, the existing technology has the following significant drawbacks: (1) Prominent Subjectivity in Localization: Although the rendered image enhances visual contrast through edge enhancement, shadow compensation, and other algorithms, thereby improving the visibility of mucosal redness and swelling to some extent, the boundaries of peri-implant mucosal redness and swelling are inherently ambiguous, and there are natural differences in the clinical experience, observation perspectives, and judgment criteria of different doctors. Some doctors use the deviation of mucosal color from the normal range as the core judgment criterion, while others focus on the morphological changes of mucosal swelling, and still others combine their own diagnostic and treatment experience to make a comprehensive judgment. This leads to obvious subjective ambiguity in the boundary definition of the same red and swollen area among different doctors, ultimately resulting in a lack of uniformity in the determination of the lesion range. This not only affects the consistency of diagnostic and treatment decisions but also brings great trouble to subsequent follow-up visits, comparisons, and efficacy evaluations.

[0005] (2) Insufficient monitoring visibility: With the rapid development of telemedicine, remote monitoring has become an important scenario for peri-implant health maintenance, especially suitable for follow-up of patients in different locations and postoperative rehabilitation guidance. At the same time, in large-sample screening scenarios such as community oral health screening and large-scale epidemiological surveys of peri-implant diseases, the need for dynamic tracking and pattern analysis of mucosal redness and swelling sites is becoming increasingly urgent. However, existing technologies lack automated lesion identification methods, making it impossible to intuitively record the specific location and related characteristics of redness and swelling sites in a standardized manner. It is also difficult to effectively track the evolution of redness and swelling sites at different time points for the same patient or within the same time period for different patients. As a result, doctors cannot accurately obtain information on lesion changes in remote monitoring, and it is difficult to form systematic lesion data in large-sample screening, which seriously limits the application depth and effectiveness of related scenarios.

[0006] (3) Low efficiency of clinical prediction: When reviewing digital models, doctors need to visually inspect the mucosal areas around the implants one by one due to the lack of automated detection and guidance mechanisms. This process often takes a lot of time. For patients with multiple implants, a single case assessment may take several minutes or even longer, which significantly reduces the efficiency of clinical diagnosis and treatment. More importantly, the visual inspection of the whole mouth lacks precise guidance for high-risk redness and swelling sites. Doctors find it difficult to quickly locate the areas that need to be focused on, which leads to subsequent exploration and examination being carried out blindly. This not only increases the patient's discomfort during diagnosis and treatment, but also easily leads to the missed diagnosis of potential redness and swelling lesions due to visual fatigue, distraction and other factors, thus delaying the best time for intervention.

[0007] In summary, we need a technical solution that can fully utilize the visual advantages of digital dental model rendering images to achieve automated identification, precise positioning, standardized recording, and quantitative feedback of peri-implant mucosal redness and swelling areas. This solution would address the core issues of subjective positioning, lack of visual monitoring, and low clinical efficiency in existing technologies, providing strong support for the precise diagnosis and treatment, remote maintenance, and large-sample screening of peri-implant diseases. Summary of the Invention

[0008] To address the aforementioned problems, the present invention aims to provide a method and system for detecting peri-implant mucosal redness and swelling areas based on deep learning. It aims to utilize deep learning object detection technology to directly process rendered screenshots of digital dental models with visual enhancement characteristics, achieving automated identification and spatial localization of peri-implant mucosal redness and swelling areas, thereby providing visual evidence for precise clinical probing and remote early warning.

[0009] The above-mentioned objective of this invention is achieved through the following technical solutions: A deep learning-based method for detecting peripapillary mucosal redness and swelling includes the following steps: S1: Perform enhanced screenshot data acquisition and precise annotation. Collect three-dimensional model data of the patient's implantation site through the intraoral scanning system, and use the rendering function of the software accompanying the intraoral scanning system to export standardized view rendering screenshots with visual enhancement effects. S2: Automatic extraction of lesion features driven by transfer learning is performed. A deep convolutional neural network based on the YOLOv8 architecture is constructed, and the model is trained using a transfer learning strategy to achieve automatic extraction of peri-implant mucosal redness and swelling lesion features. S3: Automatically generate detection boxes and perform multi-scale coordinate correction. Through the feature fusion module inside the deep convolutional neural network, suspected redness sites are automatically searched on feature maps of different resolutions, and the coordinates of candidate regions are corrected to generate accurate detection boxes. S4: Performs automated evaluation and multi-dimensional index validation. Based on the trained model, it automatically performs batch predictions on the test set, outputs detection boxes and quantitative indicators, and verifies the model performance by comparing it with the expert gold standard, providing visual guidance for clinical decision-making.

[0010] Further, in step S1, a standardized viewpoint rendering screenshot with visual enhancement effects is exported, specifically as follows: The software accompanying the intraoral scanning system first processes the collected 3D model data of the patient's implantation sites using a light and shadow rendering algorithm, which includes an edge enhancement algorithm and a shadow compensation algorithm. Then, color gain and edge sharpening are used to enhance the mucosal features, and finally, a standardized viewpoint rendering screenshot with color information and texture enhancement effect is exported. The visual contrast of this standardized rendering screenshot is higher than that of the original point cloud data, making the mucosal redness and swelling features more distinguishable at the pixel level.

