Dental image auxiliary diagnosis method, system and equipment based on artificial intelligence and medium

Abstract

The invention discloses a dental image auxiliary diagnosis method, system and device based on artificial intelligence and a medium, and relates to the technical field of artificial intelligence medical imagines.The method comprises the following steps that after dental images of different modalities are received, pre-processed images are obtained through targeted optimization guided by anatomy priori; inputting a dynamic multi-task joint segmentation network, and outputting a tooth instance, a key anatomical structure and a lesion-anatomical association segmentation result in parallel; contour correction and conflict resolution are carried out in combination with anatomical prior, it is ensured that a segmentation result conforms to anatomical logic, and finally clinical parameters are extracted and reasoned in combination with clinical rules to generate structured diagnosis suggestions and treatment references; by integrating deep learning-driven image preprocessing, multi-task joint segmentation, cross-modal enhancement and semi-supervised training technologies, precise analysis and structured diagnosis of dental images are realized, and the diagnosis efficiency and practicability are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence medical imaging technology, specifically to artificial intelligence-based dental imaging-assisted diagnostic methods, systems, devices, and media. Background Technology

[0002] Dental imaging serves as a core support for modern oral medicine diagnosis and treatment planning. The panoramic X-ray (OPG), periapical radiographs, and cone-beam computed tomography (CBCT) images it provides are crucial for dentists to accurately identify tooth anatomy and determine the type and severity of lesions. With the deep penetration of artificial intelligence (AI) technology into medical imaging, deep learning-based image processing methods are gradually replacing traditional manual analysis methods, demonstrating high efficiency in tasks such as denoising, object detection, and image segmentation in dental images. This provides a new approach to solving problems such as low clinical diagnostic efficiency and large subjective errors.

[0003] However, existing dental image segmentation technologies mostly adopt a single-task, independent processing mode, that is, segmenting tooth instances, pulp, alveolar bone, and lesion areas separately, without establishing the inherent relationship between the segmentation targets, resulting in a lack of consistency and clinical rationality in the segmentation results. On the one hand, tooth instance segmentation is disconnected from anatomical structure segmentation. For example, some algorithms only focus on tooth contour extraction, but ignore the spatial inclusion relationship between pulp and tooth, and the attachment relationship between alveolar bone and tooth. When encountering scenarios with tilted, overlapping, or partially missing teeth, problems such as misalignment of anatomical structure segmentation (e.g., pulp masking exceeding the tooth range) and blurred boundaries are prone to occur. On the other hand, lesion area segmentation lacks coordination with anatomical structure segmentation, failing to utilize the clinical logic that "lesions must be attached to specific anatomical structures" (e.g., dental caries only occurs in the hard tissue of the tooth, periapical periodontitis surrounds the root tip), leading to missed detection of small lesions and missegmentation of lesion areas (e.g., misclassifying alveolar bone resorption areas as periapical periodontitis).

[0004] Meanwhile, existing segmentation techniques generally lack a comprehensive anatomical constraint mechanism: the feature extraction stage does not fully incorporate the anatomical topology of teeth (such as the arrangement of adjacent teeth and the orientation of the pulp central axis), resulting in poor model adaptability to individual anatomical differences; the post-segmentation processing stage relies only on pixel-level thresholding or simple morphological operations, without addressing conflicts based on oral anatomical patterns (such as overlapping segmentation masks of adjacent teeth and contradictions between the segmentation boundaries of the neural tube and alveolar bone), further reducing segmentation accuracy. In addition, existing segmentation models are mostly trained for specific imaging devices or single-modality data, with fixed model parameters. When applied to other brands of equipment or different modalities of images, the generalization ability drops sharply due to feature distribution shifts, making it difficult to meet the clinical needs of collaborative diagnosis across multiple institutions and devices.

