A method and system for assessing the risk of oral disease

Abstract

The application relates to the technical field of medical image processing, and discloses an oral disease risk assessment method and system. The method comprises the following steps: performing oral cavity basic region positioning processing on a target intraoral image to obtain a region distribution result; extracting image quality characterization data within the range defined by the region distribution result to generate a region quality confidence map and an available region result; performing region differentiation image correction processing based on the image quality characterization data, the region quality confidence map and the available region result to obtain a structure fidelity enhanced image; and finally extracting credible risk feature data based on the structure fidelity enhanced image, the region quality confidence map and the available region result and outputting a screening risk result. The method can effectively reduce the interference of highlights, shadows, light and moderate detail attenuation and low-quality regions on risk assessment, avoid false abnormalities from participating in risk judgment, and improve the credibility, stability and pertinence of the screening risk result.

Description

Technical Field

[0001] This application relates to the field of medical image processing technology, and in particular to a method and system for assessing the risk of oral diseases. Background Technology

[0002] With the increasing application scenarios such as community oral health screening, initial screening at primary care clinics, remote follow-up visits, and home visits, oral disease screening methods based on intraoral images are gradually being adopted. Current applications typically involve observing and analyzing intraoral photos taken with a mobile phone or images captured by an intraoral camera to determine the tooth surface, gums, or oral mucosa, and outputting corresponding screening risk results. This method is convenient to use, low-cost, and widely applicable, and can meet the oral health screening needs in non-specialist settings to a certain extent, thus possessing strong practical application value.

[0003] However, in actual data acquisition, the oral cavity is a moist, confined, and highly reflective imaging environment. Intraoral images are easily affected by factors such as saliva highlight reflection, local shadow occlusion, changes in shooting angle, fluctuations in shooting distance, focus drift, and differences in mouth opening posture, resulting in issues like local overexposure, blurred edges, color distortion, and inconsistent field of view. Existing technologies typically employ uniform preprocessing methods such as brightness adjustment, white balance correction, image sharpening, or noise reduction, or directly input intraoral images into classification, detection, or segmentation models for screening risk assessment. However, most of these methods lack differentiated image quality analysis mechanisms for local highlights, shadows, blur, and occlusion, and also lack the ability to determine the usability of each region. This makes the subsequent screening risk assessment process susceptible to interference from low-quality areas, leading to false positives, false negatives, and significant fluctuations in screening results for the same individual over time. Summary of the Invention

[0004] This application proposes a method and system for assessing oral disease risk to address the problems mentioned in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: a method for assessing the risk of oral diseases, characterized by comprising the following steps:

[0006] Step S1: Obtain the target intraoral image, perform basic oral cavity region localization processing on the target intraoral image, and obtain the region distribution result;

[0007] Step S2: Extract image quality characterization data based on the regional distribution results, and generate regional quality confidence maps and available region results based on the image quality characterization data;

[0008] Step S3: Based on the image quality characterization data, the region quality confidence map, and the available region result, perform region-differentiated image correction processing on the target intraoral image to obtain a structure-fidelity enhanced image;

[0009] Step S4: Extract credible risk feature data based on the structure-fidelity enhanced image, the region quality confidence map, and the available region results, and output the screening risk results based on the credible risk feature data.

[0010] Furthermore, in step S1, the basic oral cavity region localization process includes:

[0011] The intraoral image of the target is normalized to obtain a normalized intraoral image;

[0012] The normalized intraoral image is input into the region localization model to obtain the initial region probabilities corresponding to the tooth surface region, gingival region and mucosal region, respectively.

[0013] Based on the relative positional relationships, adjacency relationships, and mutual exclusion relationships among the tooth surface region, the gingival region, and the mucosal region, a topological consistency correction process is performed on the initial region probability to obtain a corrected region label map;

[0014] Based on the corrected region label map, corrected region masks are generated corresponding to the tooth surface region, the gingival region, and the mucosal region, respectively;

[0015] The regional distribution results include the corrected regional mask corresponding to the tooth surface region, the corrected regional mask corresponding to the gingival region, and the corrected regional mask corresponding to the mucosal region.

[0016] Furthermore, in step S2, the image quality characterization data includes:

[0017] The local brightness shift features, highlight intensity features, edge sharpness features, texture continuity features, and occlusion indication features are extracted within the range defined by the corrected region mask corresponding to the tooth surface region, the corrected region mask corresponding to the gingival region, and the corrected region mask corresponding to the mucosa region, respectively.

[0018] The local brightness shift feature, the highlight intensity feature, the edge sharpness feature, the texture continuity feature, and the occlusion indication feature are normalized to obtain the image quality characterization data.

[0019] Furthermore, in step S2, the generation of the regional quality confidence map and the available region results includes:

[0020] Based on the local brightness shift feature, the highlight intensity feature, the edge sharpness feature, the texture continuity feature, and the occlusion indication feature, the region quality confidence value corresponding to each pixel in the tooth surface region, the gingival region, and the mucosal region is calculated respectively.

[0021] The region quality confidence map is generated based on the region quality confidence values ​​corresponding to each pixel in the tooth surface region, the gingival region, and the mucosal region.

[0022] Based on the regional quality confidence map and the preset availability determination conditions, the available region indication values ​​corresponding to the tooth surface region, the gingival region and the mucosa region are determined respectively;

[0023] The available area results include the available area indication value corresponding to the tooth surface area, the available area indication value corresponding to the gingival area, and the available area indication value corresponding to the mucosal area.

[0024] Furthermore, the region quality confidence map generated in step S2 serves as the basis for determining the correction weight in step S3 and the feature weighting weight in step S4.

[0025] Furthermore, the region-differential image correction process includes:

[0026] Based on the image quality characterization data, unsaturated highlight areas, low-frequency shadow areas, and areas with mild to moderate detail attenuation are identified.

[0027] Perform highlight suppression processing on the unsaturated highlight regions to obtain highlight suppression results;

[0028] Illumination compensation processing is performed on the low-frequency shadow region to obtain the illumination compensation result;

[0029] Perform detail compensation processing on the mild to moderate detail attenuation region to obtain the detail compensation result;

[0030] The highlight suppression processing, the illumination compensation processing, and the detail compensation processing are each limited in their execution range based on the available area results.

[0031] Furthermore, the obtained structure-fidelity enhanced image includes:

[0032] Following the order of highlight suppression processing, illumination compensation processing, and detail compensation processing, sequential gating fusion processing is performed on the highlight suppression results, illumination compensation results, and detail compensation results.

[0033] The sequential gating fusion process determines the correction weights corresponding to each processing branch based on the region quality confidence map, and limits the fusion range corresponding to the detail compensation process based on the available region results, thereby obtaining the structure-fidelity enhanced image.

[0034] Furthermore, the extraction of credible risk feature data includes:

[0035] Based on the structural fidelity enhanced image, color abnormality features, texture abnormality features, edge abnormality features, and morphological abnormality features are extracted in the tooth surface region, the gingival region, and the mucosal region, respectively.

[0036] Based on the region quality confidence map and the available region results, weighted processing is performed on the color anomaly features, the texture anomaly features, the edge anomaly features, and the morphological anomaly features to obtain candidate anomaly response values;

[0037] The candidate abnormal response values ​​are subjected to region-level aggregation processing to obtain credible risk feature vectors corresponding to the tooth surface region, the gingival region, and the mucosal region, respectively.

[0038] The credible risk feature data is obtained based on the credible risk feature vector corresponding to the tooth surface region, the credible risk feature vector corresponding to the gingival region, and the credible risk feature vector corresponding to the mucosal region.

