A model for rapid screening of cervical tumors and a method for establishing the same

Abstract

The present invention provides a model for rapid screening of cervical tumors and a method for establishing the same. The method comprises the following steps: (1) collecting serum samples from patients with cervical tumors of varying degrees of lesion, measuring and obtaining Fourier transform infrared (FTIR) spectroscopy data; and (2) using the PSO-CNN algorithm to establish a classification model for the FTIR spectroscopy data, thereby obtaining the model for rapid screening of cervical tumors. The model for rapid screening of cervical tumors and the method for establishing the same, described in the present invention, can improve recognition performance by optimizing the PSO-CNN algorithm, thereby enabling better application in rapid screening of cervical tumors.

Description

Technical Field

[0001] The present invention specifically relates to a model for rapid screening of cervical tumors and a method for establishing the model. Background Art

[0002] Cervical cancer is considered the second most common malignant tumor in women and the leading cause of cancer death worldwide, second only to breast cancer. New cases of cervical cancer in my country have reached 110,000 annually, accounting for one-third of all new cases worldwide. Cervical intraepithelial neoplasia (CIN) is closely associated with the development of cervical cancer. Persistent high-risk human papillomavirus infection can trigger low-grade cervical epithelial lesions (CIN I), which then progress to high-grade cervical precancerous lesions (CIN II and CIN III), and ultimately, cervical cancer. This process generally takes 8-15 years. However, some cases develop more rapidly. If cervical precancerous lesions can be diagnosed promptly and appropriate measures implemented during this period, timely intervention can be implemented to prevent adverse progression. Current methods for helping clinicians detect cervical precancerous lesions early include cytopathology, HPV testing, and histopathology, but these methods are limited by subjectivity, cost, and time. Therefore, an objective, rapid detection method that requires less sample preparation is needed clinically.

[0003] FT-IR is an optical spectroscopic technique that can effectively provide information on the structure and chemical composition of biomaterials at the molecular level. It is sensitive to subtle biochemical changes occurring at the molecular level and can detect spectral changes corresponding to the onset of disease. Over the past 15 years, numerous studies have demonstrated the potential of FT-IR spectroscopy and machine learning algorithms for detecting various cancers, including some in cervical cancer screening. Mo et al. used principal component analysis (PCA) and linear discriminant analysis (LDA) algorithms, combined with serum FT-IR spectra, to distinguish cervical cancer patients from healthy controls, achieving diagnostic accuracy, specificity, and sensitivity of 98%. While this experiment achieved promising results, it neglected the diagnosis of CIN. Nordstrom et al. used a multivariate algorithm based on the UV fluorescence spectral characteristics of biopsy tissue to distinguish normal tissue, CIN I, and CIN II / III. The sensitivity and specificity for CIN II / III versus normal tissue were 91% and 93%, respectively, and 86% and 87%, respectively, for CIN I versus normal tissue. While the classification results are relatively accurate, this experiment could further enhance the recognition of different stages of high-grade cervical precancerous lesions. Furthermore, Yang et al. proposed a feature fusion-based method to classify cervicitis, low-grade squamous intraepithelial lesions, high-grade squamous intraepithelial lesions, cervical squamous cell carcinoma, and cervical adenocarcinoma. The classification accuracies of the KNN, ELM, ABC-SVM, CS-SVM, PSO-SVM, and CNN-LTSM methods were 60.91%, 67.84%, 77.64%, 78.49%, 75.54%, and 70.72%, respectively, representing improvements over the original spectral classification accuracy without feature fusion. However, like most studies using CNN as a feature extractor or classifier, they did not optimize the CNN's network structure, which to some extent limited its recognition effectiveness.

[0004] In view of this, the present invention proposes a new model for rapid screening of cervical tumors to further optimize the existing early screening model. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for establishing a model for rapid screening of cervical tumors, which can improve the recognition effect by optimizing the PSO-CNN algorithm.

[0006] In order to achieve the above objectives, the technical solutions adopted are:

[0007] A method for establishing a model for rapid screening of cervical tumors comprises the following steps:

[0008] (1) After collecting serum samples from patients with cervical cancer of different lesion severity, the Fourier transform infrared spectroscopy data were measured and obtained;

[0009] (2) The Fourier transform infrared spectroscopy data is used to establish a classification model through the PSO-CNN algorithm to obtain the model for rapid screening of cervical tumors.

