Intraoral radiographic image phenotype generation method for potential malignant conditions of oral mucosa

Abstract

This invention discloses a method for generating phenotypes of potential malignant diseases of the oral mucosa from intraoral images. Relating to the field of intraoral image processing, this method performs quality perception and usability assessment on input intraoral images, suppressing low-quality images and irrelevant background interference. It extracts depth visual features, morphological features, color features, boundary texture features, and follow-up change features from effective mucosal regions and lesion candidate regions to form a structured risk phenotype. Multiple intraoral images from the same patient are weighted and aggregated based on quality, risk response, lesion visibility, and uncertainty to generate a patient-level image risk characterization. This provides standardized, quantitative, and interpretable technical support for long-term follow-up, risk management, and image verification of oral mucosal lesions.

Description

Technical Field

[0001] This invention relates to the field of intraoral image processing, and more specifically to a method for generating phenotypes of intraoral images of potential malignant diseases of the oral mucosa. Background Technology

[0002] Oral potentially malignant disorders (OPMD) are a class of diseases or lesion states occurring in the oral mucosa with varying degrees of risk of malignant progression. These include oral leukoplakia, oral erythroplakia, oral lichen planus, oral submucosal fibrosis, and oral lichenoid lesions. These lesions exhibit significant heterogeneity in morphology, color, borders, surface texture, degree of keratinization, extent of erosion, and changes during follow-up. Some lesions can undergo malignant progression under the influence of long-term stimulation, abnormal immune inflammation, abnormal epithelial proliferation, and genetic molecular alterations. Therefore, extracting structured image risk phenotypes from routine intraoral images that reflect the lesion status, heterogeneity, and progression trend, and generating risk assessment information for physician review, is a crucial technical requirement for the long-term management and intelligent follow-up of oral mucosal lesions.

[0003] Currently, risk assessment of oral mucosal lesions mainly relies on visual observation by physicians, intraoral photography, auxiliary examinations, and, when necessary, tissue biopsies and pathological analysis. While these methods provide important clinical evidence, they still have significant limitations in practical application: First, physicians' judgments on lesion color, boundaries, roughness, degree of red-white mixing, erosion, and keratinization areas are somewhat subjective, and the consistency of results among different observers is limited. Second, biopsies and pathological examinations usually only reflect the state of local sampling areas and are difficult to cover the entirety of lesions that are large, multifocal, or have significant internal heterogeneity. Third, although routine intraoral photography has advantages such as being non-invasive, low-cost, repeatable, and suitable for long-term follow-up, different equipment, lighting, exposure, focal length, shooting angle, saliva reflection, tooth obstruction, and degree of mouth opening can cause significant differences in image quality, directly affecting the stability of subsequent image analysis models. Therefore, relying solely on a single observation or a single image is insufficient to fully express the overall risk phenotype and patient-level trends of lesions.

[0004] In recent years, optical imaging techniques such as narrowband imaging, autofluorescence imaging, high-resolution intraoral endoscopy, and multispectral imaging have been used to enhance the visibility of abnormal areas in the oral mucosa. These techniques can highlight mucosal vascular texture, epithelial structure, tissue fluorescence changes, or local color differences to a certain extent, helping to detect abnormal areas that are not easily detected in ordinary white light images. However, these techniques usually rely on specialized equipment and standardized operations, and are greatly affected by light source conditions, imaging angle, patient cooperation, and operator experience. In most cases, doctors still need to subjectively interpret the image appearance. More importantly, existing optical imaging methods usually focus on enhancing local visibility and are still difficult to uniformly calculate multiple intraoral images, multi-time point follow-up images, image quality differences, lesion morphological changes, and model uncertainties. It is also difficult to form standardized, quantifiable, and interpretable image risk phenotypic information.

[0005] With the development of artificial intelligence, deep learning, and visual fundamental models, intelligent analysis of oral mucosal lesions based on intraoral images is gradually becoming a new technological direction. Existing methods mostly use convolutional neural networks, visual Transformers, or other classification models to extract features from single images or local image patches and output lesion categories, risk scores, or classification results. These methods have demonstrated the feasibility of image models in the analysis of oral mucosal lesions, but several technical bottlenecks remain: First, many methods classify the entire image directly, making them susceptible to interference from teeth, lips, instruments, shadows, reflections, and background mucosal areas; Second, while local image patch models can capture some high-risk visual cues, they easily overlook the overall morphology of the lesion, multifocal distribution, and patient-level multi-image information; Third, existing models typically lack explicit modeling of image quality, lesion visibility, and prediction uncertainty, resulting in insufficient output stability when the quality of real clinical intraoral photographs varies; Fourth, ordinary classification models struggle to simultaneously integrate multi-dimensional image phenotypic information such as depth visual features, lesion area, boundary irregularity, red-white ratio, color heterogeneity, texture roughness, and follow-up changes, leading to limited interpretability and verifiability of model results. Summary of the Invention

[0006] To address the aforementioned shortcomings in the prior art, the method for generating phenotypic images of potential malignant diseases of the oral mucosa provided by this invention solves the problem that the prior art lacks explicit modeling of image quality, depth visual features, and lesion features when generating phenotypic images of intraoral mucosa, resulting in incomplete extraction of image information.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: A method for generating intraoral image phenotypes of potentially malignant diseases of the oral mucosa is provided, comprising the following steps: Acquire intraoral images related to the oral mucosa, calculate the feature values ​​of the intraoral images related to the oral mucosa in different dimensions, and then obtain the corresponding overall image quality score based on the feature values ​​in different dimensions. Effective mucosal region masks and lesion candidate region masks are generated from intraoral images related to the oral mucosa, and then the corresponding effective mucosal region constraint images and lesion candidate region constraint images are obtained. Obtain the global features, effective mucosal region features, and lesion candidate region features corresponding to the intraoral images related to the oral mucosa, the effective mucosal region constraint images, and the lesion candidate region constraint images, respectively. Oral mucosa-related intraoral images or lesion candidate regions are divided into image patch sets of different scales. The features of each image patch are obtained, and then the risk response intensity of each image patch is calculated. The features of the K image patches with the highest risk response intensity are weighted and aggregated to obtain the local risk response features. Global features, effective mucosal region features, lesion candidate region features, and local risk response features are stitched together to form image-level risk features; Based on the image-level risk features, lesion morphology phenotype, color phenotype, boundary texture phenotype, and overall image quality score corresponding to the same oral mucosa-related intraoral image, a risk phenotype vector is constructed. This vector is then used to convert ordinary oral mucosa-related intraoral images into risk phenotype vectors that combine depth features and interpretable image features, providing input for subsequent tasks.

[0008] The beneficial effects of this invention are as follows: 1. Leading Image Quality Perception Capability: By comprehensively scoring the clarity, exposure, overexposure, saliva reflection, occlusion, effective mucosal ratio, and lesion visibility of input intraoral images, it can dynamically filter low-quality images and intelligently weight them in multi-image aggregation and risk assessment, making the overall risk phenotype extraction more stable and reliable, and significantly improving the adaptability to different acquisition devices, lighting conditions, and shooting angles in real clinical intraoral photography.

