A method and system for monitoring the state of an eye patch in jaundice phototherapy

Abstract

The application provides a jaundice phototherapy eyecup state monitoring method and system, and the method comprises the following steps: acquiring facial images of a historical target object during blue light irradiation; extracting feature information of normal facial images and abnormal facial images to obtain normal feature vectors and abnormal feature vectors, and expanding the abnormal feature vectors to generate substitute feature vectors corresponding to the normal feature vectors and the abnormal feature vectors after expansion; training an eyecup wearing state recognition model by using the substitute feature vectors, recognizing the eyecup wearing state of a current target object by using the eyecup wearing state recognition model, and taking different prompt information according to different eyecup wearing states. The application expands abnormal samples in a differentiated manner, and generates privacy-protected substitute feature vectors by using segment vector weighted summation, which is beneficial to solving the problems of rare abnormal samples and data privacy, and improves the generalization ability and safety of eyecup wearing state recognition.

Description

Technical Field

[0001] This invention relates to the field of medical technology, and more specifically, to a method and system for monitoring the status of an eye mask during phototherapy for jaundice. Background Technology

[0002] Neonatal jaundice is a common condition in newborns, and phototherapy is its primary treatment. During phototherapy, newborns must wear a light-blocking eye shield to protect their retina. However, newborns are very active, and the eye shield can easily fall off, shift, or become inadequately covering the retina. If not detected and addressed promptly, this can lead to vision damage. Current monitoring methods mainly rely on regular rounds by medical staff, which is not only inefficient and costly but also difficult to detect sudden eye shield detachment, posing a safety hazard. While video analysis-based methods can achieve automated monitoring, the number of abnormal samples of eye shield detachment in real-world scenarios is scarce, leading to a severe data imbalance problem in model training. Furthermore, directly transmitting or processing newborn facial images poses a privacy risk. Therefore, there is an urgent need for an intelligent eye shield status monitoring method that can both protect privacy and effectively address the data imbalance problem. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for monitoring the state of an eye mask during phototherapy for jaundice, so as to improve the above-mentioned problems.

[0004] To achieve the above objectives, this application provides the following technical solution: On one hand, embodiments of this application provide a method for monitoring the state of an eye patch during phototherapy for jaundice, the method comprising: Acquire facial images of the target object during blue light exposure, including normal facial images with the blindfold properly worn and abnormal facial images without the blindfold properly worn. Extract feature information from normal and abnormal facial images to obtain normal and abnormal feature vectors. Expand the abnormal feature vectors to generate alternative feature vectors corresponding to the abnormal and normal feature vectors. A blindfold wearing status recognition model is trained using alternative feature vectors. This model is then used to identify the current blindfold wearing status of the target object, and different prompts are given based on the different blindfold wearing statuses.

[0005] Secondly, this application provides an eye patch status monitoring system during jaundice phototherapy, the system comprising: The acquisition module is used to acquire facial images of historical target objects during blue light irradiation. The facial images include normal facial images with the blindfold properly worn and abnormal facial images without the blindfold properly worn. The extraction module is used to extract feature information from normal and abnormal facial images, obtain normal feature vectors and abnormal feature vectors, expand the abnormal feature vectors, and generate replacement feature vectors corresponding to the abnormal and normal feature vectors after expansion. The prompting module is used to train an eye mask wearing status recognition model using alternative feature vectors, identify the current eye mask wearing status of the target object using the eye mask wearing status recognition model, and provide different prompting information based on different eye mask wearing statuses.

[0006] Thirdly, this application provides an eye patch status monitoring device for jaundice phototherapy, the device comprising a memory and a processor. The memory stores a computer program; the processor executes the computer program to implement the steps of the above-described eye patch status monitoring method for jaundice phototherapy.

[0007] Fourthly, this application provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for monitoring the state of an eye mask during jaundice phototherapy.

[0008] The beneficial effects of this invention are as follows: 1. This invention clusters a small number of real anomaly feature vectors to obtain multiple anomaly pattern clusters, and divides these clusters into two classes based on a distance threshold calculated using contour coefficients and local density. The number of augmentation samples is then dynamically determined according to the proportion of samples in each cluster. Furthermore, augmentation samples that closely resemble real anomalies are generated by sampling perturbation vectors from a multivariate Gaussian distribution. This solves the data imbalance problem caused by scarce anomaly samples and improves the generalization ability of the eye mask wearing status recognition model.

[0009] 2. This invention divides the original feature vector into multiple segment vectors. After learning the weight coefficients through a multilayer perceptron, the segment vectors are weighted and summed to generate a substitute feature vector. This substitute feature vector is only composed of a weighted combination of segment vectors and does not contain information from the original complete feature vector, thus protecting the facial privacy of newborns during model training and application.