[0039] Furthermore, in step S4, the step of outputting the risk screening result based on the credible risk characteristic data includes:

[0040] Calculate the regional quality weights based on the regional quality confidence map;

[0041] Calculate the overall analyzability index based on the available area results;

[0042] When the overall analyzability index is lower than a preset threshold, a repeat shooting prompt message is output as the screening risk result.

[0043] When the overall analyzability index is not lower than the preset threshold, at least one of the risk level and risk area prompt information is output as the screening risk result based on the credible risk feature data and the regional quality weight.

[0044] An oral disease risk assessment system, comprising:

[0045] The region localization module is used to acquire a target intraoral image, perform basic oral region localization processing on the target intraoral image, and obtain the region distribution result;

[0046] The quality analysis module is used to extract image quality characterization data based on the regional distribution results, and generate regional quality confidence maps and available region results based on the image quality characterization data;

[0047] The image correction module is used to perform regional differential image correction processing on the target intraoral image based on the image quality characterization data, the regional quality confidence map, and the available region results, to obtain a structure-fidelity enhanced image;

[0048] The risk assessment module is used to extract credible risk feature data based on the structure-fidelity enhanced image, the regional quality confidence map, and the available region results, and output the screening risk results based on the credible risk feature data.

[0049] The beneficial effects of this invention are as follows:

[0050] By first performing basic oral cavity region localization processing on the target intraoral image to obtain the region distribution results, and then extracting image quality characterization data within the range defined by the region distribution results to generate a region quality confidence map and available region results, and further performing region differential image correction processing based on the image quality characterization data, the region quality confidence map, and the available region results to obtain a structure-fidelity enhanced image, and finally extracting reliable risk feature data based on the structure-fidelity enhanced image, the region quality confidence map, and the available region results to output the screening risk results, this method can effectively reduce the interference of highlights, shadows, mild to moderate detail attenuation, and low-quality regions on risk assessment, avoid false anomalies in risk judgment, and improve the reliability, stability, and relevance of the screening risk results. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort:

[0052] Figure 1 This is a flowchart of the method of the present invention;

[0053] Figure 2 This is a flowchart of step S2 of the present invention;

[0054] Figure 3 This is a system framework diagram of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] like Figure 1 and Figure 2 As shown, this invention discloses a method for assessing the risk of oral diseases, comprising the following steps:

[0058] Step S1: Obtain the target intraoral image, perform basic oral cavity region localization processing on the target intraoral image, and obtain the region distribution result;

[0059] Step S2: Extract image quality characterization data based on the regional distribution results, and generate regional quality confidence maps and available region results based on the image quality characterization data;

[0060] Step S3: Based on the image quality characterization data, regional quality confidence map, and available region results, perform regional differential image correction processing on the target intraoral image to obtain a structure-fidelity enhanced image;

[0061] Step S4: Extract credible risk feature data based on the structure-fidelity enhanced image, regional quality confidence map, and available region results, and output the screening risk results based on the credible risk feature data.

[0062] In this embodiment, step S1 is used to stably distinguish the tooth surface region, gingival region and mucosal region when there are local highlights, shadow occlusion, color shift and boundary blur in the target intraoral image, so as to avoid the chain reaction of region misclassification on the subsequent extraction of image quality characterization data.

[0063] In existing technologies, a common approach is to directly determine the region division results based on the output of the region localization model. This approach mainly relies on pixel appearance features for judgment. When the edges of the tooth surface region are affected by highlights, or when the gingival region and the mucosa region have similar colors under shadow conditions, region boundary drift, region adhesion, or local misclassification can easily occur. To address this issue, this implementation method introduces the relative positional relationships, adjacency relationships, and mutual exclusion relationships between the tooth surface region, gingival region, and mucosa region based on the initial region probability output by the region localization model. Topological consistency correction processing is then performed on the initial region probability to obtain a region distribution result that better conforms to the actual distribution law of the basic oral cavity regions. This topological consistency correction processing is the core technical means used in step S1 to solve the background technical problem.

[0064] In practical implementation, the first step is to acquire a target intraoral image. This image can be obtained using a mobile phone camera, an intraoral camera, or other near-field oral image acquisition devices, as long as it forms a two-dimensional intraoral image including the tooth surface region, gingival region, and mucosal region. After acquiring the target intraoral image, normalization processing is performed on it to obtain a normalized intraoral image. Normalization processing is mainly used to eliminate the influence of differences in acquisition devices, shooting distances, and input sizes on the stability of the region localization model output. This is a conventional preprocessing method supporting step S1. Preferably, the target... The short side of the intraoral image is scaled to 512-1024 pixels while maintaining the original aspect ratio. At the same time, the pixel values ​​of each channel are linearly mapped to the 0-1 range, limiting the short side to the above range. On the one hand, this preserves the local structural information of the tooth surface edge, gingival edge, and mucosal transition zone. On the other hand, it avoids excessive amplification of specular noise and local texture fluctuations due to excessively high input resolution. If the short side is less than 512 pixels, the details of the region boundary are easily lost; if the short side is greater than 1024 pixels, local highlights and noise are more likely to interfere with region localization.

[0065] After obtaining the normalized intraoral image, the normalized intraoral image is input into the region localization model to obtain the initial region probabilities corresponding to the tooth surface region, gingival region, and mucosal region, respectively. The region localization model is used to output the initial probability that each pixel position belongs to the tooth surface region, gingival region, and mucosal region. It can be implemented using existing semantic segmentation network structures, such as encoder-decoder networks, multi-scale feature fusion networks, or other network structures that can output pixel-level region probabilities. Since the region localization model itself is mainly used to support the initial region probabilities of this scheme and is not the core innovation of step S1, this implementation does not further limit the specific network layer structure of the model, only requiring that it can stably output the initial region probabilities of the tooth surface region, gingival region, and mucosal region.

[0066] After obtaining the initial region probabilities, the region category corresponding to the maximum initial region probability is not directly used as the final region division result. Instead, topological consistency correction is performed. This is because, although the region division result obtained based solely on the initial region probabilities can reflect the preliminary assignment trend of local color and texture, it is still easily affected by the following factors: 1. Local highlights on the tooth surface may cause the edge of the tooth surface region to be incorrectly expanded; 2. The gingival region and the mucosa region have similar colors under certain lighting conditions, which can easily lead to confusion near the boundary; 3. The tooth surface region and the mucosa region may have abnormal direct contact in local images due to misclassification, which does not conform to the actual distribution pattern of the basic oral cavity regions. Therefore, this implementation method re-constrains the initial region probabilities through relative positional relationships, adjacency relationships, and mutual exclusion relationships, so that the region determination no longer depends solely on local appearance features, but also satisfies the spatial structural relationship between the basic oral cavity regions.

[0067] Based on the initial region probabilities and relative positional relationships, adjacency relationships, and mutual exclusion relationships mentioned above, the topology consistency correction process can be expressed as:

[0068] ;

[0069] in, Indicates pixel position; This represents the label map of the area to be optimized; This represents the corrected region label map obtained after topology consistency correction processing. Indicates pixel position In the area label map The initial region probability in the corresponding region category; Indicates mutually exclusive conflicting terms; This represents a combination of relative positional relationships and adjacency relationships as constraint terms; The weight parameter represents the conflict term in a mutually exclusive relationship; The weight parameter represents the combined constraint terms of relative positional relationship and adjacency relationship.

[0070] In the above formula, the first term This is used to maintain consistency between the corrected region label map and the initial region probability, that is, to prioritize retaining the region attribution results that have been output with high confidence by the region localization model; the second term Used to suppress regional distributions that do not conform to mutually exclusive relationships; the third item To enhance the distribution of regions that conform to relative positional and adjacency relationships, this formula does not simply perform threshold segmentation on the output of the region localization model, but introduces structural relationship constraints between basic oral regions based on the initial region probabilities, resulting in a more stable corrected region label map.