[0005] Furthermore, the segmentation results in existing technologies are merely independent outputs for image analysis, not directly linked to clinical diagnostic needs to form a complete closed loop of segmentation results, clinical parameters, and diagnostic recommendations. On the one hand, existing models lack clinical parameter extraction modules for segmentation results, failing to automatically calculate tooth spacing and tilt angles from tooth instance segmentation results, extract pulp cavity volume and alveolar bone thickness from anatomical structure segmentation results, or quantify key clinical indicators such as lesion area and infiltration depth from lesion segmentation results. This necessitates manual measurement and analysis of segmentation results by doctors, adding to their workload. On the other hand, the lack of a reasoning mechanism based on clinical rules prevents the integration of segmentation results with oral medicine treatment guidelines (such as determining the need for root canal treatment based on caries area and pulp distance, or developing periodontal treatment plans based on alveolar bone resorption). They can only output segmentation masks and cannot directly provide diagnostic recommendations and treatment directions. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, device and medium for dental image-assisted diagnosis based on artificial intelligence, so as to solve the problems of insufficient accuracy and generalization ability of dental image segmentation, high dependence on annotation, weak cross-device adaptability and disconnect between segmentation and diagnosis in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] According to the first aspect of this disclosure, an artificial intelligence-based dental image-assisted diagnostic method is proposed, comprising the following steps:

[0009] S1. Receive the input dental image, perform anatomically prior-guided targeted optimization, including coarsely locating the tooth region based on the tooth anatomy, and adaptively enhancing the contrast of the coarsely located tooth region to obtain the preprocessed image.

[0010] S2. Input the preprocessed image into the dynamic multi-task joint segmentation network. The dynamic multi-task joint segmentation network adopts a shared feature extractor and, based on a dynamic dual attention mechanism, outputs in parallel the segmentation results of tooth instances, the segmentation results of key anatomical structures, and the segmentation results of the association between lesion areas and anatomical structures, thereby obtaining the multi-task joint segmentation results.

[0011] S3. Based on anatomical morphology priors, contour correction and conflict resolution are performed on the multi-task joint segmentation results to obtain a refined segmentation mask;

[0012] S4. Extract clinical parameters based on the refined segmentation mask, and perform reasoning based on a preset clinical rule base to generate structured diagnostic suggestions.

[0013] Furthermore, in step S1, the anatomically prior-guided targeting optimization includes the following steps:

[0014] S101. A lightweight detection model is used to coarsely locate the teeth in the input dental image to obtain a set of bounding boxes for the tooth region.

[0015] S102. Based on the bounding box of the tooth region, the region within the box is enhanced using a joint algorithm that dynamically adjusts the enhancement intensity based on the tooth density within the region, and noise is suppressed on the background outside the box.

[0016] S103. Perform grayscale and resolution normalization processing on images from different imaging devices and modalities.

[0017] Further, in step S2, the dynamic multi-task joint segmentation network is a dynamic dual-attention embedding network, including:

[0018] A feature extraction backbone network is used to extract multi-scale feature maps from the preprocessed image;

[0019] A dynamic dual attention module is used to adaptively adjust channel weights based on the multi-scale feature map and generate a spatial attention mask based on anatomical topological priors.

[0020] A cross-modal feature enhancement unit is used to generate pseudo-cross-modal features and fuse them with the original features when a single modal input is received.

[0021] The multi-task decoding head, after the shared feature extractor, is equipped with an instance segmentation decoding head, an anatomical structure segmentation decoding head, and a lesion association segmentation decoding head, respectively, to perform parallel output of the three tasks and generate tooth instance segmentation results, anatomical structure segmentation results, and lesion association segmentation results.

[0022] Furthermore, the dynamic dual-attention embedding network is trained using a multi-task joint loss function, which is expressed as follows:

[0023]

[0024] in, The loss for tooth instance segmentation is calculated based on Dice loss and generalized cross-union ratio loss, which is used to ensure the integrity and boundary accuracy of individual tooth segmentation; The anatomical structure segmentation loss is calculated based on the Focal-Dice loss and is used to handle sample class imbalance. , , These are the corresponding weighting coefficients, which can be dynamically adjusted according to clinical needs; The anatomy-association loss measures the spatial consistency between the lesion region mask and its corresponding anatomical structure mask, and is expressed as:

[0025]

[0026] in, For the first Segmentation mask for each lesion region, This is a segmentation mask corresponding to the anatomical structure. The number of lesions. The pixel area.