[0009] 2. Innovative Mucosal and Lesion Region Constraints: The system generates masks of effective mucosal regions and lesion candidate regions. By using probabilistic weighting to reduce background interference, it preserves local lesion features while taking into account the global mucosal structure. This provides accurate input for subsequent high-risk image patch perception and risk phenotype coding, improving the local lesion information capture rate and global contextual consistency.

[0010] 3. Multi-scale high-risk image patch perception and weighted aggregation: The candidate region is divided into multi-scale image patches, the risk response of each image patch is calculated, and high-risk image patches are selected for weighted aggregation to form image-level risk features. This mechanism can highlight local high-risk clues (such as mixed red and white, erosion, thickened keratinization, and abnormal boundaries) while preserving global structural information and avoiding dilution of small-area lesions.

[0011] 4. Structured risk phenotype generation and consistency constraints: Image features are encoded into structured risk phenotype vectors, including depth visual features, lesion morphology features, color features, boundary texture features, image quality features, and follow-up time features. Phenotypic consistency constraints are added to ensure that depth features are consistent with interpretable morphology, color, and texture features, thereby improving interpretability and model verifiability.

[0012] 5. Quality-Risk-Uncertainty Joint Multi-Image Aggregation: For multiple intraoral images of the same patient, the system weights and fuses the risk phenotypes of each image based on image quality, lesion visibility, high-risk response, and uncertainty to generate patient-level risk information. This allows high-quality, high-visibility, and low-uncertainty images to contribute more to the final phenotype, thereby improving the stability and reliability of patient-level risk assessment.

[0013] 6. Dynamic modeling of lesion phenotypes in multiple time periods: It can analyze the changes in lesion risk phenotypes from intraoral photographs taken at multiple time points, calculate the temporal trend and rate of change of lesion phenotypes, capture the characteristics of lesion changes over time, and improve the ability to predict long-term follow-up and risk trends. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating the method. Figure 2 This is a graph showing the perceived quality and usability of intraoral imaging using this method. Figure 3 This is a phenotypic coding diagram of lesion risk in this method; Figure 4 This is a multi-scale high-risk image patch perception map for this method; Figure 5 This is a multi-graph aggregation graph with joint weighting for quality, risk, and uncertainty in this method; Figure 6 This is a schematic diagram of the data processing flow of this system; Figure 7 The uncertainty calibration cascade evaluation diagram for this system; Figure 8 This is the output diagram of the risk evidence map and structured report of this system; Figure 9 This is a schematic diagram of the high-risk lesion identification results in Example 2; Figure 10 This is a schematic diagram illustrating the prediction of high-risk lesions in Example 2. Detailed Implementation

[0015] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0016] Example 1: like Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 As shown, the method for generating intraoral image phenotypes of this potentially malignant oral mucosal disease includes the following steps: S1. Obtain intraoral images related to the oral mucosa, calculate the feature values ​​of the intraoral images related to the oral mucosa in different dimensions, and then obtain the corresponding overall image quality score based on the feature values ​​in different dimensions. In this embodiment, intraoral imaging can Images are acquired using dental cameras, ordinary digital cameras, mobile phones, endoscopic devices, or other image acquisition devices. Each intraoral image can be associated with the patient's number, the time of capture, the location of the image, the source of the image, and follow-up information.

[0017] Suppose a certain patient have An internal view of the mouth is represented as:

[0018] in Indicates the patient The Internal image of the open mouth.

[0019] If there are follow-up images at multiple time points, it can be represented as:

[0020] in Indicates the first Each sample at time point The data set; Indicates the first The first sample A feature or observation, Indicates the corresponding time or condition label; The total number of data points for this sample represents the number of observations per sample. Linking multiple images by patient ID and time information provides a data foundation for subsequent patient-level risk phenotypic fusion and multi-temporal variation modeling.

[0021] To facilitate subsequent feature calculations, this embodiment can also perform standardized preprocessing on the intraoral images, including color space conversion (such as RGB → Lab or HSV), size uniform adjustment, oral mucosa region cropping, pixel normalization, background edge cropping, and tensor transformation, thereby obtaining intraoral images related to the oral mucosa.

[0022] It should be noted that image preprocessing can employ other standardized methods, such as adaptive histogram equalization, Retinex illumination correction, edge enhancement, or super-resolution reconstruction, to ensure the clarity of mucosal structures and lesion features, while also generating image quality scores for subsequent weighted aggregation.

[0023] Unlike traditional image preprocessing, this invention further introduces an intraoral imaging quality perception scoring mechanism. The system scores the first... Zhang Image Calculate the quality representation vector of the seven sub-dimensions : ; The calculation method for each sub-score is as follows: Clarity rating Image sharpness is measured by calculating the variance of the Laplacian operator or the Sobel edge strength. A higher variance indicates a sharper image; a lower variance indicates a more blurred image. The expression is:

[0024] in For the first Intraoral images related to the oral mucosa The image response obtained after applying Laplacian operator filtering. This represents the variance calculation function, used to measure the dispersion of the image response results.

[0025] Too dark : Calculates the difference between the mean of the grayscale or luminance channel and a standardized threshold, used to reflect whether the image is too dark. Its expression is:

[0026] in Indicates pixel coordinates The brightness value at that location; This represents the total number of pixels in the image.

[0027] Overexposure Statistical brightness exceeds threshold The pixel ratio, used to reflect the proportion of overly bright areas in an image, is expressed as: ; Saliva reflection / highlight interference The percentage of reflective areas is statistically determined by local brightness peak detection or morphological processing; its expression is: ; Occlusion rating A mask is generated using teeth, tongue, restorations, instruments, and shadow detection. The percentage of occluded pixels is calculated, and its expression is: ; Percentage of effective mucosal area The effective mucosal region is extracted using the mucosal segmentation module, and its proportion of the entire image is calculated. The expression for this is: ; Visibility of candidate lesion areas The visibility ratio is calculated based on the lesion candidate region mask, reflecting the clarity of the lesion in the image. Its expression is: ; Overall quality score The sub-scores are merged into a continuous value for subsequent patient-level multi-graph aggregation and risk phenotype assessment.

[0028] For example, the feature values ​​of the same intraoral image related to the oral mucosa in different dimensions are used to construct a quality feature vector, which is then input into a linear mapping layer to obtain a weighted result in the image quality space. The weighted result is then normalized to [0,1] using the Sigmoid function to obtain the overall image quality score corresponding to the intraoral image related to the oral mucosa. The corresponding expression is:

[0029] in For the first The overall image quality score of intraoral images related to the oral mucosa is a normalized continuous scalar; Represents the Sigmoid function; This is a linearly mapped weight vector used to map the quality feature vector. Mapped to the image quality space, reflecting the importance of each feature value to image quality; superscript Indicates transpose; For the first The quality feature vector is composed of the feature values ​​of different dimensions corresponding to the intraoral images of the oral mucosa. This is a bias term used to adjust the baseline value after linear mapping. and It can be obtained through learning or training, or it can be set manually.

