A lung image feature extraction method based on deep learning and particle swarm optimization

Abstract

The application discloses a lung image feature extraction method based on deep learning and particle swarm optimization, and relates to the technical field of medical images, and the method comprises the following steps: constructing a feature extraction network, extracting image features through a deep learning model, optimizing full connection layers of the network by using a multi-objective fitness function, then further optimizing model parameters by using a particle swarm optimization algorithm, and improving classification precision; in order to evaluate model performance, AUC, FPR and other indicators are used, and ideal values of the indicators are combined by linear weighting to optimize the method, and the method has remarkable effects in lung disease prediction, and especially shows wide application prospects in diagnosis and evaluation of new crown infection.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and more specifically, to a method for extracting lung image features based on deep learning and particle swarm optimization. Background Technology

[0002] CT is a commonly used medical imaging technique widely applied in clinical diagnosis and treatment. X-rays are one of the most common methods for lung medical imaging, revealing abnormal shadows in the lungs. These lung medical images have high resolution and high contrast, enabling doctors to more accurately diagnose diseases, assess damage, or plan surgery.

[0003] Improving the accuracy of lung medical image classification undoubtedly enhances the ability to detect and warn of related diseases, thereby effectively suppressing their spread. Traditional medical image classification methods typically rely on manually designed feature extraction and machine learning algorithms. These methods require manually selecting and extracting features relevant to specific diseases, and then using a classifier for classification. However, due to the complexity and high dimensionality of lung medical images, traditional methods are often limited by their expressive and feature generalization capabilities, making it difficult to capture the rich information within the images.

[0004] In the clinical management and diagnosis of COVID-19, CT and X-ray technologies have been widely used for lung imaging assessment in suspected or confirmed cases. These images provide more intuitive and detailed information about lung structure, which can help doctors determine the characteristics and extent of lung lesions and assist in clinical diagnosis and treatment decisions. Due to the infectiousness and clinical characteristics of COVID-19, hospitals and medical institutions face the challenge of rapidly and accurately classifying a large amount of medical images of the lungs. Accurate classification of medical images of COVID-19 infection is crucial for early diagnosis, tracking disease progression, and guiding treatment. Therefore, deep learning-based lung image classification methods have become a research hotspot. Deep learning methods automatically learn feature representations from large amounts of image data by constructing deep neural networks with strong nonlinear modeling and generalization capabilities. Applying deep learning to COVID-19 lung image classification has become a promising solution. By classifying COVID-19 lung images, doctors can more accurately describe the pathological features of the infection, assess the severity of the disease, and predict patient prognosis. Furthermore, medical image classification can help monitor the spread of the epidemic, assess epidemiological characteristics, and provide necessary decision support for public health departments.

[0005] In recent years, there has been much research on deep learning algorithms for CT images or chest X-ray images. Current methods mainly involve building models for image classification or segmentation using CNN networks, such as ResNet, Inception, and U-Net. Existing methods have limited research on reducing the false positive rate (FPR). They primarily focus on accuracy, AUC (area under the curve), or F1 score, while neglecting FPR. Experiments have shown that excessively pursuing AUC during training leads to a decrease in FPR. Data-driven experimental methods treat each metric as equally important without weighting them, and in actual disease surveillance, it is more acceptable for negative patients to be misclassified as positive, while misclassifying positive patients as negative results in more losses. High FPR increases unnecessary detection costs, thereby reducing the efficiency of disease detection. Summary of the Invention

[0006] The purpose of this invention is to improve the accuracy and efficiency of medical image classification.

[0007] The technical solution of this invention is: to provide a lung image feature extraction method based on deep learning and particle swarm optimization, the method comprising:

[0008] S1. Collect CXR images of several COVID-19 positive and negative patients to construct a COVID-19 radiology dataset as dataset 1; collect lung CT scan images of several COVID-19 positive and negative patients to construct a COVID-CT dataset as dataset 2;

[0009] S2. Construct a feature extraction network model. The feature extraction network module includes convolutional layers, residual network layers, pooling layers, and fully connected layers.

[0010] S3. Use the dataset 1 and dataset 2 constructed in step S1 as training sets to train the feature extraction network model constructed in step S2.

[0011] S4. Use the particle swarm optimization algorithm to optimize the fully connected layers of the feature extraction network model, and use the optimized feature extraction network model to extract lung image features.