[0027] Furthermore, in step S3, the process of obtaining the refined segmentation mask includes the following steps:

[0028] S301. An active contour model combining anatomical morphology prior parameters is used to smooth and correct the edges of each segmentation mask in the multi-task joint segmentation result.

[0029] S302. Based on preset anatomical logic rules, resolve the boundary conflict between the tooth instance segmentation result and the key anatomical structure segmentation result in the multi-task joint segmentation result.

[0030] S303. Start the local re-segmentation process for regions whose segmentation quality is lower than the preset threshold.

[0031] Furthermore, the training method for the dynamic multi-task joint segmentation network adopts a hybrid training framework combining self-supervised and semi-supervised training, including:

[0032] In the self-supervised training phase, unlabeled dental image data is used to pre-train the model through mask reconstruction and cross-scale feature contrast learning tasks.

[0033] In the semi-supervised fine-tuning stage, the pre-trained model is fine-tuned using a teacher-student framework with a small amount of labeled data and high-confidence pseudo-labels generated by the teacher model on a large amount of unlabeled data.

[0034] Furthermore, image data from different imaging devices and hospitals are added to the training set of the hybrid training framework, and this is achieved through a domain adaptation layer. Dynamically adjusting feature distribution and adaptively aligning cross-domain features reduces the impact of domain offset, as expressed as:

[0035]

[0036] in, The feature mean of the target domain, For adaptive domain offset coefficients, For multi-scale feature maps, Feature map after domain adaptation.

[0037] According to a second aspect of this disclosure, an artificial intelligence-based dental image-assisted diagnostic system is proposed, comprising:

[0038] The data input module is used to receive and parse dental image data from different imaging devices and modalities;

[0039] An image preprocessing module is used to perform anatomically prior-guided targeted optimization processing on the input dental image data to obtain preprocessed dental images.

[0040] The dynamic multi-task segmentation module includes a segmentation network constructed based on a dynamic dual attention mechanism, which is used to perform joint segmentation on the preprocessed dental images, output segmentation results of tooth instances, key anatomical structures and lesion-related regions in parallel, and perform fine-tuning processing on the segmentation results based on anatomical constraints to obtain fine-tuned segmentation results.

[0041] The diagnostic fusion module is used to extract clinical parameters based on the refined segmentation results and generate structured diagnostic annotations according to a preset clinical rule base.

[0042] The results output module is used to visualize the segmentation results and diagnostic annotations, and generate diagnostic reports that can be exported.

[0043] Furthermore, the dynamic multi-task segmentation module also includes:

[0044] The multi-task coordination unit is used to dynamically allocate computing resources and adjust the execution weights of the three sub-tasks—tooth instance segmentation, anatomical structure segmentation, and lesion association segmentation—based on the modality type of the input preprocessed dental image.

[0045] A cross-modal feature enhancement unit is used to generate features that simulate multimodal information when the input is a single image modality, and fuse them with the original features extracted from the input preprocessed dental image.

[0046] According to a third aspect of this disclosure, a computer device is provided, the device comprising:

[0047] One or more processors;

[0048] Storage device for storing one or more programs;

[0049] When the one or more programs are executed by the one or more processors, the one or more processors implement the artificial intelligence-based dental image-assisted diagnosis method mentioned in the first aspect.

[0050] According to a fourth aspect of this disclosure, a storage medium is proposed that stores a computer program thereon, which, when executed by a processor, implements the artificial intelligence-based dental image-assisted diagnostic method mentioned in the first aspect.

