Oral precancerous lesion intelligent follow-up system and method

Abstract

This invention discloses an intelligent follow-up system and method for precancerous oral lesions, relating to the fields of smart healthcare and computer vision technology. The system includes: an intraoral standardized image acquisition device, a patient smart terminal, a processing terminal, and a medical care terminal. The acquisition device provides a two-dimensional physical reference for scale and chromaticity. The processing terminal acquires standardized images from multiple time points, performs normalization, registration at the same location, and lesion segmentation to extract area and chromaticity feature sequences. By calculating the two sets of feature sequences, a temporal correlation is established representing the leading or lagging status of chromaticity and area changes. Based on this, the phase matching category is determined in conjunction with the current change, and a differential weighted fusion is performed using the corresponding weighting factor to obtain a lesion progression score. Finally, risk stratification is output, and warning information is pushed to the medical care terminal. This invention effectively overcomes the systemic recognition bottleneck in the follow-up of this type of lesion.

Description

Technical Field

[0001] This invention relates to the fields of smart healthcare and computer vision technology, specifically to a smart follow-up system and method for precancerous oral lesions. Background Technology

[0002] Oral leukoplakia, oral erythema, and other oral mucosal lesions with a risk of malignancy usually require long-term, continuous, and dynamic observation in order to identify potential signs of malignancy in the early stages of disease progression, thereby gaining a window of opportunity for subsequent diagnosis and treatment decisions.

[0003] With the popularization of smart terminals and the development of telemedicine, long-term follow-up of such lesions has gradually extended from single outpatient visits to a management path that combines home image collection and remote quantitative assessment. On this basis, a preliminary processing framework has been formed in which patients regularly collect oral lesion images at home, the backend quantitatively extracts various indicators such as lesion morphology and color reflected in the images, and assesses the degree of disease progression accordingly.

[0004] However, when patients take images of oral lesions at home, there is generally a lack of uniform perspective, distance, lighting, and measurement references, making it difficult to directly compare images collected at different times. This phenomenon has long been regarded as the main reason affecting the reliability of remote follow-up.

[0005] Even with standardized data collection and stable extraction of various lesion indicators, the assessment of disease progression at the back end has long relied on the premise that all lesion changes obtained in the same follow-up are assumed to reflect the disease progression trend simultaneously. This premise is not universally valid in clinical practice. Changes in some indicators often appear significantly earlier than others, while in other cases, the order of changes in indicators deviates from the previous pattern. In the former case, the early change signals that appear first are diluted by other indicators that have not yet followed, resulting in long-term insensitivity of the assessment results to early progression. In the latter case, combinations of indicators with abnormal order cannot be effectively identified because they do not conform to the existing assessment assumptions, causing atypical progression signals that should be given attention to be continuously missed.

[0006] Therefore, how to ensure that early progression signals appearing in individual dimensions are not masked by the lag in other dimensions during remote follow-up of such lesions, thereby avoiding the dulling of the systematic identification of early evolution, is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides an intelligent follow-up system and method for oral precancerous lesions.

[0008] To achieve the above objectives, the technical solution of the present invention is as follows:

[0009] In a first aspect, the present invention discloses an intelligent follow-up system for precancerous oral lesions, comprising:

[0010] Standardized intraoral data collector, patient smart terminal, processing terminal and medical care terminal used in conjunction with the standardized intraoral data collector;

[0011] The intraoral normalized acquisition unit includes at least a scale structure for providing scale reference in the acquired image and a chromaticity reference structure for providing chromaticity reference;

[0012] The processing terminal communicates with the patient's smart terminal and the medical staff's terminal, and the processing terminal is configured as follows:

[0013] Multiple standardized images of the same patient were acquired at different follow-up time points using a patient smart terminal with the assistance of an intraoral standardized image acquisition device. Each of the multiple standardized images carries scale reference information and colorimetric reference information.

[0014] Based on the scale reference information and color reference information it carries, normalization, same-site registration and lesion segmentation are performed on multiple frames of standardized images to obtain the lesion area feature sequence and lesion color feature sequence of the same patient at different follow-up time points.

[0015] Based on the lesion area feature sequence and the lesion color feature sequence, the temporal correlation is determined. The temporal correlation is used to characterize the leading or lagging status of the time point of change of lesion color feature relative to the time point of change of lesion area feature.

[0016] Obtain the changes in lesion color and lesion area at the current follow-up time point, and determine the phase matching category at the current follow-up time point by combining the temporal correlation.

[0017] Using weighting factors corresponding to the phase-matching category, weighted fusion processing was performed on the changes in lesion color and lesion area to obtain the lesion progression score of the same patient at the current follow-up time.

[0018] Risk stratification results are output based on lesion progression scores, and warning information corresponding to risk stratification results that meet preset warning conditions is sent to medical staff terminals.

[0019] Secondly, this invention discloses an intelligent follow-up method for precancerous lesions of the oral cavity, comprising the following steps:

[0020] Multiple standardized images of the same patient were acquired at different follow-up time points using a patient smart terminal with the assistance of an intraoral standardized image acquisition device. Each of the multiple standardized images carries scale reference information and colorimetric reference information.

[0021] Based on the scale reference information and color reference information it carries, normalization, same-site registration and lesion segmentation are performed on multiple frames of standardized images to obtain the lesion area feature sequence and lesion color feature sequence of the same patient at different follow-up time points.

[0022] Based on the lesion area feature sequence and the lesion color feature sequence, the temporal correlation is determined. The temporal correlation is used to characterize the leading or lagging status of the time point of change of lesion color feature relative to the time point of change of lesion area feature.

[0023] Obtain the changes in lesion color and lesion area at the current follow-up time point, and determine the phase matching category at the current follow-up time point by combining the temporal correlation.

[0024] Using weighting factors corresponding to the phase-matching category, weighted fusion processing was performed on the changes in lesion color and lesion area to obtain the lesion progression score of the same patient at the current follow-up time.

[0025] Risk stratification results are output based on lesion progression scores, and warning information corresponding to risk stratification results that meet preset warning conditions is sent to medical staff terminals.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. By using an intraoral standardized acquisition device to simultaneously carry both scale and color references, images acquired at different follow-up points are directly comparable in terms of geometry and color. The correlation calculation under time misalignment is performed on the lesion area feature sequence and lesion color feature sequence formed by each follow-up of the patient to determine the temporal correlation relationship. Then, the phase matching category is determined by combining the color change and area change of the current follow-up. The change is weighted and fused using a weight factor corresponding to the phase matching category, so that the progression signal that appears in individual dimensions in the early stage is no longer diluted by other dimensions that have not yet been followed up. This overcomes the long-standing systematic recognition passivation bottleneck in remote follow-up of this type of lesion.

[0028] 2. The phase matching category determined by this invention can identify atypical combinations that violate the patient's existing time sequence pattern by not synchronously predicting the color at the time point when the area has changed significantly as an independent category. This allows such combinations to obtain a higher weight factor than typical cases that evolve steadily according to existing patterns in differentiated weighted fusion. In addition, with the bypass triggering mechanism that bypasses the conventional scoring threshold determination under this category, early warning information containing biopsy review or referral trigger prompts is directly generated. This avoids the continuous omission of atypical signals that may indicate a fundamental change in the disease progression pattern due to the reliance on a unified threshold determination in existing technologies. This significantly improves the ability to capture early signs of malignant transformation of such diseases.