[0071] Among them, mutually exclusive conflict terms The purpose of this method is to suppress abnormal direct contact between the tooth surface and the mucosa. In the normal distribution of the basic oral cavity regions, the gingival region is usually located between the tooth surface and the mucosa, playing a transitional role. If a large area of ​​the tooth surface and the mucosa are directly adjacent in a local area, it usually means that the region localization model has been misclassified due to interference from highlight edges, shadow boundaries, or color shifts. Therefore, when performing topology consistency correction, a penalty is imposed on the distribution of regions with abnormal direct contact between the tooth surface and the mucosa to reduce the probability of such unreasonable region layouts appearing in the optimization results. Preferably, The value range is 0.5 to 2.0, more preferably 0.8 to 1.5. If the mutual exclusion constraint is too small, it will not be effective in inhibiting abnormal adhesion between the tooth surface and the mucosa; if... If it is too large, it may excessively suppress the local boundary transition area, affecting the preservation of the true boundary position.

[0072] In this embodiment, the joint constraint terms of relative position relationship and adjacency relationship are... The calculation can be achieved as follows: First, determine the approximate boundary position of the tooth surface region based on the initial region probability; then, statistically analyze the consistency between the candidate region and the allowed adjacent regions within the local neighborhood; simultaneously, based on the distance relationship between the candidate region and the boundary of the tooth surface region, determine whether it conforms to the preferred distribution hierarchy of the gingival and mucosal regions. Preferably, the local neighborhood window size is 3×3 to 9×9, more preferably 5×5. If the local neighborhood window is too small, the highlight edges and noise boundaries will cause large fluctuations in the adjacency relationship statistics; if the local neighborhood window is too large, The boundaries of the regions are easily over-smoothed. For relative positional relationships, it is preferable to regard the gingival region as a transitional region close to the boundary of the tooth surface region and the mucosa region as a region located outside the gingival region. The boundary bandwidth used to distinguish the layers of the gingival region and the mucosa region is preferably 0.01 to 0.08 times the length of the short side of the normalized intraoral image, more preferably 0.02 to 0.05 times. When the above range is adopted, it can cover the common transition bandwidth of the gingival region relative to the tooth surface region under different shooting scales, and can also avoid the mucosa region being mistakenly included in the gingival region.

[0073] It should be noted that the innovation in step S1 does not lie in constructing a new region localization model, but rather in: after the region localization model outputs the initial region probability, the relative positional relationships, adjacency relationships, and mutual exclusion relationships between the tooth surface region, gingival region, and mucosal region are uniformly introduced into the basic oral cavity region localization processing. The initial region probability is further optimized using the aforementioned core formula to obtain a corrected region label map. Compared with the existing technology that directly performs region division based on the output of the region localization model, this implementation method can significantly reduce region misclassification caused by highlights and shadows, especially suppressing abnormal direct adhesion between the tooth surface region and the mucosal region, and enhancing the stable recognition ability of the gingival region as a transition region. The resulting region distribution is more suitable as the region basis for the subsequent step S2 to extract image quality characterization data in terms of region boundary integrity and region distribution rationality.

[0074] After obtaining the corrected region labeling image, corrected region masks corresponding to the tooth surface region, gingival region, and mucosa region are generated based on the corrected region labeling image. Specifically, the pixel positions belonging to the tooth surface region in the corrected region labeling image are marked as the corrected region mask corresponding to the tooth surface region, the pixel positions belonging to the gingival region are marked as the corrected region mask corresponding to the gingival region, and the pixel positions belonging to the mucosa region are marked as the corrected region mask corresponding to the mucosa region. The corrected region masks corresponding to the tooth surface region, the gingival region, and the mucosa region are then used to create the corrected region masks. The corrected region masks together constitute the region distribution result. To improve the stability of the region distribution result, connected component filtering and hole filling can be performed on each corrected region mask. This type of processing is a conventional mask processing method, which is only used to remove obviously discrete noise regions and does not change the core technical concept of step S1. Preferably, only connected regions with an area not less than 0.05% to 2% of the total area of ​​the normalized image are retained. If the area of ​​the connected region is too small, it is usually a noise misclassification region. If all of them are retained directly, it will affect the regional statistical stability of the image quality characterization data in the subsequent step S2.

[0075] In this embodiment, based on the regional distribution results obtained in step S1, the usability of the images at various locations within the tooth surface region, gingival region, and mucosal region is quantified. This avoids the situation in the prior art where only a single quality score is given for the entire intraoral image, or where the original image is directly input into the subsequent processing flow without distinguishing the impact of local highlights, local shadows, local blur, and local occlusion on different regions. This results in subsequent regional differential image correction processing and credible risk feature data extraction being interfered with by low-quality regions. Therefore, the core technical means of step S2 is not to simply evaluate whether the image is clear, but to extract local brightness shift features, highlight intensity features, edge sharpness features, texture continuity features, and occlusion indication features within the region based on the regional distribution results. These image quality features are then uniformly mapped to regional quality confidence values, further generating regional quality confidence maps and usable region results, so that subsequent steps S3 and S4 are based on the same quality constraints.

[0076] In practical implementation, the corrected region masks corresponding to the tooth surface region, the gingival region, and the mucosa region are first obtained using the region distribution results output in step S1. For ease of description, the region categories are denoted as follows: ,in Take one of the three regions: the tooth surface area, the gingival area, or the mucosal area; record the pixel position as... ; categorize the regions The corresponding corrected region mask is denoted as When pixel position Located in the region category When the corresponding corrected region is within the mask, ;otherwise, In this way, the extraction range of various image quality features in step S2 is strictly limited to the region corresponding to the regional distribution result, thereby avoiding interference from pixels outside the region or boundary misclassification residues on the calculation of image quality characterization data.

[0077] In this embodiment, local brightness offset features are used to characterize pixel positions. The degree of deviation between the nearby brightness distribution and the overall brightness level of its area is used to identify brightness anomalies caused by local shadows, local overexposure, and uneven lighting, thus providing a basis for subsequent usability quantification. In practice, the image within the target aperture can first be converted into a brightness characterization channel, and then the brightness can be represented by pixel position. Establish a local statistical window centered on the region, calculate the local average brightness, and correlate it with the region category. The reference brightness within the corresponding area is compared to obtain the local brightness shift characteristics. The window side length used to calculate the local average brightness is preferably 5 to 15 pixels, more preferably 9 pixels. If the window is too small, the highlight noise will cause the brightness shift fluctuation to be too large. If the window is too large, the brightness change near the real boundary will be over-smoothed. The regional reference brightness is preferably determined by the median brightness within the region to reduce the influence of abnormal highlights on the reference value.

[0078] Highlight intensity features are used to characterize the degree of specular reflection interference caused by wetted surfaces and reflected light sources in intraoral images. This feature is introduced because highlight areas typically exhibit both abnormally high brightness and distorted color information. If these are not quantified separately, image correction in step S3 may misprocess highlight boundaries, and anomaly feature extraction in step S4 may easily identify false highlight boundaries as abnormal structures. Specifically, candidate highlight locations can be determined first based on the normalized brightness distribution, and then normal bright white areas can be filtered out using saturation information, allowing only locations with significantly higher brightness and significantly lower color saturation to be considered. Candidate locations for highlights are identified, and their local ranges are expanded to obtain highlight intensity characteristics. The brightness threshold of the candidate highlight location is preferably the 85th to 98th percentile of the brightness distribution of the region, more preferably the 90th to 95th percentile. The saturation threshold is preferably 0.05 to 0.20, more preferably 0.08 to 0.15. If the brightness threshold is too low, normal bright areas are easily misidentified as highlights. If the brightness threshold is too high, edge highlights are easily missed. If the saturation threshold is too low, highlight recognition is too conservative. If the saturation threshold is too high, bright areas on normal tooth surfaces may be misidentified as highlights.