[0012] In any of the above technical solutions, further, the method of optimizing the fully connected layer using the particle swarm optimization algorithm in step S4 includes:

[0013] S41. Initialize the particle swarm: Straighten and merge the weight parameters of the fully connected layer into an array, with each parameter serving as the initial position and velocity of the particle;

[0014] S42. Calculate fitness: For each particle, calculate the fitness value based on its position, i.e., the objective function. ;

[0015] S43. Update the individual optimal position: For each particle, update... That is, the best solution that the particle has found so far based on the current fitness value;

[0016] S44. Update the optimal position of the population: Find the particle with the best fitness value in the population and set its position as the optimal position of the population. The optimal position;

[0017] S45, Update velocity and position: For each particle, according to and and current speed Update the particle's velocity and position;

[0018] S46. Determine the termination condition: Determine whether to terminate the algorithm based on the predefined termination condition;

[0019] S47. Return Result: Use As the output of the algorithm, the optimal solution is found, and then the optimal solution is loaded into the original model.

[0020] In any of the above technical solutions, further, the objective function of the feature extraction network model constructed in step S2 is... The expression is as follows:

[0021] ;

[0022] In the formula, , The weights of each indicator, For the ideal point of FPR, For the ideal point of AUC, It should be close to 0. It should be close to 1.

[0023] In step S4, a linear weighting method is used to adaptively change the weights as the iterations proceed. The formula for the linear weighting method is as follows:

[0024] ;

[0025] Indicates the initial inertia weight. The inertia weight represents the number of iterations up to the maximum. M represents the maximum number of iterations, and M represents the current number of iterations.

[0026] In any of the above technical solutions, the method further includes:

[0027] S5. The optimized feature extraction network model is evaluated using PREC, SPEC, and F1 scores as evaluation metrics:

[0028] PREC accuracy measures the reliability of model predictions. The expression for PREC accuracy is:

[0029] ;

[0030] SPEC specificity measures a model’s ability to distinguish true negatives. The expression for SPEC specificity is:

[0031] ;

[0032] The F1 score measures the overall performance of a model in identifying positive and negative examples. The F1 score is expressed as follows:

[0033] ;

[0034] In the formula, FP represents false positives, TN represents true negatives, TP represents true positives, and FN represents the number of false negatives.

[0035] The beneficial effects of this invention are:

[0036] The optimized neural network model proposed in this invention integrates deep learning and particle swarm optimization algorithms. This invention can significantly reduce the false positive rate (FPR) while ensuring a high detection rate, thereby improving the accuracy and reliability of diagnosis.

[0037] This model can be applied to actual medical equipment to assist doctors in the detection and diagnosis of lung diseases, and it has significant practical value, especially in the diagnosis and assessment of COVID-19 infection.

[0038] This model outperforms other traditional methods on multiple evaluation metrics, especially in reducing the false positive rate and improving AUC and F1 scores. Attached Figure Description

[0039] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0040] Figure 1 This is a schematic diagram of the framework of a lung image feature extraction method based on deep learning and particle swarm optimization according to an embodiment of the present invention;

[0041] Figure 2 This is a dataset of CT scan images of COVID-19 based on a lung image feature extraction method using deep learning and particle swarm optimization according to an embodiment of the present invention.

[0042] Figure 3 This is a structural diagram of a lung image feature extraction method based on deep learning and particle swarm optimization according to an embodiment of the present invention;

[0043] Figure 4 It is a confusion matrix on dataset 1 of a lung image feature extraction method based on deep learning and particle swarm optimization according to an embodiment of the present invention;

[0044] Figure 5 It is a confusion matrix on dataset 2 of a lung image feature extraction method based on deep learning and particle swarm optimization according to an embodiment of the present invention;

[0045] Figure 6 The ROC curves of different methods of lung image feature extraction based on deep learning and particle swarm optimization according to an embodiment of the present invention are plotted in dataset 1.

