Coronary heart disease preliminary screening method and device based on facial feature analysis and medium

Abstract

The invention discloses a coronary heart disease preliminary screening method and device based on facial feature analysis and a medium, and relates to the technical field of image understanding, and the method comprises the steps: extracting each part sub-region of a human face, and obtaining a tensor image of each sample data through the stacking of the sub-regions; obtaining the flattened feature maps of the tensor images, and integrating the feature maps to obtain a feature matrix of the corresponding sample data; calculating a feature mean value in the image dimension direction according to the feature matrix, and taking the feature mean value as a global feature tensor of a corresponding sample; carrying out dimension reduction processing and nonlinear transformation on the global feature tensor, and carrying out binary classification probability distribution output of relevance between the subregions and the coronary heart disease; performing updating iteration of the classification weight of the sub-region through cross entropy loss function constraint and a stochastic gradient descent optimizer; and performing primary screening of the coronary heart disease through the iterated primary screening model of the coronary heart disease. According to the method, the key feature capturing capability of the model is improved, and then the judgment capability on the coronary heart disease risk is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of image understanding, and in particular to a method, device and medium for initial screening of coronary heart disease based on facial feature analysis. Background Art

[0002] Coronary heart disease detection mostly relies on invasive methods, such as blood tests and coronary angiography. Although these methods are accurate, they are costly and complex, making them difficult to be widely used in daily health monitoring. Studies have shown that some head and facial features (such as frontal baldness, earlobe creases, etc.) are associated with coronary heart disease, but existing methods are difficult to accurately extract these features and rely on manual annotation, which is inefficient and has limited accuracy. Traditional deep learning models are not good at extracting subtle features related to coronary heart disease. Therefore, there is an urgent need for a new method based on advanced deep learning networks that can efficiently and accurately extract and analyze head and facial features related to coronary heart disease, so as to achieve non-invasive and convenient prediction of coronary heart disease risk. Summary of the invention

[0003] In order to achieve non-invasive and convenient risk prediction of coronary heart disease, the present invention proposes a method for initial screening of coronary heart disease based on facial feature analysis, comprising the steps of: S1: Extract sub-regions of various parts of the face by cropping a facial image sample dataset containing healthy individuals and patients with coronary heart disease, and obtain tensor images of each sample data by stacking the sub-regions; S2: Obtain the flattened feature map of each tensor image by layer-by-layer feature extraction, and obtain the feature matrix of the corresponding sample data by integrating the feature map; S3: Calculate the feature mean in the image dimension direction according to the feature matrix, and use the feature mean as the global feature tensor of the corresponding sample; S4: By performing dimensionality reduction and nonlinear transformation on the global feature tensor, a binary classification probability distribution output of the association between the sub-region and coronary heart disease is obtained; S5: Update the sub-region classification weights iteratively through the cross entropy loss function constraint and the stochastic gradient descent optimizer, and return to step S1 before the classification accuracy reaches the standard; S6: A coronary heart disease screening model with a satisfactory accuracy after iteration is used to perform initial screening for coronary heart disease based on the target patient's facial image.

[0004] Furthermore, in the step S1, before performing the sub-region extraction, the following steps are also included: The face boundary box in the face image is selected using the YOLO target detection algorithm.

[0005] Furthermore, in the step S2, the feature maps of all subsequent layers are obtained based on the output of each layer's feature map as input under the dense connection mechanism of the DenseNet121 network.

[0006] Furthermore, in the S2 step, the importance of each feature channel is dynamically adjusted through the channel attention mechanism to highlight the key features related to coronary heart disease and suppress the interference of irrelevant channels; the facial area related to coronary heart disease is highlighted and the background interference is weakened according to the spatial distribution of the feature map through the spatial attention mechanism.

[0007] Furthermore, by embedding the attention mechanism in each Dense Block of the DenseNet121 network, the feature channel weights are dynamically acquired.

[0008] Furthermore, in the DenseNet121 network, a fully connected layer is used to replace the original classifier part for classification judgment.

[0009] Furthermore, in the step S5, the loss is calculated by forward propagation in the update iteration process, and the model parameters are updated by back propagation, and the area under the ROC curve and the classification accuracy of the training set are calculated in each update cycle.