[0011] Furthermore, in step S1, the standardized viewpoint rendering screenshot can completely preserve the local color shift features of the mucosa, providing a stable and unified input feature space for the subsequent training of the deep learning model. At the same time, its enhanced visual features conform to the extraction logic of visual texture by the target detection algorithm, laying the foundation for accurate labeling of the peri-implant mucosal redness and swelling area.

[0012] Further, in step S2, a deep convolutional neural network based on the YOLOv8 architecture is constructed, and the model is trained using a transfer learning strategy, specifically as follows: First, an end-to-end deep convolutional neural network based on the YOLOv8 architecture is constructed. Then, the backbone network freezing technique is used to lock the first 10 backbone layers of the pre-trained model in the early stage of model training, so that this part of the network maintains the basic texture extraction capability learned on a large-scale general dataset. Only the parameters of the back-end detection head of the deep convolutional neural network are fine-tuned to adapt to the detection needs of peri-implant mucosal redness and swelling lesions.

[0013] Furthermore, in step S2, the automatic extraction of the characteristics of peri-implant mucosal redness and swelling lesions is achieved, specifically as follows: By using a finely tuned deep convolutional neural network, the algorithm accurately identifies the combined features of local color and texture roughness changes and anatomical structures caused by inflammation in the rendered screenshot. These combined features are then converted into coordinate information that can be recognized by a computer, enabling the algorithm to have an expert-level ability to identify peri-implant mucosal redness and swelling lesions. At the same time, the backbone network freezing technology effectively avoids the overfitting problem caused by the medical specialties of implant mucosal images and the relatively limited sample size, significantly improving the robustness of feature extraction.

[0014] Further, in step S3, automatic generation and multi-scale coordinate correction of detection boxes are performed. Through the feature fusion module within the deep convolutional neural network, suspected erythema sites are automatically retrieved on feature maps of different resolutions, and the coordinates of candidate regions are corrected to generate accurate detection boxes. Specifically: First, the feature fusion module inside the deep convolutional neural network comprehensively searches for suspected inflammatory sites on feature maps of different resolutions. Then, a regression algorithm is used to iteratively refine the retrieved candidate regions in multiple rounds. During training, a specific hyperparameter optimization strategy is adopted simultaneously, combined with an early stopping mechanism, to continuously guide the detection box to approach the true boundary of the lesion. This process, through precise tuning of regression parameters, specifically solves the technical problems of blurred boundaries and irregular shapes of inflammatory areas, ultimately enabling the generated detection box to surround the inflammatory lesions in the image with pixel-level precision.

[0015] Furthermore, in step S4, automated evaluation and multi-dimensional index validation are performed. Based on the trained model, batch prediction is automatically executed on the test set, outputting detection boxes and quantitative indicators. The model performance is validated by comparison with expert gold standards, providing visual guidance for clinical decision-making. Specifically: The model automatically performs batch predictions on the test set, simultaneously outputting bounding boxes and quantitative metrics including mean AP50, precision, and recall. Subsequently, the automatically generated bounding boxes are compared with the expert gold standard using Intersection over Union (IoU) calculation, and a confusion matrix is ​​automatically generated, forming a standardized automated metric aggregation report. This report ensures that the model's generalization ability in locating lesion boundaries meets the reliability requirements of clinical medicine. Simultaneously, the bounding boxes serve as visual guides for clinical decision-making, accurately directing probing pressure and probing sites within a digital workflow, achieving standardized, visualized, and automated monitoring of peri-implant health status.

[0016] A deep learning-based system for detecting peri-implant mucosal redness and swelling, used to perform the deep learning-based method for detecting peri-implant mucosal redness and swelling as described above, includes: The enhanced screenshot acquisition and annotation module is used for enhanced screenshot data acquisition and precise annotation. It acquires three-dimensional model data of the patient's implantation site through the intraoral scanning system and uses the rendering function of the software accompanying the intraoral scanning system to export standardized view rendering screenshots with visual enhancement effects. The transfer learning lesion feature extraction module is used to automatically extract lesion features driven by transfer learning. It constructs a deep convolutional neural network based on the YOLOv8 architecture, trains the model using a transfer learning strategy, and realizes the automatic extraction of lesion features of peri-implant mucosal redness and swelling. The detection box generation coordinate correction module is used to automatically generate detection boxes and correct their coordinates at multiple scales. Through the feature fusion module inside the deep convolutional neural network, it automatically searches for suspected redness sites on feature maps of different resolutions, corrects the coordinates of candidate regions, and generates accurate detection boxes. The automated evaluation index verification module is used for automated evaluation and multi-dimensional index verification. Based on the trained model, it automatically performs batch prediction on the test set, outputs detection boxes and quantitative indicators, and verifies the model performance by comparing it with the expert gold standard, providing visual guidance for clinical decision-making.

[0017] A computer device, characterized in that it includes a memory and one or more processors, wherein the memory stores computer code, and when the computer code is executed by the one or more processors, causes the one or more processors to perform the method as described above.

[0018] A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer code, which, when executed, is performed as described above.