[0051] Compared with existing technologies, the artificial intelligence-based dental image-assisted diagnosis method, system, device, and medium provided by this invention, through a dynamic dual-attention mechanism and a multi-task joint segmentation framework, deeply integrates the segmentation of tooth instances, key anatomical structures, and lesion areas, and introduces anatomical constraints for full-process optimization. This effectively solves the problems of insufficient segmentation accuracy and poor generalization ability caused by significant differences in dental equipment and complex tooth morphology. The solution employs a hybrid self-supervised and semi-supervised training strategy, ensuring segmentation accuracy while reducing the model's dependence on professionally labeled data, significantly lowering application costs. Furthermore, through cross-modal feature enhancement and clinical rule reasoning, it can adapt to various imaging devices and generate structured diagnostic reports directly linked to the segmentation results within seconds, improving doctors' diagnostic efficiency and achieving an efficient closed loop from precise image analysis to direct clinical decision support, significantly improving the accuracy, robustness, and clinical applicability of assisted diagnosis. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0053] Figure 1 The overall flowchart of the artificial intelligence-based dental image-assisted diagnosis method provided in the embodiments of the present invention;

[0054] Figure 2 A flowchart of the preprocessing steps provided in an embodiment of the present invention;

[0055] Figure 3 A flowchart of post-processing steps provided for embodiments of the present invention;

[0056] Figure 4 A block diagram of an artificial intelligence-based dental image-assisted diagnostic system provided in an embodiment of the present invention;

[0057] Figure 5 Actual dental images provided for embodiments of the present invention; wherein, Figure 5 a is a local image of the original state of the teeth; Figure 5 b is a visualization of the tooth instance segmentation results, with different color masks used to distinguish each tooth instance and labeled with a number; Figure 5 c is a visualization of the lesion-anatomical correlation diagnosis results, with the lesion area marked by a red mask and associated with the corresponding anatomical structure, and the diagnosis description displayed simultaneously. Detailed Implementation

[0058] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0059] Example 1:

[0060] This invention provides an artificial intelligence-based dental image-assisted diagnostic method, as shown in the appendix. Figure 1 As shown, it includes the following steps:

[0061] S1. Receive the input dental image, perform anatomically prior-guided targeted optimization, including coarsely locating the tooth region based on the tooth anatomy, and adaptively enhancing the contrast of the coarsely located tooth region to obtain the preprocessed image.

[0062] Building upon traditional denoising and contrast enhancement, this method employs anatomically guided targeted optimization for precise region preprocessing, reducing background interference and enhancing segmented target features; as shown in the attached figure. Figure 2 As shown, the specific steps include:

[0063] S101. A lightweight detection model is used to coarsely locate the teeth in the input dental image, obtaining a set of tooth region bounding boxes. The lightweight detection model is a lightweight YOLOv8-tiny model, which quickly locates the approximate regions of all teeth in the image based on the anatomical features of teeth (such as curved contours and size range), and outputs a set of coarsely located bounding boxes. , The number of teeth.

[0064] S102. Based on the bounding box of the tooth region, the region within the box is enhanced using a joint algorithm that dynamically adjusts the enhancement intensity based on the tooth density within the region, while noise suppression is applied to the background outside the box; that is, the enhancement intensity is adaptively adjusted by using a joint enhancement algorithm of "CLAHE + multi-scale Retinex" on the pixels within the coarse localization region. It is based on the distribution of tooth density within the region. Dynamic calculation: For background pixels outside the region, Gaussian blur is used to suppress noise interference.

[0065] S103. For images from different modalities (X-ray, CBCT) and devices, use the modal feature mapping function. The image grayscale range and resolution are standardized into a standard format to eliminate differences across devices.

[0066] S2. Input the preprocessed image into the dynamic multi-task joint segmentation network. The dynamic multi-task joint segmentation network adopts a shared feature extractor and, based on a dynamic dual attention mechanism, outputs in parallel the segmentation results of tooth instances, the segmentation results of key anatomical structures, and the segmentation results of the association between lesion areas and anatomical structures, thereby obtaining the multi-task joint segmentation results.

[0067] Among them, the dynamic multi-task joint segmentation network is a dynamic dual-attention embedding network that enables joint training and inference of three tasks: "tooth instance segmentation - anatomical structure segmentation - lesion association segmentation". It shares feature extractors to improve segmentation efficiency and result correlation, specifically including:

[0068] A feature extraction backbone network is used to extract multi-scale feature maps from the preprocessed image; an improved version of ResNet-50 with an added reversible bottleneck module is employed to enhance the depth of feature extraction and gradient propagation efficiency, outputting multi-scale feature maps. .