[0010] The present invention also includes a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method for initial screening of coronary heart disease based on facial feature analysis are implemented.

[0011] Also included is a device for processing data, comprising: a memory having a computer program stored thereon; A processor is used to execute the computer program in the memory to implement the steps of the method for initial screening of coronary heart disease based on facial feature analysis.

[0012] Compared with the prior art, the present invention has at least the following beneficial effects: (1) The method, device and medium for initial screening of coronary heart disease based on facial feature analysis described in the present invention realize efficient feature extraction and classification prediction of specific areas of the face by combining the DenseNet121 network structure with multi-scale feature fusion technology, which enables the model to more accurately identify subtle facial features related to coronary heart disease; (2) By accurately cropping specific facial areas and performing multi-scale feature extraction, this method improves the model's ability to capture key features, thereby enhancing the ability to discriminate the risk of coronary heart disease; (3) Dynamic learning rate adjustment ensures the robustness and accuracy of the model under different data distributions and lighting conditions, which helps to improve the reliability of the model in practical application scenarios; (4) It provides an efficient and deployable solution for low-cost, non-invasive screening of coronary heart disease, which has broad clinical and health management application value, especially for the initial screening of large populations. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A step diagram of a method for initial screening of coronary heart disease based on facial feature analysis; Figure 2 A schematic diagram of the model structure. DETAILED DESCRIPTION

[0014] The following are specific embodiments of the present invention and the accompanying drawings to further describe the technical solution of the present invention, but the present invention is not limited to these embodiments.

[0015] The relationship between coronary artery disease (CAD) and facial features is a relatively new area of ​​research, especially after the use of artificial intelligence and deep learning technology to analyze facial photos, research in this field has made significant progress. Studies have shown that certain facial features may be associated with the occurrence and development of CAD. Although these associations are not sufficient to serve as independent diagnostic evidence, they can provide valuable clues for early screening and risk assessment. In order to find facial features that are more associated with CAD and screen for CAD based on these features, Figure 1 As shown, the present invention proposes a method for initial screening of coronary heart disease based on facial feature analysis, comprising the steps of: S1: Extract sub-regions of various parts of the face by cropping a facial image sample dataset containing healthy individuals and patients with coronary heart disease, and obtain tensor images of each sample data by stacking the sub-regions; S2: Obtain the flattened feature map of each tensor image by layer-by-layer feature extraction, and obtain the feature matrix of the corresponding sample data by integrating the feature map; S3: Calculate the feature mean in the image dimension direction according to the feature matrix, and use the feature mean as the global feature tensor of the corresponding sample; S4: By performing dimensionality reduction and nonlinear transformation on the global feature tensor, a binary classification probability distribution output of the association between the sub-region and coronary heart disease is obtained; S5: Update the sub-region classification weights iteratively through the cross entropy loss function constraint and the stochastic gradient descent optimizer, and return to step S1 before the classification accuracy reaches the standard; S6: A coronary heart disease screening model with a satisfactory accuracy after iteration is used to perform initial screening for coronary heart disease based on the target patient's facial image.

[0016] In order to screen facial features related to coronary heart disease, the present invention first collects data sets containing facial images, which include facial image sample data collected from patients and healthy individuals who have been medically diagnosed with coronary heart disease, and are divided into a coronary heart disease group and a negative group according to the diagnosis results. Each image is labeled (by LabelImg such as automatic labeling or manual correction) and cropped, including sub-regions such as earlobes, foreheads, noses, eyes, etc. that are relevant to coronary heart disease. Of course, in order to reduce the workload and avoid misidentification, YOLO target detection will be used to infer each image to obtain the face bounding box in the image. Then, a cropping operation is applied to each detected face area to obtain sub-regions of different parts. In the cropping process, an automatic adjustment mechanism can be combined to ensure that the area size is consistent and reduce the impact of uneven cropping.

[0017] Since each image contains multiple sub-regions after cropping, in order to better compare and screen, we will also stack these sub-regions into a tensor image. And improve data diversity and model robustness through data enhancement methods (such as brightness adjustment, random rotation, shearing, etc.). This data organization method simultaneously inputs information about local facial features into the deep learning model for joint learning, retaining the independence and overall relevance of key features. In addition, each image is labeled and cropped, including areas associated with coronary heart disease such as earlobes, foreheads, noses, and eyes, which can highlight feature information related to risk prediction while reducing interference from background and irrelevant areas. In addition, the cropped data not only retains key features, but also improves data processing efficiency, laying the foundation for subsequent analysis.