[0019] This invention, by integrating core technical solutions and key technical features, and through the synergistic effect of its main inventive points, secondary inventive points, and essential features, brings about multi-dimensional and groundbreaking beneficial effects, as detailed below: (1) Achieving automated and standardized positioning, eliminating subjective bias: This invention takes "using a digital dental model rendering screenshot with visual enhancement as input" as its core invention point, combined with necessary features such as "target detection model, automatic extraction of inflammatory features, automatic generation and positioning of detection boxes", which completely changes the existing technology's reliance on manual observation and annotation. By automating the detection process to automatically select lesion sites, it not only avoids the subjective ambiguity of different doctors' judgment of the redness and swelling boundary, but also eliminates the non-systematic bias of manual delineation of boundaries, making the definition of the lesion range standardized and unified, providing an objective and consistent positioning basis for follow-up comparison and efficacy evaluation, and effectively solving the core defect of subjective positioning in the existing technology.

[0020] (2) Guiding precise diagnosis and treatment, significantly improving the efficiency of clinical and remote maintenance: Based on the invention point of "establishing a detection network specifically for the characteristics of peri-implant mucosal inflammation", the detection network of this invention can automatically identify and select clinically significant inflammatory boundaries from the rendered image. The generated precise detection box can serve as a "visual guide" for clinical decision-making. This detection box can directly guide doctors to conduct targeted probing examinations of high-risk sites, eliminating the need for doctors to spend a lot of time on full-mouth visual examination. This avoids the inefficiency and missed diagnosis problems caused by blind probing, and provides an efficient means of lesion identification for scenarios such as remote monitoring and large-sample screening, significantly improving the efficiency of outpatient diagnosis and treatment and remote maintenance, and perfecting the clinical diagnosis and treatment guidance mechanism.

[0021] (3) High consistency of detection results, ensuring the reliability of clinical applications: Thanks to the input advantages of visually enhanced rendering screenshots (high pixel-level feature discrimination) and the accurate recognition capabilities of the dedicated detection network, the detection results of this invention are highly consistent with the annotations of clinical experts. Clinical validation data fully corroborates this advantage: on an internal dataset containing 200 patients and 265 implants, the mAP50 reaches 0.8968; on an external validation dataset containing 23 patients and 30 implants, the mAP50 is 0.6820. This high consistency ensures the reliability and generalization ability of the model's detection results, fully meeting the stringent requirements of clinical medicine for diagnostic techniques, and laying a solid foundation for the clinical translation and widespread application of the detection results.

[0022] In summary, this invention, through the organic combination of its main inventive points, secondary inventive points, and essential features, achieves automation, standardization, precision, and reliability in the detection of peri-implant mucosal redness and swelling areas. It not only solves the core problems of existing technologies, such as subjective positioning, lack of visual monitoring, and low efficiency, but also provides strong support for precise clinical diagnosis, remote early warning, and large-scale screening, thus propelling the diagnosis and treatment of peri-implant diseases into a new stage of digital and intelligent assistance. Attached Figure Description

[0023] Figure 1 This is a flowchart of the deep learning-based method for detecting periplasmic mucosal redness and swelling areas according to the present invention. Figure 2 This is a diagram illustrating the overall algorithm logic framework of the automatic detection system for peri-implant mucosal redness and swelling areas of the present invention. Figure 3 This diagram illustrates a comparison between the feature map of inflammatory visual features extracted using the YOLO deep learning detection network of this invention, the automatic detection results, and the results manually annotated by clinical experts. Figure 3 (1) A schematic diagram of images, expert annotations, and model prediction results for an internal dataset (200 patients, 265 implants). Figure 3 (2) is a schematic diagram of the precision-recall curve for the internal dataset (mAP@0.5 is 0.897). Figure 3 (3) is a schematic diagram of the confusion curve of the internal dataset. Figure 3 (4) A schematic diagram of images, expert annotations, and model prediction results for the external validation dataset (23 patients, 30 implants). Figure 3 (5) is a schematic diagram of the precision-recall curve for the external validation dataset (mAP@0.5 is 0.682). Figure 3 (6) A schematic diagram of the confusion curve for the external validation dataset. Figure 4 This is a structural diagram of the deep learning-based peri-implant mucosal redness and swelling area detection system of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0026] The basic concept definitions involved in this invention are as follows: (1) Rendered screenshot Rendered screenshots are two-dimensional or three-dimensional model projection images generated by the software accompanying an intraoral scanning system after image optimization processing using specific lighting and shadow rendering algorithms such as edge enhancement and shadow compensation. The core characteristic of these images is that their visual contrast is significantly higher than the original point cloud data. Through algorithmic enhancement and optimization of detailed information in the original scan data, they can more clearly present the color variations, texture features, and anatomical details of the peri-implant mucosa, providing a more discriminative image basis for subsequent lesion identification and analysis. This is a key image format for mucosal condition assessment in oral digital imaging processing.

[0027] (2) Target detection Object detection is a deep learning-based technology whose core function is to automatically identify specific lesions in an input image and accurately pinpoint their spatial location using graphical methods such as bounding boxes. This technology extracts, analyzes, and matches pixel-level features, enabling precise location of target lesions from complex image backgrounds without manual delineation or judgment. It not only automates lesion identification but also visually presents the spatial extent and coordinates of the lesion, providing objective evidence for subsequent quantitative analysis and clinical decision-making. It has core application value in the automated detection of specific lesions such as peri-implant mucosal redness and swelling.