[0029] 3. The temporal correlation determined by this invention is calculated based on two feature sequences formed from the patient's previous follow-ups. The corresponding preset change threshold can also be adaptively determined based on the patient's stable period fluctuation level, so that the scoring parameters and judgment criteria are adapted to the individual patient's lesion evolution pattern. This overcomes the problem of systematic mistreatment of both ends of the patient's condition caused by the use of uniform scoring parameters for all patients in the prior art, which results in small lesions being submerged and large lesions being frequently falsely alarmed. Without increasing the patient's data collection burden and hardware costs, it achieves individualized adaptation of remote follow-up assessment. Attached Figure Description

[0030] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a flowchart illustrating the system architecture of Embodiment 1 of the present invention;

[0032] Figure 2 This is a schematic diagram illustrating the principle of the risk stratification and bypass triggering mechanism in Embodiment 1 of the present invention;

[0033] Figure 3 This is a timing diagram of the data flow in Embodiment 1 of the present invention;

[0034] Figure 4 This is a schematic diagram of the intraoral standardized data collector in Embodiment 1 of the present invention;

[0035] Figure 5 This is an overall block diagram of the method in Embodiment 2 of the present invention.

[0036] The diagram shows: 1. Terminal fixing clamp; 2. Color reference structure; 3. Tongue pressure auxiliary component; 4. Annular supplementary light component; 5. Mouth corner opening structure; 6. Scale structure. Detailed Implementation

[0037] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0038] Example 1:

[0039] like Figures 1-4As shown, an intelligent follow-up system for precancerous lesions of the oral cavity includes: an intraoral standardized acquisition device, a patient smart terminal used in conjunction with the intraoral standardized acquisition device, a processing terminal, and a medical care terminal; the intraoral standardized acquisition device includes at least a scale structure 6 for providing scale reference in the acquired images and a colorimetric reference structure 2 for providing colorimetric reference; the processing terminal is communicatively connected to the patient smart terminal and the medical care terminal.

[0040] The intraoral standardized image acquisition device, serving as the physical basis for scale and colorimetric references in this scheme, structurally includes at least a scale structure 6 for providing scale references in the acquired images and a colorimetric reference structure 2 for providing colorimetric references. The scale structure 6 can be a non-stretchable millimeter scale strip with millimeter-level graduations. During use, it is located on the same photographic plane as the lesion to be acquired, ensuring that the ratio between the pixel length of the scale structure 6 and its actual physical length in the acquired image directly reflects the scale information of the current capture. The colorimetric reference structure 2 can be a color chart printed with multiple standard color blocks. The colors of the color blocks can include typical colors such as red, green, blue, yellow, cyan, magenta, white, and black. During use, it is also located on the same photographic plane as the lesion to be acquired, ensuring that the pixel values ​​of each color block in the acquired image directly reflect the lighting conditions and white balance state of the current capture.

[0041] Furthermore, the standardized intraoral image acquisition device may also include at least one of the following structures: a terminal fixing clamp 1 for fixing the patient's smart terminal, a ring-shaped supplementary lighting part 4 for providing ring illumination during acquisition, a mouth corner retraction structure 5 for retracting the corners of the patient's mouth during acquisition, and a tongue pressure auxiliary component 3 for pressing down on the patient's tongue during acquisition. The terminal fixing clamp 1 is used to stably clamp the patient's smart terminal in a fixed position, avoiding image blurring caused by hand shake during shooting; the ring-shaped supplementary lighting part 4 is used to provide uniform ring illumination around the camera of the patient's smart terminal, improving the problem of poor lighting conditions inside the oral cavity; the mouth corner retraction structure 5 is used to retract the corners of the patient's mouth to both sides during acquisition, expanding the visible range inside the oral cavity; the tongue pressure auxiliary component 3 is used to press down or push the patient's tongue to the side during acquisition, preventing the tongue from obscuring the lesion to be acquired. The above structures can be flexibly combined according to specific application scenarios, and this embodiment does not limit this.

[0042] The patient smart terminal is used to acquire images of intraoral lesions with the assistance of a standardized intraoral imager. It can be a portable smart device with a camera, such as a smartphone or tablet. The medical staff terminal is used to receive and process warning information pushed by the processing terminal, allowing medical staff to view risk trends and generate manual review opinions. It can be a doctor's workstation in the hospital information system or a mobile terminal connected to that workstation. The processing terminal is used to perform subsequent data processing on the images acquired by the patient smart terminal. It can be a server cluster deployed in the cloud or a workstation deployed locally in the medical institution; this embodiment does not limit this. The processing terminal communicates with the patient smart terminal and the medical staff terminal via wired or wireless network to transmit image data, feature data, and warning information.

[0043] Specifically, the processing terminal is configured to perform the processing operations of steps S1 to S6 below.

[0044] For ease of understanding, this example illustrates the process using a remote follow-up of a patient with oral leukoplakia in a department specializing in oral mucosal diseases. Assume patient P was diagnosed with homogeneous oral leukoplakia of the left buccal mucosa in January 2024, and clinical recommendations included remote follow-up every 90 days. Using a smartphone and the standardized intraoral data acquisition device provided by this system, the patient completed six follow-up data collections in January, April, July, and October 2024, and January and April 2025. The lesions remained clinically stable in the first four follow-ups. In the fifth follow-up, the buccal lesion area showed a slight darkening of color but no significant change in size. In the sixth follow-up, the buccal lesion area showed a significant increase in size. This example will use this scenario as a specific case to explain the implementation process of each step in detail.

[0045] Step S1: Acquire multiple standardized images of the same patient at different follow-up time points using the patient's smart terminal with the assistance of an intraoral standardized image acquisition device. Each of the multiple standardized images carries scale reference information and colorimetric reference information.

[0046] In this step, the standardized image refers to the oral cavity image acquired by the patient's smart terminal with the lesion to be acquired as the main subject, and simultaneously presenting the scale structure 6 and the colorimetric reference structure 2 within the same frame; the scale reference information carried refers to the calculable proportional relationship between the pixel length occupied by the scale structure 6 in the standardized image and the actual physical length of the scale structure 6, and this proportional relationship constitutes the scale reference information for this acquisition; the colorimetric reference information carried refers to the calculable deviation relationship between the pixel value of each color block in the standardized image and the standard color value of that color block, and this deviation relationship constitutes the colorimetric reference information for this acquisition.

[0047] The core purpose of this step is to strongly bind physical scale and color references to the images collected at each follow-up visit, so that the lesion features extracted from images collected at different follow-up time points in subsequent steps have geometric and color comparability across time points, laying the foundation for collaborative analysis of feature sequences formed by multiple follow-ups in the time dimension.

[0048] Specifically, according to the follow-up plan, at each follow-up visit, the patient inserts a standardized intraoral image acquisition device into their mouth, ensuring that the lesion to be acquired, the scale structure 6, and the colorimetric reference structure 2 are simultaneously within the field of view of the patient's smart terminal camera, triggering the capture of a standardized image for that follow-up visit. The patient's smart terminal uploads the acquired standardized images to a processing terminal via the network. The processing terminal archives these images into the patient's follow-up file according to the patient's identifier and follow-up timestamp. When the processing terminal needs to perform subsequent analysis on the images from this follow-up visit, it retrieves multiple frames of standardized images acquired during this follow-up visit and previous follow-ups from the patient's follow-up file.

[0049] Using the follow-up example of patient P mentioned above, the processing terminal retrieves six standardized images from patient P's follow-up records collected during six follow-up visits in January, April, July, and October 2024, and January and April 2025. to Each frame of the image simultaneously displays the lesion area on the patient's left buccal mucosa, a millimeter ruler, and a color card composed of 8 color blocks.

[0050] Step S2: Based on the scale reference information and color reference information carried, normalization, same-site registration and lesion segmentation are performed on the multi-frame standardized images to obtain the lesion area feature sequence and lesion color feature sequence corresponding to the same patient at different follow-up time points.

[0051] In this step, normalization refers to the process of using the scale reference information and color reference information carried by each frame of the normalized image to normalize the image in two dimensions: geometric scale and color presentation. This process ensures that images acquired at different follow-up points achieve a unified benchmark in terms of geometric measurement and color presentation. Local registration refers to the process of spatially aligning lesion regions in multiple frames of normalized images, enabling pixels at the same physical location in images acquired at different follow-up points to be retrieved in a unified coordinate system. Lesion segmentation refers to the process of extracting the image region occupied by the lesion from the normalized and registered image.