[0046] Figure 7 The graph shows the ROC curves of different methods for lung image feature extraction based on deep learning and particle swarm optimization according to an embodiment of the present invention in dataset 2. Detailed Implementation

[0047] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0048] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0049] like Figure 1 As shown, this embodiment provides a lung image feature extraction method based on deep learning and particle swarm optimization. The extracted lung image features can be used as an aid to physicians in disease diagnosis. The method includes:

[0050] S1. Collect CXR images of several COVID-19 positive and negative patients to construct a COVID-19 radiology dataset as dataset 1; collect lung CT scan images of several COVID-19 positive and negative patients to construct a COVID-CT dataset as dataset 2.

[0051] In this embodiment, 18,479 CXR images from 15,000 patients were collected. From these, 427 positive and 498 negative images were selected, totaling 925 CXR images, to construct dataset 1; (as shown in the image) Figure 2 As shown, dataset 2 contains CT scan images of the lungs of 349 positive and 463 negative patients.

[0052] It is hereby noted that the two datasets used in this study and its embodiments—the open-source COVID-19 infection radiographic imaging dataset and the open-source COVID-19 infection lung CT scan dataset—are anonymized to ensure data security and personal privacy, including sensitive information handling, encrypted storage and transmission, and compliance with privacy laws. Furthermore, their long-standing public availability and use in numerous publications underscore their transparency, validity, and representativeness, supporting the reproducibility and verifiability of the research results.

[0053] Because the image quality of the collected data varied, all images were converted into RGB images of size (299, 299) to facilitate subsequent loading into the model for processing.

[0054] S2. Construct a feature extraction network model, such as Figure 3 As shown, the feature extraction network module includes convolutional layers (CNN) and residual network layers (ResNet).

[0055] The convolutional layer consists of multiple 3×3 convolutional kernels. As the number of layers increases, the number of kernels gradually increases, and the convolutional stride is 2 to reduce the size of the output feature map. Dropout and Pooling layers have been removed to ensure that the data features passed to the residual network layer are extracted to the maximum extent.

[0056] The residual network layer consists of multiple convolutional kernels of different dimensions, such as 1×1 and 3×3, to increase the network's adaptability to different scales and extract more global features. 1×1 convolutional kernels can reduce the size of the input features, thus reducing the computational workload of larger convolutional kernels. Pooling layers are then connected to the residual network layer to extract image features.

[0057] The feature extraction network model uses four convolutional layers: one with a kernel size of 3, a stride of 1, and zero padding; three convolutional layers with a stride of 2 and a kernel size of 1; and three convolutional layers with a stride of 1. The residual layers use convolutional layers with a kernel size of 3, a stride of 1, and zero padding, as well as convolutional layers with a kernel size of 1 and a stride of 1. The data normalization function BatchNorm2d uses default parameters, the batch size is 16, the initial learning rate is 0.001, and the number of training epochs is 100. The MultiStepLR function is used to dynamically adjust the learning rate. The number of adjustment epochs is 30. The optimizer used is AdamW with a weight decay coefficient of 1e-6.

[0058] Detailed information about hyperparameter values ​​is shown in Table 1 below:

[0059] Table 1 provides detailed information on hyperparameter values;

[0060] The input of each convolutional kernel passes through a linear layer, and the output of each convolutional kernel is stitched together into a one-dimensional feature vector. The residual is calculated using the input feature data, which reduces network overfitting and solves the problems of gradient vanishing and gradient exploding.

[0061] Finally, after the pooling layer, a fully connected layer (Linear) is used to predict the category of each feature.

[0062] The Area of ​​Receiver Operating Curve (AUC), False Positive Rate (FPR), and True Positive Rate (TPR) are three interrelated metrics. Generally, overemphasizing one metric will lead to changes in the other two. In this paper, after extracting features using a deep learning network, we reduce the FPR by optimizing the weight parameters of the fully connected layers. However, overemphasizing the FPR may lead to a decrease in either AUC or TPR; therefore, these two interrelated metrics should be optimized simultaneously. The goal is to maximize AUC and minimize FPR while maintaining high accuracy.