[0010] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of the eye mask status monitoring method during jaundice phototherapy as described in this embodiment of the invention; Figure 2 This is a schematic diagram of the eye mask status monitoring system in jaundice phototherapy as described in this embodiment of the invention; Figure 3 This is a schematic diagram of the eye mask status monitoring device in jaundice phototherapy as described in this embodiment of the invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0014] It should be noted that similar reference numerals or letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0015] Example 1 like Figure 1 As shown, this embodiment provides a method for monitoring the state of an eye mask during phototherapy for jaundice, which includes steps S1, S2 and S3.

[0016] Step S1: Obtain facial images of the target object during blue light irradiation. The facial images include normal facial images with the eye mask properly worn and abnormal facial images without the eye mask properly worn. In this step, the target group is the historical recipients, namely newborns who have previously received phototherapy.

[0017] Step S2: Extract feature information from normal and abnormal facial images to obtain normal and abnormal feature vectors. Expand the abnormal feature vectors to generate alternative feature vectors corresponding to the abnormal and normal feature vectors. In this step, a ResNet-50 convolutional neural network pre-trained on the ImageNet dataset is used as the feature extractor. Each normal facial image is input into this feature extractor, and it outputs a normal feature vector. Each abnormal facial image is similarly input into the feature extractor, and it outputs an abnormal feature vector. Since the number of abnormal samples with detached eye masks in real-world scenarios is generally small, it is necessary to augment the abnormal feature vectors. In other words, augmenting the abnormal feature vectors can address the data imbalance problem caused by the scarcity of abnormal samples.

[0018] In this step, the specific implementation steps for expanding the abnormal feature vector include steps S21 and S22; Step S21: Cluster all the abnormal feature vectors to obtain multiple abnormal feature clusters, calculate the center of each abnormal feature cluster, and obtain multiple cluster centers; count the number of samples in each abnormal feature cluster and calculate the average of all sample counts. This step can use the K-means algorithm for clustering. For each anomalous feature cluster, calculate the average of all anomalous feature vectors within that cluster, and use this average as the cluster center. Count the number of samples in each anomalous feature cluster (i.e., the number of anomalous feature vectors contained in each cluster), and calculate the average number of samples across all anomalous feature clusters. Step S22: Determine the abnormal feature cluster with the largest number of samples as the reference feature cluster, and denote the remaining abnormal feature clusters as non-reference feature clusters; calculate the feature distance between each non-reference feature cluster and the reference feature cluster, using the Euclidean distance between the cluster centers; set a distance threshold based on the feature distance, and determine the non-reference feature clusters with a feature distance greater than the distance threshold as the first cluster, and determine the non-reference feature clusters with a feature distance less than or equal to the distance threshold as the second cluster; complete the expansion of the abnormal feature vector based on the first and second clusters.

[0019] In this step, the purpose of dividing the non-reference feature clusters into a first cluster and a second cluster is to achieve differentiated anomalous feature vector augmentation. Specifically: The second cluster represents anomaly types that are relatively close to the reference cluster in feature space. These anomalies occur relatively frequently in real-world scenarios and belong to common anomaly patterns. For these patterns, the augmentation quantity should be increased (e.g., multiplied by an enhancement factor) to strengthen their representativeness, enabling the model to fully learn the characteristics of common anomalies.

[0020] The first cluster represents anomalies that differ significantly from the reference cluster. These anomalies occur infrequently in real-world scenarios and are considered rare anomaly patterns. For these patterns, the augmentation amount should be reduced (e.g., by multiplying by a decay factor) to avoid oversampling and introducing unrealistic samples.

