Newborn surface feature rare disease auxiliary diagnosis system based on AI image recognition

Abstract

The invention discloses a newborn face feature rare disease auxiliary diagnosis system based on AI image recognition, relates to the technical field of rare disease auxiliary diagnosis, and aims to solve the technical problem that dynamic feature analysis is insufficient in existing newborn rare disease auxiliary diagnosis. Detection, segmentation, feature classification and dynamic change analysis are completed, and abnormal textures, pigment distribution or blood vessel features are identified; the face key point analysis module is used for marking face key anatomical points, calculating three-dimensional space coordinates and analyzing relative positions and geometrical relationships of the key points and evolution trends of the key points along with time; the dynamic modeling module is used for constructing a dynamic evolution model of the facial features of the newborns along with the change of day ages, and correlating the time dependence of growth and development stages and rare disease features; and the feature fusion unit is used for integrating output results of the skin analysis module, the face key point analysis module and the dynamic modeling module. The method has the advantage of improving the diagnosis accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of auxiliary diagnosis of rare diseases, and more specifically, to an auxiliary diagnosis system for rare diseases of neonates based on AI image recognition of facial features. Background Art

[0002] In the field of auxiliary diagnosis of rare diseases in newborns, existing technologies primarily rely on static facial feature analysis or single-modality data, lacking systematic modeling of the dynamic changes in newborn facial features as they grow and develop. Specifically, traditional auxiliary diagnosis systems struggle to effectively integrate multi-dimensional features such as skin texture and facial geometry, and are unable to accurately capture the time-dependent evolution of rare disease features with age. This results in low recognition rates for rare diseases with atypical early symptoms, making missed diagnoses and misdiagnoses more likely. In light of this, we propose a newborn facial feature-based rare disease auxiliary diagnosis system based on AI image recognition. Summary of the Invention

[0003] The purpose of the present invention is to provide a neonatal facial feature rare disease auxiliary diagnosis system based on AI image recognition to solve the problem of insufficient dynamic feature analysis technology in the existing neonatal rare disease auxiliary diagnosis.

[0004] To solve the above technical problems, the present invention provides the following technical solutions: a neonatal facial feature rare disease auxiliary diagnosis system based on AI image recognition, comprising: The skin analysis module performs detection, segmentation, feature classification, and dynamic change analysis on the facial skin area of ​​newborns, identifying abnormal textures, pigment distribution, or vascular characteristics. Facial key point analysis module, which marks key facial anatomical points and calculates their 3D spatial coordinates, analyzing the relative positions, geometric relationships, and temporal evolution trends of key points; A dynamic modeling module builds a dynamic evolution model of newborn facial features as they change with age, correlating the temporal dependence of growth and development stages with rare disease characteristics; The feature fusion unit integrates the output results of the skin analysis module, facial key point analysis module and dynamic modeling module to generate a comprehensive feature vector containing static features and dynamic evolution features; The comprehensive diagnosis unit combines clinical information, multimodal feature models and growth and development standards, cross-validates the fused feature vectors, and generates a diagnostic report that includes disease recommendations, feature matching basis and follow-up plan.

[0005] Preferably, it also includes: An image acquisition module continuously collects clinical video streams of newborns at different time points after birth, wherein the clinical video streams contain multiple frames of continuous images to record the dynamic changes of facial features as the newborn grows and develops; The image preprocessing module performs facial region positioning, geometric correction, and brightness optimization on each frame of the video stream to eliminate the effects of shooting angle deviation and uneven lighting on image quality; The image recognition model, built on a deep learning framework, automatically extracts multi-dimensional features of the texture, contour, and color of newborn faces through a convolutional neural network, forming a standardized feature expression. The feature matching module calculates the similarity between the facial features extracted in real time and the pre-built rare disease facial feature library, and outputs the disease type with the highest matching degree and the confidence score.