[0010] Furthermore, in step (1), different degrees of lesions include cervical cancer, precancerous lesions grade I, precancerous lesions grade II, precancerous lesions grade III, and uterine fibroids.

[0011] Furthermore, in step (2), the PSO-CNN algorithm for searching for the optimal CNN structure includes the following steps in sequence: initialization of the particle swarm, fitness evaluation of a single particle, measurement of the difference between two particles, velocity calculation, and particle update.

[0012] Furthermore, in the step (2), the process of initializing the particle swarm is as follows: first, the range of the number of network layers is set, then the convolution layer, the maximum pooling layer, the average pooling layer and the fully connected layer are randomly selected as the current layer, and their parameters are configured before performing the particle swarm initialization calculation;

[0013] Among them, the first layer of each particle is a convolutional layer, the last layer is a fully connected layer, and at least one of the convolutional layers, maximum pooling layers, average pooling layers and fully connected layers is randomly interspersed in the middle.

[0014] Furthermore, in the step (2), the architecture obtained before the particle swarm initialization calculation during the particle swarm initialization process is:

[0015] The first layer is a convolutional layer with a kernel size of 6×1 and 136 output channels;

[0016] The second layer is the convolution layer with a kernel size of 6×1 and 90 output channels;

[0017] The third layer is the convolutional layer with a kernel size of 5×1 and 217 output channels;

[0018] The fourth layer is a convolutional layer with a kernel size of 6×1 and 141 output channels;

[0019] The fifth layer is a convolutional layer with a kernel size of 3×1 and 197 output channels;

[0020] The sixth layer is a fully connected layer with 82 core neurons;

[0021] The seventh layer is a fully connected layer with 5 core neurons.

[0022] Furthermore, in the step (2), the architecture obtained before the particle swarm initialization calculation during the particle swarm initialization process is:

[0023] The first layer is a convolutional layer with a kernel size of 6×1 and 51 output channels;

[0024] The second layer is average pooling with a pooling size of 3×1 and a stride of 2;

[0025] The third layer is the convolution layer with a kernel size of 4×1 and 97 output channels;

[0026] The fourth layer is the convolutional layer with a kernel size of 5×1 and an output channel number of 228;

[0027] The fifth layer is a convolutional layer with a kernel size of 5×1 and an output channel number of 228;

[0028] The sixth layer is a fully connected layer with 279 core neurons;

[0029] The seventh layer is a fully connected layer with 5 core neurons.

[0030] Furthermore, the parameters used in the PSO-CNN algorithm include: particle swarm optimization parameters, CNN architecture initialization parameters, and CNN training parameters.

[0031] Furthermore, the particle swarm algorithm parameters: number of iterations, population size and Cg are set to 30, 10 and 0.5 respectively;

[0032] The CNN architecture is initialized as follows: the maximum number of network layers is set to 15, the number of convolution kernels is selected in the range of [3, 256], the convolution kernel size is selected in the range of [3×1, 7×1], the number of neurons in the fully connected layer is selected in the range of [1, 300], the number of neurons in the output layer is 5, the convolution layer generation probability is 0.6, the pooling layer generation probability is 0.3, and the fully connected layer generation probability is 0.1;

[0033] The CNN training parameters are as follows: the number of training epochs for particle evaluation is 1, the number of training epochs for the global best particle is 200, and the Dropout is 0.3.

[0034] Furthermore, in step (2), the particles are updated using the following formula:

[0035]

[0036]

[0037] Where k is the current iteration number; and They represent the velocity and position of particle i in the dth dimension in the next iteration respectively; Cg is the threshold; r is a number between 0 and 1 randomly generated at each position of the particle.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] This invention uses an optimization algorithm to adaptively select the optimal CNN framework, enabling rapid and efficient early cervical cancer screening based on FT-IR technology. Genetic algorithms and particle swarm optimization (PSO) are commonly used in the field of CNN parameter optimization. Generally speaking, PSO is simpler and converges faster than GA because it does not involve crossover and mutation operations. Furthermore, due to the communication between particles and the random nature of the PSO algorithm, it also achieves excellent results.