[0052] The lesion area feature refers to the physical size of the lesion after scale normalization, representing the lesion image region obtained from lesion segmentation. In this embodiment, the lesion area feature is the actual area value obtained by multiplying the total number of pixels in the lesion image region by the actual physical area of ​​each pixel, with the unit being square millimeters. The lesion chromaticity feature refers to the color of the lesion after chromaticity normalization, representing the lesion image region obtained from lesion segmentation. In this embodiment, the lesion chromaticity feature is represented by the mean value of the red channel of all pixels in the RGB color space within the lesion image region. In other optional embodiments, the CIELab color space can also be used. The mean value of the channels or the proportion of the red channel in the sum of the three channels is not limited in this embodiment.

[0053] Furthermore, when the processing terminal performs normalization processing, it specifically includes: performing scale normalization processing on the normalized image to eliminate shooting distance differences based on the proportional relationship between the pixel length occupied by the scale structure 6 in the normalized image and the actual physical length of the scale structure 6; and performing chromaticity normalization processing on the normalized image to eliminate shooting illumination and white balance differences based on the deviation between the color block pixel values ​​presented by the chromaticity reference structure 2 in the normalized image and the pre-stored chromaticity baseline values.

[0054] Specifically, the processing terminal first performs image recognition on the ruler structure 6 in the standardized image. This recognition can be achieved through the following process: HSV color space conversion is performed on the standardized image. Based on the significant color distinguishability of the ruler structure 6 (e.g., the ruler bars use a specific hue with high saturation, contrasting with the pinkish-red of the oral mucosa), a preliminary candidate region mask for the ruler structure 6 is obtained using preset hue and saturation thresholds. After performing morphological opening operations on this mask to remove noise, the largest connected region is extracted as the ruler structure region based on a contour detection operator (such as findContours in OpenCV). This region is then fitted using a minimum bounding rectangle, with the length of the long side of the rectangle used as the pixel length occupied by the ruler structure 6 in the standardized image. Due to the actual physical length of scale structure 6 Given a known quantity (e.g., the total length of the millimeter scale bar in this embodiment is 50mm), the scale normalization coefficient for this shooting can be obtained: ;

[0055] Where k represents the actual physical length of each pixel under the shooting conditions, in millimeters per pixel; This indicates the actual physical length of scale structure 6, in millimeters. This represents the pixel length of scale structure 6 in the normalized image, in pixels. The processing terminal binds and stores this scale normalization coefficient k with the current normalized image. This is used in subsequent lesion segmentation to obtain the lesion image region. Convert the total number of pixels in the lesion image area into the actual physical area.

[0056] Pre-stored chromaticity baseline values ​​refer to the standard color values ​​that should be presented by each color block in chromaticity reference structure 2 under standard lighting conditions, which are pre-stored in the processing terminal. These pre-stored chromaticity baseline values ​​can be obtained during the initial system deployment by taking a baseline photo of chromaticity reference structure 2 under a standard D65 light source and extracting the average pixel value of each color block in the RGB color space, denoted as... ,in L represents the color patch number, and L is the total number of color patches contained in the chromaticity reference structure 2 (L=8 in this embodiment).

[0057] The processing terminal performs color block positioning on the chromaticity reference structure 2 in the standardized image: based on the preset geometric positional relationship of the chromaticity reference structure 2 relative to the scale structure 6 (pre-fixed on the collector structure), it circles the rectangular area where the color card is located in the standardized image; within this rectangular area, it follows the format of the color card (e.g., in this embodiment). The image is divided into eight equal sub-regions (in a checkerboard pattern). A square sampling window with a side length of 20 pixels is taken at the center of each sub-region. The average value of all pixels within this sampling window across the RGB three channels is taken as the pixel value of that color patch in the standardized image, denoted as . .

[0058] The processing terminal will analyze the observed pixel values ​​of each color block. With the corresponding pre-stored chromaticity baseline value Compare and solve a set of linear transformation matrix This minimizes the deviation between the transformed color block pixel values ​​and the baseline values.

[0059] ;

[0060] in, This indicates the distance from the color space of this shooting to the standard baseline color space. Linear transformation matrix; Representing vectors The square of the norm. The closed-form solution to this least-squares problem can be obtained by stacking the pixel values ​​of each color block into a matrix and then calculating the pseudo-inverse:

[0061] ;

[0062] in, and These are formed by stacking the baseline pixel values ​​and observed pixel values ​​of each color patch column by column. matrix; Indicates matrix transpose; This represents matrix inversion. The processing terminal will calculate the linear transformation matrix obtained from the solution. The color value applied to all pixels in this standardized image, that is, the color value of the pixel located at coordinates in the image. pixels at Perform transformation The color normalization process was completed to eliminate the differences in lighting and white balance between this shooting and standard lighting conditions.

[0063] Furthermore, when the processing terminal performs registration at the same point, it specifically includes: determining one frame from the multi-frame standardized images as the baseline image; for each frame in the multi-frame standardized images other than the baseline image, determining the geometric transformation relationship between the frame and the baseline image based on the dental feature points located at the edge of the dental arch in the frame and the scale reference information provided by the scale structure 6, and mapping the frame to a coordinate system unified with the baseline image based on the geometric transformation relationship.

[0064] Baseline images are typically standardized images acquired during the patient's first follow-up visit, serving as a unified reference frame for registration of all subsequent follow-up images. Dental arch feature points refer to relatively stable image features located at the incisal edges of the maxillary or mandibular dentition, such as the cusps of incisors or contact points between adjacent teeth. These feature points can be extracted from images using the SIFT or ORB feature detection operators. Because dentition typically exhibits high geometric stability over follow-up periods of several months, dental arch feature points are suitable as a reference for site-specific registration of oral images.

[0065] Specifically, the processing terminal processes the frames to be registered. With baseline image Extract feature point sets from the dental arch regions respectively. and The delineation of the dentition region can be accomplished using a pre-trained oral anatomy recognition model, or by performing edge detection on the image and then using a brightness threshold to coarsely locate the region based on the higher brightness of the tooth region compared to the mucosa region. The processing terminal performs SIFT descriptor matching on two feature point sets to obtain a set of candidate feature point pairs. The Random Sample Consensus Algorithm (RANSAC) iteratively selects the subset of feature point pairs with the highest consistency, with 1000 iterations and an interior point distance threshold of 3 pixels. Based on the interior point pair set output by RANSAC, the least squares method is used to solve for... arrive affine transformation matrix This makes it possible for any pair of feature points Approximately:

[0066] ;

[0067] in, and These are the homogeneous pixel coordinates of the feature points in the frame to be registered and the baseline image, respectively. The processing terminal is based on... Perform coordinate transformation on all pixels in the frame to be registered to complete the registration at the same point.

[0068] If the RANSAC algorithm fails to converge to a sufficient number of inliers within the upper limit of iteration (e.g., fewer than 8 pairs of inliers), the registration of the same locus is deemed a failure. At this time, the processing terminal prompts the patient's smart terminal to re-acquire the image, and does not include the failed image in subsequent feature extraction.

[0069] After normalization and registration at the same site, the processing terminal performs segmentation processing on the lesion region in each frame of the normalized image. In this embodiment, lesion segmentation is implemented using a semantic segmentation model based on a convolutional neural network, specifically using the U-Net network structure. The structure and training process of the U-Net network are described in detail in this embodiment as follows:

[0070] The overall structure of the U-Net network consists of three parts: the encoder path, the bottleneck layer, and the decoder path. The specific connection relationships are as follows:

[0071] The encoder path contains four downsampling stages, each consisting of two consecutive downsampling stages. A convolutional layer (with a stride of 1, padding of 1, followed by a batch normalization layer and a ReLU activation function) and a The system consists of a max pooling layer (with a step size of 2); the number of output channels in the four stages are 64, 128, 256, and 512, respectively.