[0006] Preferably, the skin analysis module includes: The skin detection unit builds a skin color model in the YCbCr color space, uses a Gaussian mixture clustering algorithm to fit the color distribution of normal infant skin, and combines it with the Otsu threshold method to automatically determine the skin color segmentation threshold, locate the facial skin area, and exclude non-skin pixels; The skin segmentation unit uses the GrabCut interactive segmentation algorithm to refine the boundaries of the detected skin area. It iteratively optimizes the probability distribution of foreground / background pixels and introduces a conditional random field model to optimize the continuity of the segmentation boundary to ensure the complete extraction of the skin lesion area. The skin classification unit builds a skin feature classification network based on the DenseNet architecture. It inputs the segmented skin region image and outputs the classification results of skin phenotypes such as pigmentation, vascular malformations, and texture abnormalities. It also analyzes the change rate of each phenotype at different time points. The regional analysis unit quantifies the spatial and temporal characteristics of skin lesions and determines whether they conform to the evolutionary patterns of skin phenotypes associated with rare diseases by comparing the growth curves with those of normal skin characteristics. The positioning auxiliary unit delineates the anatomical sub-regions around the eyes, nose, and mouth based on the coordinates of the corners of the eyes, wings of the nose, and corners of the mouth output by the facial key point analysis module, and specifically analyzes the skin characteristics of each anatomical sub-region to improve the detection accuracy of local abnormalities.

[0007] Preferably, the facial key point analysis module includes: The key point labeling unit, based on a cascaded convolutional neural network, automatically detects and labels several key points on the face, forming a two-dimensional coordinate set reflecting the facial morphology; The noise filtering unit uses the DBSCAN density clustering algorithm to process the key point coordinates of different frames of the same newborn, identifying and removing abnormal displacement points caused by crying expressions. It also performs mean filtering on the coordinate sequence through a spatiotemporal sliding window to improve the stability of the key point trajectory. The 3D reconstruction unit combines multi-view videos at the same time point and uses triangulation to calculate the 3D spatial coordinates of key points , construct a facial three-dimensional geometric model to analyze the three-dimensional structural characteristics of nose bridge height and facial symmetry; The trend analysis unit calculates the Euclidean distance change, angle change, and relative position offset of key points at adjacent time points, generates the dynamic evolution trajectory of facial key points, and identifies the evolution pattern of abnormal facial proportions or contour deformities related to rare diseases; The temporal modeling unit uses a long short-term memory network to model the three-dimensional coordinate sequence of key points, learns the time dependence of the key point trajectories during normal growth, and establishes an anomaly detection model for pathological evolution trajectories.

[0008] Preferably, the dynamic modeling module includes: The growth stage division unit divides the postnatal evolution process into three stages according to the development speed of the newborn's facial features: early, middle, and late. Each stage corresponds to a different sensitivity to feature changes and the probability of pathological feature manifestation; The dynamic feature extraction unit calculates the change amplitude, fluctuation frequency and abnormal feature occurrence frequency of skin features and key point features for each growth stage, and generates quantitative indicators reflecting the dynamic nature of the features; The standard model library stores the facial feature evolution standards of known rare diseases at various growth stages, including the normal fluctuation range of feature parameters, the typical appearance time and evolution curve of pathological features, for staged comparison of real-time detection data.

[0009] Preferably, the feature fusion unit includes: The static feature fusion module combines the skin texture features in a single frame image with the three-dimensional structural features of key points in series, or performs weighted fusion through an attention mechanism to form a multi-dimensional static feature vector that includes surface features and geometric features; The dynamic feature fusion module integrates feature change data at different time points and generates dynamic evolution features based on growth stage weights. The growth stages are clearly divided into early, middle and late stages based on the newborn's age. The growth stage weights are automatically adjusted according to the newborn's age to highlight the key diagnostic features of each growth stage. The normalization processing module standardizes the fused static and dynamic features, uses the Z-score normalization method to unify the feature dimensions and dimensions, eliminates the scale differences of different feature types, and provides input data with consistent format for subsequent diagnostic models.