[0040] Based on the particle swarm optimization algorithm's low parameter count and fast convergence, this paper uses the algorithm to automatically construct a CNN structure with a variable number of layers and layer type parameters. This paper designed a CNN model for early cervical cancer screening based on FT-IR technology. Using validation set accuracy as the evaluation basis for network training, experimental results show that the CNN structure obtained by the PSO-CNN algorithm achieves better detection results for early cervical cancer compared to several other classic CNN structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is the average FT-IR spectrum of five types of serum samples; the shaded area represents the standard deviation;

[0042] Figure 2 Updated schematics for particle architecture;

[0043] Figure 3 is the optimal architecture obtained by the PSO-CNN algorithm; where a is the CNN framework obtained for the second time, and b is the CNN framework obtained for the fifth time. DETAILED DESCRIPTION

[0044] To further illustrate the present invention's model for rapid screening of cervical tumors and its establishment method, and to achieve its intended purpose, the following describes in detail, in conjunction with preferred embodiments, a model for rapid screening of cervical tumors and its establishment method, including its specific implementation, structure, features, and efficacy. In the following description, different references to "one embodiment" or "embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0045] Before elaborating in detail the model for rapid screening of cervical tumors and the method for establishing the same, it is necessary to further explain the relevant materials mentioned in the present invention in order to achieve better results.

[0046] The following is a detailed description of a model for rapid screening of cervical cancer and its establishment method according to the present invention, with reference to specific embodiments:

[0047] Based on the automatic selection of CNN network layer architecture and hyperparameters using particle swarm optimization (PSO), this paper proposes an early tumor screening model for cervical cancer based on one-dimensional FT-IR spectral data and PSO-CNN, providing a reference for the selection of CNN frameworks in this field in the future.

[0048] The technical solution of the present invention is:

[0049] A method for establishing a model for rapid screening of cervical tumors comprises the following steps:

[0050] (1) After collecting serum samples from patients with cervical cancer of different lesion severity, the Fourier transform infrared spectroscopy data were measured and obtained;

[0051] (2) The Fourier transform infrared spectroscopy data is used to establish a classification model through the PSO-CNN algorithm to obtain the model for rapid screening of cervical tumors.

[0052] Furthermore, in step (1), different degrees of lesions include cervical cancer, precancerous lesions grade I, precancerous lesions grade II, precancerous lesions grade III, and uterine fibroids.

[0053] Preferably, in step (2), the PSO-CNN algorithm for searching for the optimal CNN structure includes the following steps in sequence: initialization of the particle swarm, fitness evaluation of a single particle, measurement of the difference between two particles, velocity calculation, and particle update.

[0054] Further preferably, in the step (2), the process of initializing the particle swarm is as follows: first setting the range of the number of network layers, then randomly selecting the convolutional layer, the maximum pooling layer, the average pooling layer and the fully connected layer as the current layer, and configuring their parameters before performing particle swarm initialization calculation;

[0055] Among them, the first layer of each particle is a convolutional layer, the last layer is a fully connected layer, and at least one of the convolutional layers, maximum pooling layers, average pooling layers and fully connected layers is randomly interspersed in the middle.

[0056] Further preferably, in the step (2), the architecture obtained before the particle swarm initialization calculation during the particle swarm initialization process is:

[0057] The first layer is a convolutional layer with a kernel size of 6×1 and 136 output channels;

[0058] The second layer is the convolution layer with a kernel size of 6×1 and 90 output channels;

[0059] The third layer is the convolutional layer with a kernel size of 5×1 and 217 output channels;

[0060] The fourth layer is a convolutional layer with a kernel size of 6×1 and 141 output channels;

[0061] The fifth layer is a convolutional layer with a kernel size of 3×1 and 197 output channels;

[0062] The sixth layer is a fully connected layer with 82 core neurons;

[0063] The seventh layer is a fully connected layer with 5 core neurons.