[0072] The bottleneck layer is located after the encoder path and consists of two consecutive layers. It consists of convolutional layers (with the same convolutional parameters as above) and has 1024 output channels.

[0073] The decoder path consists of four upsampling stages, each consisting of one The deconvolutional layer (with a stride of 2), a cross-layer concatenation operation (concatenating the upsampled output with the output of the layer at the same resolution as the encoder path along the channel dimension), and two consecutive... The system consists of convolutional layers (with the same convolutional parameters as above); the number of output channels in the four stages are 512, 256, 128, and 64, respectively.

[0074] The network end is one The convolutional layer has 1 output channel, followed by a sigmoid activation function to output the probability value of each pixel belonging to the lesion.

[0075] The input to the U-Net network is a normalized image that has undergone normalization and same-site registration processing, and is uniformly scaled to... Pixel resolution; the output is a probability mask with the same resolution as the input image, where the value of each pixel is in the range of [pixel resolution]. Within the interval, it represents the probability that the pixel belongs to a lesion.

[0076] The training process for the U-Net network is as follows:

[0077] Training dataset construction: Standardized images of several patients with oral leukoplakia, oral erythema, and other oral mucosal lesions were collected (the number of images was determined based on data availability, usually no less than 500 cases). Physicians with oral mucosal disease diagnosis and treatment qualifications manually annotated the lesion areas in each frame of the image at the pixel level to obtain a binary annotation mask with the same resolution as the original image (lesion pixels were marked as 1, and other pixels were marked as 0). The dataset was divided into training set, validation set, and test set in a ratio of 7:2:1.

[0078] Data augmentation: During the training phase, perform random horizontal flipping and random rotation (angle range) on each training image. Random brightness adjustment (range) This is to improve the robustness of the model.

[0079] Loss function: The weighted sum of binary cross-entropy loss and Dice loss is used as the optimization objective.

[0080] ;

[0081] in, For pixel-level binary cross-entropy loss, This is a loss for Dice.

[0082] Training hyperparameters: The Adam optimizer is used, and the initial learning rate is set to... The batch size was set to 8, and the total number of training rounds was 100. The learning rate adopted a cosine annealing strategy. During the training process, the Dice coefficient on the validation set was monitored, and the training was terminated early when the Dice coefficient on the validation set no longer increased for 10 consecutive rounds. Finally, the model weight with the highest Dice coefficient on the validation set was retained as the deployment model.

[0083] Post-processing of output during deployment: The probability mask of the network output is binarized with a threshold of 0.5 to obtain a binary mask of the lesion; morphological opening and closing operations are performed on the mask to remove isolated small regions and fill small holes to obtain the final lesion image region.

[0084] The processing terminal extracts the lesion area features from this follow-up based on the binary mask of the lesion segmentation. Color characteristics of lesions Among them, the characteristics of lesion area The calculation method is as follows:

[0085] ;

[0086] in, This represents the total number of pixels with a value of 1 in the binary mask of the lesion; k is the scale normalization coefficient for this capture determined in the preceding step S2. The unit is square millimeters.

[0087] Lesion color characteristics The calculation method is as follows: For the color-normalized image, at the pixel position indicated by the binary mask of the lesion, calculate the mean value of the red channel in the RGB color space:

[0088] ;

[0089] in, This represents the set of pixel locations indicated by the binary mask of the lesion; This indicates the image after chroma normalization in coordinates. The red channel pixel value at that location ranges from 0 to 255; This indicates the total number of pixels in the set of pixel locations.

[0090] The processing terminal will extract data from each follow-up visit. The lesion area characteristic sequence of the patient is stored in the patient's characteristic file according to the follow-up time. Color feature sequence of lesions .

[0091] Furthermore, the lesion area feature sequence is composed of the lesion area features of the same patient at the most recent N follow-up times, and the lesion color feature sequence is composed of the lesion color features of the same patient at the most recent N follow-up times, where N is a preset positive integer not less than 4, and N expands as the number of historical follow-ups of the same patient accumulates.

[0092] Specifically, the lower limit of N is set to 4 because the subsequent step S3 requires estimating the correlation between two feature sequences under time misalignment. Statistically, when the sequence length is less than 4, the estimated time offset has a large error and is difficult to stably reflect the true temporal relationship. The upper limit of N can be set according to actual computing resources and the clinically meaningful retrospective period, such as 8 or 12, that is, only the most recent 8 or 12 follow-ups are taken to form the feature sequence, avoiding interference from the current correlation estimation due to fundamental changes in the lesion state in earlier follow-up data. When the patient's historical follow-up number has not yet reached the set upper limit of N, N is taken as the actual cumulative number of follow-ups for the patient; when the historical follow-up number exceeds the set upper limit, N is maintained at the upper limit value, and a sliding window method is used to retain only the data from the most recent N follow-ups.

[0093] Following the previous follow-up example of patient P, the upper limit of N is set to 8. The processing terminal acquires standardized images during the 6th follow-up visit of patient P. During processing, both the lesion area feature sequence and the lesion color feature sequence are composed of features corresponding to 6 follow-up visits, denoted as... and Assuming that after normalization, registration at the same site, and lesion segmentation, the processing terminal extracts the lesion area feature sequence as {120.0, 121.5, 119.8, 120.3, 122.1, 138.5} (unit: square millimeters), and the lesion color feature sequence (characterized by the mean R channel value within the lesion ROI) as {158.2, 159.0, 157.8, 158.6, 172.4, 178.6}, it can be observed that in the first four follow-ups, both sequences remained within a small fluctuation range. In the fifth follow-up, the color feature showed a significant increase (jumping from approximately 158 to 172.4), while the area feature remained stable. In the sixth follow-up, the area feature significantly increased (jumping from approximately 122 to 138.5), while the color feature continued to remain at a high level.

[0094] Using the above approach, lesion area feature sequences and lesion color feature sequences with cross-time point geometric comparability and color comparability are extracted from standardized images collected from multiple follow-up visits, laying a feature foundation for subsequent time-dimensional collaborative analysis.

[0095] Step S3: Based on the lesion area feature sequence and the lesion color feature sequence, determine the temporal correlation relationship. The temporal correlation relationship is used to characterize the leading or lagging status of the change point of the lesion color feature relative to the change point of the lesion area feature.

[0096] In this step, the temporal correlation refers to the quantitative representation of the relative order of the lesion color feature sequence and the lesion area feature sequence in the time dimension. It reflects whether the change point of the color feature appears earlier (leading), appears simultaneously (synchronous), or appears later (lagging) than the change point of the area feature for the patient, and the number of follow-up cycles spanned by the leading or lagging.

[0097] The core purpose of this step is to extract the temporal sequence of changes in lesion color and lesion area from the two feature sequences formed by the patient's previous follow-ups. This aims to break the existing technical premise that assumes all lesion change indicators synchronously reflect the disease progression, which is not in line with clinical practice. This provides a basis for subsequent steps to make an assessment of the indicator changes observed at the current follow-up time point that is consistent with this temporal sequence.

[0098] Traditional multi-indicator comprehensive assessment methods only perform weighted summation of various indicators extracted at the same follow-up time point, completely ignoring the objective law that changes in various indicators often have an inherent temporal sequence during long-term follow-up of the same patient. Therefore, they cannot identify early signals that appear earlier in individual dimensions, nor can they respond to atypical combinations that violate this temporal sequence. This invention performs correlation calculations under temporal misalignment on two feature sequences formed from the patient's various follow-ups, extracting the temporal sequence relationship between the two as a reference for the patient's subsequent assessment, thus achieving explicit modeling of this objective law.

[0099] Furthermore, when determining the temporal correlation, the processing terminal specifically includes: performing correlation calculations on the lesion area feature sequence and the lesion color feature sequence under time misalignment, obtaining the correlation strength between the lesion color feature sequence and the lesion area feature sequence at different time offsets, and the correlation strength is used to characterize the degree of synchronous change of the two feature sequences at the corresponding time offsets; the time offset in which the correlation strength reaches its maximum value and exceeds the preset correlation strength threshold is used as the quantification value of the temporal correlation.