[0021] If this classification is not performed, applying the same expansion strategy to all non-reference feature clusters will lead to insufficient representativeness of common patterns or overfitting of rare patterns. Therefore, dividing the clusters into two categories based on a distance threshold is a prerequisite for achieving subsequent differentiated expansion. In this step, the specific implementation steps for setting the distance threshold based on the feature distance include steps S221 and S222; Step S221: Obtain the minimum and maximum values ​​among all feature distances. Within the interval between the minimum and maximum values, perform equal-interval sampling according to a preset fixed step size to obtain multiple candidate distance thresholds. Calculate the overall contour coefficient corresponding to each candidate distance threshold and select the candidate distance threshold that maximizes the overall contour coefficient as the first candidate threshold. In this step, the interval is (minimum value, maximum value); the specific implementation steps for calculating the overall contour coefficient corresponding to each candidate distance threshold include steps S2211 and S2212; Step S2211: For each candidate distance threshold, divide all non-reference feature clusters into two groups: the first group consists of non-reference feature clusters whose feature distance is greater than the candidate distance threshold, and the second group consists of non-reference feature clusters whose feature distance is less than or equal to the candidate distance threshold; after grouping, calculate the silhouette coefficient corresponding to each non-reference feature cluster, wherein the first average distance between each non-reference cluster and other non-reference feature clusters in the same group is calculated, and the second average distance between each non-reference cluster and other non-reference feature clusters in the other group is also calculated; calculate the difference between the first average distance and the second average distance, and divide the difference by the larger of the first average distance and the second average distance to obtain the silhouette coefficient; In this step, the first average distance between each non-reference cluster and other non-reference feature clusters in the same group is calculated, and the second average distance between each non-reference cluster and another group of non-reference feature clusters is also calculated. This can be understood as: The specific calculation steps are as follows: Calculation of the first average distance: For the non-reference feature cluster to be calculated (denoted as cluster A), find all other non-reference feature clusters in its group. Calculate the Euclidean distance between the cluster center of cluster A and the cluster centers of other non-reference feature clusters respectively. Take the average of all the obtained Euclidean distances, and the result is the first average distance.

[0022] Calculation of the second average distance: For cluster A, find all non-reference feature clusters in the other group (i.e., non-reference feature clusters not in the same group as cluster A). Calculate the Euclidean distance between the cluster center of cluster A and the cluster center of each non-reference feature cluster in the other group. Take the average of all the obtained Euclidean distances, and the result is the second average distance.

[0023] In addition, when the number of non-reference feature clusters in a group is 1, the first average distance of the non-reference feature cluster is defined as 0, and its silhouette coefficient is set to 0 accordingly. Step S2212: After calculating the contour coefficient corresponding to each non-reference feature cluster, average all the contour coefficients to obtain the overall contour coefficient corresponding to each candidate distance threshold.

[0024] Step S222: Calculate the local density of the reference feature cluster and the non-reference feature cluster; multiply the feature distance of each non-reference feature cluster by the ratio of the local density of the reference feature cluster to its own local density to obtain the adjusted feature distance; calculate the mean and standard deviation of all adjusted feature distances, add twice the standard deviation to the mean to obtain the second candidate threshold; obtain the distance threshold based on the first candidate threshold and the second candidate threshold.

[0025] In this step, the local density is: the number of samples in the cluster divided by the average of the squared distances from all samples in the cluster to the cluster center. The specific calculation steps are as follows: Obtain the cluster center; calculate the Euclidean distance from each sample in the cluster to the cluster center; square each Euclidean distance to obtain the squared distance; calculate the average of all squared distances to obtain the average squared distance value; divide the number of samples in the cluster by the average squared distance value to obtain the local density of the cluster.

[0026] In this step, the specific implementation steps for obtaining the distance threshold based on the first candidate threshold and the second candidate threshold include steps S2221 and S2222. Step S2221: Obtain the maximum and minimum values ​​among all feature distances, calculate the standard deviation and mean of all feature distances, and calculate the lower limit constraint value and upper limit constraint value of the distance threshold based on the maximum, minimum, mean and standard deviation. Wherein, the lower limit constraint value of the distance threshold = max(minimum value, mean - 2 × standard deviation), and the upper limit constraint value of the distance threshold = min(maximum value, mean + 2 × standard deviation). Step S2222: Calculate the average of the first candidate threshold and the second candidate threshold as a temporary threshold; Distance threshold = max(lower limit constraint value of distance threshold, min(upper limit constraint value of distance threshold, temporary threshold)).

[0027] This step utilizes a statistical interval constructed from the mean ± 2 standard deviations to adaptively reflect the central tendency and dispersion of feature distances, avoiding interference from extreme values ​​in the threshold. Minimum and maximum pruning ensures that the upper and lower limits do not exceed the actual data range, preventing a threshold that is too small from classifying all non-reference clusters into the first cluster, or too large from classifying all non-reference clusters into the second cluster. The arithmetic mean of the first and second candidate thresholds is used as a temporary threshold, combining the complementary advantages of silhouette coefficients and density weighting. Finally, the temporary threshold is constrained within the statistical interval, ensuring the rationality of the output distance threshold.