[0064] Further preferably, in the step (2), the architecture obtained before the particle swarm initialization calculation during the particle swarm initialization process is:

[0065] The first layer is a convolutional layer with a kernel size of 6×1 and 51 output channels;

[0066] The second layer is average pooling with a pooling size of 3×1 and a stride of 2;

[0067] The third layer is the convolution layer with a kernel size of 4×1 and 97 output channels;

[0068] The fourth layer is the convolutional layer with a kernel size of 5×1 and an output channel number of 228;

[0069] The fifth layer is a convolutional layer with a kernel size of 5×1 and an output channel number of 228;

[0070] The sixth layer is a fully connected layer with 279 core neurons;

[0071] The seventh layer is a fully connected layer with 5 core neurons.

[0072] Preferably, the parameters used in the PSO-CNN algorithm include: particle swarm algorithm parameters, CNN architecture initialization parameters and CNN training parameters.

[0073] Further preferably, the particle swarm algorithm parameters: number of iterations, population size and Cg are set to 30, 10 and 0.5 respectively;

[0074] The CNN architecture is initialized as follows: the maximum number of network layers is set to 15, the number of convolution kernels is selected in the range of [3, 256], the convolution kernel size is selected in the range of [3×1, 7×1], the number of neurons in the fully connected layer is selected in the range of [1, 300], the number of neurons in the output layer is 5, the convolution layer generation probability is 0.6, the pooling layer generation probability is 0.3, and the fully connected layer generation probability is 0.1;

[0075] The CNN training parameters are as follows: the number of training epochs for particle evaluation is 1, the number of training epochs for the global best particle is 200, and the Dropout is 0.3.

[0076] Preferably, in step (2), the particles are updated using the following formula:

[0077]

[0078]

[0079] Where k is the current iteration number; and They represent the velocity and position of particle i in the dth dimension in the next iteration respectively; Cg is the threshold; r is a number between 0 and 1 randomly generated at each position of the particle.

[0080] Example 1.

[0081] (1) Serum sample acquisition and spectral collection

[0082] The case samples included 38 cases of cervical cancer, 21 cases of grade I precancerous lesions, 27 cases of grade II precancerous lesions, 29 cases of grade III precancerous lesions, and 30 cases of uterine fibroids. All serum samples were stored in a freezer at -20°C before testing. During the testing process, serum samples were naturally thawed at a constant temperature of 22°C. Subsequently, 50 μL of each serum sample was placed on a ZnSe crystal and dried to form a uniform film before infrared testing. This study was approved by the regional ethics committee.

[0083] When measuring the spectrum, a VERTEX 70 infrared spectrometer produced by Bruker, Germany, and an attenuated total reflection (ATR) sample measurement accessory from SPECAC were used. The sample slot of the ATR is a ZnSe crystal, with an incident angle of 45° and three reflections. The beam splitter is KBr. Before recording each spectral data, the background data is measured using OPUS 65 software on a Windows XP system, and then each sample data is measured. After the test is completed, a Fourier transform is automatically performed in the software to obtain the Fourier infrared spectral data. The scanning range is 600-4000cm -1 , with a resolution of 8cm -1 The number of scans was 32. All patient serum samples were measured three times, resulting in a total of 435 data points. CO2 compensation was selected as the compensation parameter.

[0084] (2) Data processing and sample division

[0085] The experiment used a random partitioning method to divide each dataset into training, validation, and test sets in a ratio of 6:2:2. To ensure the reliability of the experimental results, the spectrum of a single sample was only included in the training set, validation set, or test set.

[0086] (3) Spectral analysis

[0087] Figure 1 The average spectra of the spectral data for five serum samples before modeling are shown. The shaded area represents the spectral standard deviation. Because different spectral absorption peaks correspond to different molecular structures, spectral differences can reflect changes in the substances in the serum. Based on existing FT-IR spectroscopy research, the different spectral characteristic peaks corresponding to infrared absorption bands and their attributions are listed in Table 1.