[0069] A dynamic dual attention module is used to adaptively adjust channel weights based on the multi-scale feature map and generate a spatial attention mask based on anatomical topological priors; thereby replacing the traditional fixed attention module and adaptively matching the feature distribution of different images; specifically, it includes adaptive channel attention and dynamic spatial attention.

[0070] Adaptive channel attention is based on the channel response variance of the feature map. Dynamically adjusting channel weights to highlight key feature channels and suppress redundant channels is represented as follows:

[0071]

[0072] in, The number of feature channels, These are the initial channel weights. To adaptively adjust the channel weights;

[0073] Dynamic spatial attention is based on the anatomical topology of teeth (such as the positional relationship between adjacent teeth and the central axis position of the dental pulp) to dynamically generate spatial attention masks, enhancing the spatial correlation of the target region, represented as:

[0074]

[0075] in, This is the result of multi-scale feature fusion. This is an anatomical topological prior feature map. For dynamic spatial attention mask, for Activation function It is a two-dimensional convolutional layer.

[0076] A cross-modal feature enhancement unit is used to generate pseudo-cross-modal features and fuse them with the original features when the input is a single modality; even if the input is a single modality (such as only X-ray images), it is enhanced by the modal feature generator. Simulating cross-modal information (such as 3D structural features from CBCT), 2D features are enhanced into "2D + pseudo-3D" fused features, improving segmentation robustness, as represented by: ;in For cross-modal fusion features, This is a channel-level splicing operation. For original 2D modal features, It is a pseudo-3D structural feature.

[0077] The multi-task decoding head, after the shared feature extractor, is equipped with an instance segmentation decoding head, an anatomical structure segmentation decoding head, and a lesion association segmentation decoding head, respectively, to perform parallel output of the three tasks and generate tooth instance segmentation results, anatomical structure segmentation results, and lesion association segmentation results.

[0078] The dynamic dual-attention embedding network is trained using a multi-task joint loss function to balance the training priorities of the three tasks and strengthen the association constraints between lesions and anatomical structures; the multi-task joint loss function is expressed as follows:

[0079]

[0080] in, The loss for tooth instance segmentation is calculated based on Dice loss and generalized cross-union ratio loss, which is used to ensure the integrity and boundary accuracy of individual tooth segmentation; For anatomical structure segmentation loss, based on Focal-Dice loss calculation, it is used to improve the segmentation accuracy of small structures such as dental pulp, alveolar bone, and nerve tube for unbalanced samples. , , These are the corresponding weighting coefficients, which can be dynamically adjusted according to clinical needs; The anatomy-association loss measures the spatial consistency between the lesion region mask and its corresponding anatomical structure mask (e.g., the distance constraint between the caries and the dental pulp), and is expressed as:

[0081]

[0082] in, For the first Segmentation mask for each lesion region, This is a segmentation mask corresponding to the anatomical structure. The number of lesions. The pixel area.

[0083] The segmentation output includes:

[0084] Instance segmentation results: pixel-level mask, FDI number, and segmentation confidence (≥0.95) for each tooth;

[0085] Anatomical structure segmentation results: pixel-level mask and anatomical parameters (such as pulp cavity volume and alveolar bone thickness) of dental pulp, alveolar bone, and nerve canal.

[0086] Lesion association segmentation results: mask of lesion area (dental caries, periapical periodontitis, etc.), lesion type, spatial relationship with corresponding anatomical structure (e.g., "dental caries 2mm from pulp").

[0087] S3. Based on anatomical morphology priors, contour correction and conflict resolution are performed on the multi-task joint segmentation results to obtain a refined segmentation mask; as shown in the appendix. Figure 3 As shown, the specific steps include:

[0088] S301. An active contour model incorporating prior anatomical morphology parameters is used to smooth the edges of each segmentation mask in the multi-task joint segmentation result; that is, based on prior anatomical morphology of the teeth (e.g., the crown of the tooth is arc-shaped, and the root is conical), the Active Contour model is used to smooth the segmentation edges and eliminate jagged artifacts, as shown below:

[0089]

[0090] in, This is the original segmentation mask. These are the a priori parameters for anatomical morphology.