[0030] In practical applications, image quality can be categorized into three types: high quality, usable quality, and low-confidence quality. Low-quality images are not necessarily discarded directly; instead, their weight is dynamically reduced in patient-level aggregation or cascade assessments. When image quality falls below a preset threshold, a low-confidence alert or a review alert can be output. Through this mechanism, image quality information not only serves preprocessing screening but also directly participates in subsequent risk phenotype generation, multi-image fusion, and cascade assessment calculations.

[0031] S2. Generate effective mucosal region masks and lesion candidate region masks from intraoral images related to oral mucosa, and then obtain the corresponding effective mucosal region constraint images and lesion candidate region constraint images. In this embodiment, real intraoral photographs often include teeth, lips, tongue, restorations, instruments, saliva reflections, shadows, and non-target backgrounds. To reduce the interference of irrelevant backgrounds on image feature extraction, this invention uses intraoral photographs related to the oral mucosa as input to a pre-trained mucosal region extraction module to obtain corresponding effective mucosal region masks; and uses intraoral photographs related to the oral mucosa as input to a pre-trained lesion segmentation model to obtain corresponding lesion candidate region masks; the process can be represented as follows:

[0032]

[0033] in For effective mucosal area masking; For the input of the first Zhang's intraoral images related to the oral mucosa; This indicates a pre-trained mucosal region extraction module; Masking candidate lesion areas; For pre-trained lesion segmentation models; Then, the intraoral images related to the oral mucosa are multiplied pixel-by-pixel with the corresponding effective mucosal region mask to obtain the effective mucosal region constraint image; the intraoral images related to the oral mucosa are then multiplied pixel-by-pixel with the corresponding lesion candidate region mask to obtain the lesion candidate region constraint image. This process can be represented as:

[0034]

[0035] in Constrained images for effective mucosal regions; Constrained image for candidate lesion regions; This indicates pixel-by-pixel multiplication.

[0036] The above process preserves the original full image, the effective mucosal region image, and the lesion candidate region image, which can then be used to extract the global context, the effective mucosal background, and the local features of the lesion.

[0037] It should be noted that the pre-trained mucosal region extraction module can employ one or more of the following: semantic segmentation networks, instance segmentation networks, salient region extraction networks, or visual base model segmentation modules. Specifically, it can use U-Net, U-Net++, DeepLabv3+, Mask R-CNN, SegFormer, SAM, and / or MobileSAM. The pre-trained lesion segmentation model can employ one or more of the following: lesion semantic segmentation networks, lesion instance segmentation networks, Transformer segmentation networks, or base model prompting segmentation networks. Specifically, it can use U-Net, Attention U-Net, U-Net++, DeepLabv3+, Mask R-CNN, SegFormer, TransUNet, Swin-UNet, SAM, and MedSAM. The training process uses intraoral images related to the oral mucosa as input and images labeled with corresponding region masks as labels. Common parameter adjustment methods can be used. The specific training process is not an innovation of this invention and will not be described in detail in this embodiment.

[0038] The mask acquisition process can employ probabilistic segmentation models, weakly supervised generative models, attention-guided models, or graph structure optimization models. Mask pixel weights can be automatically adjusted based on image feature responses, local entropy, or region saliency. Specifically, probabilistic segmentation models can include one or more of U-Net, U-Net++, Attention U-Net, DeepLabv3+, SegFormer, TransUNet, Swin-UNet, Mask R-CNN, SAM, and MedSAM; weakly supervised generative models can include one or more of weakly supervised segmentation models based on class activation maps, CAM, Grad-CAM, Score-CAM, MIL weakly supervised segmentation models, or pseudo-label iterative optimization models; attention-guided models can include one or more of channel attention networks, spatial attention networks, CBAM, SE-Net, Transformer Attention Map, or Attention Rollout models; and graph structure optimization models can include one or more of conditional random fields, graph cut models, graph convolutional networks, graph attention networks, or superpixel graph optimization models.

[0039] The core of this invention does not lie in the structural development of the above models. All of the above models can use existing models, and the specific model structures will not be described in detail.

[0040] S3. Obtain the global features, effective mucosal region features, and lesion candidate region features corresponding to the intraoral images related to the oral mucosa, the effective mucosal region constraint images, and the lesion candidate region constraint images, respectively. In this embodiment, feature acquisition can be performed using a pre-trained visual encoder. Specifically, the intraoral images related to the oral mucosa, the constrained images of the effective mucosal region, and the constrained images of the lesion candidate region are used as inputs to the pre-trained visual encoder, resulting in global features, effective mucosal region features, and lesion candidate region features, respectively. This process can be represented as:

[0041]

[0042]

[0043] in Represents a visual encoder. , and These represent global features, effective mucosal region features, and lesion candidate region features, respectively. This region constraint mechanism can effectively reduce the influence of non-lesion background on the model output, improving the specificity and interpretability of image risk phenotypes.

[0044] S4. Divide the oral mucosa-related intraoral images or lesion candidate regions into image patch sets of different scales, obtain the features of each image patch, and then calculate the risk response intensity of each image patch. The features of the K image patches with the highest risk response intensity are weighted and aggregated to obtain the local risk response features. High-risk visual cues in oral mucosal lesions often manifest as localized erythema, erosion, mixed red and white areas, thickened keratinization, coarse texture, blurred boundaries, or localized irregular changes. These areas may be small in size within a whole intraoral image and are easily diluted by global features. Therefore, this step introduces a multi-scale high-risk image patch perception mechanism, specifically including the following steps: S4-1. Divide intraoral images or candidate lesion regions related to the oral mucosa into image patch sets of different scales:

[0045] in For the first Zhang is a set of image patches of different scales corresponding to an image that has been divided into image patches; Indicates the scale of an image patch. Indicates the first Image blocks.

[0046] S4-2. The features extracted by the trained visual encoder for each image patch can be represented as follows:

[0047] in For the first The images divided into image blocks are in the first... The first scale Features of an image patch.

[0048] S4-3. Calculate the risk response intensity of each image patch based on its features. The expression is as follows:

[0049] in Indicates the first The images divided into image blocks are in the first... The first scale Risk response intensity of each image patch; This represents the Sigmoid function, which maps linear results to the [0,1] interval, making it usable as a probabilistic risk response. This is a linearly mapped weight vector used to map image patch features to the risk response space, reflecting the importance of each feature dimension to the risk score; superscript Indicates transpose; This is a bias term used to adjust the baseline value after linear mapping; and Both can be obtained through training.

[0050] S4-4. The features of the K image patches with the highest risk response intensity are weighted and aggregated to obtain the local risk response features, the expression of which is:

[0051] in For the first Local risk response characteristics of an image divided into image patches; Image patch features The weight, ; It is the set of the K image patches with the highest risk response intensity.

[0052] It should be noted that image patch segmentation can also adopt deformable patch or ROI-guided patch extraction methods; the selection of image patches with the highest risk response intensity can also use non-maximum suppression, local entropy sorting or reinforcement learning strategies to automatically determine the top K; the aggregation weight can be adaptively weighted by combining patch risk response, attention weight, image quality and historical change information.