[0100] Specifically, the processing terminal first processes the lesion area feature sequence. Color feature sequence of lesions Calculate the difference sequence between adjacent follow-up time points:

[0101] ;

[0102] in, and Let represent the area change and chromaticity change at the i-th follow-up relative to the (i-1)-th follow-up, respectively. The purpose of calculating the difference sequence is to eliminate the difference in baseline levels between the two feature sequences, so that the correlation calculation focuses on the correspondence between the changing trends.

[0103] Then, the processing terminal targets a set of preset time offsets. Each of them Calculate the offset between the lesion color difference sequence and the lesion area difference sequence. Normalized cross-correlation coefficients:

[0104] ;

[0105] in, Indicates that at a time offset of At that time, the normalized cross-correlation coefficient of the lesion color difference sequence relative to the lesion area difference sequence, i.e. the correlation strength at that time offset; This represents the offset chromaticity difference value, requiring an index. Both i and i fall within the effective index range of the difference sequence; and Let represent the mean values ​​of the two difference numerator sequences involved in the summation within the overlapping interval; and These represent the standard deviations of the two difference numerator sequences involved in the summation within the overlapping interval; This represents the number of sample pairs within the overlapping interval, i.e., the number of i that satisfy the above index constraints. This is the upper limit for the search of time offsets, and its value usually does not exceed [a certain value]. This ensures that there are enough samples in the overlapping intervals to participate in the calculation, for example, when N=8. Option 3 is acceptable.

[0106] The range of values ​​is .when When the value is close to +1, it indicates that at that time offset, the chromaticity difference sequence and the area difference sequence are highly synchronized in the direction of change, that is, the changes in chromaticity features and area features exhibit strong positive synchronization at that time misalignment; when When the value is close to 0, it indicates that there is no significant correlation between the trends of the two sequences at that time offset; when... When the value is close to -1, it indicates that the two sequences change in opposite directions at that time offset.

[0107] The processing terminal calculates the correlation strength at each time offset. Then, extract from it Time offset for obtaining the maximum value :

[0108] ;

[0109] The processing terminal then verifies the maximum value. Whether it exceeds the preset correlation strength threshold . The preset correlation strength threshold is the lowest threshold for determining whether the peak of the correlation strength is statistically significant. Its value usually ranges from 0.4 to 0.7 and can be adjusted according to the noise level of the clinical data. If , then is used as the quantization value of the temporal correlation relationship; if , it is determined that the historical data of this patient is not sufficient to stably estimate the temporal correlation relationship, and the processing terminal sets the quantization value of the temporal correlation relationship to 0, that is, it is processed according to synchronization.

[0110] The physical meaning of is that when , it means that the change in chromaticity characteristics leads the change in area characteristics in time by follow-up time points (that is, the chromaticity changes first, and the area follows up after follow-up cycles); when , it means that the change in chromaticity characteristics lags behind the change in area characteristics in time by follow-up time points; when

[0111] Following the previous follow-up example of patient P, when the processing terminal processes the 6th follow-up of patient P, it performs the correlation calculation of the area difference sequence and the chromaticity difference sequence (the lengths of the difference sequences are both 5) under the time offset . The calculation results show that , , , . It can be seen that the correlation strength reaches a peak of 0.82 at and exceeds the preset threshold . Therefore, the processing terminal determines the quantization value of the temporal correlation relationship of this patient as . Its physical meaning is that for this patient, the change time point of the lesion chromaticity characteristics leads the change time point of the lesion area characteristics by 1 follow-up cycle (that is, about 90 days) in time.

[0112] Through the above scheme, the processing terminal extracts the inherent time sequence relationship of chromaticity change - area change of this patient from the two feature sequences formed by the patient's历次 follow-ups, laying a key basis for making an adaptive evaluation of the index changes observed in the current follow-up in the subsequent steps.

[0113] Step S4: Obtain the changes in lesion color and lesion area at the current follow-up time point, and determine the phase matching category at the current follow-up time point by combining the temporal correlation.

[0114] In this step, the change in lesion color intensity refers to the magnitude of change in the lesion color intensity characteristics at the current follow-up time point relative to the lesion color intensity characteristics at the previous follow-up time point, i.e. The definition of the change in lesion area corresponds to this, namely... In other optional embodiments, the change in lesion can also be the difference between the feature value at the current follow-up time point and the feature value collected at the patient's first follow-up. This embodiment does not limit this. The phase matching category refers to the matching status between the indicator changes observed in this follow-up and the patient's existing temporal sequence patterns, determined by the processing terminal based on the chromaticity and area changes observed at the current follow-up time point, combined with the patient's estimated temporal correlation.

[0115] The core purpose of this step is to classify the multidimensional changes observed at the current follow-up point into differentiated treatment categories according to their matching status with the patient's existing time sequence. This allows subsequent steps to use weights appropriate to the clinical significance of each category to perform weighted fusion of the changes, thus preventing early progression signals in individual dimensions from being diluted by other dimensions that have not yet been followed up, and preventing atypical signals that violate the existing time sequence from being treated as routine cases.

[0116] Furthermore, when determining the phase matching category, the processing terminal specifically includes: comparing the lesion chromaticity change with a preset chromaticity change threshold, comparing the lesion area change with a preset area change threshold, and obtaining chromaticity comparison results and area comparison results; based on the chromaticity comparison results, area comparison results, and temporal correlation, determining the phase matching category of the current follow-up time point as one of the following three categories.

[0117] Specifically, a preset chromaticity change threshold is used. Compared with the preset area change threshold This is an objective threshold used to determine whether the change constitutes a "significant change," and its value is determined using an adaptive method based on the patient's stable-period fluctuation level. The specific determination process is as follows:

[0118] Stable period identification: Retrieve the chromaticity feature sequence and area feature sequence corresponding to each follow-up visit from the patient's follow-up records; calculate the patient's difference sequence. and Scan two difference sequences using a sliding window (window length 3, step size 1). If the absolute values ​​of all difference values ​​within a certain window do not exceed the preset stability window criterion (e.g., ... , If the initial value of the system is used, then the follow-up time points covered by the window are included in the stable period sample set. .

[0119] Threshold estimation: based on the stable period sample set Calculate the sample standard deviation of the chromaticity difference and area difference values ​​corresponding to each time point. and Preset color change threshold Pick Three times the preset area change threshold Pick 3 times, that is:

[0120] ;

[0121] Rollback processing: If the sample set during the stable period If the number of samples is less than 3, the system default value will be used. and As a preset threshold, using 3 times the standard deviation as the threshold multiple corresponds to the judgment of "exceeding the range of fluctuations in the stable period" with a confidence level of about 99.7% under the assumption of normal distribution, which is statistically significant.

[0122] The processing terminal will record the chromaticity change during this follow-up. and The colorimetric comparison results (exceeding the threshold / not exceeding the threshold) are compared, and the area change during this follow-up is recorded. and The area comparison results are obtained (exceeding the threshold / not exceeding the threshold), and the chromaticity comparison results are combined with the area comparison results along with the established temporal correlation. Common input phase matching category determination logic.

[0123] Category 1: The colorimetric comparison results indicate that the colorimetric change of the lesion exceeds the preset colorimetric change threshold, while the area comparison results indicate that the area change of the lesion does not exceed the preset area change threshold at the corresponding time point indicated by the time-series correlation.

[0124] The first category, clinically speaking, corresponds to the "leading state," meaning that according to the patient's existing chronological pattern, chromaticity characteristics should have changed before area characteristics. Under the first category, the chromaticity changes observed in this follow-up are significant, while the corresponding time point at which area changes should have occurred, inferred from the temporal correlation (i.e., at...) At a point t after a follow-up period, the area change has not yet reached a significant level, which means that this follow-up captured an early change signal that has not yet been confirmed by area change but conforms to the time-prior law. Specifically, when determining the first category, the processing terminal needs to verify whether the current follow-up point t simultaneously meets the following two conditions:

[0125] Condition one: ;

[0126] Condition 2: For (like If it has not yet arrived, then the condition is considered to be in a pending verification state; if Then skip that condition. .