[0010] Preferably, the comprehensive diagnostic unit comprises: The weight allocation module dynamically adjusts the weight coefficients of skin features, key point features, and their dynamic evolution features based on feature type and growth stage through a gradient descent algorithm, giving priority to features with high correlation with the current age. The multimodal comparison module uses support vector machines (SVMs) and random forest algorithms to classify static feature vectors and identify structural / texture matches with known rare diseases. At the same time, dynamic time warping technology is used to match the real-time feature evolution trajectory with the pathological curve in the standard model library, calculate the similarity distance of the time series, and achieve dual verification of static structure and dynamic evolution; In dynamic time warping technology, two time series are calculated and distance When we first define the local distance matrix , and then solve the cumulative distance matrix through dynamic programming : ; The final similarity distance ; in, represents the dynamic evolution time series of newborn facial features collected in real time, Represents real-time time series The feature data points, represents the index of the data point in the sequence, represents the pre-built standard evolution time series of rare disease facial features, Represents a standard time series The feature data points, represents the index of the data point in the standard sequence, Represents real-time time series With standard time series Similarity distance calculated by dynamic time warping technique; Represents real-time time series Middle data points With standard time series Middle data points The local distance between Represents the cumulative distance matrix elements in the dynamic time warping process, used to represent the starting point of the real-time sequence To data points, starting from the standard sequence To The cumulative distance of the best matching path of each data point; Indicates the cumulative distance between the previous real-time data point and the current standard data point. Indicates the cumulative distance between the current real-time data point and the previous standard data point. Indicates the cumulative distance between the previous real-time data point and the previous standard data point. Represents the final element of the cumulative distance matrix, i.e., the real-time time series With standard time series The total cumulative distance after complete matching; The report generation module generates a visual diagnostic report based on the comparison results, including: feature matching heat map, 3D facial model, dynamic evolution curve, clinical recommendations, and annotates the contribution of each feature to the diagnostic results.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention constructs a dynamic evolution model of the facial features of newborns as they change with age through a dynamic modeling module, and uses a feature fusion unit to integrate the outputs of multiple modules such as skin analysis and key point analysis to form a comprehensive vector containing static and dynamic features, thereby accurately capturing the time dependence of rare disease features, significantly improving diagnostic accuracy, and effectively solving the problem of insufficient dynamic feature analysis in the existing technology.

[0012] 2. The skin analysis module of the present invention uses a positioning auxiliary unit combined with the coordinates of facial key points to delineate anatomical sub-regions, and specifically analyzes local skin features such as those around the eyes and mouth. Simultaneously, the facial key point analysis module improves the stability of key point trajectories through three-dimensional reconstruction and noise filtering technology. The synergistic effect of the two can accurately identify subtle lesions such as wide-set eyes and localized pigmentation abnormalities, further enhancing the ability to capture the characteristics of rare diseases.

[0013] 3. The present invention also generates a diagnostic report including a follow-up plan by combining clinical information with growth and development standards through a comprehensive diagnostic unit. By matching real-time features with standard pathological curves through dynamic time warping technology, it not only provides disease recommendations, but also automatically adjusts the follow-up frequency and examination items according to age, thereby achieving early intervention and disease course monitoring for rare diseases. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0015] To facilitate those skilled in the art to understand the technical solution of the present invention, the technical solution of the present invention is further described with reference to the accompanying drawings.

[0016] Example 1, as Figure 1 As shown, the present invention provides a neonatal facial feature rare disease auxiliary diagnosis system based on AI image recognition, comprising: The image acquisition module is equipped with high-definition video acquisition equipment to continuously collect clinical video streams of newborns at different time points after birth, and simultaneously obtain clinical information including basic physiological indicators, family genetic disease history, and pregnancy risk factors. The clinical video stream contains multiple frames of continuous images to record the dynamic changes of facial features as they grow and develop; The image preprocessing module performs facial region positioning, geometric correction, and brightness optimization on each frame of the video stream to eliminate the effects of shooting angle deviation and uneven lighting on image quality; The image recognition model, built on a deep learning framework, automatically extracts multi-dimensional features of a newborn's face, such as texture, contour, and color, through a convolutional neural network to form a standardized feature representation. The feature matching module calculates the similarity between the facial features extracted in real time and the pre-built rare disease facial feature library, and outputs the disease type with the highest matching degree and the confidence score; The skin analysis module performs detection, segmentation, feature classification, and dynamic change analysis on the facial skin area of ​​newborns, identifying abnormal textures, pigment distribution, or vascular characteristics. Facial key point analysis module, which marks key facial anatomical points and calculates their 3D spatial coordinates, analyzing the relative positions, geometric relationships, and temporal evolution trends of key points; A dynamic modeling module builds a dynamic evolution model of newborn facial features as they change with age, correlating the temporal dependence of growth and development stages with rare disease characteristics; The feature fusion unit integrates the output results of the skin analysis module, facial key point analysis module and dynamic modeling module to generate a comprehensive feature vector containing static features and dynamic evolution features; The comprehensive diagnosis unit combines clinical information, multimodal feature models and growth and development standards, cross-validates the fused feature vectors, and generates a diagnostic report that includes disease recommendations, feature matching basis and follow-up plan.