[0063] Therefore, a multi-objective fitness function with linear weighting as the objective was designed, and then another multi-objective fitness function was designed as the objective. In linear weighting, each evaluation index is assigned a weight, and the sum of these weights is usually 1. The weights can be determined based on the importance of the task, the optimization objective, and relevant requirements. A higher weight indicates a greater contribution of that index to the overall performance. The basic idea of ​​linear weighting is to determine the indices AUC and Fpr to be evaluated, with each index having a weight of... , , , ∈(0,1), the ideal value of Fpr is 0%, and the ideal value of AUC is 100%. Since these two indicators cannot perfectly reach their ideal values, a reasonable trade-off between the indicators is evaluated, and a trade-off point is introduced to ensure that all indicators have effective changes. In addition, to reduce overfitting, L2 regularization is introduced as a penalty term. The objective function of the feature extraction network model is... The expression is as follows:

[0064] ;

[0065] In the formula, , The weights of each indicator, For the ideal point of FPR, For the ideal point of AUC, It should be close to 0. It should be close to 1.

[0066] S3. Use the dataset 1 and dataset 2 constructed in step S1 as training sets to train the feature extraction network model constructed in step S2.

[0067] S4. Use the particle swarm optimization algorithm to optimize the fully connected layers of the feature extraction network model.

[0068] Particle Swarm Optimization (PSO) is a heuristic optimization algorithm inspired by the behavior of biological groups (such as flocks of birds or schools of fish). It finds the optimal solution to a problem by simulating the cooperation and information exchange among individuals within a group. In PSO, the solution to a problem is represented by the position of a particle. Each particle has its own position and velocity and searches based on its current position and velocity. Particles continuously update their positions and velocities to search for the optimal solution.

[0069] Particle swarm optimization is defined as:

[0070] ;

[0071] in, This represents the velocity vector of particle i in the d-th dimension at generation k+1. This represents the position vector of particle i in the d-th dimension at generation k. This represents the best historical position of particle i in the d-th dimension at generation k; industry professionals commonly refer to it as the individual optimal solution. This represents the population-optimal solution. This represents the inertia weight, used to adjust the influence of the current speed. Represents individual learning factors. Represents the population learning factor. and The speed at which particles move towards their individual optimal positions and the global optimal positions are controlled separately. and This represents a random number within the range [0,1], used to introduce randomness. Indicates the number of iterations.

[0072] In this invention, the method for optimizing fully connected layers using the particle swarm optimization algorithm includes:

[0073] S41. Initialize the particle swarm: Straighten and merge the weight parameters of the fully connected layer into an array, with each parameter serving as the initial position and velocity of the particle.

[0074] S42. Calculate fitness: For each particle, calculate the fitness value based on its position, i.e., the objective function. .

[0075] S43. Update the individual optimal position: For each particle, update... That is, the best solution that the particle has found so far based on its current fitness value.

[0076] S44. Update the optimal position of the population: Find the particle with the best fitness value in the population and set its position as the optimal position of the population. The optimal position.

[0077] S45, Update velocity and position: For each particle, according to and and current speed Update the particle's velocity and position.

[0078] S46. Determine the termination condition: Determine whether to terminate the algorithm based on the predefined termination condition (such as reaching the maximum number of iterations or meeting a specific fitness requirement).

[0079] S47. Return Result: Use As the output of the algorithm, the optimal solution is found, and then the optimal solution is loaded into the original model to verify the effectiveness of the algorithm.

[0080] When solving practical problems, a global search followed by a local search is usually required. Therefore, the weights need to be adaptively adjusted as the iterations progress. In this paper, we use a linear weighted method, as shown in the following formula:

[0081] ;

[0082] Indicates the initial inertia weight. The inertia weight represents the number of iterations up to the maximum. M represents the maximum number of iterations, and M represents the current number of iterations. In this paper, the maximum number of iterations is set to 100. Set it to 0.9. Set to 0.4, the initial learning factor and Set it to 1.5.

[0083] S5. Evaluate the optimized feature extraction network model.

[0084] Optimal evaluation metrics are crucial for assessing model performance. Currently, most experimental methods use traditional confusion matrix calculation metrics, such as accuracy, precision, AUC, and F1, to evaluate the overall model performance. However, in medical image processing tasks, practical applications tend to focus more on evaluating the AUC metric, often neglecting the impact of FPR.

[0085] AUC is a commonly used metric for evaluating the performance of binary classification models. It represents the area under the ROC curve (Receiver Operating Characteristic curve) and ranges from 0 to 1. The closer the AUC is to 1, the better the model performs at different thresholds. FPR, on the other hand, represents the proportion of negative samples that the model misclassifies as positive. Especially in high-sensitivity applications such as medical imaging, reducing FPR is crucial.