[0079] Edge sharpness features are used to characterize pixel location. The sharpness of the surrounding boundary structure is a feature mainly used to identify local defocusing, mild to moderate blurring, or boundary degradation, thus providing a basis for detail compensation processing in the subsequent step S3. In specific implementation, local gradient magnitude, local Laplacian response, or other existing sharpness evaluation methods can be used to quantify edge sharpness. Since the calculation method of edge sharpness feature is a conventional image analysis support method for implementing this scheme, this implementation does not further limit its algorithm form, only requiring that it can stably characterize the sharpness of the local structural boundaries within the region.

[0080] Texture continuity features are used to characterize the spatial continuity and stability of textures within a region. Unlike edge sharpness features, which mainly reflect the sharpness of local boundaries, texture continuity features primarily reflect whether the texture of tissues within a region is continuous and whether there are abrupt breaks. The reason for introducing this feature is that noise, shadow boundaries, reflection breaks, or slight occlusion in intraoral images often destroy texture continuity. However, this type of degradation does not necessarily manifest as completely blurred edges. In specific implementation, the consistency of the main direction of texture, the stability of texture variance, or the similarity between adjacent texture descriptions can be statistically analyzed within a local statistical window to obtain texture continuity features. The window side length used to calculate texture continuity features is preferably 7 to 21 pixels, more preferably 11 to 15 pixels. If the window is too small, local noise will cause excessive fluctuations in texture continuity; if the window is too large, the real texture changes are easily masked.

[0081] Occlusion indication features are used to characterize whether a local area is affected by occlusion from the tongue, lips, instruments, or severe shadows. The reason for introducing this feature is that occluded areas usually do not have stable and recoverable structural information. If they are not identified in advance, the regional differential image correction processing in step S3 may perform unnecessary enhancement on the occluded areas, and the extraction of credible risk feature data in step S4 may also generate false responses. In specific implementation, factors such as local brightness being significantly lower than the regional reference brightness, edge sharpness being significantly reduced, texture continuity being significantly reduced, and contact with the boundary of non-oral areas can be comprehensively considered to form occlusion indication features. Preferably, when a pixel position in the local window simultaneously meets the conditions of significantly reduced brightness, edge sharpness feature being lower than a preset sharpness threshold, and texture continuity feature being lower than a preset continuity threshold, its occlusion indication feature is improved. The continuity threshold is preferably 0.20 to 0.60, more preferably 0.30 to 0.50. This feature helps to distinguish between shadow areas that can be improved by illumination compensation and occluded areas that lack effective information.

[0082] After extracting the above five types of image quality features, normalization is performed on the local brightness shift feature, highlight intensity feature, edge sharpness feature, texture continuity feature, and occlusion indication feature to obtain image quality characterization data. Preferably, the five types of features are mapped to the interval of 0 to 1. Among them, the larger the value of the local brightness shift feature, highlight intensity feature, and occlusion indication feature, the worse the quality. The larger the value of the edge sharpness feature and texture continuity feature, the better the quality. Through unified normalization, the comparability and dimensional consistency of various image quality features in the calculation of regional quality confidence values ​​can be guaranteed, and the abnormal dominance of a certain type of feature with a large value range in the final result can be avoided.

[0083] Based on the above image quality characterization data, in order to uniformly quantify the image usability of each pixel location in the tooth surface region, gingival region, and mucosal region, the region quality confidence value is preferably calculated using the following formula:

[0084] ;

[0085] in, Indicates the region category At pixel position The regional quality confidence value at that location; Indicates local brightness shift characteristics; Indicates the characteristic of highlight intensity; Indicates edge sharpness features; Indicates texture continuity characteristics; Indicates occlusion indication features; This represents the edge sharpness balance parameter; This represents the texture continuity balance parameter; This represents the penalty coefficient for local brightness shift; Indicates the specular intensity penalty coefficient; This indicates the penalty coefficient for occlusion indication.

[0086] In this formula, Used to limit the calculation of regional quality confidence values ​​to regional categories The corresponding corrected region inside the mask; fractional terms and These respectively reflect the positive contributions of edge sharpness features and texture continuity features to the region quality confidence value; that is, the sharper the edges and the more continuous the texture, the more likely the pixel location is to retain true structural information; the exponential term... This reflects the negative impact of local brightness shift features, highlight intensity features, and occlusion indication features on the confidence value of regional quality. That is, the greater the brightness shift, the stronger the highlight, and the heavier the occlusion, the less suitable the corresponding location is as a reliable input for subsequent processing.

[0087] in, The sensitivity of the contribution of edge sharpness features and texture continuity features to the region quality confidence value is preferably adjusted as follows: The value range is 0.05 to 0.40, more preferably 0.10 to 0.25; The value range is 0.05 to 0.40, more preferably 0.10 to 0.25. If the value is too small, slight changes in edge sharpness features or texture continuity features will cause significant fluctuations in the region quality confidence value, which is detrimental to stable use in complex intraorbital scenes; if... If the size is too large, the ability to distinguish between different quality regions will be weakened.

[0088] The penalty intensities used to characterize local brightness shift features, specular intensity features, and occlusion indication features, respectively, are preferably... The value range is 0.5 to 2.5, and more preferably 0.8 to 1.8; The value range is 1.0 to 4.0, and more preferably 1.5 to 3.0; The value range is 0.8 to 3.5, more preferably 1.2 to 2.8. The reason for this is... The preferred upper limit is set higher than This is because, in intraoral images, highlights and occlusions typically cause greater damage to local structural information than general brightness shifts. While local brightness shifts can often be improved through illumination compensation in subsequent step S3, highlights and occlusions are more likely to cause irrecoverable or unusable information loss. Therefore, in preferred implementation, [the following can be taken as an example]. When using this set of parameters, it is possible to better distinguish between normal usable areas and areas with highlights, shadows, and occlusion interference.

[0089] After calculating the regional quality confidence values, the regional quality confidence values ​​corresponding to each pixel location in the tooth surface region, gingival region, and mucosa region are backfilled to their respective regional locations to form a regional quality confidence map. Unlike the prior art, which only gives a single quality score for the entire image, this implementation generates a pixel-level, region-restricted regional quality confidence map, which can clearly indicate which locations in the same target intraoral image are more suitable as reliable inputs for subsequent processing.

[0090] After generating the region quality confidence map, based on the region quality confidence map and preset availability criteria, the available region indicator values ​​for the tooth surface region, gingival region, and mucosal region are determined respectively, thus obtaining the available region results. Specifically, the region quality confidence value of each pixel position can be compared with a preset quality threshold. When the region quality confidence value of a pixel position is not lower than the preset quality threshold, its corresponding available region indicator value is set to 1; when the region quality confidence value is lower than the preset quality threshold, its corresponding available region indicator value is set to 0. Furthermore, to reduce dispersion... The impact of noise points on the usable area results can be addressed by filtering connected regions after pixel-level determination. Only connected regions with an area not less than 0.05% to 2% of the total area of ​​the normalized image within the aperture are retained as the final usable region. The preset quality threshold is preferably 0.45 to 0.75, more preferably 0.55 to 0.65. If the preset quality threshold is too low, highlight edges, shadow boundaries, and slightly occluded areas may still be retained in the usable region results. If the preset quality threshold is too high, some areas that have slight brightness shifts but still retain effective structural information will be excessively removed.