[0018] After obtaining the tensor image after stacking multiple sub-regions, we input it into the features part of the DenseNet121 network, using the pre-trained DenseNet121 as the basic network (the classifier part is replaced by a custom fully connected layer). The DenseNet121 network extracts multi-level feature maps from the input image through multiple convolutional layers, pooling layers, and jump connections, and gradually enhances the extraction from low-dimensional features (such as texture, edges) to high-dimensional features (such as shape and object structure). Based on the multiple Dense Blocks that make up the backbone of the DenseNet121 network, the Dense Blocks are used to densely connect the features of each layer, that is, to connect the features of the current layer with the features of all previous layers, thereby enhancing feature transfer and avoiding information loss.

[0019] Furthermore, if Figure 2As shown in the figure, we also dynamically adjust the importance of each feature channel through the channel attention mechanism, highlight the key features related to coronary heart disease, and suppress the interference of irrelevant channels. Using the spatial attention mechanism, based on the spatial distribution of the feature map, the facial area that may be associated with coronary heart disease is highlighted and the background interference is weakened. Combining the two for feature enhancement enables the model to more accurately capture the global and local features required for coronary heart disease screening, thereby improving the accuracy and reliability of screening.

[0020] Among them, the channel attention mechanism calculates the global features of each channel through global average pooling and global maximum pooling, adds the two pooling results, and sends them to a two-layer fully connected network (implemented by 1x1 convolution) to generate channel weights, and finally maps them to [0, 1] through Sigmoid activation, multiplies the weights with the input features channel by channel, adjusts the weights of each channel, and highlights the spatial attention of important feature channels. Through average pooling and maximum pooling in the channel dimension, two single-channel feature maps are obtained, and the two pooling results are spliced ​​in the channel dimension. The spatial weight map is generated through convolution operation, and then adjusted to [0, 1] through Sigmoid activation. The spatial weight is multiplied with the input features position by position to adjust the importance of different spatial positions in the feature map.

[0021] Then, the ReLU activation function is used to perform a nonlinear transformation on the feature values, returning negative values ​​to zero and retaining positive values, and combined with global average pooling to compress the spatial dimension of the features of each channel to obtain the representative value of each channel. This method not only reduces the computational complexity but also retains the global semantic information.

[0022] The extracted features are integrated into an overall feature matrix. Finally, the feature matrices of multiple cropped regions are averaged along the regional dimension to generate a single feature vector representing the entire sample. This method fuses the multi-region features of each sample into a global feature, enhancing the model's understanding of the overall sample features. Finally, the classifier layer (a layer composed of fully connected layers) in the DenseNet121 model generates the prediction results of the binary classification task. Output the binary classification result (coronary heart disease or negative).

[0023] Among them, the core of feature fusion is to perform some kind of merging between the features of all multiple images so that the model can make the final classification decision based on the features of these images. The fusion method adopted in this embodiment is to average the features of multiple images: average the first dimension of the feature matrix (that is, the feature dimension of multiple images). The feature vectors of multiple images of each sample are fused to obtain an overall feature vector containing all image information. Specifically, the model averages the features of multiple images in each sample, and obtains the global features of each sample through the feature dimensions extracted by DenseNet121. The model can make the final classification decision based on the joint features of multiple parts.

[0024] On the basis of taking the average, the present invention further embeds an attention mechanism (such as SE module or channel weighting mechanism) in the Dense Block to dynamically assign weights according to the importance of each cropped area to the final classification. This method not only retains the feature integrity of multiple regions, but also highlights the model's ability to capture features of sub-regions (such as earlobes or noses), thereby improving the accuracy of feature fusion.

[0025] Finally, the fused features generate a global feature representing each sample.