[0086] AUC measures a model's classification ability at different thresholds; a higher value indicates better classification performance. However, optimizing AUC can sometimes lead the model to tend to reduce false positives, thus compressing the prediction range for positive samples and increasing the risk of false negatives. In this case, although the model's prediction accuracy on negative samples improves, true positive samples may be misclassified as negative, increasing the number of false negatives.

[0087] In disease surveillance and medical diagnosis, the risk of misclassifying a sick patient as a healthy patient (false negative) is far greater than the risk of misclassifying a healthy patient as a sick patient (false positive). False negatives can lead to delayed treatment, increased health risks for patients, and unnecessary testing and treatment costs. Therefore, in model training, excessively pursuing an increase in AUC may reduce FPR, but it may also lead to an increase in false negatives.

[0088] Balancing the risks of false positives and false negatives is crucial during model optimization. Choosing an appropriate classification threshold, combined with domain expert knowledge and the specific application requirements, allows for better management of the model's predictive accuracy and risk. In disease surveillance, simply pursuing higher AUC is insufficient; it's necessary to comprehensively consider different evaluation metrics, such as FPR, and adjust the optimization objective according to actual needs.

[0089] In this study, since the dataset labels were set to 0 for positive patients and 1 for negative patients, the FPR (Free Precipitation Rate) metric was of particular interest. To reduce the FPR while maintaining a certain AUC (Average Underlying Value), the network model was optimized.

[0090] In this experiment, FPR was defined as the proportion of COVID-19 patients who were misdiagnosed as not having the disease, and was calculated using the following formula:

[0091] ;

[0092] PREC accuracy is the proportion of samples predicted to be positive that are actually positive. It measures the reliability of the model's predictions and is calculated using the following formula:

[0093] ;

[0094] SPEC specificity measures the model's ability to distinguish true negatives, and is calculated using the following formula:

[0095] ;

[0096] The F1 score is a key indicator for evaluating the effectiveness of binary classification models, measuring the model's overall performance in identifying positive and negative examples. The calculation formula is:

[0097] ;

[0098] In the formula, FP represents false positives, TN represents true negatives, TP represents true positives, and FN represents the number of false negatives.

[0099] In another embodiment of the invention, the dataset is manually divided into a training set and a test set in a 7:3 ratio, and 100 rounds of training are performed. In each round, the model results are validated using the test set data, and finally, an accuracy analysis graph of the validation set and test set for each model is generated. The deep learning models proposed in this paper and the deep learning models involved in the experimental comparison are all implemented using the open-source deep learning framework PyTorch, in a Python 3.7 environment, with 100 iterations.

[0100] The confusion matrix of the experimental results is as follows Figure 4 and Figure 5 As shown in the diagram, the confusion matrix is ​​an essential tool in classification models, providing a visual representation of the algorithm's performance. It comprises four components: true positives (TP), indicating correct disease identification; false positives (FP), indicating incorrect detection of a non-existent disease; true negatives (TN), confirming the absence of a certain disease; and false negatives (FN), indicating that the model failed to detect an existing disease. These components allow us to calculate key metrics such as accuracy, precision, recall, and F1 score. As the matrix shows, the results demonstrate excellent identification capabilities, with accuracy exceeding 85%.

[0101] In another embodiment of the invention, the proposed network model was validated on the COVID-19 radiology dataset and the COVID-19 dataset, and then comparative experiments were conducted using RESNET-50, INCEPTION, a five-layer deep convolutional neural network (5LCNN), and EfficientNetV2. RESNET-50, with 50 layers, achieves high performance in tasks such as image classification by utilizing skip connections. INCEPTION uniquely combines convolutions of different sizes for cost-effective parameter use, drawing inspiration from the retina. 5LCNN introduces a random pooling method, providing robustness against overfitting and computational efficiency. Finally, EfficientNetV2, part of the Google model series, introduces scalable improvements to meet speed and accuracy requirements, optimizing performance across various devices. The ROC curves between different methods on datasets 1 and 2 are shown below. Figure 6and Figure 7 As shown in Tables 3 and 4 below:

[0102] Table 3. Analysis of the model proposed in Dataset 1;

[0103] Table 4. Analysis of the model proposed in dataset 2;

[0104] In performance evaluations on the COVID-19 infection radiographic imaging dataset (Dataset 1) and the COVID-CT dataset (Dataset 2), the proposed method outperforms established methods such as RENet-50, Initial, 5LCNN, and EfficientNetV2 on most evaluation metrics, including accuracy, area under the ROC curve, false positive rate, precision, and specificity. The only deviation is a slightly lower F1 score than 5LCNN on Dataset 1. However, on Dataset 2, the proposed method outperforms all six evaluation metrics, surpassing four of the compared methods. These results clearly validate the effectiveness of the proposed method in the COVID-19 infection detection imaging task and its superior and consistent performance across different datasets.