[0028] Why this patent came about: In the routine monitoring of peri-implant disease, the core challenge faced in clinical practice is the difficulty in locating the affected area. (1) Fragmentation of clinical observation: When doctors conduct visual examinations using rendering models from intraoral scanning software, they can detect mucosal congestion, but due to the lack of spatial calibration tools, they cannot accurately record the exact extent and location of redness and swelling, making it difficult to assess the progression or resolution of inflammation through comparison during follow-up visits.

[0029] (2) Non-standardization of manual annotation: In remote consultations or large-scale screenings, different doctors have subjective differences in their understanding of the redness and swelling edges. Manually drawing the lesion boundaries is not only time-consuming and laborious, but also has extremely inconsistent standards, making it impossible to form standardized diagnostic data.

[0030] (3) Lack of diagnostic guidance: Doctors need an automated "navigation" tool for rendering models that patients self-examine or upload remotely. If the system can automatically select high-risk redness and swelling areas, it will greatly assist doctors in making accurate subsequent diagnostic decisions and avoid the inefficiency caused by blind exploration.

[0031] The implementation idea of ​​this patent is as follows: Utilizing the visual advantages of rendered images, a deep learning model is used to achieve automated spatial labeling of lesions. The specific technical logic is as follows: (1) Data source utilization: Explicitly utilize rendered screenshots with visual enhancement effects. These images undergo software color enhancement and edge sharpening, making the mucosal redness and swelling features more distinguishable at the pixel level, which is more in line with the object detection algorithm's logic for extracting visual textures.

[0032] (2) Spatial feature extraction: Construct an end-to-end deep learning object detection network. By training the model to identify the combination features of specific tones, textures and anatomical structures in the rendered image, the algorithm can "identify lesions at a glance" like an expert, and convert them into coordinate information that can be recognized by a computer.

[0033] (3) Clinical closed-loop application: The automatically generated detection box is used as a "visual guide" for clinical decision-making. In the digital workflow, the detection box can guide the probing pressure and site, and realize standardized, visualized and automated monitoring of the peri-implant health status.

[0034] The following is an illustration through specific examples: First Embodiment like Figure 1 and 2 As shown, this embodiment provides a deep learning-based method for detecting peri-implant mucosal redness and swelling areas, including the following steps: S1: Perform enhanced screenshot data acquisition and precise annotation. Collect three-dimensional model data of the patient's implantation site through the intraoral scanning system, and use the rendering function of the software accompanying the intraoral scanning system to export standardized view rendering screenshots with visual enhancement effects.

[0035] In step S1, a standardized viewpoint rendering screenshot with visual enhancement effects is exported, specifically: The software accompanying the intraoral scanning system first processes the collected 3D model data of the patient's implantation sites using a light and shadow rendering algorithm, which includes an edge enhancement algorithm and a shadow compensation algorithm. Then, color gain and edge sharpening are used to enhance the mucosal features, and finally, a standardized viewpoint rendering screenshot with color information and texture enhancement effect is exported. The visual contrast of this standardized rendering screenshot is higher than that of the original point cloud data, making the mucosal redness and swelling features more distinguishable at the pixel level.

[0036] In step S1, the standardized viewpoint rendering screenshot can completely preserve the local color shift features of the mucosa, providing a stable and unified input feature space for the subsequent training of the deep learning model. At the same time, its enhanced visual features conform to the extraction logic of visual texture by the target detection algorithm, laying the foundation for accurate labeling of the peri-implant mucosal redness and swelling area.

[0037] The core of step S1 lies in accurately addressing the source data quality issues in detecting peri-implant mucosal redness and swelling. While the raw intraoral 3D point cloud data can completely record the spatial structure of the implant site, it suffers from defects such as blurred mucosal texture, unclear boundaries between redness and swelling and normal tissue, and insufficient visual contrast when directly used for lesion detection, making it difficult for target detection algorithms to effectively identify. However, the software accompanying the intraoral scanning system has natural compatibility with the raw data format. Its built-in edge enhancement algorithm can accurately enhance the anatomical boundaries between the mucosa and implant, as well as the contour differences between the red and swollen areas and normal mucosa. The shadow compensation algorithm can eliminate local shadow interference generated during the 3D model projection process, preventing shadows from being misjudged as red and swollen lesions. Subsequent color gain processing can specifically amplify the color shift features caused by mucosal redness and swelling (such as the difference between the pink of healthy mucosa and the deep red of the red and swollen areas). Edge sharpening further improves the recognition of pixel-level details. The resulting standardized viewpoint rendering screenshot not only solves the visual defects of the raw data at its root, but also ensures that screenshots from different patients and different implant sites have consistent feature distribution patterns through standardized processing such as fixed viewpoints and unified scaling ratios. Crucially, the screenshot fully preserves the local color shift features of the mucosa, which is the core basis for distinguishing between red and swollen lesions and normal tissue. Its enhanced visual features perfectly match the target detection algorithm's extraction logic for core features such as texture, edge, and color. This not only provides annotators with clear annotation criteria and reduces subjective errors in manual annotation, but also allows subsequent deep learning models to learn efficiently in a stable and unified input feature space, laying a solid data foundation for the accuracy of the entire detection process.