[0017] In an embodiment of the present invention, the image preprocessing module includes: The camera calibration unit obtains the camera's internal parameters (focal length, principal point coordinates, distortion coefficient) and external parameters (rotation matrix, translation vector) by photographing a checkerboard calibration object, and establishes a transformation relationship between the camera coordinate system and the image coordinate system to correct image distortion caused by differences in camera position and angle; Face correction unit, which calculates the homography transformation matrix based on the camera calibration results , through perspective transformation, facial images from side, top or bottom angles are converted into standard frontal perspectives, eliminating the influence of geometric distortion on feature analysis; Homography transformation matrix Calculated by the following formula: ; in, is the pixel coordinate in the image coordinate system, is the coordinate in the world coordinate system; The brightness enhancement unit uses adaptive histogram equalization technology to adjust the local contrast of the corrected image. By dividing the image into sub-blocks and performing equalization processing separately, it enhances the texture details and color differences of the baby's skin, improving the sensitivity of subsequent feature detection; For the pixel values ​​within the image sub-block , pixel values ​​after adaptive histogram equalization Calculated by the following formula: ; in, is the pixel value The histogram statistics of is the total number of gray levels in the image.

[0018] In an embodiment of the present invention, the training method of the image recognition model includes: During the data annotation phase, a clinical video dataset containing newborns of different ages was annotated. The annotations included the facial contours, key point coordinates of the corners of the eyes, nose tips, and mouth corners, and the location and range of skin lesions (such as pigmented spots and hemangiomas). In the facial segmentation stage, the U-Net semantic segmentation network architecture is used to train the facial segmentation model. After inputting the original image, it outputs a binary mask (the facial area is 0 and the background is 1), separating the facial area from the non-facial area (such as hair, clothing, and medical equipment), and generating a clean facial image dataset. During U-Net network training, the cross entropy loss function is used. For optimization, the formula is as follows: ; in, is the sample size, is the number of categories, For samples Belong to category The true label, Predict samples for the model Belong to category probability; In the feature extraction stage, ResNet-50 is used as the backbone network to extract deep semantic features of facial images. A feature pyramid network is then used to fuse feature maps at different levels, capturing multi-scale features and generating feature vectors that include geometric features (facial feature proportions), texture features (skin roughness), and color features (skin tone evenness). In the ResNet-50 network, the output feature map of a certain layer Calculated by the following formula: ; in, is the weight matrix, is the feature map of the previous layer, is the bias vector, is the activation function; During the model training phase, the cross entropy loss function is used and triplet loss function Jointly optimize network parameters and, through end-to-end training, enable the model to accurately distinguish normal facial features from abnormal feature patterns associated with rare diseases; Triplet loss function The formula is as follows: ; in, is the anchor sample, is a positive sample, is a negative sample, is the feature extraction function, is the spacing parameter, , the formula means taking and the maximum value of 0, in the triple loss function It is used to ensure that the loss value is non-negative. Specifically, when calculating the triple loss, a difference plus the spacing parameter is obtained. The result after ,if A negative number means that the distance between the anchor sample and the positive sample is small enough and smaller than the distance between the anchor sample and the negative sample. In this case, the loss is 0. Only when Only when it is a positive number will it be used as part of the loss to participate in the optimization of network parameters, thereby prompting the model to learn more discriminative feature expressions, making the feature vectors of samples of the same category closer and the feature vectors of samples of different categories farther apart.

[0019] In an embodiment of the present invention, the skin analysis module includes: The skin detection unit builds a skin color model in the YCbCr color space, uses a Gaussian mixture clustering algorithm to fit the color distribution of normal infant skin, and combines it with the Otsu threshold method to automatically determine the skin color segmentation threshold, locate the facial skin area, and exclude non-skin pixels (such as lips and hair); In Gaussian mixture clustering, for pixel points , which belongs to The probability of a Gaussian distribution Calculated by the following formula: ; in, For the Gaussian distribution weights, The mean is , the covariance is The Gaussian probability density function, is the number of Gaussian distributions; The skin segmentation unit uses the GrabCut interactive segmentation algorithm to refine the boundaries of the detected skin area. It iteratively optimizes the probability distribution of foreground / background pixels and introduces a conditional random field model to optimize the continuity of the segmentation boundary to ensure the complete extraction of the skin lesion area. In the CRF model, the energy function Calculated by the following formula: ; in, is a unary potential energy function, describing the pixel The probability of belonging to a certain category, is a binary potential energy function describing the pixel and the relationship between; The skin classification unit builds a skin feature classification network based on the DenseNet architecture. It inputs the segmented skin region image and outputs the classification results of skin phenotypes such as pigmentation, vascular malformations, and texture abnormalities. It also analyzes the change rate of each phenotype at different time points. In the classification layer of the DenseNet network, the output category probability Calculated by the following formula: ; in, and are the weights and biases of the fully connected layer, is the input feature of the fully connected layer; The regional analysis unit quantifies the spatial characteristics (location coordinates, area size, shape complexity) and temporal characteristics (first appearance time, duration, and magnitude of change) of skin lesions. By comparing the growth curve of normal skin characteristics, it determines whether the skin phenotype evolution pattern associated with rare diseases is consistent. For the area of ​​skin lesions In time The rate of change , calculated using the following formula: ; in, is the time interval; The positioning assistance unit delineates the anatomical sub-regions of the eye area (from the inner canthus to the outer canthus), nose (from the bridge of the nose to the wing of the nose), and mouth area (from the corner of the mouth to the vermilion edge of the lips) based on the coordinates of the eye corners, nose wings, and mouth corners output by the facial key point analysis module. It then conducts targeted analysis of the skin features of each anatomical sub-region to improve the detection accuracy of local abnormalities (such as wide-set eyes accompanied by abnormal eyelid pigmentation).