[0028] In step S22, the specific implementation steps for expanding the abnormal feature vector based on the first cluster and the second cluster include steps S223 and S224. Step S223: For each first cluster, divide its sample count by the average sample count of all abnormal feature clusters to obtain a scaling factor. Multiply the preset baseline generation count by this scaling factor, and then multiply by the attenuation factor λ to obtain the anomalous feature vector expansion count corresponding to the first cluster. For each second cluster, divide its sample count by the average sample count of all abnormal feature clusters to obtain a scaling factor. Multiply the preset baseline generation count by this scaling factor, and then multiply by the enhancement factor μ to obtain the anomalous feature vector expansion count corresponding to the second cluster. Wherein, 0 < λ < 1, μ > 1. In this step, the preset number of baseline generated data can be, for example, 100, λ can be 0.5, and μ can range from 1 to 3, for example, μ is 1.2. In this way, the number of generated data for the first cluster (rare anomaly patterns) is appropriately reduced to avoid oversampling; the number of generated data for the second cluster (common anomaly patterns) is appropriately increased to enhance its representativeness, thereby more reasonably balancing the data distribution and improving the model training effect. Step S224: For each first cluster and each second cluster, perform the following steps respectively: calculate the first covariance matrix of the cluster; independently sample perturbation vectors multiple times from a multivariate Gaussian distribution with zero vector as mean and the first covariance matrix as covariance matrix; add the perturbation vector obtained from each sampling to the cluster center of the cluster to obtain the expanded abnormal feature vector.

[0029] In this step, the number of samplings is equal to the number of anomalous feature vectors expanded for each cluster; This step generates new, reasonable anomalous features by adding random perturbations near the cluster centers, while preserving the original distribution characteristics of the anomalous features. First, the first covariance matrix of all anomalous feature vectors within the cluster is calculated. Then, a multivariate Gaussian distribution is constructed using the zero vector as the mean and the first covariance matrix as the covariance matrix. Perturbation vectors are independently sampled multiple times from this distribution; the number of samples is the augmentation quantity calculated in step S223. Since the covariance matrix reflects the distribution shape of the original data, the sampled perturbation vectors naturally inherit the covariance structure of the cluster. Finally, each perturbation vector is added to the cluster center to obtain the augmented anomalous feature vector. The vectors generated in this way both revolve around the cluster center and maintain the variance and correlation within the original clusters, thus closely resembling the true anomalous features.

[0030] After generating the expanded abnormal feature vector, you can also choose to filter the generated abnormal feature vector. The specific steps include steps S2241 and S2242. Step S2241: Denote the expanded abnormal feature vector as the expanded abnormal feature vector. For all expanded abnormal feature vectors corresponding to each abnormal feature cluster, calculate the cosine similarity between each expanded abnormal feature vector and the cluster center of the abnormal feature cluster. Determine the expanded abnormal feature vectors with a cosine similarity greater than the first similarity threshold and less than the second similarity threshold as the preliminary expanded abnormal feature vectors corresponding to each abnormal feature cluster. Step S2242: Group the abnormal feature clusters according to the number of samples in each abnormal feature cluster, where the group number is the logarithmic value of each cluster. 10 (Sample count) is rounded down, and abnormal feature clusters with the same group number are grouped together. The preliminary expanded abnormal feature vectors corresponding to each abnormal feature cluster in each group are collected to obtain the preliminary expanded abnormal feature vector set for each group. In the candidate expanded abnormal feature vector set, all preliminary expanded abnormal feature vectors are sorted in descending order according to the cosine similarity corresponding to each preliminary expanded abnormal feature vector. After sorting, the first preset number of preliminary expanded abnormal feature vectors is selected for output to obtain the filtered expanded abnormal feature vectors for each group. If the number of preliminary expanded abnormal feature vectors in the preliminary expanded abnormal feature vector set is less than the first number, then all preliminary expanded abnormal feature vectors in the preliminary expanded abnormal feature vector set are output to obtain the filtered expanded abnormal feature vectors for each group.

[0031] In the above screening steps, by setting a first similarity threshold and a second similarity threshold, expanded samples that are too close to or too different from the cluster center are eliminated, and only high-quality features that are highly correlated with the original abnormal pattern and have moderate diversity are retained. The logarithmic grouping mechanism is used to group clusters with similar sample numbers together, which avoids small sample clusters being overwhelmed by large sample clusters in the sorting and selection process, and ensures that the expanded features of each type of cluster have a reasonable proportion of output opportunities. Finally, by sorting in descending order within the group and selecting a preset number of vectors, the output candidate expanded abnormal feature vectors maintain representativeness while having a controllable number, thereby improving the quality of the training data.