[0091] It should be noted that the regional quality confidence map generated in step S2 is not only used for usability determination within this step, but also serves as the basis for determining the correction weight in step S3 and the feature weighting weight in step S4. In other words, this implementation does not perform quality evaluation independently first, and then perform image correction processing and credible risk feature data extraction independently. Instead, the regional quality confidence map is uniformly passed to steps S3 and S4 as a shared intermediate result of quality constraints across steps, so that regional differentiated image correction processing and credible risk feature data extraction are based on the same quality judgment standard.

[0092] In summary, step S2 extracts local brightness shift features, highlight intensity features, edge sharpness features, texture continuity features, and occlusion indication features within the range defined by the regional distribution results to form image quality characterization data. It further calculates regional quality confidence values, generates regional quality confidence maps and usable region results, thereby achieving unified quantification of the usability of images within the tooth surface region, gingival region, and mucosal region. Thus, step S2 not only provides a correction weight basis for the regional differential image correction processing in step S3, but also provides a feature weighting basis for the extraction of credible risk feature data in step S4, thereby establishing a consistent quality constraint foundation across steps.

[0093] In this embodiment, based on the image quality characterization data, region quality confidence map, and available region results obtained in step S2, differential correction is performed on different types of local quality degradation in the target mouth image. This improves the structural distortion caused by high-light reflection, low-frequency shadows, and mild to moderate detail attenuation, while avoiding the false boundary magnification, shadow noise enhancement, and over-recovery of low-confidence regions caused by uniform enhancement of the entire image in the prior art. In other words, the core technical means of step S3 is not to simply increase the brightness or sharpness of the entire image, but rather: first, based on the image quality characterization data, distinguish unsaturated highlight regions, low-frequency shadow regions, and mild to moderate detail attenuation regions, and then perform highlight suppression processing, illumination compensation processing, and detail compensation processing respectively. Furthermore, the region quality confidence map and available region results are used to perform sequential gated fusion processing on the three processing branches to obtain a structure-fidelity enhanced image.

[0094] In practical implementation, the target intraoral image is first used as the basic input for step S3. Combined with the image quality characterization data, region quality confidence map, and available region results output from step S2, the locations in the target intraoral image where different correction strategies need to be implemented are identified. For ease of description, the target intraoral image is represented in color channels... and pixel position The normalized pixel value at that location is denoted as ,in, Choose one of the three: red channel, green channel, or blue channel; categorize the region as follows: ,in, Take one of the following three regions: the tooth surface region, the gingival region, or the mucosal region; record the region quality confidence value obtained in step S2 as... The available area indication value obtained in step S2 is recorded as In this embodiment, the execution scope of highlight suppression processing, illumination compensation processing and detail compensation processing is based on the image quality characterization data and available area results obtained in step S2, rather than re-judging based on experience without going through step S2.

[0095] Unsaturated highlight regions are used to represent areas that, despite significant specular reflection interference, still retain some color levels and local structural information. In intraoral images, tooth surfaces and moist mucosa areas are prone to highlights. If uniform sharpening or uniform brightness enhancement is directly applied to these regions, the highlight edges are easily magnified into pseudo-abnormal boundaries. Therefore, this implementation first determines unsaturated highlight regions based on the highlight intensity features in the image quality characterization data, and then performs highlight suppression processing only on these regions. Preferably, when the highlight intensity feature of a pixel location is higher than a preset highlight candidate threshold and its brightness has not reached the saturation upper limit, the image is considered unsaturated. The pixel location is determined to be an unsaturated highlight region. The highlight candidate threshold is preferably 0.55-0.90, more preferably 0.65-0.80; the brightness saturation upper limit is preferably 0.92-0.99, more preferably 0.95-0.98. If the highlight candidate threshold is too low, normal bright areas are easily mistakenly included in the highlight suppression processing range. If the highlight candidate threshold is too high, edge highlights and weak highlight regions may be missed. If the brightness saturation upper limit is too low, highlight regions that still retain color levels will be eliminated prematurely. If the brightness saturation upper limit is too high, areas that are close to saturation will still be mistakenly included in the highlight suppression processing range, affecting the stability of the suppression results.

[0096] The purpose of highlight suppression is to reduce the interference of unsaturated highlight areas on subsequent structural analysis, rather than simply reducing the brightness of the area. In practice, within the unsaturated highlight area, the brightness can be bounded back based on the highlight intensity characteristics, while maintaining the color continuity and edge transition between the area and the surrounding areas. Preferably, the local brightness statistics of similar areas around the unsaturated highlight area are used as the backshortage reference to restrict and compress the brightness of the unsaturated highlight area. For positions with higher highlight intensity characteristics and lower regional quality confidence values, the backshortage amplitude is appropriately increased; for positions with lower highlight intensity characteristics or higher regional quality confidence values, the backshortage amplitude is appropriately decreased. The highlight suppression amplitude is preferably 0.20 to 0.80 times the difference between the current brightness and the reference brightness, more preferably 0.35 to 0.60 times. If the highlight suppression amplitude is too small, the highlight pseudo-boundary will still be difficult to eliminate; if the highlight suppression amplitude is too large, the normal bright white texture may be excessively darkened, affecting the true structural expression of the tooth surface and mucosa areas.

[0097] Low-frequency shadow regions are used to represent areas of slow brightness decay caused mainly by insufficient local lighting, shooting angle deviation, or obstruction of the light path inside the mouth. The main problem with these regions is not excessive reflection, but low brightness and limited dynamic range. If the brightness of the entire image is uniformly increased indiscriminately, non-shadow areas will be raised synchronously, which will weaken the contrast between regions. Therefore, this embodiment first determines low-frequency shadow regions based on local brightness shift features in image quality characterization data, combined with edge sharpness features and occlusion indication features. Preferably, when the local brightness shift feature of a certain pixel position is higher than a preset brightness shift threshold, the occlusion indication feature is lower than a preset occlusion threshold, and the brightness change around it shows a slow and continuous trend, it is determined to be a low-frequency shadow region. The brightness shift threshold is preferably 0.35 to 0.75, more preferably 0.45 to 0.60; the occlusion threshold is preferably 0.20 to 0.55, more preferably 0.25 to 0.40. By introducing occlusion indication features, it is possible to avoid misjudging occluded areas that do not have effective information as low-frequency shadow regions that can be improved by lighting compensation.

[0098] The purpose of illumination compensation processing is to enhance the visible structural information of low-frequency shadow areas, rather than simply brightening all low-brightness areas. In practice, the illumination compensation magnitude can be determined based on the difference between the low-frequency shadow area and the reference brightness of its surrounding area, and a bounded increase in brightness is performed on the area. Preferably, the local reference brightness of the area where the low-frequency shadow area is located is used as the target level, and the current brightness is compensated by a factor of 0.15 to 0.70, more preferably by a factor of 0.25 to 0.50. If the illumination compensation magnitude is too small, the local shadow area will still remain obviously dark and the effective structural information cannot be recovered. If the illumination compensation magnitude is too large, the noise and color deviation in the shadow area will be amplified simultaneously, and may form an unnatural brightness discontinuity with the surrounding area. In order to reduce the abrupt change of the compensation boundary, it is preferable to set a transition band with a width of 3 to 11 pixels at the boundary of the low-frequency shadow area, and perform gradually attenuating illumination compensation within the transition band.