[0105] In another embodiment of the present invention, to verify the effectiveness of different steps of the proposed method, we constructed ablation experiments, a method without residual elements (RES), a method without PSO, and the proposed method, respectively. Tables 5 and 6 show the analysis of the models proposed in datasets 1 and 2.

[0106] Tables 5 and 6 show ablation studies using the COVID-19 radiology and COVID-CT datasets, demonstrating that our method outperforms other methods and improves key performance metrics. Specifically, omissions in the residual module and particle swarm optimization led to a nearly 23% decrease in model accuracy (ACC), area under the ROC curve (AUC), and false positive rate (FPR) on the radiology dataset, as well as a decrease in F1 score, precision (PREC), and specificity (SPEC). Implementing our complete method not only recovers these metrics but also improves them. The performance on the COVID-CT dataset also exhibits this decline pattern, and the subsequent improvements using the complete method highlight the crucial role of the included components in improving model effectiveness.

[0107] Table 5. Analysis of the model proposed in dataset 2;

[0108] Table 6. Ablation experiments in dataset 2;

[0109] Initially, multi-layer convolutions were used to efficiently extract image features. This technique captures a range of features at different depths, giving the model enhanced representational capabilities. Subsequently, the architecture integrated residual modules to alleviate the vanishing gradient problem prevalent in traditional deep networks, thereby achieving more efficient training and improved model performance. Finally, Particle Swarm Optimization (PSO) was used to fine-tune the false positive rate (FPR). PSO is a swarm intelligence-based algorithm that can quickly and accurately find the global optimum of the current optimization problem. Through this optimization process, the model cleverly minimizes the false positive rate, significantly reducing false positives while maintaining a high detection rate, thus improving the model's practicality and value.

[0110] The significant results obtained in Tables 3 and 4 demonstrate that, on the COVID-19 radiographic and COVID-CT datasets, removing residual modules (without RES) and particle swarm optimization (without PSO) reduces performance metrics such as ACC, AUC, FPR, F1 score, PREC, and SPEC; while fully implementing the proposed method significantly improves the results, particularly reducing FPR by nearly 23%, highlighting its effectiveness in reducing false positives.

[0111] The evaluation results of the model on two datasets (Dataset 1 and Dataset 2) show the values ​​of various performance metrics and confidence intervals at a 95% confidence level. In Dataset 1, the model's precision is 0.8957 with a confidence interval of [0.8598, 0.9316]. Values ​​and confidence intervals are also provided for AUC, FPR, F1 score, precision, and specificity. For Dataset 2, the confidence interval for precision is [0.7813, 0.8795], the AUC is 0.9137 with a confidence interval of [0.8769, 0.9505], and values ​​and confidence intervals are provided for FPR, F1 score, precision, and specificity.

[0112] While the recommended approach is highly effective, it cannot replace clinical trial results. The model provides physicians with insights based on clinical trials and patient symptoms, identifying high-risk individuals for early intervention to reduce morbidity and customizing personalized treatment predictions; it also simplifies patient care by rapidly filtering low-risk cases and reducing wait times. However, the model's results should only supplement, not replace, professional medical judgment.