[0026] Based on the global features extracted above, we first initialize the classifier layer of the DenseNet121 network and select the cross-entropy loss function (CrossEntropyLoss) as the optimization target to measure the error between the model prediction value and the true label. The optimizer selects SGD and combines the weight decay strategy (weight decay) to prevent the model from overfitting. At the same time, the dynamic learning rate scheduler (ReduceLROnPlateau) is used to automatically adjust the learning rate according to the loss change of the validation set to improve the convergence efficiency of the model. During the training process, the model calculates the loss through forward propagation and updates the model parameters through backpropagation. In each training cycle (epoch), the area under the ROC curve (AUC) and classification accuracy of the training set are calculated to evaluate the model performance. In the validation stage, gradient calculation is disabled and the AUC and accuracy of the validation set are evaluated. If the AUC or accuracy of the validation set is better than the historical best value, the parameters of the current model are saved in TorchScript format for easy deployment.

[0027] During the training and validation process, the learning rate is adjusted dynamically, and the loss, AUC, and accuracy of each cycle are recorded. When the AUC or accuracy of the validation set reaches the best, the weight parameters of the model are saved to ensure the best performance when deployed.

[0028] Finally, the performance of the model under different data distributions is comprehensively evaluated through AUC and accuracy, and the best model is saved in TorchScript format to support rapid deployment on resource-constrained devices (such as embedded devices).

[0029] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0030] In addition, in the present invention, descriptions such as "first", "second", "one", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0031] In the present invention, unless otherwise clearly specified and limited, the terms "connection", "fixation", etc. should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0032] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in the field can implement it. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

Claims

1. A method for initial screening of coronary heart disease based on facial feature analysis, characterized in that: Includes steps: S1: Extract sub-regions of various parts of the face by cropping a facial image sample dataset containing healthy individuals and patients with coronary heart disease, and obtain tensor images of each sample data by stacking the sub-regions; S2: Obtain the flattened feature map of each tensor image by layer-by-layer feature extraction, and obtain the feature matrix of the corresponding sample data by integrating the feature map; S3: Calculate the feature mean in the image dimension direction according to the feature matrix, and use the feature mean as the global feature tensor of the corresponding sample; S4: By performing dimensionality reduction and nonlinear transformation on the global feature tensor, a binary classification probability distribution output of the association between the sub-region and coronary heart disease is obtained; S5: Update the sub-region classification weights iteratively through the cross entropy loss function constraint and the stochastic gradient descent optimizer, and return to step S1 before the classification accuracy reaches the standard; S6: A coronary heart disease screening model with a satisfactory accuracy after iteration is used to perform initial screening for coronary heart disease based on the target patient's facial image.

2. A method for initial screening of coronary heart disease based on facial feature analysis as claimed in claim 1, characterized in that: In the step S1, before performing sub-region extraction, the following steps are also included: The face boundary box in the face image is selected using the YOLO target detection algorithm.

3. A method for initial screening of coronary heart disease based on facial feature analysis as claimed in claim 1, characterized in that: In the step S2, the feature maps of all subsequent layers are obtained based on the output of each layer's feature map as input under the dense connection mechanism of the DenseNet121 network.

4. A method for initial screening of coronary heart disease based on facial feature analysis as claimed in claim 3, characterized in that: In the S2 step, the importance of each feature channel is dynamically adjusted through the channel attention mechanism to highlight the key features related to coronary heart disease and suppress the interference of irrelevant channels; Through the spatial attention mechanism, the facial areas related to CHD are highlighted according to the spatial distribution of feature maps, while the background interference is weakened.

5. A method for initial screening of coronary heart disease based on facial feature analysis as claimed in claim 4, characterized in that: By embedding the attention mechanism in each Dense Block of the DenseNet121 network, the feature channel weights are dynamically acquired.

6. A method for initial screening of coronary heart disease based on facial feature analysis as claimed in claim 3, characterized in that: In the DenseNet121 network, a fully connected layer is used to replace the original classifier part for classification judgment.

7. A method for initial screening of coronary heart disease based on facial feature analysis as claimed in claim 1, characterized in that: In the step S5, the loss is calculated by forward propagation in the update iteration process, and the model parameters are updated by back propagation, and the area under the ROC curve and the classification accuracy of the training set are calculated in each update cycle.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the preliminary screening method described in any one of claims 1 to 7 are implemented.

9. A device for processing data, characterized in that: include: a memory having a computer program stored thereon; A processor, configured to execute the computer program in the memory to implement the steps of the preliminary screening method according to any one of claims 1 to 7.

Images