[0099] Mild to moderate detail attenuation regions are used to represent areas where edge sharpness features are reduced but certain local texture contours are still retained. Unlike severely blurred or severely occluded areas, mild to moderate detail attenuation regions usually still possess recoverable texture and boundary information, making them suitable for detail compensation processing. If the entire image is sharpened uniformly without distinguishing region quality, while mild to moderate detail attenuation regions will be enhanced, highlight edges, shadow edges, and occluded edges will also be magnified simultaneously, resulting in a large number of false details. Therefore, this implementation first determines mild to moderate detail attenuation regions based on edge sharpness features and texture continuity features in the image quality characterization data, combined with the available region results. Preferably, when a certain pixel... When the edge sharpness feature of a location is lower than the preset upper limit of sharpness but higher than the preset lower limit of sharpness, the texture continuity feature is not lower than the preset continuity threshold, and its usable area indicator value is 1, it is determined to be a region with mild to moderate detail attenuation. The lower limit of sharpness is preferably 0.15 to 0.40, more preferably 0.20 to 0.30; the upper limit of sharpness is preferably 0.40 to 0.75, more preferably 0.50 to 0.65; and the continuity threshold is preferably 0.30 to 0.70, more preferably 0.40 to 0.55. By using the above conditions, severely out-of-focus areas, occluded areas, and highlight edges can be excluded from the scope of detail compensation processing, and only mild to moderate detail attenuation areas with recovery value are retained.

[0100] The purpose of detail compensation processing is to enhance the structural boundaries and local textures in areas of mild to moderate detail attenuation, rather than simply increasing overall sharpness. Specifically, it involves constrained enhancement of local high-frequency information within these areas based on edge sharpness features, texture continuity features, and region quality confidence values. Preferably, the difference between the local smoothing result and the original image is used as the detail component, and the injection intensity of the detail component is adjusted according to the region quality confidence value. The detail compensation intensity is preferably 0.10–0.60, more preferably 0.20–0.40. If the detail compensation intensity is too low, the mild to moderate detail attenuation areas will be difficult to improve effectively; if the detail compensation intensity is too high, texture noise, edge ringing, and specular false boundaries will be amplified. It should be noted that detail compensation processing is not used to recover severely blurred areas, but only to perform constrained detail enhancement on areas of mild to moderate detail attenuation.

[0101] It is important to emphasize that the innovation of step S3 does not lie in using any one of the following processes alone: ​​highlight suppression, illumination compensation, or detail compensation. Rather, it lies in: first, distinguishing the three types of degraded regions based on image quality characterization data and available region results; then, performing sequential gated fusion processing on the three processing branches based on the region quality confidence map. This allows the correction results of different degradation types to be superimposed in an orderly manner on the same quality constraint, rather than simply stitching them together in parallel or performing uniform enhancement on the entire image. This processing method can significantly reduce the problems of false boundary expansion and over-processing of low-confidence regions caused by uniform enhancement in existing technologies.

[0102] Based on the above results of highlight suppression, illumination compensation, and detail compensation, in order to ensure that each processing branch acts sequentially on the intraorbital image under unified quality constraints, the sequential gated fusion processing is preferably performed according to the following formula:

[0103] ;

[0104] ;

[0105] ;

[0106] in, This indicates the highlight suppression results in the color channel. Pixel position Pixel value at; Indicates the lighting compensation results in the color channel Pixel position Pixel value at; This indicates the detail compensation result in the color channel. Pixel position Pixel value at; This represents the first-stage fusion result after specular suppression processing; This indicates the second-stage fusion result after performing lighting compensation processing; This indicates the final structure-fidelity enhanced image in the color channels. Pixel position Pixel value at; This indicates the correction weights corresponding to the highlight suppression processing; This indicates the correction weights corresponding to the lighting compensation processing; This indicates the correction weights corresponding to the detailed compensation processing; This means that the result is restricted to the range of 0 to 1.

[0107] In the above formula, the first equation represents gated fusion that first performs specular suppression processing on the intraoral image of the target mouth, that is, according to the correction weights corresponding to specular suppression processing. The first equation determines the intensity of the effect of the highlight suppression result on the original image; the second equation represents gated fusion that performs illumination compensation processing on the first-stage fusion result, i.e., based on the correction weights corresponding to the illumination compensation processing. The third equation determines the intensity of the illumination compensation result; it represents the gated fusion that performs detail compensation processing on the second-stage fusion result, i.e., based on the correction weights corresponding to the detail compensation processing. The strength of the detail compensation result is determined by truncation to ensure that the pixel values ​​of the structure-fidelity enhanced image are still within the effective range. The order of "highlight suppression processing - illumination compensation processing - detail compensation processing" is adopted because suppressing reflection artifacts first helps to reduce the risk of false enhancement at the highlight edges, restoring low-frequency brightness helps to improve the structural visibility of shadow areas, and finally performing detail enhancement on the positions with restoration value helps to avoid the detail compensation result being diluted by the previous brightness correction, and also helps to reduce the mutual interference caused by parallel stacking.

[0108] Among them, the correction weights corresponding to the highlight suppression processing Correction weights corresponding to lighting compensation processing Correction weights corresponding to detail compensation processing All are determined based on the region quality confidence map and are positively correlated with the region quality confidence value at the corresponding pixel location. That is, at locations with high region quality confidence values, the three processing branches tend to retain and utilize their corresponding correction results; at locations with low region quality confidence values, the strength of the three processing branches is correspondingly weakened. Preferably, The values ​​range from 0 to 1. The correction weight for highlight suppression is preferably 0.20 to 0.90, the correction weight for illumination compensation is preferably 0.15 to 0.85, and the correction weight for detail compensation is preferably 0.10 to 0.70. The reason for setting the upper limit of the correction weight for detail compensation lower than that for highlight suppression and illumination compensation is that detail compensation is more likely to amplify noise and false boundaries, and therefore should be subject to stricter quality constraints.

[0109] Furthermore, the reason for limiting the fusion range corresponding to detail compensation processing based on the available region results is that detail compensation processing has the characteristic of enhancing local high-frequency information. If it is applied to a location with a low region quality confidence value or a location that has been determined to be unusable in step S2, it is easy to perform unnecessary enhancement on occluded edges, highlight edges, and noise textures, which will reduce the stability of the structure-fidelity enhanced image. Therefore, in this embodiment, in addition to determining the correction weight based on the region quality confidence map, detail compensation processing is also constrained by the available region results. That is, it only participates in sequential gating fusion processing within the range of locations marked as usable by the available region results. For unusable locations, even if there is some detail attenuation, it is not preferable to perform detail compensation processing.

[0110] In summary, step S3 distinguishes between unsaturated highlight regions, low-frequency shadow regions, and regions with mild to moderate detail attenuation based on image quality characterization data, and performs highlight suppression processing, illumination compensation processing, and detail compensation processing respectively. Furthermore, it performs sequential gated fusion processing based on the region quality confidence map and the available region results to obtain a structure-fidelity enhanced image. Thus, step S3 can suppress highlight artifacts and shadow interference while avoiding excessive detail enhancement in low-confidence regions, thereby providing a more stable image foundation for the extraction of credible risk feature data in step S4.

[0111] In this embodiment, based on the structurally enhanced image obtained in step S3, abnormal information that truly contributes to the screening risk results is extracted from the tooth surface region, gingival region, and mucosa region. Using the regional quality confidence map and available region results established in step S2, the abnormal information is filtered, weighted, and aggregated under quality constraints. This avoids two problems caused by directly inputting the entire image or region into the black-box classification model in existing technologies: firstly, false boundaries, false textures, and false color differences in low-quality regions are incorrectly included in the risk; secondly, local real anomalies are masked by low-quality interference in the overall regional statistics. Therefore, the core technical means of step S4 is not simply outputting a classification result, but rather: firstly, color abnormality features, texture abnormality features, edge abnormality features, and morphological abnormality features are extracted from the tooth surface region, gingival region, and mucosa region respectively; then, candidate abnormality response values ​​are constructed by combining the regional quality confidence map and available region results; and finally, a credible risk feature vector is obtained through regional-level aggregation processing. Finally, the screening risk results are output by combining regional-level quality weights and overall analyzability indicators.