[0020] In an embodiment of the present invention, the facial key point analysis module includes: The key point annotation unit, based on a cascaded convolutional neural network, automatically detects and annotates several key points on the face (including anatomical landmarks such as the brow peak, eye corners, nose tip, mouth corners, and mandibular margin), forming a two-dimensional coordinate set reflecting the facial morphology; The noise filtering unit uses the DBSCAN density clustering algorithm to process the key point coordinates of different frames of the same newborn, identifying and removing abnormal displacement points caused by crying expressions. It also performs mean filtering on the coordinate sequence through a spatiotemporal sliding window to improve the stability of the key point trajectory. In the DBSCAN algorithm, for the key point coordinates , its core distance Calculated by the following formula: ; in, for point of The point set in the neighborhood, The minimum number of points required for the core point, for point To Neighboring Points distance; The 3D reconstruction unit combines multi-view videos at the same time point and uses triangulation to calculate the 3D spatial coordinates of key points , construct a facial three-dimensional geometric model to analyze the three-dimensional structural characteristics of nose bridge height and facial symmetry; For the key points corresponding to the coordinates under two perspectives and , calculate the three-dimensional coordinates by triangulation , the formula is as follows: ; in, and are the pseudo-inverses of the projection matrices of the two view cameras, is the homography matrix between the two views; The trend analysis unit calculates the Euclidean distance change, angle change (such as the angle between the line connecting the two eyes and the horizontal line), and relative position offset (such as the displacement of the nose tip relative to the center of the eyebrows) of key points at adjacent time points, generating dynamic evolution trajectories of facial key points and identifying the evolutionary patterns of facial feature proportion abnormalities (such as low-set ears and a short philtrum) or contour deformities (such as a small mandible) associated with rare diseases; For two time points and The key point coordinates and , Euclidean distance change Calculated by the following formula: ; The temporal modeling unit uses a long short-term memory network to model the three-dimensional coordinate sequence of key points, learn the time dependence of the key point trajectories during normal growth (such as the normal rate of increase in nose bridge height with age), and establish an anomaly detection model for pathological evolution trajectories; In the LSTM network, the memory unit The update formula is as follows: ; in, is the output of the forget gate, is the input gate output, is a candidate memory unit, It is the memory unit of the previous moment.

[0021] In an embodiment of the present invention, the dynamic modeling module includes: The growth stage division unit divides the postnatal evolution process into three stages based on the development speed of newborn facial features: early stage (0-7 days, facial edema subsides), middle stage (8-28 days, facial features stabilize), and late stage (29 days and above, contour features appear). Each stage corresponds to different feature change sensitivities and pathological feature manifestation probabilities. The dynamic feature extraction unit calculates the change amplitude (the difference between the current value and the previous stage mean), fluctuation frequency (the number of times the feature value exceeds the normal range), and the frequency of abnormal features (such as persistent skin vascular abnormalities) for skin features (such as pigment spot area) and key point features (such as interocular distance) at each growth stage, generating quantitative indicators reflecting the dynamic nature of the features. For the eigenvalue The range of change in a certain growth stage , calculated using the following formula: ; in, is the characteristic mean of the previous growth stage; The standard model library stores the evolution standards of facial features of known rare diseases (such as trisomy 21 syndrome and Noonan syndrome) at various growth stages, including the normal fluctuation range of feature parameters (based on healthy newborn cohort data), the typical appearance time of pathological features (such as the low ear position feature of a certain syndrome usually appears in the middle stage) and the evolution curve (the reference trajectory of the feature value changing with age), which is used for staged comparison of real-time detection data.