[0088] As cancer occurs, cancer cells will divide rapidly, which is consistent with the spectrum of CINI, CINII, CIN III, cervical cancer and uterine fibroids at 1056cm -1 The absorption of the nucleic acid band at 1400 cm -1 In addition, studies have shown that there are significant differences in phosphatidylcholine, phosphatidylethanolamine, diglyceride and free fatty acids between healthy controls, CINI, CINII / III and cervical cancer patients, indicating that the evolution of cervical precancerous lesions to cervical cancer is accompanied by changes in lipids. The spectral absorption peak associated with lipids corresponds to 2932 cm -1 、2962cm -1 1539cm -1 and 1647cm -1 The strong absorption peaks at the 400 nm and 400 nm correspond to the amide II and amide I bands, respectively, originating from the vibration of amide groups in proteins. As the severity of precancerous lesions decreases, the intensity of the spectral absorption peaks increases. Serum spectral intensities in patients with cervical cancer are lower than those in patients with precancerous lesions, consistent with the findings of Jusman et al. In summary, the five sample types differ primarily in the nucleic acid, lipid, and protein bands, corresponding to variations in spectral content. This provides both feasibility and biological evidence for combining FT-IR spectroscopy with machine learning algorithms to identify the five sample types.

[0089] Table 1 Peak positions and attributions of the main infrared bands of human serum

[0090]

[0091] (4) Optimizing the construction of convolutional neural network model based on particle swarm algorithm

[0092] CNN is a multi-layer neural network that can not only analyze the intrinsic information of signal sequences, but also overcome the nonlinearity and complexity of signals. In recent years, it has achieved remarkable achievements in the field of computer vision. In particular, in the study of the application of molecular spectroscopy technology in medical diagnosis, CNN has been widely used as a feature extractor and classifier. However, the structure of CNN is very complex, and the performance of the algorithm depends largely on the structure of CNN. Therefore, how to select a suitable CNN network structure has become an urgent problem that scholars in this field need to solve. Guo et al. used the PSO algorithm to optimize the hyperparameters of CNN, such as convolution kernel parameters, pooling type, activation function, and learning rate, under the premise of fixing the number and arrangement of network layers. Although they further increased the number of optimized hyperparameters, they still did not get rid of the premise of fixed network layer architecture.

[0093] In light of this, the present invention employs the PSO-CNN algorithm, a particle swarm optimization algorithm that uses variable-length particles to search for the optimal architecture of a convolutional neural network without size restrictions. Experimental results demonstrate that the optimal model discovered using PSO-CNN can achieve comparable results to complex convolutional neural network architectures without using any data augmentation techniques. Based on this, the present invention further proposes a model for rapid cervical tumor screening that incorporates the principles of the PSO-CNN algorithm, which can be used to assist in the early diagnosis of cervical cancer.

[0094] ①Algorithm thinking

[0095] In CNN, layers are stacked together so that the output of any given layer becomes the input of the next layer. The corresponding relationship between input and output can be expressed as formula (1). i represents the output of the i-th layer; f i (·) represents the activation function of the i-th layer; g i (·) represents the weighted operation of the i-th layer; Z i is the output of the weighted operation of the i-th layer before the activation function; W i is the weight of the i-th layer, and b represents the bias of the layer. CNNs are mainly composed of convolutional layers, pooling layers, and fully connected layers. For different network layers, the output corresponds to different weighted operations, as shown in Equation (2). The output of the convolutional layer is the convolution operation of its input and weights; the pooling layer simplifies the information in the convolutional layer output, including maximum pooling and average pooling; and the fully connected layer is similar to traditional artificial neural networks, and its output is a function of the weight multiplied by the input.

[0096]

[0097]

[0098] The basic idea of the particle swarm optimization algorithm is to find the optimal solution through collaboration and information sharing among individuals in the group. Each particle has a fitness value determined by the objective function and knows its best position (pbest) found so far and its current position x i This can be seen as the particle's own flight experience. In addition, each particle also knows the best position (gbest) found by all particles in the entire group so far (gbest is the best value in pbest). This can be seen as the experience of its companions. Particles track the two extreme values pbest and gbest by updating their own speed and position:

[0099]

[0100]

[0101] Where k is the current iteration number, and They represent the velocity and position of particle i in the dth dimension in the next iteration respectively; ω is the inertia weight, so It reflects the ability of particles to inherit the speed of the previous iteration; c1 and c2 are learning factors, c1 expresses the degree of dependence of the particle on its own memory, and c2 determines the influence of other particles in the particle group on the particle itself, which makes each particle move closer to the positions of pbest and gbest respectively; ranf() represents a random number between 0 and 1 that follows a uniform distribution, which is used to simulate slight disturbances in group behavior in nature.