[0038] S2: Automatic extraction of lesion features driven by transfer learning is performed. A deep convolutional neural network based on the YOLOv8 architecture is constructed, and the model is trained using a transfer learning strategy to achieve automatic extraction of peri-implant mucosal redness and swelling lesion features.

[0039] In step S2, a deep convolutional neural network based on the YOLOv8 architecture is constructed, and the model is trained using a transfer learning strategy, specifically as follows: First, an end-to-end deep convolutional neural network based on the YOLOv8 architecture is constructed. Then, the backbone network freezing technique is used to lock the first 10 backbone layers of the pre-trained model in the early stage of model training, so that this part of the network maintains the basic texture extraction capability learned on a large-scale general dataset. Only the parameters of the back-end detection head of the deep convolutional neural network are fine-tuned to adapt to the detection needs of peri-implant mucosal redness and swelling lesions.

[0040] In step S2, the automatic extraction of the characteristics of peri-implant mucosal redness and swelling lesions is achieved, specifically as follows: By using a finely tuned deep convolutional neural network, the algorithm accurately identifies the combined features of local color and texture roughness changes and anatomical structures caused by inflammation in the rendered screenshot. These combined features are then converted into coordinate information that can be recognized by a computer, enabling the algorithm to have an expert-level ability to identify peri-implant mucosal redness and swelling lesions. At the same time, the backbone network freezing technology effectively avoids the overfitting problem caused by the medical specialties of implant mucosal images and the relatively limited sample size, significantly improving the robustness of feature extraction.

[0041] The core logic of step S2 is to accurately match the specific scenario requirements of peri-implant mucosal redness and swelling detection, achieving efficient and robust feature extraction with small sample medical data. The YOLOv8 architecture was chosen as the foundation because of its end-to-end detection efficiency, anchor-free design flexibility, and decoupled head structure accuracy, perfectly meeting the dual requirements of speed and accuracy in medical imaging lesion detection. Its native support for multi-scale feature fusion and rapid inference adapts to the actual needs of batch processing rendered screenshots and quickly outputting detection results in clinical scenarios. The transfer learning strategy essentially addresses the core pain points of "scarce samples and high annotation costs" in the medical imaging field: the basic feature extraction capabilities such as edge and texture learned by the pre-trained model on large-scale general datasets are the universal foundation for identifying peri-implant mucosal anatomical structures and distinguishing tissue boundaries, eliminating the need for training from scratch for medical scenarios and significantly reducing dependence on the number of medical samples.

[0042] The first 10 backbone layers of the pre-trained model were deliberately frozen, a precise design based on the architectural characteristics of YOLOv8 and the specific needs of medical images. The first few layers of the backbone network are responsible for extracting global basic features of the image. These features are universal in general scenarios and oral imaging, and freezing them can prevent the destruction of existing effective features during training. By only fine-tuning the back-end detection head, the model can focus its learning on the specific lesion features of the peri-implant mucosa. These include local color shifts caused by inflammation (such as the color difference between the inflamed area and the healthy mucosa), changes in texture roughness, and spatial correlation features with the anatomical structures around the implant, achieving a precise adaptation of "general basic features + specific features". Meanwhile, implant mucosal images have unique medical characteristics such as blurred lesion boundaries and large individual differences, and it is difficult to expand the labeled samples on a large scale. The backbone network freezing technology effectively avoids the risk of overfitting due to excessive memorization of training sample noise by limiting unnecessary parameter updates. Combined with the detection head's deep learning of specific features, it ultimately achieves expert-level accurate recognition capabilities. This not only ensures the specificity of feature extraction but also improves the model's generalization robustness on data from different patients and implantation sites, laying the core technical foundation for subsequent detection box generation and clinical applications.

[0043] S3: Automatically generate detection boxes and perform multi-scale coordinate correction. Through the feature fusion module inside the deep convolutional neural network, suspected redness sites are automatically searched on feature maps of different resolutions, and the coordinates of candidate regions are corrected to generate accurate detection boxes.

[0044] In this embodiment, step S3 specifically includes: First, the feature fusion module inside the deep convolutional neural network comprehensively searches for suspected inflammatory sites on feature maps of different resolutions. Then, a regression algorithm is used to iteratively refine the retrieved candidate regions in multiple rounds. During training, a specific hyperparameter optimization strategy is adopted simultaneously, combined with an early stopping mechanism, to continuously guide the detection box to approach the true boundary of the lesion. This process, through precise tuning of regression parameters, specifically solves the technical problems of blurred boundaries and irregular shapes of inflammatory areas, ultimately enabling the generated detection box to surround the inflammatory lesions in the image with pixel-level precision.

[0045] The core logic of step S3 is to address the clinical imaging characteristics of peri-implant mucosal lesions—characterized by "blurred boundaries, irregular shapes, and significant scale differences"—by employing a multi-module collaborative approach to achieve precise generation and optimization of detection bounding boxes. First, the feature fusion module within the deep convolutional neural network utilizes a multi-scale feature complementarity strategy. High-resolution feature maps accurately capture the edge details and local textures of small-area lesions, while low-resolution feature maps effectively extract global semantic and anatomical structural association information for large-area lesions. Through cross-scale fusion of bidirectional feature flows, this approach avoids missing small lesions and solves the problem of inaccurate localization of large lesions, achieving comprehensive coverage and accurate retrieval of suspected lesion sites at different scales.