[0032] In step S2, the specific implementation steps for generating the abnormal feature vector and the replacement feature vector corresponding to the normal feature vector after expansion include steps S23-S25; Step S23: Record each normal feature vector and each abnormal feature vector as the first feature vector; In this step, if steps S2241 to S2242 have been performed, then in this step, the filtered expanded abnormal feature vector, the original normal feature vector, and the abnormal feature vector are all recorded as the first feature vector. Step S24: Divide each first feature vector uniformly into K segment vectors, where the length of the segment vector is calculated as L = ceil(d / K), and d is the dimension of the first feature vector; pad the end of the first feature vector with zeros to extend its length to L×K; divide the extended vector into K segment vectors of length L in sequence; input each segment vector into a multilayer perceptron, and use the output of the multilayer perceptron as the initial weight coefficient for each segment vector. The multilayer perceptron includes an input layer, two hidden layers, and an output layer. The number of nodes in the input layer is equal to the dimension of the segment vector. The number of nodes in the first hidden layer is the number of nodes in the input layer divided by 2 and then rounded down. The number of nodes in the second hidden layer is the number of nodes in the first hidden layer divided by 2 and then rounded down. The output layer has one node and does not contain an activation function. In this step, ceil(d / K) means rounding up the result of dividing d by K; Step S25: Normalize the initial weight coefficients corresponding to all segment vectors to obtain the association weights corresponding to each segment vector. The normalization process includes: calculating the exponential function value with the natural constant as the base for each initial weight coefficient, summing all the exponential function values ​​to obtain the total exponential sum, and then dividing each exponential function value by the total exponential sum to obtain the corresponding association weight; multiplying each segment vector by its corresponding association weight to obtain a weighted vector, and then adding all the weighted vectors element by element according to their corresponding positions to obtain the alternative feature vector corresponding to the first feature vector.

[0033] This step can be understood as: After expansion, a large number of abnormal feature vectors (including original and generated ones) are obtained. The next step is to convert the normal feature vectors and abnormal feature vectors (collectively referred to as the first feature vectors) into corresponding alternative feature vectors for subsequent model training.

[0034] The purpose of generating alternative feature vectors is to protect privacy: the original feature vectors are deep features directly extracted from newborn facial images, which poses a risk of leaking the newborn's biological privacy. This step involves uniformly dividing the first feature vector into multiple segment vectors, then using a multilayer perceptron to learn weight coefficients for each segment vector, and finally summing the segment vectors according to their weights. The resulting alternative feature vector is only composed of a weighted combination of segment vectors and no longer contains information from the original complete feature vector, thus effectively protecting the newborn's facial privacy during model training and application.

[0035] Step S3: Train the blindfold wearing status recognition model using alternative feature vectors, identify the current blindfold wearing status of the target object using the blindfold wearing status recognition model, and provide different prompts based on different blindfold wearing statuses.

[0036] In this step, the alternative feature vectors corresponding to the normal feature vectors and the alternative feature vectors corresponding to the abnormal feature vectors are used as training samples, labeled as the state of wearing the eye mask, normal or abnormal. The alternative feature vectors corresponding to the normal feature vectors are labeled as normal, and the alternative feature vectors corresponding to the abnormal feature vectors are labeled as abnormal. During training, the alternative feature vectors can be used to train the convolutional neural network to obtain the eye mask wearing state recognition model.

[0037] After the model is trained, it is used for real-time monitoring. When a newborn begins receiving blue light therapy, the current facial image of the newborn is acquired in real time and input into the feature extractor to obtain the current first feature vector. Then, following the same method as in the training phase, the alternative feature vector of the current first feature vector is calculated, with the calculation steps being the same as steps S24-S25. This alternative feature vector is then input into the trained eye mask wearing status recognition model. The model outputs the result of the current eye mask wearing status, and the prompts taken according to different statuses may include: no prompt or only logging in the normal state, and sound and light alarms and push messages to the nurse's mobile terminal in the abnormal state.

[0038] Example 2 like Figure 2 As shown in the figure, this embodiment provides an eye mask status monitoring system during jaundice phototherapy. The system includes an acquisition module 1, an extraction module 2, and a prompting module 3.

[0039] The acquisition module 1 is used to acquire facial images of historical target objects during blue light irradiation. The facial images include normal facial images with the blindfold properly worn and abnormal facial images without the blindfold properly worn. Extraction module 2 is used to extract feature information from normal facial images and abnormal facial images, obtain normal feature vectors and abnormal feature vectors, expand the abnormal feature vectors, and generate replacement feature vectors corresponding to the abnormal feature vectors and normal feature vectors after expansion. Prompt module 3 is used to train an eye mask wearing status recognition model using alternative feature vectors, identify the current eye mask wearing status of the target object using the eye mask wearing status recognition model, and provide different prompt information according to different eye mask wearing statuses.

[0040] In one specific embodiment of this disclosure, the extraction module 2 further includes a clustering unit 21 and an expansion unit 22.