[0053] In some embodiments, in addition to the existing visual Transformer + CNN combined structure, the visual encoder may adopt ViT-Large, Swin-Large, ConvNeXt-L, EfficientNet-L, MAE, CLIP image encoder or multimodal fusion Transformer-CNN structure to support multi-scale global feature modeling and local high-risk region capture.

[0054] The training method for a visual encoder includes the following steps: A1. Use image-level risk features as input to the phenotypic prediction head to obtain the corresponding predicted phenotypic vector. ; A2. The vector composed of lesion morphology phenotype, color phenotype, and boundary texture phenotype is denoted as the reference phenotype vector. ; A3. Constructing phenotypic consistency loss based on predicted phenotypic vectors and reference phenotypic vectors. Its expression is:

[0055] in Indicates the L1 distance; These are the weighting coefficients; It is a correlation function; A4. Based on Phenotypic Consistency Loss The visual encoder is trained using backpropagation until it converges, so that the features extracted by the trained visual encoder are consistent with the interpretable morphological, color and texture phenotypes.

[0056] S5. Combine global features, effective mucosal region features, lesion candidate region features, and local risk response features to form image-level risk features; In this embodiment, image-level risk features are defined as: ; This mechanism preserves the global structural information of the entire image while highlighting the contribution of local high-risk image patches, thereby preventing small-area high-risk cues from being diluted by the normal mucosal background.

[0057] S6. Construct a risk phenotype vector based on the image-level risk features, lesion morphology phenotype, color phenotype, boundary texture phenotype, and overall image quality score corresponding to the same oral mucosa-related intraoral image. Then, convert ordinary oral mucosa-related intraoral images into risk phenotype vectors that have both depth features and interpretable image features, providing input for subsequent tasks.

[0058] In this embodiment, the lesion morphology phenotype is defined as:

[0059] in Indicates the first The area of ​​candidate lesions in intraoral images related to the oral mucosa. Indicates the first The proportion of candidate lesion areas to the effective mucosal area in intraoral images related to the oral mucosa (lesion area percentage). Indicates the first The maximum diameter of the candidate lesion region in intraoral images related to the oral mucosa. Indicates the first Perimeter of candidate lesion regions in intraoral images related to the oral mucosa. Indicates the first Irregularity of lesion candidate region boundaries in intraoral images related to the oral mucosa. Indicates the first The degree of fragmentation of lesions in intraoral images related to the oral mucosa.

[0060] The percentage of lesion area is defined as: ; The irregularity of the boundary of the lesion candidate region is defined as:

[0061] in To prevent extremely small constants with a denominator of zero, Pi is the mathematical constant of a circle. The larger the value, the more irregular the boundaries of the candidate lesion region.

[0062] The degree of fragmentation of lesions is defined as:

[0063] in Indicates the first The number of connected regions in the candidate regions of lesions in intraoral images related to the oral mucosa.

[0064] Color phenotype is defined as:

[0065] in Indicates the first The proportion of white keratinized areas in intraoral images related to the oral mucosa. Indicates the first The proportion of red erosions or erythema areas in intraoral images related to the oral mucosa. Indicates the first The proportion of red and white mixed in intraoral images related to the oral mucosa. Indicates the first Color heterogeneity in intraoral images related to the oral mucosa Indicates the first The proportion of dark red or hemorrhagic areas in intraoral images related to the oral mucosa.

[0066] The red-white mixing ratio is defined as:

[0067] in For the first The area of ​​the red region in intraoral images related to the oral mucosa; For the first The area of ​​white keratinized region in intraoral images related to the oral mucosa.

[0068] Color heterogeneity is defined as:

[0069] in This indicates the calculation of color variance; This indicates that the pixels are converted to the Lab color space; Indicates the first In intraoral images related to the oral mucosa Pixel value at the location; For the first Mask of candidate lesion regions in intraoral images related to the oral mucosa.

[0070] Boundary texture phenotype is defined as:

[0071] in Indicates the first Surface roughness of intraoral images related to the oral mucosa Indicates the first Blurred or gradient boundary features in intraoral images related to the oral mucosa. Indicates the first Scoring of erosion-like texture in intraoral images of the oral mucosa. Indicates the first Keratin-like texture scoring of intraoral images related to the oral mucosa.

[0072] Surface roughness is defined as:

[0073] in For the first Masking of candidate lesion regions in intraoral images related to the oral mucosa; This indicates that the texture contrast is calculated based on the gray-level co-occurrence matrix (GLCM), and the higher the value, the more obvious the surface texture undulation of the lesion candidate area.

[0074] The boundary blurring or boundary gradient feature is defined as:

[0075] in A larger value indicates a clearer boundary, while a smaller value indicates a blurry boundary. For the first Gradient amplitude of the lesion candidate region mask in intraoral images related to the oral mucosa; This indicates the calculation of the boundary length.

[0076] The definition of erosion-like texture score is:

[0077] in A higher value indicates a more pronounced degree of erosion or depression. For the first The total number of pixels in the lesion candidate region mask of intraoral images related to the oral mucosa; To use the candidate lesion region as a mask for pixels Centered on the grayscale or luminance channel image, the radius of a local neighborhood window is selected; For the first Coordinates within the mask of candidate lesion regions in intraoral images related to the oral mucosa; This represents the grayscale or luminance channel pixel value at the corresponding location; This refers to the StdDev function.

[0078] The formula for calculating the keratinized texture score is:

[0079] in A higher value indicates a more pronounced keratinized texture; Indicates the first The high-frequency response of the candidate lesion region mask in intraoral images related to the oral mucosa is obtained by applying high-frequency filtering or wavelet transform.

[0080] Through the above-mentioned structured phenotypic encoding, this invention converts ordinary intraoral images into risk phenotypic vectors that combine depth features and interpretable imaging features, providing input for subsequent patient-level fusion and cascade assessment.

[0081] In practice, the same patient may have multiple intraoral images, and these images differ significantly in quality, shooting angle, lesion visibility, and risk response. To avoid information loss or interference from low-quality images caused by simple averaging or maximum value aggregation, this invention designs a quality-risk-uncertainty joint weighted multi-image aggregation mechanism to obtain patient-level risk phenotypes. The specific method includes the following steps: B1. Based on the patient's highest risk response intensity For each image patch, calculate the image-level risk response score for the patient's images that are divided into image patches. The expression is as follows:

[0082] in, For the first The image-level risk response score of an image is divided into image patches and used to characterize the overall response level of high-risk image patches in the image. The aggregation weights are for the corresponding image patches; B2. Based on at least one feature value and / or at least one lesion morphology phenotype of intraoral images related to the oral mucosa, construct an image-level lesion visibility feature vector, and then calculate the lesion visibility score, the expression of which is:

[0083] in For the first The lesion visibility score of oral mucosa-related intraoral images is used to characterize the observability of lesion candidate regions in the oral mucosa-related intraoral images. For the first Image-level lesion visibility feature vectors corresponding to intraoral images of the oral mucosa; This is the visibility mapping weight vector; For bias terms; The Sigmoid function is represented; the selected eigenvalues ​​and / or lesion morphology phenotypes can be one or more of the following: lesion candidate area ratio, effective mucosal area ratio, image sharpness, occlusion score, and reflection interference score; lesion visibility score. This is used to adjust the contribution weight of the intraoral image during patient-level multi-image aggregation, and is different from the image block-level risk response strength. ; B3. Obtain the first [prediction] from prediction entropy, model integration variance, or MC Dropout. Uncertainty score of intraoral images related to the oral mucosa ; B4. Define the joint weighted multi-graph aggregation weight of quality, risk, and uncertainty. :

[0084] in Indicates the natural index; A learnable weight vector; For patients Index of intraoral images related to the oral mucosa; For the first Overall image quality score of intraoral images related to the oral mucosa The weights; For the first Image-level risk response score of intraoral images related to the oral mucosa. The weights; No. Visibility of lesions in intraoral images related to the oral mucosa The formula assigns weights to images with high quality, strong lesion visibility, high risk response, and low uncertainty, thus contributing more to patient-level risk phenotypes, while reducing the weight of blurred, occluded, reflective, or highly uncertain images.