[0112] In practical implementation, the structure-fidelity enhanced image obtained in step S3 is first used as the basic input for step S4. Since step S3 has already performed region-differentiated image correction processing on highlights, shadows, and mild to moderate detail attenuation, the feature extraction in step S4 is not performed directly on the original target intraoral image, but on the structure-fidelity enhanced image that has suppressed highlight artifacts, improved low-frequency shadows, and avoided over-enhancement of low-confidence regions. For ease of description, the region categories are denoted as... ,in Take one of the three regions: the tooth surface area, the gingival area, or the mucosal area; record the pixel position as... The region quality confidence value obtained in step S2 is recorded as follows: The available area indication value obtained in step S2 is recorded as... ; The region categories obtained in step S1 The corresponding corrected region mask is denoted as In step S4, the extraction of all abnormal features, the formation of candidate abnormal response values, and the regional-level aggregation of credible risk feature vectors are all limited to the corresponding regional categories. The corresponding corrected area is within the mask range.

[0113] In practical implementation, the structure-fidelity enhanced image obtained in step S3 is first used as the basic input for step S4. Since step S3 has already performed region-differentiated image correction processing on highlights, shadows, and mild to moderate detail attenuation, the feature extraction in step S4 is not performed directly on the original target intraoral image, but on the structure-fidelity enhanced image that has suppressed highlight artifacts, improved low-frequency shadows, and avoided over-enhancement of low-confidence regions. For ease of description, the region categories are denoted as... ,in Take one of the three regions: the tooth surface area, the gingival area, or the mucosal area; record the pixel position as... The region quality confidence value obtained in step S2 is recorded as follows: The available area indication value obtained in step S2 is recorded as... ; The region categories obtained in step S1 The corresponding corrected region mask is denoted as In step S4, the extraction of all abnormal features, the formation of candidate abnormal response values, and the regional-level aggregation of credible risk feature vectors are all limited to the corresponding regional categories. The corresponding corrected area is within the mask range.

[0114] ;

[0115] in, Indicates the region category At pixel position Candidate anomaly response values ​​at the location; Indicates the region category At pixel position The regional quality confidence value at that location; Indicates the region category At pixel position Available area indicator value at; Indicates abnormal color characteristics; Indicates texture anomaly features; Indicates edge anomaly features; Indicates abnormal morphological features; Indicates the weight of color anomaly features; Indicates the weight of texture anomaly features; Indicates the weight of edge anomaly features; This represents the weight of morphological anomaly features.

[0116] In the above formula, Together, they constitute the quality constraints on anomalous features, among which, Used to characterize the reliability of this pixel location, This setting is used to determine whether a pixel location belongs to the available area confirmed in step S2. In other words, when a pixel location has strong anomalous features but its region quality confidence value is low, or its available area indication value is 0, the candidate anomalous response value at that location will be significantly suppressed. On the other hand, when a pixel location has a high region quality confidence value and its available area indication value is 1, the real anomalous information at that location can be more fully preserved in the candidate anomalous response value. The purpose of this setting is to prevent false boundaries, false textures, and false color differences in low-quality areas from being directly transmitted to the credible risk feature data.

[0117] in, These are used to adjust the contribution intensity of the four types of abnormal features in the candidate abnormal response values, preferably, The value range is 0.10 to 0.45, more preferably 0.15 to 0.30; The value range is 0.10 to 0.40, and more preferably 0.15 to 0.28; The value range is 0.10 to 0.35, more preferably 0.12 to 0.25; The value range is 0.05 to 0.30, more preferably 0.08 to 0.20. In the preferred embodiment, The reason for including in the preferred range The upper limit is set slightly higher than This is because, in intraoral image scenes, local color and texture changes typically reflect structural anomalies earlier and more stably, while edge and morphological anomalies depend more on the fact that the anomalous region already has a certain boundary size and connectivity. Therefore, in the early stages of candidate anomaly response value formation, it is advisable to give color and texture anomaly features a slightly higher contribution weight. In preferred implementation, this can be taken as... Under this set of parameters, the contribution of local color changes, texture changes, edge changes and morphological changes to the candidate anomaly response value can be well balanced.

[0118] After obtaining candidate anomaly response values, a region-level aggregation process is performed on these values ​​to obtain reliable risk feature vectors corresponding to the tooth surface region, gingival region, and mucosal region, respectively. This region-level aggregation process does not simply take the average; instead, it uses the candidate anomaly response values ​​as weights to weight and summarize the local anomaly features at each location within the region. This ensures that locations with high anomaly severity and reliable image quality occupy a higher proportion in the region-level representation. Specifically, this can be achieved by considering region categories... Within the corresponding corrected region mask, color anomaly features, texture anomaly features, edge anomaly features, and morphological anomaly features are weighted and aggregated using candidate anomaly response values ​​to obtain the region category. The corresponding credible risk feature vectors are generated in this way, and the tooth surface region, gingival region, and mucosa region each form their own credible risk feature vectors. Then, based on the credible risk feature vectors corresponding to the tooth surface region, the credible risk feature vectors corresponding to the gingival region, and the credible risk feature vectors corresponding to the mucosa region, credible risk feature data is obtained.

[0119] After obtaining credible risk characteristic data, the risk level is not immediately output. Instead, regional quality weights are calculated based on the regional quality confidence map, and overall analyzability indicators are calculated based on the available area results. Regional quality weights are used to reflect the credible contribution of the tooth surface area, gingival area, and mucosal area in this analysis. Preferably, the regional quality weights are determined jointly based on the average level of regional quality confidence value and the proportion of available area within the corresponding region. That is, the higher the average level of regional quality confidence value and the larger the proportion of available area, the higher the regional quality weight of the corresponding region. The reason for this setting is that if a certain region has a local abnormal response, but its overall regional quality confidence value is low or its available area is too small, it should not occupy too high a weight in the final screening risk results.

[0120] The overall analyzability index is used to characterize whether the current target intraoral image has a sufficient number of usable areas to support risk assessment. Preferably, the overall analyzability index is obtained by weighting the proportion of pixels with a usable area indicator value of 1 in the tooth surface region, gingival region, and mucosa region to the area of ​​their respective corrected region mask. The overall analyzability index threshold is preferably 0.35 to 0.80, more preferably 0.45 to 0.65. If the overall analyzability index threshold is too low, a large number of low-quality images may still be incorrectly included in the normal risk assessment process. If the overall analyzability index threshold is too high, some images that, although of average overall quality, can still support preliminary screening may be over-judged as unanalyzable. In preferred implementation, the overall analyzability index threshold can be set to 0.55. Under this condition, it can ensure that the effective structural information reaches a certain coverage range without over-compressing the actual applicable scenarios of the system.

[0121] After obtaining credible risk feature data, regional quality weights, and overall analyzability indicators, when the overall analyzability indicator is lower than a preset threshold, a re-image prompt is output as the screening risk result. This setting aims to prevent the system from providing a seemingly clear but actually unreliable risk level when the overall quality of the target intraoral image is insufficient. Conversely, when the overall analyzability indicator is not lower than the preset threshold, at least one of the risk level and risk area prompt information is output as the screening risk result based on the credible risk feature data and regional quality weights. Specifically, the credible risk feature vectors corresponding to the tooth surface region, gingival region, and mucosal region can be weighted and fused according to their respective regional quality weights, and the risk level and risk area prompt information are output based on the fusion result. Since the risk level mapping rules and risk area prompt information generation methods can be implemented using existing statistical mapping or rule-based judgment methods, this implementation does not elaborate on them as the core innovation of step S4, but emphasizes the processing logic of outputting the screening risk result based on the quality-constrained credible risk feature data only when the overall analyzability indicator meets the requirements.