[0041] Clustering unit 21 is used to cluster all abnormal feature vectors to obtain multiple abnormal feature clusters, calculate the center of each abnormal feature cluster, and obtain multiple cluster centers; count the number of samples in each abnormal feature cluster and calculate the average of the number of samples. The expansion unit 22 is used to determine the abnormal feature cluster with the largest number of samples as the reference feature cluster, and the remaining abnormal feature clusters as non-reference feature clusters; calculate the feature distance between each non-reference feature cluster and the reference feature cluster, and the feature distance is the Euclidean distance between the cluster centers; set a distance threshold based on the feature distance, determine the non-reference feature clusters with a feature distance greater than the distance threshold as the first cluster, and determine the non-reference feature clusters with a feature distance less than or equal to the distance threshold as the second cluster; and complete the expansion of the abnormal feature vector based on the first cluster and the second cluster.

[0042] In one specific embodiment of this disclosure, the expansion unit 22 further includes a selection unit 221 and an adjustment unit 222.

[0043] Selection unit 221 is used to obtain the minimum and maximum values ​​among all feature distances. Within the interval between the minimum and maximum values, it performs equal-interval sampling according to a preset fixed step size to obtain multiple candidate distance thresholds. It calculates the overall contour coefficient corresponding to each candidate distance threshold and selects the candidate distance threshold that maximizes the overall contour coefficient as the first candidate threshold. The adjustment unit 222 is used to calculate the local density of the reference feature cluster and the non-reference feature cluster; multiply the feature distance of each non-reference feature cluster by the ratio of the local density of the reference feature cluster to its own local density to obtain the adjusted feature distance; calculate the mean and standard deviation of all adjusted feature distances, add twice the standard deviation to the mean to obtain the second candidate threshold; and obtain the distance threshold based on the first candidate threshold and the second candidate threshold.

[0044] In one specific embodiment of this disclosure, the selection unit 221 further includes a partitioning unit 2211 and a first calculation unit 2212.

[0045] The partitioning unit 2211 is used to divide all non-reference feature clusters into two groups for each candidate distance threshold: the first group consists of non-reference feature clusters whose feature distance is greater than the candidate distance threshold, and the second group consists of non-reference feature clusters whose feature distance is less than or equal to the candidate distance threshold. After grouping, the silhouette coefficient corresponding to each non-reference feature cluster is calculated, wherein the first average distance between each non-reference cluster and other non-reference feature clusters in the same group is calculated, and the second average distance between each non-reference cluster and other non-reference feature clusters in the other group is also calculated. The difference between the first average distance and the second average distance is calculated, and the difference is divided by the larger of the first average distance and the second average distance to obtain the silhouette coefficient. The first calculation unit 2212 is used to calculate the contour coefficient corresponding to each non-reference feature cluster, and then average all the contour coefficients to obtain the overall contour coefficient corresponding to each candidate distance threshold.

[0046] In one specific embodiment of this disclosure, the adjustment unit 222 further includes an acquisition unit 2221 and a second calculation unit 2222.

[0047] The acquisition unit 2221 is used to acquire the maximum and minimum values ​​among all feature distances, calculate the standard deviation and mean of all feature distances, and calculate the lower limit constraint value and upper limit constraint value of the distance threshold based on the maximum, minimum, mean and standard deviation. The lower limit constraint value of the distance threshold is max(minimum value, mean - 2 × standard deviation), and the upper limit constraint value of the distance threshold is min(maximum value, mean + 2 × standard deviation). The second calculation unit 2222 is used to calculate the average of the first candidate threshold and the second candidate threshold as a temporary threshold; distance threshold = max(lower limit constraint value of distance threshold, min(upper limit constraint value of distance threshold, temporary threshold)).

[0048] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0049] Example 3 Corresponding to the above method embodiments, this disclosure also provides an eye mask status monitoring device for jaundice phototherapy. The eye mask status monitoring device for jaundice phototherapy described below can be referred to in correspondence with the eye mask status monitoring method for jaundice phototherapy described above.

[0050] Figure 3 This is a block diagram illustrating an eye patch status monitoring device 300 during jaundice phototherapy, according to an exemplary embodiment. Figure 3 As shown, the eye patch status monitoring device 300 for jaundice phototherapy may include: a processor 301 and a memory 302. The eye patch status monitoring device 300 for jaundice phototherapy may also include one or more of a multimedia component 303, an I / O interface 304, and a communication component 305.