[0113] In summary, this invention designs a lung image feature extraction method based on deep learning and particle swarm optimization. This method uses an optimized feature extraction network model to extract features from CT scans and X-ray images of the lungs, thereby assisting physicians in diagnosing lung diseases. The ACC and AUC of the COVID-19 chest CT scan dataset validation set are 83.04% and 91.37%, respectively, and the accuracy and AUC of the COVID-19 chest X-ray dataset validation set are 89.57% and 97.01%, respectively, both higher than other comparative experiments. The low FPR deep neural network model designed in this paper achieves FPR indices of 12.75% and 7.52% on the CT dataset and chest X-ray dataset, respectively. This indicates that the proposed method outperforms other comparative models in relevant metrics, showing a more significant improvement compared to the metrics before optimization. Experiments demonstrate that the proposed model can simplify the COVID-19 disease detection process and, when applied to practical medical devices, can accelerate disease detection, providing doctors and patients with faster and more accurate diagnoses. The next step will be to validate the model on multiple datasets across different pathologies, and to continue optimizing and improving the existing model to enhance its generalization ability to a wider range of diseases. Researchers should collect and organize medical imaging datasets for various diseases and use this data to further validate the model. The model's performance in different disease scenarios should be thoroughly evaluated, and its limitations and areas for improvement should be identified through extensive testing under various pathological conditions. In this process, the existing model can be continuously optimized and enhanced by adjusting the network structure, optimizing training strategies, and introducing more advanced data augmentation techniques. This will further promote the development of the field of medical imaging and provide better diagnostic solutions for a wider range of diseases.

[0114] This invention is not limited to the specific embodiments described above. Those skilled in the art can implement this invention using various other specific embodiments based on the disclosed content of the embodiments and accompanying drawings. Therefore, any design that adopts the design structure and concept of this invention and makes some simple changes or modifications falls within the protection scope of this invention.

Claims

1. A lung image feature extraction method based on deep learning and particle swarm optimization, characterized in that, The method includes: S1. Collect CXR images of several COVID-19 positive and negative patients to construct a COVID-19 radiology dataset as dataset 1; collect lung CT scan images of several COVID-19 positive and negative patients to construct a COVID-CT dataset as dataset 2; S2. Construct a feature extraction network model. The feature extraction network module includes convolutional layers, residual network layers, pooling layers, and fully connected layers. S3. Use the dataset 1 and dataset 2 constructed in step S1 as training sets to train the feature extraction network model constructed in step S2. S4. Use the particle swarm optimization algorithm to optimize the fully connected layers of the feature extraction network model, and use the optimized feature extraction network model to extract lung image features. The objective function of the feature extraction network model constructed in step S2 is... The expression is as follows: ； In the formula, , The weights of each indicator, For the ideal point of FPR, For the ideal point of AUC, It should be close to 0. It should be close to 1.

2. The lung image feature extraction method based on deep learning and particle swarm optimization as described in claim 1, characterized in that, The method for optimizing the fully connected layer using the particle swarm optimization algorithm in step S4 includes: S41. Initialize the particle swarm: Straighten and merge the weight parameters of the fully connected layer into an array, with each parameter serving as the initial position and velocity of the particle; S42. Calculate fitness: For each particle, calculate the fitness value based on its position, i.e., the objective function. ; S43. Update the single optimal position: For each particle, update... That is, the best solution that the particle has found so far based on the current fitness value; S44. Update the optimal position of the population: Find the particle with the best fitness value in the population and set its position as the optimal position of the population. The optimal position; S45, Update velocity and position: For each particle, according to and and current speed Update the particle's velocity and position; S46. Determine the termination condition: Determine whether to terminate the algorithm based on the predefined termination condition; S47. Return Result: Use As the output of the algorithm, the optimal solution is found, and then the optimal solution is loaded into the original model.

3. The lung image feature extraction method based on deep learning and particle swarm optimization as described in claim 1, characterized in that, In step S4, a linear weighting method is used to adaptively change the weights as the iteration progresses. The formula for the linear weighting method is as follows: ； Indicates the initial inertia weight. The inertia weight represents the number of iterations up to the maximum. M represents the maximum number of iterations, and M represents the current number of iterations.

4. The lung image feature extraction method based on deep learning and particle swarm optimization as described in claim 1, characterized in that, The method further includes: S5. The optimized feature extraction network model is evaluated using PREC, SPEC, and F1 scores as evaluation metrics: PREC accuracy measures the reliability of model predictions. The expression for PREC accuracy is: ； SPEC specificity measures a model’s ability to distinguish true negatives. The expression for SPEC specificity is: ； The F1 score measures the overall performance of a model in identifying positive and negative examples. The F1 score is expressed as follows: ； In the formula, FP represents false positives, TN represents true negatives, TP represents true positives, and FN represents the number of false negatives.

Images