[0085] B5. Define patient-level risk phenotypes : .

[0086] Should It can be used to generate current high-risk anomaly risk information, generate future progress risk information, and output structured reports.

[0087] In cases where intraoral photography is performed at multiple time points, this embodiment further calculates the temporal changes in patient-level risk phenotypes. Let the patient... At the point of time Patient-level risk phenotype is The phenotypic change between adjacent time points is defined as: ; The rate of phenotypic change is defined as:

[0088] in, This indicates the phenotypic change trend per unit time. This trend can reflect dynamic image information such as lesion area expansion, changes in the red-white ratio, increased color heterogeneity, enhanced boundary irregularity, or changes in texture roughness.

[0089] In the absence of multi-time point images, it can be and Set as an empty vector or a zero vector, and subsequent assessments can be based on a single point-in-time risk phenotype.

[0090] It should be noted that, in addition to weighted averaging, patient-level risk phenotype aggregation can also employ graph convolutional networks (GCN), graph attention networks (GAT), or Transformer layers to fuse multi-graph features, dynamically weighting based on image quality, uncertainty, and lesion visibility. Furthermore, multi-temporal phenotypic changes can be captured using sequence modeling methods (LSTM, Temporal Transformer, or autoregressive models) to achieve dynamic feature capture during long-term follow-up.

[0091] Example 2: This embodiment is a further extension based on Embodiment 1. Specifically, as follows: Figure 6 , Figure 7 and Figure 8 As shown, based on the method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa, this embodiment also provides an intraoral image phenotype evaluation system, which includes: The image acquisition module is used to receive one or more intraoral images and their associated information; The image preprocessing module is used to perform image preprocessing, including reading, resizing, color space conversion, normalization, and enhancement. The quality perception module is used to calculate the overall image quality score; The mucosal region and lesion candidate region extraction module is used to obtain image patch features; The risk phenotype encoding module is used to generate image-level risk features (depth features), lesion morphology phenotypes, color phenotypes, and boundary texture phenotypes, and then construct risk phenotype vectors. The multi-image aggregation module is used to dynamically weight and fuse multiple intraoral images to obtain patient-level risk phenotypes; The current risk information generation module outputs current high-risk abnormal risk information based on image-level risk features or patient-level risk phenotypes. The future progression risk information generation module is used to generate image-level risk features or patient-level risk phenotypes and output future progression risk information. The interpretable evidence mapping module is used to generate category-aware saliency maps, risk heatmaps, and lesion candidate region overlay maps; The structured report output module is used to save risk phenotypes, image quality scores, multi-image weights, risk scores, uncertainty information, and interpretable image reports.

[0092] For example, the current high-risk anomaly information includes the current risk probability and the current risk uncertainty. The current risk probability is defined as:

[0093] in, For the current risk information generation module, a trained fully connected neural network or classification head can be used to perform the corresponding data processing. Its weights are automatically optimized through supervised learning and used to convert image-level risk features or patient-level risk phenotypes. Mapped to risk scores, This represents the probability of image risk in the current high-risk abnormal state.

[0094] Current risk uncertainty is defined as:

[0095] Alternatively, a multi-model integrated variance representation can be used:

[0096] in, Indicates the number of ensemble models. This represents the ensemble average probability. Indicates the first The ensemble probability of an ensemble model.

[0097] The future progression risk information generation module outputs progression risk assessment information calculated based on intraoral images and follow-up data, which does not directly replace clinical diagnosis or treatment decisions.

[0098] When multiple time-point images exist, the risk of future progress is defined as:

[0099] in, The module for generating future risk information can perform corresponding data processing through fully connected neural networks (MLP), Transformer encoders, or sequence prediction networks (such as LSTM / GRU) to model the relationship between phenotypic changes over time and future risks. Indicates the follow-up time interval or time code. Indicates the amount of phenotypic change. Indicates the rate of phenotypic change.

[0100] When only a single-time-point image exists, it can be simplified to:

[0101] Meanwhile, there is uncertainty regarding future progress:

[0102] Alternatively, model integration variance, temperature calibration error, or MC Dropout variance can be used for estimation.

[0103] This embodiment employs an uncertainty calibration cascade evaluation mechanism, rather than a simple threshold classification. Let the current risk threshold be... The risk threshold for future progress is The current uncertainty threshold is The threshold for uncertainty in future progress is Output of cascaded evaluation state Defined as:

[0104] in This indicates current high-risk and abnormal information. This indicates high-risk information regarding future developments. This indicates low-risk and stable information. This indicates a low-confidence output or a verification prompt.

[0105] To provide continuous and sortable risk information, this embodiment constructs a comprehensive risk score calibrated with quality and uncertainty. The basic condition risk score is defined as follows: ; Further define joint uncertainty:

[0106] in This represents the combined weight of current risk uncertainty and future development risk uncertainty.

[0107] The patient-level mean quality score is defined as: ; The final quality and uncertainty calibration comprehensive risk score is defined as follows:

[0108] in and These are the weights for the uncertainty penalty term and the image quality calibration term, respectively. This embodiment can... Used for patient-level risk ranking, report display, and priority prompts for follow-up review.

[0109] In this embodiment, when training the model of the intraorbital imaging phenotypic generation method and its subsequent output modules, a multi-objective joint training strategy is designed to address the problem of a limited number of high-risk samples, rather than using only ordinary cross-entropy loss.

[0110] The overall loss function is defined as:

[0111] in This indicates the loss generated by the current risk information. This indicates the potential loss from generating risk information regarding future progress. This indicates loss of phenotypic consistency at risk. Indicates risk ranking loss. This indicates a counterfactual risk phenotype-enhanced loss. Indicates probability calibration loss. arrive These are learnable or preset hyperparameters used to control the relative weight of each auxiliary loss in the total loss.

[0112] For imbalanced tasks, Focal loss can be used:

[0113] in For category weights, Focusing coefficient for difficult samples This represents the predicted probability corresponding to the true category.

[0114] Risk ranking loss is:

[0115] in and These represent the risk scores for high-risk and low-risk samples, respectively. This represents the sorting interval.