[0127] Category 2: The colorimetric comparison results and area comparison results indicate that the colorimetric change of the lesion exceeds the preset colorimetric change threshold and the area change of the lesion exceeds the preset area change threshold, respectively, and the occurrence sequence and temporal correlation of the two match.

[0128] The second category corresponds clinically to the "synchronous state," meaning that both the color and area changes observed in this follow-up are significant, and the chronological relationship between their occurrence is consistent with the patient's existing temporal correlation (e.g., temporal correlation). At that time, the color change occurred at the previous follow-up point. The area change occurs at the current time point t). The second category indicates that the progress observed in this follow-up is a typical case of steady evolution according to existing patterns. Specifically, when determining the second category, the processing terminal needs to verify:

[0129] Condition one: ;

[0130] Condition 2: Exists , making ;

[0131] Condition 3: The time of occurrence of the color change is relative to the time of occurrence of the area change. The order of events.

[0132] Category 3: Area comparison results indicate that the change in lesion area exceeds the preset area change threshold, while color comparison results indicate that the change in lesion color does not exceed the preset color change threshold at the prior time point indicated by the temporal correlation relationship.

[0133] The third category corresponds to "atypical" in a clinical sense, meaning that according to the patient's existing chronological pattern, the chromatic characteristics should have changed before the area characteristics. However, what was observed in this follow-up was that the area change was already significant, while at the corresponding time point inferred from the time-series correlation, where the color change should have appeared first (i.e., The color change was not significant. Category 3 indicates that the combination of indicators observed in this follow-up study violates the patient's existing chronological pattern and exhibits an abnormal progression pattern. Specifically, when determining Category 3, the processing terminal needs to verify:

[0134] Condition one: ;

[0135] Condition 2: For (Require and Within the observed follow-up range, .

[0136] It should be noted that when none of the above three judgment conditions are met (for example, the chromaticity change and area change in this follow-up do not exceed the corresponding threshold), the processing terminal can determine the phase matching category as "no significant change". In this case, it will not enter the subsequent differentiated weighted fusion processing, but will be processed directly according to the normal stable state.

[0137] Using the follow-up example of patient P mentioned above, the processing terminal has determined the temporal correlation of this patient. This means that the color change leads the area change by one follow-up period. Based on the difference sequences {0.8, -1.2, 0.8} and {1.5, -1.7, 0.5} of the patient's first four stable periods, the following was calculated: Therefore, a preset chromaticity change threshold is established. Preset area change threshold For ease of explanation, this embodiment will be referred to thereafter as... , The values ​​are provided as illustrative examples.

[0138] The processing terminal was first observed at the 5th follow-up ( Phase matching category determination is performed during this follow-up period. ,Exceed This follow-up Not exceeding ; and inferred from the temporal correlation The time point has not yet arrived, so it is considered to be in a state pending verification. Therefore, the processing terminal determines the phase match category of the 5th follow-up to be Category 1 (leading state), that is, it has captured an early chromaticity change signal that has not yet been confirmed by area changes.

[0139] Subsequently, the processing terminal performed a phase-matching category determination at the 6th follow-up (t=6). This follow-up... ,Exceed This follow-up far exceeding The terminal then verifies the timing correlation. ,exist At a certain point in time Exceeded Overall assessment: This follow-up study meets the conditions for the second category (synchronous state)—color change appeared first at the 5th follow-up, and area change lagged by one follow-up cycle at the 6th follow-up. The timing and temporal correlation of these two changes are discussed. Exact match.

[0140] Through the above scheme, the processing terminal accurately classifies the multidimensional change indicators observed at the current follow-up time point into processing categories that match their temporal patterns, providing a category basis for the differentiated weighted fusion of subsequent steps.

[0141] Step S5: Using the weighting factor corresponding to the phase matching category, perform weighted fusion processing on the lesion color change and lesion area change to obtain the lesion progression score of the same patient at the current follow-up time point.

[0142] In this step, the differential weighted fusion processing refers to the process by which the processing terminal selects a weighting factor corresponding to the phase matching category determined in step S4, and performs a weighted summation on the changes in lesion color and lesion area observed in this follow-up to obtain a comprehensive score reflecting the overall progression intensity of this follow-up. The lesion progression score is a scalar value obtained through differential weighted fusion processing, reflecting the overall progression intensity of the same patient at the current follow-up time point; the larger the value, the stronger the progression signal observed in this follow-up.

[0143] The core objective of this step is to assign different weighting factors to the same set of chromaticity changes and area changes under different phase matching categories, so that the progression signals observed in this follow-up can be amplified or suppressed differentially according to their matching status with the patient's existing time sequence. This allows for the amplification of early signals that have not yet been confirmed by area changes in the leading state, the treatment of routine progression in the synchronous state, and the independent response to abnormal combinations that violate existing patterns in the atypical state.

[0144] Specifically, the processing terminal pre-configures a set of weighting factors for each phase-matching category. For the first category (leading state), the weighting factors are configured... For the second category (synchronous state), configure the weighting factor. For the third category (atypical), configure weighting factors. .

[0145] Furthermore, in the differentiated weighted fusion process, the weight factor corresponding to the third category is greater than the weight factor corresponding to the first category, and the weight factor corresponding to the first category is greater than the weight factor corresponding to the second category. Specifically, the weight factor configured for each phase-matching category can be determined by a set of category coefficients. To reflect their relative size relationship, among which The overall weight levels corresponding to the first, second, and third categories respectively, and satisfying the following conditions: Final weighting factors for each category and It can be obtained by multiplying the category coefficient by the basic weight:

[0146] ;

[0147] in, This represents the category coefficient corresponding to the current phase-matching category; and The preset base weights are all set to 0.5 in this embodiment.

[0148] The weighting relationship used in this step The design principle is as follows: The second category (synchronous state) corresponds to the typical situation of steady evolution according to existing rules, which has been fully confirmed by the significant changes in both indicators. Therefore, it is treated with a relatively low weight and routine processing. The first category (leading state) corresponds to the situation where chroma changes first but area has not followed. If the early signal captured in this situation is not amplified, it is easily diluted by the area change that has not yet reached the target. Therefore, it is amplified with a relatively high weight. The third category (atypical state) corresponds to the situation where area has changed significantly but chroma has not been synchronously predicted at the time point that should have occurred. This situation violates the patient's inherent time sequence and often indicates an abnormal switch in the progression pattern in clinical practice. Therefore, it is responded to with the highest weight and the highest level of response.

[0149] After selecting the weighting factor corresponding to the phase-matching category of this follow-up, the processing terminal performs weighted fusion processing on the lesion color change and lesion area change according to the following formula to obtain the lesion progression score P:

[0150] ;

[0151] Where P represents the lesion progression score of the same patient at the current follow-up time point; and These represent the weighting factors corresponding to the chromaticity change and area change under phase matching category c, respectively. and These represent the changes in chromaticity and area, respectively, during this follow-up period. This indicates the phase matching category determined in step S4.

[0152] In a specific example, the base weights can be set to The category coefficient can be set to... That is, the weighting factor for the second category is (0.5, 0.5), the weighting factor for the first category is (0.75, 0.75), and the weighting factor for the third category is (1.0, 1.0). The specific values ​​of the coefficients for each category can be adjusted based on clinical data, but should be maintained. The relative size relationship.

[0153] Using the follow-up example of patient P mentioned above, the processing terminal uses the category coefficient at the 5th follow-up (when the patient is classified as category 1). The corresponding weighting factors are (0.75, 0.75); this follow-up... Therefore, the lesion progression score Subsequently, at the 6th follow-up (when the terminal was classified as category 2), the category coefficient was used. The corresponding weighting factors are (0.5, 0.5); this follow-up... Therefore, the lesion progression score .