[0102] The PSO algorithm has a fast convergence speed and is simple to implement, so each particle is regarded as a possible solution to the optimal CNN structure. Each particle is represented by a discrete group of network layer function blocks, for example: conv|conv|max_pool|conv|fc|fc|, which conform to the pre-set arrangement rules of the CNN network layer. The fitness function in the PSO optimization process of the CNN network structure is set to the classification accuracy of the validation set. The construction of the CNN architecture of the present invention is implemented using the Keras framework. Due to the high encapsulation of the Keras framework, it is simple and easy to use the Keras framework to construct a convolutional neural network during particle initialization and particle update.

[0103] ②Algorithm process

[0104] The algorithm consists of a particle swarm optimization algorithm framework, which includes the following five steps for searching the optimal CNN structure in addition to the effective CNN representation: particle swarm initialization, fitness evaluation of a single particle, measuring the difference between two particles, velocity calculation and particle update.

[0105] When the particle swarm initializes the network structure, it is necessary to set the range of network layers. In this embodiment, the minimum and maximum number of network layers are set to 3 and 20 respectively. After determining the number of network layers, the convolution layer, maximum pooling layer, average pooling layer and fully connected layer are randomly selected as the current layer. Here, the set network arrangement rules must be met: the first layer of each particle must be a convolution layer, the last layer must be a fully connected layer, and the convolution layer, maximum pooling layer, average pooling layer and fully connected layer are randomly interspersed in the middle. In addition, once the fully connected layer is selected as the current layer, it can only be followed by a fully connected layer. The probabilities of randomly selecting a convolution layer, a pooling layer and a fully connected layer are 0.6, 0.3 and 0.1 respectively. The implementation steps of the particle swarm initialization algorithm are shown in Table 2. While selecting the convolution layer, the number of convolution kernels is randomly configured in the range of [3,256] and the convolution kernel size is generated in the range of [3×1,7×1]. When selecting the pooling layer, the pooling filter size is set to 3×1 and the sampling step is 2. The maximum number of neurons in the fully connected layer is set to 300, and the number of neurons in the last output layer is 5.

[0106] Table 2 Particle swarm initialization of PSO-CNN algorithm

[0107]

[0108]

[0109] Furthermore, a function must be defined to evaluate the fitness of the corresponding particle architecture, using the validation set accuracy as the basis for evaluation. When directly compiling a complete CNN architecture from a single particle, this example sets the activation function for the convolutional and fully connected layers to ReLU, and the activation function for the final output layer to Softmax. To stabilize the intermediate output values of the neural network at each layer, batch normalization is applied to both the convolutional and fully connected layers. The dropout ratio between fully connected layers is set to 0.5.

[0110] In order to find the position of the global optimal solution, the particle needs to update its position when pbest, gbest, and its current position are known. From formulas (3) and (4), we can see that the first thing to solve here is the calculation of particle differences, namely pbest-x and gbest-x. Therefore, we need to specify the particle difference calculation rules. In order to avoid the appearance of fully connected layers between the convolutional layer and the pooling layer, when solving the particle difference, the fully connected layer will be separated separately, and then the layer types between the particles will be compared. If the particle layer types are the same, it is 0; if they are different, it is the layer type of the first particle. In addition, if the number of layers of the two particles is different, and the number of layers of the first particle is more than that of the second particle, the layer type of the first particle with more layers is added to the corresponding position. Conversely, if the number of layers of the first particle is less than that of the second particle, the layer type corresponding to the second particle is removed.

[0111] After clarifying the calculation rules of particle differences, the speed calculation can be performed to update the position. Since it is necessary to optimize the CNN architecture and update the particles without using real-value coding. Therefore, it is necessary to improve formula (3). This embodiment discards the inertia weight part and sets a selection threshold Cg. A number r between 0 and 1 is randomly generated at each position of the particle. When r≤Cg, pbest-x is taken as the updated speed. Otherwise, the speed is updated to gbest-x. Formula (3) can be rewritten as (5). We update the particles according to formula (5) and formula (4). It can be seen from formula (4) that the new particle is composed of the original particle and the speed. When the speed is P, C or F, the functional block of the original particle is replaced; when the speed is 0, the functional block of the original particle is retained; when the speed is R, the functional block corresponding to the original particle is removed. The particle architecture update process is as follows Figure 2 shown.