[0091] S302. Based on preset anatomical logic rules, resolve the boundary conflict between the tooth instance segmentation result and the key anatomical structure segmentation result in the multi-task joint segmentation result, and retain the mask region that conforms to the physiological structure.

[0092] S303. By calculating the Dice coefficient and edge matching degree of the segmentation results, a local re-segmentation process is initiated for regions whose segmentation quality is lower than a preset threshold.

[0093] Meanwhile, the training method for the dynamic multi-task joint segmentation network adopts a hybrid training framework combining self-supervised and semi-supervised training, including:

[0094] In the self-supervised training phase, unlabeled dental image data was used for model pre-training through mask reconstruction and cross-scale feature contrast learning tasks. The unlabeled data included:

[0095] Task 1: Tooth mask reconstruction. Randomly occlude 20% of the tooth region and train a model to reconstruct the mask of the occluded region based on contextual information.

[0096] Task 2: Cross-scale feature contrast learning, which compares features at different scales of the same image to enhance the scale consistency of features;

[0097] Pre-training dataset: 50,000 unlabeled dental images (covering different modalities and imaging devices).

[0098] In the semi-supervised fine-tuning stage, the pre-trained model is fine-tuned using a teacher-student framework, utilizing a small amount of labeled data and high-confidence pseudo-labels generated by the teacher model on a large amount of unlabeled data. The small amount of labeled data includes:

[0099] A teacher-student framework is adopted, in which the teacher model is trained based on labeled data, and the student model is trained using pseudo-labels generated by the teacher model (adaptive threshold filtering: only pseudo-labels with confidence ≥ 0.98 are retained);

[0100] Iterative optimization of pseudo-labels: Every 5 epochs of training, the teacher model is updated, pseudo-labels are regenerated, and low-confidence pseudo-labels (<0.95) are removed.

[0101] Labeling data requirements: Only 5% of the labeled data is needed to achieve the same performance.

[0102] Furthermore, image data from different imaging devices and hospitals are added to the training set of the hybrid training framework, and this is achieved through a domain adaptation layer. Dynamically adjusting feature distribution and adaptively aligning cross-domain features reduces the impact of domain offset, as expressed as:

[0103]

[0104] in, The feature mean of the target domain, For adaptive domain offset coefficients, For multi-scale feature maps, Feature map after domain adaptation.

[0105] S4. Extract clinical parameters based on the refined segmentation mask, and perform reasoning based on a preset clinical rule base to generate structured diagnostic suggestions.

[0106] Example 2:

[0107] This invention provides an artificial intelligence-based dental image-assisted diagnostic system, as shown in the attached figure. Figure 4 As shown, it includes:

[0108] The data input module is used to receive and parse dental image data from different imaging devices and modalities;

[0109] An image preprocessing module is used to perform anatomically prior-guided targeted optimization processing on the input dental image data to obtain preprocessed dental images.

[0110] The dynamic multi-task segmentation module includes a segmentation network constructed based on a dynamic dual-attention mechanism, used to jointly segment the preprocessed dental images, and output segmentation results of tooth instances, key anatomical structures, and lesion-related regions in parallel. The segmentation results are then refined based on anatomical constraints to obtain refined segmentation results. The dynamic multi-task segmentation module also includes:

[0111] The multi-task coordination unit is used to dynamically allocate computing resources and adjust the execution weights of the three sub-tasks—tooth instance segmentation, anatomical structure segmentation, and lesion association segmentation—based on the modality type of the input preprocessed dental image.

[0112] A cross-modal feature enhancement unit is used to generate features that simulate multimodal information when the input is a single image modality, and fuse them with the original features extracted from the input preprocessed dental image.

[0113] The diagnostic fusion module is used to extract clinical parameters based on the refined segmentation results and generate structured diagnostic annotations according to a preset clinical rule base.

[0114] The results output module is used to visualize the segmentation results and diagnostic annotations, and generate diagnostic reports that can be exported.