[0046] Secondly, the choice of a regression algorithm for multi-round iterative correction is based on YOLOv8's core advantage of directly transforming object detection into a coordinate regression problem. By directly predicting the coordinate parameters of the detection box, no additional candidate region selection steps are needed, enabling efficient response to the irregular shapes of inflammatory lesions. The multi-round iterative design specifically addresses the problem of blurred boundaries: each iteration dynamically adjusts the coordinate parameters based on the error between the previous round's prediction results and the actual annotations, gradually reducing the deviation between the detection box and the actual boundary of the lesion, ultimately achieving pixel-level fit.

[0047] The simultaneous use of hyperparameter optimization strategies and early stopping mechanisms further ensures the robustness and effectiveness of the correction process: fine-tuning of hyperparameters (such as learning rate and IoU threshold) allows the regression algorithm to converge to the optimal solution faster, avoiding inefficient correction due to improper parameter settings; the early stopping mechanism monitors the performance of the validation set and terminates training when the model reaches its best generalization ability, preventing overfitting of the detection boxes to the noise of the training set due to overtraining, and ensuring stable and accurate output of boundaries on the image data of different patients.

[0048] The entire technical process forms a closed loop of comprehensive retrieval, precise correction, and robust optimization. It not only solves the problem of comprehensive lesion retrieval by utilizing multi-scale feature fusion, but also overcomes the problem of accurate boundary localization through regression iteration correction. Furthermore, it ensures the generalization ability of the model by using hyperparameter optimization and early stopping mechanism. The final generated detection box can not only surround the red and swollen lesion with pixel-level accuracy, but also adapt to the complex morphology and individual differences of lesions in clinical images, providing a reliable spatial localization basis for subsequent clinical guidance.

[0049] S4: Performs automated evaluation and multi-dimensional index validation. Based on the trained model, it automatically performs batch predictions on the test set, outputs detection boxes and quantitative indicators, and verifies the model performance by comparing it with the expert gold standard, providing visual guidance for clinical decision-making.

[0050] In this embodiment, step S4 specifically includes: The model automatically performs batch predictions on the test set, simultaneously outputting bounding boxes and quantitative metrics including mean AP50, precision, and recall. Subsequently, the automatically generated bounding boxes are compared with the expert gold standard using Intersection over Union (IoU) calculation, and a confusion matrix is ​​automatically generated, forming a standardized automated metric aggregation report. This report ensures that the model's generalization ability in locating lesion boundaries meets the reliability requirements of clinical medicine. Simultaneously, the bounding boxes serve as visual guides for clinical decision-making, accurately directing probing pressure and probing sites within a digital workflow, achieving standardized, visualized, and automated monitoring of peri-implant health status.

[0051] The core logic of step S4 is to build a closed loop of quantitative verification and clinical implementation. This ensures the medical reliability of the model's detection through a multi-dimensional indicator system and achieves seamless integration of technical results with clinical workflow. The selection of mean precision (mAP50), precision, and recall as core quantitative indicators is based on the stringent requirements of medical imaging detection. mAP50, as an authoritative indicator in the field of target detection, can simultaneously evaluate the model's accuracy in identifying lesions and the precision of boundary localization. It is particularly suitable for detection scenarios with blurred boundaries of peri-implant mucosal redness and swelling. Its characteristic of using IoU ≥ 50% as the criterion perfectly matches the core clinical need for coarse lesion localization. Precision can minimize false positives (such as misclassifying normal mucosa as redness and swelling), avoiding unnecessary probing that causes discomfort to patients and reducing the waste of clinical resources. Recall can effectively reduce the risk of missed diagnoses, ensuring that potential red and swollen lesions are not missed, which is crucial for early disease intervention.

[0052] The Intersection over Union (IoU) calculation and confusion matrix generation provide a deep breakdown and verification of model performance: IoU directly quantifies the spatial overlap between the detection bounding box and the expert gold standard, intuitively reflecting the accuracy of localization; its value directly determines whether the test results can provide effective reference for clinical practice. The confusion matrix clearly presents the model's classification error patterns, accurately locating whether the model has problems of "oversensitivity" or "underrecognition" through the distribution of true positives, false positives, true negatives, and false negatives, providing a clear direction for model optimization and avoiding the one-sidedness that may result from single-indicator evaluation. The standardized indicator aggregation report generated on this basis integrates scattered quantitative data into a systematic and interpretable evaluation conclusion, meeting the core requirements of "verifiable and reproducible" medical technology, and enabling clinicians to quickly determine whether the model performance meets the diagnostic and treatment standards, solving the problem of the lack of unified quantitative basis in traditional manual evaluation.