[0051] The processor 301 controls the overall operation of the eye patch status monitoring device 300 for jaundice phototherapy to complete all or part of the steps in the aforementioned eye patch status monitoring method for jaundice phototherapy. The memory 302 stores various types of data to support the operation of the eye patch status monitoring device 300 for jaundice phototherapy. This data may include, for example, instructions for any application or method operating on the eye patch status monitoring device 300 for jaundice phototherapy, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 302 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 303 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 302 or transmitted via the communication component 305. The audio component also includes at least one speaker for outputting audio signals. I / O interface 304 provides an interface between processor 301 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 305 is used for wired or wireless communication between the eye patch status monitoring device 300 and other devices in the jaundice phototherapy. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 305 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0052] In an exemplary embodiment, the eye patch status monitoring device 300 for jaundice phototherapy may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described eye patch status monitoring method for jaundice phototherapy.

[0053] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided, which, when executed by a processor, implement the steps of the above-described method for monitoring the state of an eye patch during jaundice phototherapy. For example, the computer-readable storage medium may be the memory 302 including the program instructions, which may be executed by the processor 301 of the eye patch state monitoring device 300 during jaundice phototherapy to complete the above-described method for monitoring the state of an eye patch during jaundice phototherapy.

[0054] Example 4 Corresponding to the above method embodiments, this disclosure also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the eye mask status monitoring method in jaundice phototherapy described above.

[0055] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the eye patch status monitoring method in jaundice phototherapy as described in the above method embodiments.

[0056] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring the state of an ocular mask in jaundice phototherapy, characterized by, include: Acquire facial images of the target object during blue light exposure, including normal facial images with the blindfold properly worn and abnormal facial images without the blindfold properly worn. Extract feature information from normal and abnormal facial images to obtain normal and abnormal feature vectors. Expand the abnormal feature vectors to generate alternative feature vectors corresponding to the abnormal and normal feature vectors. A blindfold wearing status recognition model is trained using alternative feature vectors. The blindfold wearing status recognition model is then used to identify the current blindfold wearing status of the target object, and different prompts are given according to different blindfold wearing statuses. This includes expanding the abnormal feature vector, including: Cluster all the abnormal feature vectors to obtain multiple abnormal feature clusters, calculate the center of each abnormal feature cluster, and obtain multiple cluster centers; count the number of samples in each abnormal feature cluster, and calculate the average of the number of samples in all clusters. The abnormal feature cluster with the largest number of samples is identified as the reference feature cluster, and the remaining abnormal feature clusters are designated as non-reference feature clusters. The feature distance between each non-reference feature cluster and the reference feature cluster is calculated using the Euclidean distance between the cluster centers. A distance threshold is set based on the feature distance, and non-reference feature clusters with a feature distance greater than the threshold are identified as the first cluster, while non-reference feature clusters with a feature distance less than or equal to the threshold are identified as the second cluster. The abnormal feature vectors are then expanded based on the first and second clusters. Among them, setting a distance threshold based on feature distance includes: Obtain the minimum and maximum values ​​among all feature distances. Within the interval between the minimum and maximum values, perform equal-interval sampling according to a preset fixed step size to obtain multiple candidate distance thresholds. Calculate the overall contour coefficient corresponding to each candidate distance threshold and select the candidate distance threshold that maximizes the overall contour coefficient as the first candidate threshold. Calculate the local density of the reference feature cluster and the non-reference feature cluster; multiply the feature distance of each non-reference feature cluster by the ratio of the local density of the reference feature cluster to its own local density to obtain the adjusted feature distance; calculate the mean and standard deviation of all adjusted feature distances, add twice the standard deviation to the mean to obtain the second candidate threshold; obtain the distance threshold based on the first candidate threshold and the second candidate threshold.

2. The jaundice phototherapy eye pad state monitoring method according to claim 1, characterized by, Calculate the overall contour coefficient corresponding to each candidate distance threshold, including: For each candidate distance threshold, all non-reference feature clusters are divided into two groups: the first group consists of non-reference feature clusters whose feature distance is greater than the candidate distance threshold, and the second group consists of non-reference feature clusters whose feature distance is less than or equal to the candidate distance threshold. After grouping, the silhouette coefficient corresponding to each non-reference feature cluster is calculated. Specifically, the first average distance between each non-reference cluster and other non-reference feature clusters in the same group is calculated, and the second average distance between each non-reference cluster and other non-reference feature clusters in the other group is also calculated. The difference between the first average distance and the second average distance is calculated, and the difference is divided by the larger of the first average distance and the second average distance to obtain the silhouette coefficient. After calculating the contour coefficients corresponding to each non-reference feature cluster, the contour coefficients are averaged to obtain the overall contour coefficient corresponding to each candidate distance threshold.