[0115] The system's workflow is as follows:

[0116] ① Data input: Dental images are acquired via device interface or manually imported, and the format parsing unit converts them into a standard format;

[0117] ②Preprocessing: First, the tooth region is blindly denoised using 3D CNN, and then the tooth region is selected by anatomical prior localization units for targeted enhancement and normalization;

[0118] ③ Joint segmentation: The dynamic multi-task segmentation module outputs instance segmentation, anatomical structure segmentation, and lesion association segmentation results, which are then optimized through post-processing;

[0119] ④ Diagnostic fusion: Extract clinical parameters, combine them with the lesion classification results of the detection module, and generate diagnostic suggestions through clinical rule reasoning;

[0120] ⑤ Output Results: Visualizes segmentation masks and diagnostic annotations, and supports report export and data storage.

[0121] Example 3:

[0122] This invention also provides an artificial intelligence-based dental image-assisted diagnostic device, the device comprising:

[0123] One or more processors;

[0124] Storage device for storing one or more programs;

[0125] When one or more programs are executed by one or more processors, the one or more processors implement the artificial intelligence-based dental image-assisted diagnosis method proposed in Embodiment 1.

[0126] Example 4:

[0127] The present invention also provides an artificial intelligence-based dental image-assisted diagnostic medium, on which a computer program is stored, which, when executed by a processor, implements the artificial intelligence-based dental image-assisted diagnostic method proposed in Embodiment 1.

[0128] The computer storage medium of this invention can be any combination of one or more computer-readable media. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. Computer-readable storage media include, but are not limited to: electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROMD), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0129] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of sending, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0130] Program code contained on a computer-readable medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0131] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0132] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A dental image-assisted diagnostic method based on artificial intelligence, characterized in that, Includes the following steps: S1. Receive the input dental image, perform anatomically prior-guided targeted optimization, including coarsely locating the tooth region based on the tooth anatomy, and adaptively enhancing the contrast of the coarsely located tooth region to obtain the preprocessed image. S2. Input the preprocessed image into the dynamic multi-task joint segmentation network. The dynamic multi-task joint segmentation network adopts a shared feature extractor and, based on a dynamic dual attention mechanism, outputs in parallel the segmentation results of tooth instances, the segmentation results of key anatomical structures, and the segmentation results of the association between lesion areas and anatomical structures, thereby obtaining the multi-task joint segmentation results. S3. Based on anatomical morphology priors, contour correction and conflict resolution are performed on the multi-task joint segmentation results to obtain a refined segmentation mask; S4. Extract clinical parameters based on the refined segmentation mask, and perform reasoning based on a preset clinical rule base to generate structured diagnostic suggestions.

2. The artificial intelligence-based dental image-assisted diagnosis method according to claim 1, characterized in that, In step S1, the anatomically prior-guided targeting optimization includes the following steps: S101. A lightweight detection model is used to coarsely locate the teeth in the input dental image to obtain a set of bounding boxes for the tooth region. S102. Based on the bounding box of the tooth region, the region within the box is enhanced using a joint algorithm that dynamically adjusts the enhancement intensity based on the tooth density within the region, and noise is suppressed on the background outside the box. S103. Perform grayscale and resolution normalization processing on images from different imaging devices and modalities.

3. The artificial intelligence-based dental image-assisted diagnosis method according to claim 1, characterized in that, In step S2, the dynamic multi-task joint segmentation network is a dynamic dual-attention embedding network, including: A feature extraction backbone network is used to extract multi-scale feature maps from the preprocessed image; A dynamic dual attention module is used to adaptively adjust channel weights based on the multi-scale feature map and generate a spatial attention mask based on anatomical topological priors. A cross-modal feature enhancement unit is used to generate pseudo-cross-modal features and fuse them with the original features when a single modal input is received. The multi-task decoding head, after the shared feature extractor, is equipped with an instance segmentation decoding head, an anatomical structure segmentation decoding head, and a lesion association segmentation decoding head, respectively, to perform parallel output of the three tasks and generate tooth instance segmentation results, anatomical structure segmentation results, and lesion association segmentation results.