[0112]

[0113] (5) Experimental modeling and model comparison

[0114] The parameters used in the PSO-CNN algorithm can be divided into three categories: particle swarm algorithm parameters, CNN architecture initialization, and CNN training parameters.

[0115] The main parameters of the particle swarm algorithm include the number of iterations, population size, and the threshold set for the update rate. A larger number of iterations increases the probability of finding the global optimal solution. Population size also affects algorithm performance, and this effect is not a simple linear relationship. After reaching a certain size, further increases in size have limited performance improvements and instead increase computational complexity. However, the population size should not be too small. A too small size will not fully demonstrate the intelligence of the swarm intelligence optimization algorithm, severely impairing performance. Taking into account both algorithm performance and computational complexity, the number of iterations and population size are set to 30 and 10, respectively.

[0116] CNN architecture initialization parameters control the size of the particle search. To prevent the generation of complex particle architectures, the maximum number of network layers is set to 15 before particle initialization. The number of convolutional output channels, convolution kernel size, and number of neurons in the fully connected layer are initialized within a given range, and convolutional layers, pooling layers, and fully connected layers are generated with a certain probability. The last category of parameters includes the number of epochs for particle evaluation and the number of epochs for the global best particle training. To reduce overall training time, the number of epochs for particle evaluation training is set to 1. Dropout is used to randomly discard neurons in the fully connected layer, with each neuron being discarded with a probability of 0.3 to prevent overfitting. All parameter settings for the PSO-CNN algorithm are shown in Table 3.

[0117] Table 3 PSO-CNN algorithm parameter settings

[0118]

[0119]

[0120] The modeling results of the gbest network architecture found by running the PSO-CNN algorithm five times are shown in Table 4. The average accuracy of the test set of the CNN framework obtained in the second and fifth runs is higher, with the average accuracy of the five categories being 87.2% and 82.6% respectively. The optimal frameworks obtained in these two runs are shown in Table 4. Figure 3 shown.

[0121] The structure of the PSO-CNN algorithm used in this invention was discovered without any domain knowledge about the problem. Although the algorithm was tested on a small dataset, the experimental results of this example demonstrate the feasibility of the CNN network architecture found by the PSO-CNN algorithm for early diagnosis of cervical cancer patients using a serum FT-IR spectroscopy dataset.

[0122] Table 4. PSO-CNN algorithm five-time modeling results

[0123]

[0124] In this example, the network architectures obtained by the second and fifth iterations of the PSO-CNN algorithm were compared in accuracy with the classic LeNet, AlexNet, VGG16, and GoogLeNet network structures. The experimental results are shown in Table 5. Table 5 shows that the PSO-CNN algorithm's test set accuracy is higher than that of the other four classic network architectures, demonstrating the superiority of the network architecture found by this algorithm. Furthermore, the PSO-CNN used should have achieved even better results. Table 3 shows that the fully connected layer generated by the PSO-CNN algorithm has a maximum of 300 neurons, a significantly smaller number than the 4096 neurons used in the VGG16 fully connected layer. This is primarily due to the limited hardware available to search for more complex networks. Furthermore, it is impossible to run the algorithm with a large number of particles for more iterations. More particles would help the algorithm explore more CNN architectures. Therefore, this significantly limits the performance of the PSO-CNN algorithm. However, it also demonstrates that the technical solution of the present invention can achieve results with less data.

[0125] Table 5 Comparison of PSO-CNN model with other classic models

[0126]

[0127] Based on the particle swarm optimization algorithm's low parameter count and rapid convergence, this paper uses the algorithm to automatically construct a CNN structure with a variable number of layers and layer type parameters, designing a CNN model for early cervical cancer screening based on FT-IR technology. Using validation set accuracy as an evaluation basis during network training, experimental results show that the CNN structure obtained by the PSO-CNN algorithm achieves better detection results for early cervical cancer compared to several other classic CNN structures. This demonstrates the effectiveness of the PSO-CNN structure optimization method used in this paper for FT-IR technology and can provide a valuable reference for selecting CNN structures for subsequent cancer diagnosis.