[0053] More importantly, this step is not merely a technical verification, but rather a transformation of technical results into clinical value: the automatically generated detection frame serves as a "visual guide" for clinical decision-making, accurately matching the digital diagnosis and treatment process. Its precise spatial positioning directly guides doctors to adjust the probing pressure and lock onto key examination sites, eliminating the need for blind full-mouth visual examination and significantly improving outpatient efficiency and probing accuracy. Simultaneously, standardized test results enable the visual recording and dynamic tracking of the peri-implant health status, facilitating comparison of inflammation changes during follow-up visits and providing a unified lesion labeling basis for telemedicine and large-sample screening. This truly achieves a complete closed loop of "technical verification - clinical application - standardized monitoring," effectively transforming the detection results of the deep learning model into a powerful tool for improving the quality and efficiency of clinical diagnosis and treatment.

[0054] like Figure 3 The figure shows the visualization results of the automated evaluation and multi-dimensional index verification process in step S4. It clearly presents the system verification data of the detection performance of the detection method of the present invention in detecting peri-implant mucosal redness and swelling areas. It not only shows the precision-recall curves and confusion curves corresponding to the internal dataset (200 patients, 265 implants) and the external validation dataset (23 patients, 30 implants), but also marks the mean precision (mAP50), the core quantitative index of model detection, as 0.8968 and 0.6820, respectively, under the two datasets. At the same time, the comparison diagram intuitively presents the overlap between the automatic detection box of the present invention and the results of manual annotation by clinical experts (gold standard), clearly reflecting the high consistency between the model detection results and the expert annotations. The validation data and analysis results presented in this figure jointly prove that the detection method of the present invention has stable generalization ability under different datasets. Its detection accuracy and reliability meet the stringent requirements of clinical medicine for the detection of peri-implant mucosal redness and swelling areas, providing key data support for the clinical application of the method.

[0055] Second Embodiment like Figure 4 As shown, this embodiment provides a deep learning-based peri-implant mucosal redness and swelling region detection system for performing the deep learning-based method for detecting peri-implant mucosal redness and swelling regions as described in the first embodiment, comprising: The enhanced screenshot acquisition and annotation module is used for enhanced screenshot data acquisition and precise annotation. It acquires three-dimensional model data of the patient's implantation site through the intraoral scanning system and uses the rendering function of the software accompanying the intraoral scanning system to export standardized view rendering screenshots with visual enhancement effects. The transfer learning lesion feature extraction module is used to automatically extract lesion features driven by transfer learning. It constructs a deep convolutional neural network based on the YOLOv8 architecture, trains the model using a transfer learning strategy, and realizes the automatic extraction of lesion features of peri-implant mucosal redness and swelling. The detection box generation coordinate correction module is used to automatically generate detection boxes and correct their coordinates at multiple scales. Through the feature fusion module inside the deep convolutional neural network, it automatically searches for suspected redness sites on feature maps of different resolutions, corrects the coordinates of candidate regions, and generates accurate detection boxes. The automated evaluation index verification module is used for automated evaluation and multi-dimensional index verification. Based on the trained model, it automatically performs batch prediction on the test set, outputs detection boxes and quantitative indicators, and verifies the model performance by comparing it with the expert gold standard, providing visual guidance for clinical decision-making.

[0056] A computer-readable storage medium stores computer code that, when executed, performs the methods described above. Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. This program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0057] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

[0058] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0059] It should be noted that the above embodiments can be freely combined as needed. The above description is only a preferred embodiment of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for detecting peripapillary mucosal redness and swelling, characterized in that, Includes the following steps: S1: Perform enhanced screenshot data acquisition and precise annotation. Collect three-dimensional model data of the patient's implantation site through the intraoral scanning system, and use the rendering function of the software accompanying the intraoral scanning system to export standardized view rendering screenshots with visual enhancement effects. S2: Automatic extraction of lesion features driven by transfer learning is performed. A deep convolutional neural network based on the YOLOv8 architecture is constructed, and the model is trained using a transfer learning strategy to achieve automatic extraction of peri-implant mucosal redness and swelling lesion features. S3: Automatically generate detection boxes and perform multi-scale coordinate correction. Through the feature fusion module inside the deep convolutional neural network, suspected redness sites are automatically searched on feature maps of different resolutions, and the coordinates of candidate regions are corrected to generate accurate detection boxes. S4: Performs automated evaluation and multi-dimensional index validation. Based on the trained model, it automatically performs batch predictions on the test set, outputs detection boxes and quantitative indicators, and verifies the model performance by comparing it with the expert gold standard, providing visual guidance for clinical decision-making.

2. The method for detecting peri-implant mucosal redness and swelling area based on deep learning according to claim 1, characterized in that, In step S1, a standardized viewpoint rendering screenshot with visual enhancement effects is exported, specifically: The software accompanying the intraoral scanning system first processes the collected 3D model data of the patient's implantation sites using a light and shadow rendering algorithm, which includes an edge enhancement algorithm and a shadow compensation algorithm. Then, color gain and edge sharpening are used to enhance the mucosal features, and finally, a standardized viewpoint rendering screenshot with color information and texture enhancement effect is exported. The visual contrast of this standardized rendering screenshot is higher than that of the original point cloud data, making the mucosal redness and swelling features more distinguishable at the pixel level.

3. The method for detecting peri-implant mucosal redness and swelling based on deep learning according to claim 1, characterized in that, In step S1, the standardized viewpoint rendering screenshot can completely preserve the local color shift features of the mucosa, providing a stable and unified input feature space for the subsequent training of the deep learning model. At the same time, its enhanced visual features conform to the extraction logic of visual texture by the target detection algorithm, laying the foundation for accurate labeling of the peri-implant mucosal redness and swelling area.