4. The artificial intelligence-based dental image-assisted diagnosis method according to claim 3, characterized in that, The dynamic dual-attention embedding network is trained using a multi-task joint loss function, which is expressed as follows: in, The loss for tooth instance segmentation is calculated based on Dice loss and generalized cross-union ratio loss, which is used to ensure the integrity and boundary accuracy of individual tooth segmentation; The anatomical structure segmentation loss is calculated based on the Focal-Dice loss and is used to handle sample class imbalance. , , These are the corresponding weighting coefficients, which can be dynamically adjusted according to clinical needs; The anatomy-association loss measures the spatial consistency between the lesion region mask and its corresponding anatomical structure mask, and is expressed as: in, For the first Segmentation mask for each lesion region, This is a segmentation mask corresponding to the anatomical structure. The number of lesions. The pixel area.

5. The artificial intelligence-based dental image-assisted diagnostic method according to claim 1, characterized in that, In step S3, the process of obtaining the refined segmentation mask includes the following steps: S301. An active contour model combining anatomical morphology prior parameters is used to smooth and correct the edges of each segmentation mask in the multi-task joint segmentation result. S302. Based on preset anatomical logic rules, resolve the boundary conflict between the tooth instance segmentation result and the key anatomical structure segmentation result in the multi-task joint segmentation result. S303. Start the local re-segmentation process for regions whose segmentation quality is lower than the preset threshold.

6. The artificial intelligence-based dental image-assisted diagnosis method according to claim 1, characterized in that, The training method for the dynamic multi-task joint segmentation network adopts a hybrid training framework combining self-supervised and semi-supervised training, including: In the self-supervised training phase, unlabeled dental image data is used to pre-train the model through mask reconstruction and cross-scale feature contrast learning tasks. In the semi-supervised fine-tuning stage, the pre-trained model is fine-tuned using a teacher-student framework with a small amount of labeled data and high-confidence pseudo-labels generated by the teacher model on a large amount of unlabeled data. Furthermore, image data from different imaging devices and hospitals are added to the training set of the hybrid training framework, and this is achieved through a domain adaptation layer. Dynamically adjusting feature distribution and adaptively aligning cross-domain features reduces the impact of domain offset, as expressed as: in, The feature mean of the target domain, For adaptive domain offset coefficients, For multi-scale feature maps, Feature map after domain adaptation.

7. An artificial intelligence-based dental image-assisted diagnostic system, employing the dental image-assisted diagnostic method as described in any one of claims 1-6, characterized in that, include: The data input module is used to receive and parse dental image data from different imaging devices and modalities; An image preprocessing module is used to perform anatomically prior-guided targeted optimization processing on the input dental image data to obtain preprocessed dental images. The dynamic multi-task segmentation module includes a segmentation network constructed based on a dynamic dual attention mechanism, which is used to perform joint segmentation on the preprocessed dental images, output segmentation results of tooth instances, key anatomical structures and lesion-related regions in parallel, and perform fine-tuning processing on the segmentation results based on anatomical constraints to obtain fine-tuned segmentation results. The diagnostic fusion module is used to extract clinical parameters based on the refined segmentation results and generate structured diagnostic annotations according to a preset clinical rule base. The results output module is used to visualize the segmentation results and diagnostic annotations, and generate diagnostic reports that can be exported.

8. The artificial intelligence-based dental image-assisted diagnostic system according to claim 7, characterized in that, The dynamic multi-task segmentation module also includes: The multi-task coordination unit is used to dynamically allocate computing resources and adjust the execution weights of the three sub-tasks—tooth instance segmentation, anatomical structure segmentation, and lesion association segmentation—based on the modality type of the input preprocessed dental image. A cross-modal feature enhancement unit is used to generate features that simulate multimodal information when the input is a single image modality, and fuse them with the original features extracted from the input preprocessed dental image.

9. A computer device, characterized in that, The device includes: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the artificial intelligence-based dental image-assisted diagnosis method as described in any one of claims 1-6.

10. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the artificial intelligence-based dental image-assisted diagnostic method as described in any one of claims 1-6.

Images