[0128] The above is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the embodiments of the present invention. Any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the embodiments of the present invention are still within the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for establishing a model for rapid screening of cervical tumors, characterized in that , including the following steps: (1) After collecting serum samples from patients with cervical cancer of different lesion severity, the Fourier transform infrared spectroscopy data were measured and obtained; (2) The Fourier transform infrared spectroscopy data is used to establish a classification model through the PSO-CNN algorithm to obtain the model for rapid screening of cervical tumors; The PSO-CNN algorithm for searching for the optimal CNN structure includes the following steps in sequence: initialization of the particle swarm, fitness evaluation of a single particle, measurement of the difference between two particles, velocity calculation, and particle update; The particle swarm initialization process is as follows: first, the range of the number of network layers is set, then the convolution layer, the maximum pooling layer, the average pooling layer, and the fully connected layer are randomly selected as the current layer, and their parameters are configured before performing particle swarm initialization calculation; wherein, the first layer of each particle is a convolution layer, the last layer is a fully connected layer, and at least one of the convolution layer, the maximum pooling layer, the average pooling layer, and the fully connected layer is randomly interspersed in the middle; The parameters used in the PSO-CNN algorithm include: particle swarm optimization parameters, CNN architecture initialization parameters, and CNN training parameters; The particle swarm algorithm parameters: number of iterations, population size and Cg are set to 30, 10 and 0.5 respectively; The CNN architecture is initialized as follows: the maximum number of network layers is set to 15, the number of convolution kernels is selected in the range of [3, 256], the convolution kernel size is selected in the range of [3×1, 7×1], the number of neurons in the fully connected layer is selected in the range of [1, 300], the number of neurons in the output layer is 5, the convolution layer generation probability is 0.6, the pooling layer generation probability is 0.3, and the fully connected layer generation probability is 0.1; The CNN training parameters are: the number of training epochs for particle evaluation is 1, the number of training epochs for the global best particle is 200, and the Dropout is 0.3; The particle update is performed by the following formula: Where k is the current iteration number; and They represent the velocity and position of particle i in the dth dimension in the next iteration respectively; Cg is the threshold; r is a number between 0 and 1 randomly generated at each position of the particle.

2. The establishment method according to claim 1, characterized in that In the step (1), different degrees of lesions include cervical cancer, precancerous lesions grade I, precancerous lesions grade II, precancerous lesions grade III, and uterine fibroids.

3. The establishment method according to claim 1, characterized in that In the step (2), the architecture obtained before the particle swarm initialization calculation during the particle swarm initialization process is: The first layer is a convolutional layer with a kernel size of 6×1 and 136 output channels; The second layer is the convolution layer with a kernel size of 6×1 and 90 output channels; The third layer is the convolutional layer with a kernel size of 5×1 and 217 output channels; The fourth layer is a convolutional layer with a kernel size of 6×1 and 141 output channels; The fifth layer is a convolutional layer with a kernel size of 3×1 and 197 output channels; The sixth layer is a fully connected layer with 82 core neurons; The seventh layer is a fully connected layer with 5 core neurons.

4. The establishment method according to claim 1, characterized in that In the step (2), the architecture obtained before the particle swarm initialization calculation during the particle swarm initialization process is: The first layer is a convolutional layer with a kernel size of 6×1 and 51 output channels; The second layer is average pooling with a pooling size of 3×1 and a stride of 2; The third layer is the convolution layer with a kernel size of 4×1 and 97 output channels; The fourth layer is the convolutional layer with a kernel size of 5×1 and an output channel number of 228; The fifth layer is a convolutional layer with a kernel size of 5×1 and an output channel number of 228; The sixth layer is a fully connected layer with 279 core neurons; The seventh layer is a fully connected layer with 5 core neurons.

5. A model for rapid screening of cervical tumors, characterized in that: Obtained by the establishment method according to any one of claims 1 to 4.

Images