4. The method for detecting peri-implant mucosal redness and swelling area based on deep learning according to claim 1, characterized in that, In step S2, a deep convolutional neural network based on the YOLOv8 architecture is constructed, and the model is trained using a transfer learning strategy, specifically as follows: First, an end-to-end deep convolutional neural network based on the YOLOv8 architecture is constructed. Then, the backbone network freezing technique is used to lock the first 10 backbone layers of the pre-trained model in the early stage of model training, so that this part of the network maintains the basic texture extraction capability learned on a large-scale general dataset. Only the parameters of the back-end detection head of the deep convolutional neural network are fine-tuned to adapt to the detection needs of peri-implant mucosal redness and swelling lesions.

5. The method for detecting peri-implant mucosal redness and swelling area based on deep learning according to claim 1, characterized in that, In step S2, the automatic extraction of the characteristics of peri-implant mucosal redness and swelling lesions is achieved, specifically as follows: By using a finely tuned deep convolutional neural network, the algorithm accurately identifies the combined features of local color and texture roughness changes and anatomical structures caused by inflammation in the rendered screenshot. These combined features are then converted into coordinate information that can be recognized by a computer, enabling the algorithm to have an expert-level ability to identify peri-implant mucosal redness and swelling lesions. At the same time, the backbone network freezing technology effectively avoids the overfitting problem caused by the medical specialties of implant mucosal images and the relatively limited sample size, significantly improving the robustness of feature extraction.

6. The method for detecting peri-implant mucosal redness and swelling area based on deep learning according to claim 1, characterized in that, In step S3, automatic generation and multi-scale coordinate correction of detection boxes are performed. Through the feature fusion module within the deep convolutional neural network, suspected erythema sites are automatically retrieved on feature maps of different resolutions. The coordinates of candidate regions are corrected to generate accurate detection boxes. Specifically: First, the feature fusion module inside the deep convolutional neural network comprehensively searches for suspected redness and swelling sites on feature maps of different resolutions. Then, the regression algorithm is used to perform multiple rounds of iterative correction on the retrieved candidate regions. During the training process, a specific hyperparameter optimization strategy is adopted simultaneously, and combined with an early stopping mechanism, the detection box is continuously guided to approach the true boundary of the lesion. This process addresses the technical challenges of blurred boundaries and irregular shapes in inflammatory areas through precise tuning of regression parameters, ultimately enabling the generated detection box to surround the red and swollen lesions in the image with pixel-level precision.

7. The method for detecting peri-implant mucosal redness and swelling area based on deep learning according to claim 1, characterized in that, In step S4, automated evaluation and multi-dimensional index validation are performed. Based on the trained model, batch prediction is automatically executed on the test set, outputting detection boxes and quantitative indicators. The model performance is validated by comparison with expert gold standards, providing visual guidance for clinical decision-making. Specifically: The model automatically performs batch predictions on the test set, simultaneously outputting bounding boxes and quantitative metrics including mean AP50, precision, and recall. Subsequently, the automatically generated bounding boxes are compared with the expert gold standard using Intersection over Union (IoU) calculation, and a confusion matrix is ​​automatically generated, forming a standardized automated metric aggregation report. This report ensures that the model's generalization ability in locating lesion boundaries meets the reliability requirements of clinical medicine. Simultaneously, the bounding boxes serve as visual guides for clinical decision-making, accurately directing probing pressure and probing sites within a digital workflow, achieving standardized, visualized, and automated monitoring of peri-implant health status.

8. A deep learning-based peri-implant mucosal redness and swelling region detection system for performing the deep learning-based method for detecting peri-implant mucosal redness and swelling regions as described in any one of claims 1-7, characterized in that, include: The enhanced screenshot acquisition and annotation module is used for enhanced screenshot data acquisition and precise annotation. It acquires three-dimensional model data of the patient's implantation site through the intraoral scanning system and uses the rendering function of the software accompanying the intraoral scanning system to export standardized view rendering screenshots with visual enhancement effects. The transfer learning lesion feature extraction module is used to automatically extract lesion features driven by transfer learning. It constructs a deep convolutional neural network based on the YOLOv8 architecture, trains the model using a transfer learning strategy, and realizes the automatic extraction of lesion features of peri-implant mucosal redness and swelling. The detection box generation coordinate correction module is used to automatically generate detection boxes and correct their coordinates at multiple scales. Through the feature fusion module inside the deep convolutional neural network, it automatically searches for suspected redness sites on feature maps of different resolutions, corrects the coordinates of candidate regions, and generates accurate detection boxes. The automated evaluation index verification module is used for automated evaluation and multi-dimensional index verification. Based on the trained model, it automatically performs batch prediction on the test set, outputs detection boxes and quantitative indicators, and verifies the model performance by comparing it with the expert gold standard, providing visual guidance for clinical decision-making.

9. A computer device, characterized in that, The device includes a memory and one or more processors, wherein the memory stores computer code that, when executed by the one or more processors, causes the one or more processors to perform the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer code, and when the computer code is executed, the method as described in any one of claims 1 to 7 is performed.

Images