[0022] In an embodiment of the present invention, the feature fusion unit includes: The static feature fusion module combines skin texture features (such as roughness and pigment distribution vectors) in a single frame with key point 3D structural features (such as facial feature distance ratio vectors) in series, or performs weighted fusion through an attention mechanism to form a multi-dimensional static feature vector that includes surface features and geometric features. When fused through the attention mechanism, the fused feature vector Calculated by the following formula: ; in, For the class feature vector, is the weight calculated by the attention mechanism, ; The dynamic feature fusion module integrates feature change data at different time points (such as the weekly growth rate of skin lesion area and the monthly change in the spatial position of key points). It combines growth stage weights (early onset gives higher weight to skin edema-related features, while later stages prioritize contour features) to generate dynamically evolving features. The growth stage weights are automatically adjusted based on the newborn's age, highlighting key diagnostic features at each growth stage. For dynamic feature changes , the dynamic features after fusion Calculated by the following formula: ; in, Indicates the growth stage, is the total number of growth stages, For the The weight of each growth stage, For the Dynamic characteristic changes in each growth stage; The normalization processing module standardizes the fused static and dynamic features, uses the Z-score normalization method to unify the feature dimensions and dimensions, eliminates the scale differences of different feature types, and provides consistent input data for subsequent diagnostic models; The Z-score normalization formula is as follows: ; in, is the original eigenvalue, is the characteristic mean, is the characteristic standard deviation.

[0023] In an embodiment of the present invention, the comprehensive diagnostic unit comprises: The weight allocation module dynamically adjusts the weight coefficients of skin features, key point features, and their dynamic evolution features using a gradient descent algorithm based on feature type (static features reflect the current state, dynamic features reflect the evolutionary trend) and growth stage (skin features are more weighted than structural features in the early stages, and vice versa in the later stages), giving priority to features with a high correlation with the current age. For the weight coefficient , the update formula through the gradient descent algorithm is as follows ; in, is the learning rate, is the loss function; The multimodal comparison module uses support vector machines (SVMs) and random forest algorithms to classify static feature vectors and identify structural / texture matches with known rare diseases. At the same time, dynamic time warping technology is used to match the real-time feature evolution trajectory with the pathological curve in the standard model library, calculate the similarity distance of the time series, and achieve dual verification of static structure and dynamic evolution; In the SVM algorithm, for the linearly separable case, the decision function is: ; in, is the weight vector, is the input feature vector, is the bias, is a symbolic function; For the random forest algorithm, the final prediction result It is obtained by voting or averaging the prediction results of multiple decision trees. For classification problems, the voting method is used: ; in, It is The prediction results of the decision tree, is the number of decision trees, It means taking the majority; In dynamic time warping technology, two time series are calculated and distance When we first define the local distance matrix , and then solve the cumulative distance matrix through dynamic programming : ; The final similarity distance ; in, represents the dynamic evolution time series of newborn facial features collected in real time, Represents real-time time series The feature data points, represents the index of the data point in the sequence, represents the pre-built standard evolution time series of rare disease facial features, Represents a standard time series The feature data points, represents the index of the data point in the standard sequence, Represents real-time time series With standard time series Similarity distance calculated by dynamic time warping technique; Represents real-time time series Middle data points With standard time series Middle data points The local distance between Represents the cumulative distance matrix elements in the dynamic time warping process, used to represent the starting point of the real-time sequence To data points, starting from the standard sequence To The cumulative distance of the best matching path of each data point; Indicates the cumulative distance between the previous real-time data point and the current standard data point. Indicates the cumulative distance between the current real-time data point and the previous standard data point. Indicates the cumulative distance between the previous real-time data point and the previous standard data point. Represents the final element of the cumulative distance matrix, i.e., the real-time time series With standard time series The total cumulative distance after complete matching; The report generation module generates a visual diagnostic report based on the comparison results. The report includes: a feature matching heat map (marking the facial location of abnormal features), a 3D facial model (showing abnormal spatial distribution of key points), a dynamic evolution curve (comparing real-time data with standard pathology curves), clinical recommendations (follow-up frequency, further examination items), and annotated contribution of each feature to the diagnostic result. Feature contribution in the report The calculation of the score is based on feature matching and the total score of all features : ; In an embodiment of the present invention, it further includes: The model optimization module regularly imports clinical data of newly confirmed cases in the hospital's neonatal department, updates the rare disease feature model library and dynamic evolution standards through transfer learning technology, adjusts feature weights and comparison algorithms for newly discovered rare disease phenotypes, and improves the system's generalized recognition capabilities for rare diseases.