[0154] It should be noted that if the differentiated weighted fusion processing of this invention is not used, but instead traditional linear weighting with equal weights (assuming a uniform weight of 0.5) is used, the score corresponding to the 5th follow-up will be... The value is significantly lower than the 11.70 obtained in this invention. It is evident that by assigning a higher weight to the first category than to the second category, this invention effectively amplifies the early chromaticity change signal captured in the 5th follow-up, which has not yet been confirmed by area changes. This avoids the dilution of this early signal by area changes that have not yet been followed up in the same follow-up, thereby overcoming the bottleneck of the prior art's long-term insensitivity to early progress.

[0155] Using the above approach, the processing terminal performed differentiated weighted fusion on the multidimensional change indicators observed in this follow-up according to their matching status with the patient's existing time sequence, and obtained a comprehensive score that can reflect the true intensity of the progress in this follow-up.

[0156] Step S6: Output risk stratification results based on lesion progression score, and send the warning information corresponding to the risk stratification results that meet the preset warning conditions to the medical terminal.

[0157] In this step, the risk stratification result refers to the graded judgment result of the overall risk level of the patient corresponding to this follow-up based on the lesion progression score of this follow-up and the risk factor information in the patient's file. It is divided into three levels: low risk, medium risk, and high risk. The preset warning conditions refer to the objective conditions used by the processing terminal to determine whether the risk stratification result of this follow-up needs to trigger a warning. These include, but are not limited to, "the risk stratification result is high risk" or "the lesion progression score exceeds the preset progression score threshold". The warning information refers to the information generated by the processing terminal when the preset warning conditions are met, which is used to prompt medical staff to intervene in the patient. It includes the patient identifier, the risk stratification result of this follow-up, the phase matching category corresponding to this follow-up, the values ​​of key change indicators, and the recommended treatment methods.

[0158] This embodiment provides two optional implementation methods for the risk stratification rules, and the processing terminal can choose either one for deployment:

[0159] Implementation Method 1 (Rule-Based Judgment): The processing terminal presets two levels of thresholds—a high-risk scoring threshold. With medium risk scoring threshold (satisfy Typical value The risk factor count F is defined as the number of "yes" items among the four indicators in the patient's file: "whether the patient smokes," "whether the patient drinks alcohol," "whether the patient has a history of betel nut exposure," and "whether the patient's previous pathology is mild dysplasia or above." The processing terminal outputs the risk stratification results according to the following rules:

[0160] like and If so, then output high risk;

[0161] like and ,or and If so, the output will be medium risk;

[0162] In other cases, output low risk.

[0163] Implementation Method Two (Decision Based on Machine Learning Model): The processing terminal pre-trains and deploys a binary classification model based on Gradient Boosting Decision Tree (GBDT), whose input is a 6-dimensional feature vector. Where P is the lesion progression score and c is the phase matching category (integer 1, 2, or 3). This is the quantified value of the patient's temporal correlation. The variables are binary indicator variables (0 or 1) representing smoking, drinking, and betel nut exposure history, respectively; the model output is the probability value of belonging to the "high-risk" category. The model is trained using historical labeled data. Labels are provided by physicians qualified to diagnose and treat oral mucosal diseases, who assign a binary "high-risk / non-high-risk" rating to each historical follow-up record. Training employs 5-fold cross-validation, with a binary cross-entropy loss function, a maximum tree depth of 5, a tree count of 100, and a learning rate of 0.1. After deployment, the processing terminal inputs the feature vectors from each follow-up visit into the model, and then... Output high risk, Output of medium risk Output low risk.

[0164] After receiving the risk stratification results of this follow-up visit, the processing terminal verifies whether the results meet the preset warning conditions. If they do, it generates a warning message containing key information about this follow-up visit and sends it to the medical staff terminal via the communication network for medical staff to view and make subsequent treatment decisions. If the conditions are not met, the results are recorded in the patient's follow-up file according to the regular follow-up process, awaiting the next follow-up visit.

[0165] Furthermore, the processing terminal is further configured to generate an early warning message containing a biopsy review prompt or a referral trigger prompt when the phase matching category at the current follow-up time point is determined to be the third category, regardless of whether the lesion progression score meets the preset early warning conditions, and send it to the medical staff terminal.

[0166] The design principle of this bypass triggering mechanism is as follows: the combination of indicators corresponding to the third category (atypical state) violates the patient's existing temporal sequence, that is, the area has changed significantly but the color has not been synchronously predicted at the point when it should have been. This abnormal combination often indicates a fundamental shift in the disease progression pattern in clinical practice, possibly corresponding to early signs of deep submucosal lesions or malignant transformation. Its clinical warning value is significantly higher than that reflected by the lesion progression score itself. Therefore, even if the lesion progression score in this follow-up has not yet reached the numerical threshold required by the preset warning conditions, as long as the phase matching category is determined to be the third category, the processing terminal directly generates a warning message containing a biopsy review prompt or referral trigger prompt, bypassing the conventional scoring threshold determination process, ensuring that such atypical progression signals can be noticed by medical staff as soon as possible, avoiding missed diagnosis due to conventional threshold determination.

[0167] Using the follow-up example of patient P mentioned above, let's assume that the lesion progression score obtained by the processing terminal at the 5th follow-up is... The judgment rule of implementation method one is adopted, and the risk factor count of patient P is F=2 (smoking + drinking); because and The processing terminal outputs the risk stratification result of this follow-up as high risk, and generates a data set including the patient's P identifier, the risk stratification result of this follow-up as "high risk", the phase matching category as "Category 1 (Leading State)", and key change indicators. Warning messages, including "It is recommended to have a follow-up outpatient visit as soon as possible," are sent to medical staff terminals via the communication network. After receiving the warning, medical staff can manually review it on the medical staff terminal and initiate further treatment arrangements accordingly.

[0168] Furthermore, suppose that during a follow-up visit, the processing terminal identifies a patient Q whose phase-matching category is classified as Category 3, but whose lesion progression score is only 8.0, failing to reach the high-risk score threshold of 10.0. According to standard procedures, this follow-up visit should be classified as medium-risk and only routinely recorded. However, due to the bypass triggering mechanism of this invention, the processing terminal ignores the fact that the score does not reach the threshold and directly generates a warning message containing a "biopsy review recommended as soon as possible" prompt, sending it to the medical staff terminal. This clinically avoids the missed detection of an atypical progression signal that may indicate malignant transformation.

[0169] Through the above scheme, the processing terminal, based on conventional risk stratification, sets up a bypass trigger response for atypical progression signals that violate the existing time sequence of patients. This enables the dual accurate capture of early and atypical signals during remote follow-up of such lesions, overcoming the bottleneck of existing technologies that default to multidimensional lesion change indicators as synchronously reflecting the disease progression, leading to the dulling of early identification and the omission of atypical cases.

[0170] In summary, by extracting two independent temporal feature sequences—area and chroma—from the patient's standardized images and performing temporal misalignment correlation calculations, the inherent leading or lagging pattern between lesion chroma changes and area changes is explicitly defined. Based on this, for the multidimensional changes observed at the current follow-up point, the current phase matching category is first determined according to this inherent pattern, and then differentiated weighting factors are applied to the conventional leading state, synchronous state, and atypical state for weighted scoring. This approach ensures that early chroma changes that conform to the leading pattern but whose area has not yet been verified are not diluted by increasing their weight in the fusion score. For atypical patterns of sudden area increases without prior chroma prediction, which violate existing temporal patterns, a bypass triggering mechanism directly pushes a biopsy review warning to the medical staff to prevent the missed detection of malignant transformation patterns, thereby achieving accurate capture of early signs of malignant evolution.