[0116] If three risk states are used, then the following can be defined:

[0117] in, , , These represent the risk scores of current high-risk anomalies, high-risk future developments, and low-risk stable samples, respectively.

[0118] This embodiment aims to enhance the learning of a small number of high-risk phenotypes by constructing counterfactual risk phenotype samples. Let the low-risk sample phenotype be... High-risk sample centers are Low-risk sample centers are The high-risk phenotypes are: ; Constructing counterfactual phenotypes:

[0119] in Let be the perturbation coefficient. The counterfactual consistency loss is defined as:

[0120] in This is the original risk phenotype vector. This is a counterfactual risk phenotype generated along the high-risk direction. This is a comprehensive risk scoring function used to map risk phenotype vectors to continuous risk scores, which can be generated by the current risk information generation module. Future progress risk information generation module Or a combination of the two, To improve the expected risk interval, this loss ensures that when the lesion phenotype changes along the high-risk direction, the output risk score also increases accordingly, thereby improving the system's ability to identify a small number of high-risk samples and its predictive stability.

[0121] To improve the consistency between the output probability and the true risk label, the probability calibration loss can be adjusted using the BrierScore calibration loss.

[0122] in Indicates a label, This represents the model's output probability. The probability calibration loss can improve the reliability of the output risk probability.

[0123] It should be noted that when performing counterfactual enhancement and hard sample focusing operations, high-risk synthetic samples can be constructed using generative adversarial networks (GANs), variational autoencoders (VAEs), or hybrid sampling strategies. These samples can be combined with hard sample weighting, ranking constraints, and contrastive learning to enhance the model's ability to identify a small number of high-risk phenotypes.

[0124] To enhance the interpretability of system results, this embodiment constructs a category-aware image patch saliency risk evidence map. For image patches... Its characteristics are The target risk category is The category prototype or classification weight is The image patch attention weights are The contribution of the image patch to the target category is then defined as:

[0125] in This represents the similarity function.

[0126] The category-aware saliency map is defined as:

[0127] in Represents pixels Is it located in an image block? Inside.

[0128] Based on this, the system can generate current high-risk anomaly risk evidence maps, future progression risk evidence maps, and low-risk stable evidence maps, respectively, to display key image regions of interest to the model under different risk states. This saliency map can be output together with the original intraoral image, lesion candidate regions, image quality maps, and multi-image aggregation weights to form an interpretable risk evidence map.

[0129] It should be noted that, in addition to Class-Aware Patch Saliency, Grad-CAM++, Layer-wise Relevance Propagation (LRP), Attention Rollout, or TransformerAttention Map can be combined with patches to generate multi-scale, category-aware, patient-level interpretable heatmaps, while also supporting the overlay of original images and visualization of risk phenotypes.

[0130] In practice, the phenotypic vector can be expanded into a multidimensional vector, including local texture statistics, color distribution entropy, boundary curvature, lesion fragmentation index, image quality, and uncertainty information. The output is not limited to risk scores, but can also generate sortable continuous risk indices, image-level confidence heatmaps, patient-level comprehensive reports, and exportable data files.

[0131] The system can be deployed on local GPU workstations, edge devices, cloud servers, or hybrid architecture environments; inference can adopt streaming batch, real-time online, or interactive semi-automatic modes; it supports single image, batch image, and patient-level multi-image input, ensuring robustness across devices, environments, and shooting conditions.

[0132] For example, this system can ultimately output a structured risk information report. The report content includes: (1) Input basic image information, including image number, patient number, shooting time and shooting location; (2) Image quality score, including sharpness, exposure, reflection, occlusion, mucosal proportion and lesion visibility; (3) Lesion risk phenotype, including area, area percentage, maximum diameter, perimeter, boundary irregularity, red and white mixing ratio, color heterogeneity and texture roughness; (4) Multi-image aggregation weighting, showing the contribution of each image in patient-level assessment; (5) Current high-risk and abnormal information ; (6) Information on future development risks ; (7) Comprehensive risk score ; (8) Uncertainty information , and ; (9) Low confidence review prompt; (10) Category perception saliency map, risk response heat map and lesion candidate region overlay map; (11) Structured tables, image reports and exportable result files.

[0133] This system not only outputs current and future risk probabilities but also generates uncertainty information for model predictions and performs cascaded evaluations based on uncertainty calibration. For images with low confidence or high uncertainty, the system can output verification prompts, improving the verifiability and safety of the model in clinical applications.

[0134] This system combines image patch feature attention and classification weights to generate a saliency risk map, which can separately display current high-risk abnormalities, high-risk areas for future progression, and low-risk stable areas, providing interpretable image evidence for doctors or researchers to assist in clinical review and research analysis.

[0135] In one application example of the present invention, such as Figure 9 As shown, this demonstrates the performance and interpretability of identifying current high-risk lesions in oral mucosal samples. Figure 9 Part A shows the ROC curve, reflecting the classification ability of this invention in distinguishing between current high-risk lesion samples and low-risk samples. The area under the curve is AUC = 0.9645. Part B is a table of quantitative indicators, including Accuracy, F1, Sensitivity, Specificity, AUC, and AP, and provides 95% confidence intervals for comprehensively quantifying the classification performance of the model. Part C shows the original image of a typical oral mucosa sample overlaid with a Class-Aware Patch Saliency (CAPS) image. The attention heatmap shows the contribution intensity of the model to each patch when identifying the current high-risk state, with the yellow area representing the key lesion area that the model pays high attention to.

[0136] In one application example of the present invention, such as Figure 10As shown, this section presents the performance and interpretability of predicting future risks for non-high-risk oral mucosal samples. Part A (ROC curve) demonstrates the model's classification ability in distinguishing between high-risk and low-risk lesion samples, with the area under the curve (AUC = 0.8882), used to quantify the method's performance on the example data. Part B (quantification index table) includes Accuracy, F1, Sensitivity, Specificity, AUC, and AP, providing 95% confidence intervals to comprehensively quantify the model's classification performance in the example application. Part C (Class-Aware Patch Saliency interpretability heatmap) displays the original image and CAPS overlay of typical current non-high-risk oral mucosal samples. The attention heatmap shows the model's contribution to each patch when predicting future risks, with yellow areas representing key lesion areas of high model interest.

[0137] In summary, this invention innovatively introduces a multi-dimensional image quality scoring system to comprehensively quantify the sharpness, brightness, exposure, reflection, occlusion, effective mucosal ratio, and lesion visibility of intraoral images. The system can dynamically weight low-quality images during multi-image aggregation and risk phenotype generation, enabling the model to adapt to inputs under different devices, lighting conditions, and shooting angles, thereby improving the stability and reliability of feature extraction.

[0138] By generating multi-scale mucosal region and lesion candidate region masks and weighting input features in a probabilistic manner, the model can effectively ignore non-target background and occluded areas when extracting risk phenotypes, preserving local lesion features while also taking into account global mucosal structure. This module innovatively achieves priority focus on high-information-density regions in complex intraoral images, providing accurate input for subsequent high-risk patch risk perception.