[0024] The visualization module generates a dynamic feature analysis report with a timeline, allowing doctors to view facial images of newborns of different ages, skin feature segmentation results, and changes in the three-dimensional coordinates of key points by sliding the timeline, and compare normal developmental curves with pathological evolution patterns.

[0025] The embodiments disclosed in the present invention are preferred embodiments, but are not limited to them. Ordinary technicians in this field can easily understand the spirit of the present invention based on the above embodiments and make different extensions and changes. As long as they do not deviate from the spirit of the present invention, they are all within the scope of protection of the present invention.

Claims

1. A neonatal facial feature rare disease auxiliary diagnosis system based on AI image recognition, characterized by: include: The skin analysis module performs detection, segmentation, feature classification, and dynamic change analysis on the facial skin area of ​​newborns, identifying abnormal textures, pigment distribution, or vascular characteristics. Facial key point analysis module, which marks key facial anatomical points and calculates their 3D spatial coordinates, analyzing the relative positions, geometric relationships, and temporal evolution trends of key points; A dynamic modeling module builds a dynamic evolution model of newborn facial features as they change with age, correlating the temporal dependence of growth and development stages with rare disease characteristics; The feature fusion unit integrates the output results of the skin analysis module, facial key point analysis module and dynamic modeling module to generate a comprehensive feature vector containing static features and dynamic evolution features; The comprehensive diagnosis unit combines clinical information, multimodal feature models and growth and development standards, cross-validates the fused feature vectors, and generates a diagnostic report that includes disease recommendations, feature matching basis and follow-up plan.

2. The AI ​​image recognition-based neonatal facial feature rare disease auxiliary diagnosis system according to claim 1, characterized in that: Also includes: An image acquisition module continuously collects clinical video streams of newborns at different time points after birth, wherein the clinical video streams contain multiple frames of continuous images to record the dynamic changes of facial features as the newborn grows and develops; The image preprocessing module performs facial region positioning, geometric correction, and brightness optimization on each frame of the video stream to eliminate the effects of shooting angle deviation and uneven lighting on image quality; The image recognition model, built on a deep learning framework, automatically extracts multi-dimensional features of the texture, contour, and color of newborn faces through a convolutional neural network, forming a standardized feature expression. The feature matching module calculates the similarity between the facial features extracted in real time and the pre-built rare disease facial feature library, and outputs the disease type with the highest matching degree and the confidence score.

3. The AI ​​image recognition-based neonatal facial feature rare disease auxiliary diagnosis system according to claim 1, characterized in that: The skin analysis module includes: The skin detection unit builds a skin color model in the YCbCr color space, uses a Gaussian mixture clustering algorithm to fit the color distribution of normal infant skin, and combines it with the Otsu threshold method to automatically determine the skin color segmentation threshold, locate the facial skin area, and exclude non-skin pixels; The skin segmentation unit uses the GrabCut interactive segmentation algorithm to refine the boundaries of the detected skin area. It iteratively optimizes the probability distribution of foreground / background pixels and introduces a conditional random field model to optimize the continuity of the segmentation boundary to ensure the complete extraction of the skin lesion area. The skin classification unit builds a skin feature classification network based on the DenseNet architecture. It inputs the segmented skin region image and outputs the classification results of skin phenotypes such as pigmentation, vascular malformations, and texture abnormalities. It also analyzes the change rate of each phenotype at different time points. The regional analysis unit quantifies the spatial and temporal characteristics of skin lesions and determines whether they conform to the evolutionary patterns of skin phenotypes associated with rare diseases by comparing the growth curves with those of normal skin characteristics. The positioning auxiliary unit delineates the anatomical sub-regions around the eyes, nose, and mouth based on the coordinates of the corners of the eyes, wings of the nose, and corners of the mouth output by the facial key point analysis module, and specifically analyzes the skin characteristics of each anatomical sub-region to improve the detection accuracy of local abnormalities.