[0171] Example 2:

[0172] like Figure 5 As shown, a smart follow-up method for precancerous lesions of the oral cavity includes the following steps:

[0173] Multiple standardized images of the same patient were acquired at different follow-up time points using a patient smart terminal with the assistance of an intraoral standardized image acquisition device. Each of the multiple standardized images carries scale reference information and colorimetric reference information.

[0174] Based on the scale reference information and color reference information it carries, normalization, same-site registration and lesion segmentation are performed on multiple frames of standardized images to obtain the lesion area feature sequence and lesion color feature sequence of the same patient at different follow-up time points.

[0175] Based on the lesion area feature sequence and the lesion color feature sequence, the temporal correlation is determined. The temporal correlation is used to characterize the leading or lagging status of the time point of change of lesion color feature relative to the time point of change of lesion area feature.

[0176] Obtain the changes in lesion color and lesion area at the current follow-up time point, and determine the phase matching category at the current follow-up time point by combining the temporal correlation.

[0177] Using weighting factors corresponding to the phase-matching category, weighted fusion processing was performed on the changes in lesion color and lesion area to obtain the lesion progression score of the same patient at the current follow-up time.

[0178] Risk stratification results are output based on lesion progression scores, and warning information corresponding to risk stratification results that meet preset warning conditions is sent to medical staff terminals.

[0179] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0180] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0181] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A smart follow-up system for precancerous oral lesions, characterized in that, include: The standardized intraoral data collector, a patient smart terminal used in conjunction with the standardized intraoral data collector, a processing terminal, and a medical care terminal; The intraoral normalized acquisition device includes at least a scale structure for providing scale reference in the acquired image and a chromaticity reference structure for providing chromaticity reference. The processing terminal is communicatively connected to the patient's smart terminal and the medical staff's terminal, and the processing terminal is configured as follows: The patient's smart terminal acquires multiple standardized images of the same patient at different follow-up time points with the assistance of the intraoral standardized image acquisition device. Each of the multiple standardized images carries scale reference information and colorimetric reference information. Based on the scale reference information and the color reference information carried, normalization, same-site registration and lesion segmentation are performed on the multi-frame standardized images to obtain the lesion area feature sequence and lesion color feature sequence corresponding to the same patient at different follow-up time points. Based on the lesion area feature sequence and the lesion color feature sequence, a temporal correlation is determined. The temporal correlation is used to characterize the leading or lagging status of the time point of change of lesion color feature relative to the time point of change of lesion area feature. Obtain the changes in lesion color and lesion area at the current follow-up time point, and determine the phase matching category at the current follow-up time point by combining the aforementioned temporal correlation. Using a weighting factor corresponding to the phase-matching category, a weighted fusion process is performed on the lesion color change and the lesion area change to obtain the lesion progression score of the same patient at the current follow-up time point; Based on the lesion progression score, risk stratification results are output, and the warning information corresponding to the risk stratification results that meet the preset warning conditions is sent to the medical terminal.

2. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: The intraoral standardized collector further includes at least one of the following structures: a terminal fixing clamp for fixing the patient's smart terminal, an annular supplementary light for providing annular illumination during collection, an oral corner retraction structure for retracting the corners of the patient's mouth during collection, and a tongue pressure auxiliary component for pressing the patient's tongue during collection.

3. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: When the processing terminal performs normalization processing, it specifically includes: Based on the proportional relationship between the pixel length occupied by the scale structure in the standardized image and the actual physical length of the scale structure, scale normalization processing is performed on the standardized image to eliminate shooting distance differences. Based on the deviation between the pixel values ​​of the color blocks presented in the normalized image by the chromaticity reference structure and the pre-stored chromaticity baseline values, chromaticity normalization processing is performed on the normalized image to eliminate differences in shooting lighting and white balance.

4. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: When performing co-location registration, the processing terminal specifically includes: One frame is selected from the multi-frame standardized images as the baseline image; For each frame in the multi-frame standardized image other than the baseline image, based on the dental feature points located at the edge of the dental arch in the frame and the scale reference information provided by the scale structure, the geometric transformation relationship between the frame and the baseline image is determined, and the frame is mapped to a coordinate system unified with the baseline image based on the geometric transformation relationship.

5. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: The lesion area feature sequence is composed of the lesion area features of the same patient at the most recent N follow-up times, and the lesion color feature sequence is composed of the lesion color features of the same patient at the most recent N follow-up times, where N is a preset positive integer not less than 4, and N expands as the number of historical follow-ups of the same patient accumulates.

6. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: When determining the timing correlation, the processing terminal specifically includes: The correlation calculation under time misalignment is performed on the lesion area feature sequence and the lesion color feature sequence to obtain the correlation strength between the lesion color feature sequence and the lesion area feature sequence at different time offsets. The correlation strength is used to characterize the degree of synchronous change of the two feature sequences at the corresponding time offsets. The time offset in which the correlation strength reaches its maximum value and exceeds the preset correlation strength threshold is used as the quantification value of the temporal correlation relationship.

7. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: When determining the phase matching category, the processing terminal specifically includes: The amount of color change of the lesion is compared with a preset color change threshold, and the amount of color change of the lesion area is compared with a preset area change threshold to obtain the color comparison result and the area comparison result. Based on the chromaticity comparison results, the area comparison results, and the temporal correlation, the phase matching category at the current follow-up time point is determined to be one of the following three categories: Category 1: The colorimetric comparison result indicates that the colorimetric change of the lesion exceeds the preset colorimetric change threshold, while the area comparison result indicates that the area change of the lesion does not exceed the preset area change threshold at the corresponding time point indicated by the temporal correlation. Second category: The color comparison result and the area comparison result respectively indicate that the color change of the lesion exceeds the preset color change threshold and the area change of the lesion exceeds the preset area change threshold, and the occurrence sequence of the two matches the temporal correlation relationship; The third category: the area comparison result indicates that the change in the lesion area exceeds the preset area change threshold, while the color comparison result indicates that the change in the lesion color does not exceed the preset color change threshold at the prior time point indicated by the temporal correlation.

8. The intelligent follow-up system for precancerous oral lesions according to claim 7, characterized in that: In the differentiated weighted fusion process, the weight factor corresponding to the third category is greater than the weight factor corresponding to the first category, and the weight factor corresponding to the first category is greater than the weight factor corresponding to the second category.

9. The intelligent follow-up system for precancerous oral lesions according to claim 1, characterized in that: The processing terminal is configured as follows: When the phase matching category at the current follow-up time point is determined to be the third category, regardless of whether the lesion progression score meets the preset warning conditions, a warning message containing a biopsy review prompt or a referral trigger prompt is generated and sent to the medical terminal.

10. A smart follow-up method for precancerous lesions of the oral cavity, characterized in that, The method is applied to the processing terminal of the intelligent follow-up system for precancerous oral lesions according to any one of claims 1 to 9, and the method includes: The patient's smart terminal acquires multiple standardized images of the same patient at different follow-up time points with the assistance of the intraoral standardized image acquisition device. Each of the multiple standardized images carries scale reference information and colorimetric reference information. Based on the scale reference information and the color reference information carried, normalization, same-site registration and lesion segmentation are performed on the multi-frame standardized images to obtain the lesion area feature sequence and lesion color feature sequence corresponding to the same patient at different follow-up time points. Based on the lesion area feature sequence and the lesion color feature sequence, a temporal correlation is determined. The temporal correlation is used to characterize the leading or lagging status of the time point of change of lesion color feature relative to the time point of change of lesion area feature. Obtain the changes in lesion color and lesion area at the current follow-up time point, and determine the phase matching category at the current follow-up time point by combining the aforementioned temporal correlation. Using a weighting factor corresponding to the phase-matching category, a weighted fusion process is performed on the lesion color change and the lesion area change to obtain the lesion progression score of the same patient at the current follow-up time point; Based on the lesion progression score, risk stratification results are output, and the warning information corresponding to the risk stratification results that meet the preset warning conditions is sent to the medical terminal.

Images