[0139] Image patches of different scales are divided within the candidate lesion region. The risk response of each image patch is calculated, and high-risk patches are selected for weighted aggregation to form image-level risk features. This design ensures that local high-risk visual cues (such as mixed red and white, erosion, thickened keratin, and abnormal boundaries) are fully captured, while preserving global contextual information, effectively avoiding the dilution of information in small-area lesions.

[0140] Information such as image features, lesion morphology, color features, boundary texture, image quality, and follow-up time is encoded into structured risk phenotypic vectors, and phenotypic consistency constraints are introduced to ensure that deep visual features are consistent with interpretable phenotypes. This design breaks through the traditional black-box prediction method and achieves interpretable, verifiable, and quantifiable risk representation output.

[0141] For multiple intraoral images of the same patient, the system dynamically weights and fuses them based on image quality, lesion visibility, high-risk response, and model uncertainty to generate a patient-level risk phenotype. Compared to traditional averaging or maximum aggregation, this method ensures that high-quality, high-risk response, and low-uncertainty images contribute more to the final phenotype, thereby significantly improving the stability and reliability of patient-level risk assessment.

[0142] For intraoral photography during follow-up, the system can capture the temporal trends and rates of change in lesion risk phenotypes, and dynamically analyze the evolutionary characteristics of lesions. This module innovatively achieves quantitative characterization of long-term changes in lesions, providing technical support for long-term clinical follow-up, risk trend prediction, and intervention effect evaluation.

[0143] To address the problem of a scarcity of a few high-risk samples, this invention constructs a counterfactual risk phenotype and a hard sample-focused training strategy to enhance the model's ability to identify a few high-risk phenotypes and improve generalization and prediction stability in class imbalance and small sample scenarios.

[0144] This system integrates intraoral image input, quality perception, multi-scale regional constraints, high-risk patch feature extraction, risk phenotype encoding, multi-image aggregation, multi-temporal dynamic modeling, uncertainty calibration, cascade assessment, and interpretable risk evidence output. It supports single-image, batch image, and patient-level multi-image inference, forming a complete technical closed loop. It can be deployed in scientific research analysis, clinical follow-up management, and decision support systems, achieving high stability, high reliability, and verifiability.

Claims

1. A method for generating intraoral image phenotypes of potentially malignant diseases of the oral mucosa, characterized in that, Includes the following steps: Acquire intraoral images related to the oral mucosa, calculate the feature values ​​of the intraoral images related to the oral mucosa in different dimensions, and then obtain the corresponding overall image quality score based on the feature values ​​in different dimensions. Effective mucosal region masks and lesion candidate region masks are generated from intraoral images related to the oral mucosa, and then the corresponding effective mucosal region constraint images and lesion candidate region constraint images are obtained. Obtain the global features, effective mucosal region features, and lesion candidate region features corresponding to the intraoral images related to the oral mucosa, the effective mucosal region constraint images, and the lesion candidate region constraint images, respectively. Oral mucosa-related intraoral images or lesion candidate regions are divided into image patch sets of different scales. The features of each image patch are obtained, and then the risk response intensity of each image patch is calculated. The features of the K image patches with the highest risk response intensity are weighted and aggregated to obtain the local risk response features. Global features, effective mucosal region features, lesion candidate region features, and local risk response features are stitched together to form image-level risk features; Based on the image-level risk features, lesion morphology phenotype, color phenotype, boundary texture phenotype, and overall image quality score corresponding to the same oral mucosa-related intraoral image, a risk phenotype vector is constructed. This vector is then used to convert ordinary oral mucosa-related intraoral images into risk phenotype vectors that combine depth features and interpretable image features, providing input for subsequent tasks.

2. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 1, characterized in that, The specific method for acquiring intraoral images related to the oral mucosa, calculating the feature values ​​of these images in different dimensions to construct feature vectors, and then obtaining the corresponding overall image quality score based on these feature vectors includes the following steps: Acquire intraoral images using a dental camera, a regular digital camera, a mobile phone, an endoscope, or other image acquisition devices; The obtained intraoral images were subjected to standardized preprocessing, including color space conversion, size uniform adjustment, oral mucosa region cropping, pixel normalization, background edge cropping, and tensor transformation, to obtain intraoral images related to the oral mucosa. For any intraoral image related to the oral mucosa, calculate its feature values ​​in different dimensions, including sharpness score, underexposure, overexposure, saliva reflection / highlight interference, occlusion score, effective mucosal area ratio, and visibility of lesion candidate areas. The feature values ​​of the same intraoral image related to the oral mucosa in different dimensions are used to form a quality feature vector, which is then input into a linear mapping layer to obtain a weighted result in the image quality space. The weighted result is then normalized to [0,1] using the Sigmoid function to obtain the overall image quality score corresponding to the intraoral image related to the oral mucosa.

3. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 1, characterized in that, The specific method for generating effective mucosal region masks and lesion candidate region masks from intraoral images related to the oral mucosa, and then obtaining the corresponding effective mucosal region constraint images and lesion candidate region constraint images, includes the following steps: Intraoral images related to the oral mucosa are used as input to a pre-trained mucosal region extraction module to obtain the corresponding effective mucosal region mask; intraoral images related to the oral mucosa are used as input to a pre-trained lesion segmentation model to obtain the corresponding lesion candidate region mask. The intraoral images related to the oral mucosa are multiplied pixel by pixel with the corresponding effective mucosal region mask to obtain the effective mucosal region constraint image; the intraoral images related to the oral mucosa are multiplied pixel by pixel with the corresponding lesion candidate region mask to obtain the lesion candidate region constraint image.

4. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 3, characterized in that, Specific methods for obtaining global features, effective mucosal region features, and lesion candidate region features corresponding to intraoral images of oral mucosa, constrained images of effective mucosal regions, and constrained images of lesion candidate regions, respectively, include: The oral mucosa-related intraoral images, effective mucosal region constraint images, and lesion candidate region constraint images are used as inputs to the pre-trained visual encoder, respectively, to obtain global features, effective mucosal region features, and lesion candidate region features. This reduces the influence of non-lesion background on the output and improves the specificity and interpretability of image risk phenotypes.

5. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 1, characterized in that, The method for dividing intraoral images or candidate lesion regions related to the oral mucosa into image patch sets of different scales, obtaining the features of each image patch, calculating the risk response intensity of each image patch, and weighted aggregating the features of the K image patches with the highest risk response intensity to obtain the local risk response features includes the following steps: Oral mucosa-related intraoral images or lesion candidate regions are divided into image patch sets of different scales; Each image patch is used as input to a pre-trained visual encoder to obtain the features of the corresponding image patch. The risk response intensity for each image patch is calculated based on its features, and the expression is as follows: in Indicates the first The images divided into image blocks are in the first... The first scale Risk response intensity of each image patch; Represents the Sigmoid function; This is a linearly mapped weight vector used to map image patch features to the risk response space, reflecting the importance of each feature dimension to the risk score; superscript Indicates transpose; For the first The images divided into image blocks are in the first... The first scale Features of an image patch; This is a bias term used to adjust the baseline value after linear mapping; The features of the K image patches with the highest risk response intensity are weighted and aggregated to obtain the local risk response features, the expression of which is: in For the first Local risk response characteristics of an image divided into image patches; Image patch features The weights; It is the set of the K image patches with the highest risk response intensity.

6. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 1, characterized in that, The morphological phenotype of the lesion includes one or more of the following: area of ​​the candidate lesion region, proportion of the candidate lesion region to the effective mucosal area, maximum diameter of the candidate lesion region, perimeter of the candidate lesion region, irregularity of the boundary of the candidate lesion region, and degree of fragmentation of the lesion; among which: The formula for calculating the proportion of the candidate lesion area to the effective mucosal area is as follows: in Indicates the first The proportion of candidate lesion areas to effective mucosal areas in intraoral images related to the oral mucosa; For the first Masking of candidate lesion regions in intraoral images related to the oral mucosa; For the first Effective mucosal region mask for intraoral images related to the oral mucosa; To prevent extremely small constants with a denominator of zero; The expression for calculating the irregularity of the lesion candidate region boundary is as follows: in For the first Irregularity of lesion candidate region boundaries in intraoral images related to the oral mucosa; For the first Perimeter of candidate lesion regions in intraoral images related to the oral mucosa; Pi; For the first The area of ​​candidate lesions in intraoral images related to the oral mucosa; The formula for calculating the degree of fragmentation of the lesion is: in For the first The degree of fragmentation of lesions in intraoral images related to the oral mucosa; Indicates the first The number of connected regions in the candidate regions of lesions in intraoral images related to the oral mucosa.

7. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 1, characterized in that, Color phenotypes include one or more of the following: proportion of white keratotic areas, proportion of red erosion or erythema areas, proportion of mixed red and white areas, color heterogeneity, and proportion of dark red or hemorrhagic areas; among which: The formula for calculating the red-white mixing ratio is: in For the first The proportion of red and white mixed in intraoral images related to the oral mucosa; For the first The area of ​​the red region in intraoral images related to the oral mucosa; For the first The area of ​​white keratinized regions in intraoral images related to the oral mucosa; To prevent extremely small constants with a denominator of zero; The expression for calculating color heterogeneity is: in For the first Color heterogeneity in intraoral images related to the oral mucosa; This indicates the calculation of color variance; This indicates that the pixels are converted to the Lab color space; Indicates the first In intraoral images related to the oral mucosa Pixel value at the location; For the first Mask of candidate lesion regions in intraoral images related to the oral mucosa.

8. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 1, characterized in that, Boundary texture phenotypes include one or more of the following: surface roughness, boundary blurring or boundary gradient features, erosion-like texture score, and keratinization-like texture score; wherein: The expression for calculating surface roughness is: in For the first Surface roughness of intraoral images related to the oral mucosa; For the first Masking of candidate lesion regions in intraoral images related to the oral mucosa; This indicates that the texture contrast is calculated based on the gray-level co-occurrence matrix, and the higher the value, the more obvious the surface texture undulation of the lesion candidate area; The calculation expression for boundary ambiguity or boundary gradient features is as follows: in For the first The boundary blurring or boundary gradient features of intraoral images related to the oral mucosa are analyzed. The larger the value, the clearer the boundary; the smaller the value, the blurry the boundary. For the first Gradient amplitude of the lesion candidate region mask in intraoral images related to the oral mucosa; This indicates the calculation of the boundary length. The formula for calculating the erosion-like texture score is: in For the first The erosion texture score of intraoral images related to the oral mucosa is used; the higher the value, the more obvious the degree of erosion or depression. For the first The total number of pixels in the lesion candidate region mask of intraoral images related to the oral mucosa; To use the candidate lesion region as a mask for pixels Centered on the grayscale or luminance channel image, the radius of a local neighborhood window is selected; For the first Coordinates within the mask of candidate lesion regions in intraoral images related to the oral mucosa; This represents the grayscale or luminance channel pixel value at the corresponding location; For the StdDev function; The formula for calculating the keratinized texture score is: in For the first The keratinization texture score of intraoral images related to the oral mucosa is given by the larger value, the more obvious the keratinization texture is. Indicates the first The high-frequency response of the candidate lesion region mask in intraoral images related to the oral mucosa is obtained by applying high-frequency filtering or wavelet transform.

9. The method for generating intraoral image phenotypes of potential malignant diseases of the oral mucosa according to claim 5, characterized in that, When a patient has multiple intraoral images related to the oral mucosa, a patient-level risk phenotype is generated for that patient. The specific method includes the following steps: Based on the patient's highest risk response intensity For each image patch, calculate the image-level risk response score for the patient's images that are divided into image patches. The expression is as follows: in, For the first The image-level risk response score of an image is divided into image patches and used to characterize the overall response level of high-risk image patches in the image. The aggregation weights are for the corresponding image patches; Based on at least one feature value and / or at least one lesion morphology phenotype of intraoral images related to the oral mucosa, an image-level lesion visibility feature vector is constructed, and then the lesion visibility score is calculated, the expression of which is: in For the first The lesion visibility score of oral mucosa-related intraoral images is used to characterize the observability of lesion candidate regions in the oral mucosa-related intraoral images. For the first Image-level lesion visibility feature vectors corresponding to intraoral images of the oral mucosa; This is the visibility mapping weight vector; For bias terms; Represents the Sigmoid function; Get the Uncertainty score of intraoral images related to the oral mucosa ; Based on lesion visibility score Image overall quality score Image-level risk response score Uncertainty score and risk phenotype vector By performing quality-risk-uncertainty joint weighted multi-graph aggregation, the patient-level risk phenotype is obtained, and its expression is: in Patient-level risk phenotype; For patients The total number of intraoral images related to the oral mucosa; Weighted multi-graph aggregation weights based on quality, risk, and uncertainty; For patients The Risk phenotype vectors corresponding to intraoral images of the oral mucosa; Indicates the natural index; For learnable weight vectors; superscript Indicates transpose; For the first Overall image quality score of intraoral images related to the oral mucosa The weights; For the first Image-level risk response score of intraoral images related to the oral mucosa. The weights; No. Visibility of lesions in intraoral images related to the oral mucosa The weights; Part Uncertainty score of intraoral images related to the oral mucosa The weights; For patients Index of intraoral images related to the oral mucosa.

10. The method for generating intraoral image phenotypes of potentially malignant diseases of the oral mucosa according to claim 4 or 5, characterized in that, The training method for a visual encoder includes the following steps: Image-level risk features are used as input to the phenotypic prediction head to obtain the corresponding predicted phenotypic vector. ; The vector composed of lesion morphology phenotype, color phenotype, and boundary texture phenotype is denoted as the reference phenotype vector. ; Construct a phenotypic consistency loss based on predicted phenotypic vectors and reference phenotypic vectors. Its expression is: in Indicates the L1 distance; These are the weighting coefficients; It is a correlation function; Based on phenotypic consistency loss The visual encoder is trained using backpropagation until it converges, so that the features extracted by the trained visual encoder are consistent with the interpretable morphological, color and texture phenotypes.

Images