[0122] In summary, step S4 extracts color anomaly features, texture anomaly features, edge anomaly features, and morphological anomaly features based on structure-fidelity enhanced images. It further combines regional quality confidence maps and available region results to construct candidate anomaly response values ​​and form credible risk feature data. Then, it combines regional quality weights and overall analyzability indicators to output screening risk results. Thus, step S4 can avoid false anomalies in low-quality regions from directly participating in risk judgment and prioritize outputting re-image prompts when overall analyzability is insufficient, thereby improving the credibility and stability of screening risk results.

[0123] Example 2

[0124] like Figure 3 As shown, the present invention also discloses an oral disease risk assessment system, comprising:

[0125] The region localization module is used to acquire images of the target intraoral cavity, perform basic oral cavity region localization processing on the images of the target intraoral cavity, and obtain the region distribution results;

[0126] The quality analysis module is used to extract image quality characterization data based on regional distribution results, and generate regional quality confidence maps and available region results based on the image quality characterization data;

[0127] The image correction module is used to perform regional differential image correction processing on the intraorbital image of the target based on image quality characterization data, regional quality confidence map and available region results, to obtain a structure-fidelity enhanced image;

[0128] The risk assessment module is used to extract credible risk feature data based on structure-fidelity enhanced images, regional quality confidence maps, and available area results, and output risk screening results based on the credible risk feature data.

[0129] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for assessing the risk of oral diseases, characterized in that, Includes the following steps: Step S1: Obtain the target intraoral image, perform basic oral cavity region localization processing on the target intraoral image, and obtain the region distribution result; Step S2: Extract image quality characterization data based on the regional distribution results, and generate regional quality confidence maps and available region results based on the image quality characterization data; Step S3: Based on the image quality characterization data, the region quality confidence map, and the available region result, perform region-differentiated image correction processing on the target intraoral image to obtain a structure-fidelity enhanced image; Step S4: Extract credible risk feature data based on the structure-fidelity enhanced image, the region quality confidence map, and the available region results, and output the screening risk results based on the credible risk feature data.

2. The oral disease risk assessment method according to claim 1, characterized in that, In step S1, the basic oral cavity region localization process includes: The intraoral image of the target is normalized to obtain a normalized intraoral image; The normalized intraoral image is input into the region localization model to obtain the initial region probabilities corresponding to the tooth surface region, gingival region and mucosal region, respectively. Based on the relative positional relationships, adjacency relationships, and mutual exclusion relationships among the tooth surface region, the gingival region, and the mucosal region, a topological consistency correction process is performed on the initial region probability to obtain a corrected region label map; Based on the corrected region label map, corrected region masks are generated corresponding to the tooth surface region, the gingival region, and the mucosal region, respectively; The regional distribution results include the corrected regional mask corresponding to the tooth surface region, the corrected regional mask corresponding to the gingival region, and the corrected regional mask corresponding to the mucosal region.

3. The oral disease risk assessment method according to claim 2, characterized in that, In step S2, the image quality characterization data includes: The local brightness shift features, highlight intensity features, edge sharpness features, texture continuity features, and occlusion indication features are extracted within the range defined by the corrected region mask corresponding to the tooth surface region, the corrected region mask corresponding to the gingival region, and the corrected region mask corresponding to the mucosa region, respectively. The local brightness shift feature, the highlight intensity feature, the edge sharpness feature, the texture continuity feature, and the occlusion indication feature are normalized to obtain the image quality characterization data.

4. The oral disease risk assessment method according to claim 3, characterized in that, In step S2, the generated region quality confidence map and available region results include: Based on the local brightness shift feature, the highlight intensity feature, the edge sharpness feature, the texture continuity feature, and the occlusion indication feature, the region quality confidence value corresponding to each pixel in the tooth surface region, the gingival region, and the mucosal region is calculated respectively. The region quality confidence map is generated based on the region quality confidence values ​​corresponding to each pixel in the tooth surface region, the gingival region, and the mucosal region. Based on the regional quality confidence map and the preset availability determination conditions, the available region indication values ​​corresponding to the tooth surface region, the gingival region and the mucosa region are determined respectively; The available area results include the available area indication value corresponding to the tooth surface area, the available area indication value corresponding to the gingival area, and the available area indication value corresponding to the mucosal area.

5. The oral disease risk assessment method according to claim 4, characterized in that, The region quality confidence map generated in step S2 serves as the basis for determining the correction weight in step S3 and the feature weighting weight in step S4.

6. The oral disease risk assessment method according to claim 5, characterized in that, In step S3, the region-differential image correction process includes: Based on the image quality characterization data, unsaturated highlight areas, low-frequency shadow areas, and areas with mild to moderate detail attenuation are identified. Perform highlight suppression processing on the unsaturated highlight regions to obtain highlight suppression results; Illumination compensation processing is performed on the low-frequency shadow region to obtain the illumination compensation result; Perform detail compensation processing on the mild to moderate detail attenuation region to obtain the detail compensation result; The highlight suppression processing, the illumination compensation processing, and the detail compensation processing are each limited in their execution range based on the available area results.

7. The oral disease risk assessment method according to claim 6, characterized in that, In step S3, obtaining the structurally enhanced image includes: Following the order of highlight suppression processing, illumination compensation processing, and detail compensation processing, sequential gating fusion processing is performed on the highlight suppression results, illumination compensation results, and detail compensation results. The sequential gating fusion process determines the correction weights corresponding to each processing branch based on the region quality confidence map, and limits the fusion range corresponding to the detail compensation process based on the available region results, thereby obtaining the structure-fidelity enhanced image.

8. The oral disease risk assessment method according to claim 7, characterized in that, In step S4, the extraction of credible risk feature data includes: Based on the structural fidelity enhanced image, color abnormality features, texture abnormality features, edge abnormality features, and morphological abnormality features are extracted in the tooth surface region, the gingival region, and the mucosal region, respectively. Based on the region quality confidence map and the available region results, weighted processing is performed on the color anomaly features, the texture anomaly features, the edge anomaly features, and the morphological anomaly features to obtain candidate anomaly response values; The candidate abnormal response values ​​are subjected to region-level aggregation processing to obtain credible risk feature vectors corresponding to the tooth surface region, the gingival region, and the mucosal region, respectively. The credible risk feature data is obtained based on the credible risk feature vector corresponding to the tooth surface region, the credible risk feature vector corresponding to the gingival region, and the credible risk feature vector corresponding to the mucosal region.

9. The oral disease risk assessment method according to claim 8, characterized in that, In step S4, outputting the risk screening result based on the credible risk characteristic data includes: Calculate the regional quality weights based on the regional quality confidence map; Calculate the overall analyzability index based on the available area results; When the overall analyzability index is lower than a preset threshold, a repeat shooting prompt message is output as the screening risk result. When the overall analyzability index is not lower than the preset threshold, at least one of the risk level and risk area prompt information is output as the screening risk result based on the credible risk feature data and the regional quality weight.

10. An oral disease risk assessment system, employing an oral disease risk assessment method as described in any one of claims 1-9, characterized in that, include: The region localization module is used to acquire a target intraoral image, perform basic oral region localization processing on the target intraoral image, and obtain the region distribution result; The quality analysis module is used to extract image quality characterization data based on the regional distribution results, and generate regional quality confidence maps and available region results based on the image quality characterization data; The image correction module is used to perform regional differential image correction processing on the target intraoral image based on the image quality characterization data, the regional quality confidence map, and the available region results, to obtain a structure-fidelity enhanced image; The risk assessment module is used to extract credible risk feature data based on the structure-fidelity enhanced image, the regional quality confidence map, and the available region results, and output the screening risk results based on the credible risk feature data.

Images