3. The jaundice phototherapy eye pad status monitoring method according to claim 1, wherein, The distance threshold is obtained based on the first candidate threshold and the second candidate threshold, including: Obtain the maximum and minimum values ​​among all feature distances, calculate the standard deviation and mean of all feature distances, and calculate the lower limit constraint value and upper limit constraint value of the distance threshold based on the maximum, minimum, mean and standard deviation. The lower limit constraint value of the distance threshold = max(minimum value, mean - 2 × standard deviation), and the upper limit constraint value of the distance threshold = min(maximum value, mean + 2 × standard deviation). Calculate the average of the first candidate threshold and the second candidate threshold as the temporary threshold; Distance threshold = max(lower limit constraint value of distance threshold, min(upper limit constraint value of distance threshold, temporary threshold)).

4. A system for monitoring the status of an ocular mask in jaundice phototherapy, characterized by include: The acquisition module is used to acquire facial images of historical target objects during blue light irradiation. The facial images include normal facial images with the blindfold properly worn and abnormal facial images without the blindfold properly worn. The extraction module is used to extract feature information from normal and abnormal facial images, obtain normal feature vectors and abnormal feature vectors, expand the abnormal feature vectors, and generate replacement feature vectors corresponding to the abnormal and normal feature vectors after expansion. The prompt module is used to train a blindfold wearing status recognition model using alternative feature vectors, identify the current blindfold wearing status of the target object using the blindfold wearing status recognition model, and provide different prompt information according to different blindfold wearing statuses; The extraction module includes: Clustering units are used to cluster all abnormal feature vectors to obtain multiple abnormal feature clusters. The center of each abnormal feature cluster is calculated to obtain multiple cluster centers. The number of samples in each abnormal feature cluster is counted, and the average number of samples is calculated. An expansion unit is used to identify the anomaly feature cluster with the largest number of samples as the reference feature cluster, and the remaining anomaly feature clusters as non-reference feature clusters; calculate the feature distance between each non-reference feature cluster and the reference feature cluster, using the Euclidean distance between the cluster centers; set a distance threshold based on the feature distance, identify non-reference feature clusters with a feature distance greater than the distance threshold as the first cluster, and identify non-reference feature clusters with a feature distance less than or equal to the distance threshold as the second cluster; and complete the expansion of the anomaly feature vector based on the first and second clusters. The expansion unit includes: The selection unit is used to obtain the minimum and maximum values ​​among all feature distances. Within the interval between the minimum and maximum values, it performs equal-interval sampling according to a preset fixed step size to obtain multiple candidate distance thresholds. It calculates the overall contour coefficient corresponding to each candidate distance threshold and selects the candidate distance threshold that maximizes the overall contour coefficient as the first candidate threshold. An adjustment unit is used to calculate the local density of the reference feature cluster and the non-reference feature cluster; multiply the feature distance of each non-reference feature cluster by the ratio of the local density of the reference feature cluster to its own local density to obtain the adjusted feature distance; calculate the mean and standard deviation of all adjusted feature distances, add twice the standard deviation to the mean to obtain the second candidate threshold; and obtain the distance threshold based on the first candidate threshold and the second candidate threshold.

5. The eye shield state monitoring system in jaundice phototherapy according to claim 4, characterized by The selection unit includes: A partitioning unit is used to divide all non-reference feature clusters into two groups for each candidate distance threshold: the first group consists of non-reference feature clusters whose feature distance is greater than the candidate distance threshold, and the second group consists of non-reference feature clusters whose feature distance is less than or equal to the candidate distance threshold. After grouping, the silhouette coefficient corresponding to each non-reference feature cluster is calculated. Specifically, the first average distance between each non-reference cluster and other non-reference feature clusters in the same group is calculated, and the second average distance between each non-reference cluster and other non-reference feature clusters in the other group is also calculated. The difference between the first average distance and the second average distance is calculated, and the difference is divided by the larger of the first average distance and the second average distance to obtain the silhouette coefficient. The first calculation unit is used to calculate the contour coefficient corresponding to each non-reference feature cluster, and then average all the contour coefficients to obtain the overall contour coefficient corresponding to each candidate distance threshold.

6. The eye shield state monitoring system in jaundice phototherapy according to claim 4, characterized by Adjustment unit, including: The acquisition unit is used to acquire the maximum and minimum values ​​among all feature distances, calculate the standard deviation and mean of all feature distances, and calculate the lower limit constraint value and upper limit constraint value of the distance threshold based on the maximum, minimum, mean and standard deviation. The lower limit constraint value of the distance threshold is max(minimum value, mean - 2 × standard deviation), and the upper limit constraint value of the distance threshold is min(maximum value, mean + 2 × standard deviation). The second calculation unit is used to calculate the average of the first candidate threshold and the second candidate threshold as a temporary threshold; distance threshold = max(lower limit constraint value of distance threshold, min(upper limit constraint value of distance threshold, temporary threshold)).

Images