4. The AI ​​image recognition-based neonatal facial feature rare disease auxiliary diagnosis system according to claim 3, characterized in that: The facial key point analysis module includes: The key point labeling unit, based on a cascaded convolutional neural network, automatically detects and labels several key points on the face, forming a two-dimensional coordinate set reflecting the facial morphology; The noise filtering unit uses the DBSCAN density clustering algorithm to process the key point coordinates of different frames of the same newborn, identifying and removing abnormal displacement points caused by crying expressions. It also performs mean filtering on the coordinate sequence through a spatiotemporal sliding window to improve the stability of the key point trajectory. The 3D reconstruction unit combines multi-view videos at the same time point and uses triangulation to calculate the 3D spatial coordinates of key points , construct a facial three-dimensional geometric model to analyze the three-dimensional structural characteristics of nose bridge height and facial symmetry; The trend analysis unit calculates the Euclidean distance change, angle change, and relative position offset of key points at adjacent time points, generates the dynamic evolution trajectory of facial key points, and identifies the evolution pattern of abnormal facial proportions or contour deformities related to rare diseases; The temporal modeling unit uses a long short-term memory network to model the three-dimensional coordinate sequence of key points, learns the time dependence of the key point trajectories during normal growth, and establishes an anomaly detection model for pathological evolution trajectories.

5. The AI ​​image recognition-based neonatal facial feature rare disease auxiliary diagnosis system according to claim 1, characterized in that: The dynamic modeling module includes: The growth stage division unit divides the postnatal evolution process into three stages according to the development speed of the newborn's facial features: early, middle, and late. Each stage corresponds to a different sensitivity to feature changes and the probability of pathological feature manifestation; The dynamic feature extraction unit calculates the change amplitude, fluctuation frequency and abnormal feature occurrence frequency of skin features and key point features for each growth stage, and generates quantitative indicators reflecting the dynamic nature of the features; The standard model library stores the facial feature evolution standards of known rare diseases at various growth stages, including the normal fluctuation range of feature parameters, the typical appearance time and evolution curve of pathological features, for staged comparison of real-time detection data.

6. The AI ​​image recognition-based neonatal facial feature rare disease auxiliary diagnosis system according to claim 5, characterized in that: The feature fusion unit includes: The static feature fusion module combines the skin texture features in a single frame image with the three-dimensional structural features of key points in series, or performs weighted fusion through an attention mechanism to form a multi-dimensional static feature vector that includes surface features and geometric features; The dynamic feature fusion module integrates feature change data at different time points and generates dynamic evolution features based on growth stage weights. The growth stages are clearly divided into early and late stages based on the newborn's age. The growth stage weights are automatically adjusted based on the newborn's age to highlight the key diagnostic features of each growth stage. The normalization processing module standardizes the fused static and dynamic features, uses the Z-score normalization method to unify the feature dimensions and dimensions, eliminates the scale differences of different feature types, and provides input data with consistent format for subsequent diagnostic models.

7. The AI ​​image recognition-based neonatal facial feature rare disease auxiliary diagnosis system according to claim 6, characterized in that: The comprehensive diagnostic unit comprises: The weight allocation module dynamically adjusts the weight coefficients of skin features, key point features, and their dynamic evolution features based on feature type and growth stage through a gradient descent algorithm, giving priority to features with high correlation with the current age. The multimodal comparison module uses support vector machines (SVMs) and random forest algorithms to classify static feature vectors and identify structural / texture matches with known rare diseases. At the same time, dynamic time warping technology is used to match the real-time feature evolution trajectory with the pathological curve in the standard model library, calculate the similarity distance of the time series, and achieve dual verification of static structure and dynamic evolution; In dynamic time warping technology, two time series are calculated and distance When we first define the local distance matrix , and then solve the cumulative distance matrix through dynamic programming : ； The final similarity distance ; in, represents the dynamic evolution time series of newborn facial features collected in real time, Represents real-time time series The feature data points, represents the index of the data point in the sequence, represents the pre-built standard evolution time series of rare disease facial features, Represents a standard time series The feature data points, represents the index of the data point in the standard sequence, Represents real-time time series With standard time series Similarity distance calculated by dynamic time warping technique; Represents real-time time series Middle data points With standard time series Middle data points The local distance between Represents the cumulative distance matrix elements in the dynamic time warping process, used to represent the starting point of the real-time sequence To data points, starting from the standard sequence To The cumulative distance of the best matching path of each data point; Indicates the cumulative distance between the previous real-time data point and the current standard data point. Indicates the cumulative distance between the current real-time data point and the previous standard data point. Indicates the cumulative distance between the previous real-time data point and the previous standard data point. Represents the final element of the cumulative distance matrix, i.e., the real-time time series With standard time series The total cumulative distance after complete matching; The report generation module generates a visual diagnostic report based on the comparison results, including: feature matching heat map, 3D facial model, dynamic evolution curve, clinical recommendations, and annotates the contribution of each feature to the diagnostic results.

Images