CN114613494A

CN114613494A - Model for rapidly screening cervical tumors and establishing method thereof

Info

Publication number: CN114613494A
Application number: CN202210256653.0A
Authority: CN
Inventors: 马彩玲; 吕小毅; 王静; 严紫薇; 陈程; 陈晨
Original assignee: Xinjiang University; First Affiliated Hospital of Xinjiang Medical University
Current assignee: Xinjiang University; First Affiliated Hospital of Xinjiang Medical University
Priority date: 2022-03-16
Filing date: 2022-03-16
Publication date: 2022-06-10

Abstract

The invention relates to a model for rapidly screening cervical tumors and an establishing method thereof. A method of building a model for rapid screening of cervical tumors, comprising the steps of: (1) collecting serum samples of cervical tumor patients with different lesion degrees, and measuring and acquiring Fourier infrared spectrum data; (2) and establishing a classification model by the Fourier infrared spectrum data through a PSO-CNN algorithm to obtain the model for rapidly screening the cervical tumors. According to the model for rapidly screening the cervical tumors and the establishment method thereof, the PSO-CNN algorithm is optimized, so that the identification effect can be improved, and the model can be better applied to rapidly screening the cervical tumors.

Description

Model for rapidly screening cervical tumors and establishing method thereof

Technical Field

The invention particularly relates to a model for rapidly screening cervical tumors and an establishing method thereof.

Background

Cervical cancer is considered the second most common malignancy in women, and is the leading cause of cancer death worldwide next to breast cancer. The number of new cases of cervical cancer in China reaches 11 thousands every year, which accounts for 1/3 of the total new cases in the world. Cervical Intraepithelial Neoplasia (CIN) is closely related to cervical cancer, and persistent high-risk human papillomavirus infection causes low-grade lesions (CIN I) of cervical epithelial cells, which then further develop into high-grade precancerous lesions (CIN II, CIN III) of the cervix, and finally develop into cervical cancer, which generally takes 8-15 years. However, the development time is extremely fast, and if the lesion before cervical cancer can be diagnosed in time and corresponding solution measures are taken, the lesion can be interfered in time to prevent the lesion from developing towards an adverse direction. Current methods to assist clinicians in the early detection of pre-cervical cancer lesions include cytopathology, HPV detection, and histopathology, but these methods are limited in subjectivity, cost, and time. Therefore, there is a clinical need for an objective, rapid assay that requires less sample preparation.

FT-IR is an optical spectroscopic technique that can efficiently provide information on the structure and chemical composition of biological materials at the molecular level, 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, there have been numerous studies revealing the potential of FT-IR spectroscopy and machine learning algorithms for detecting various cancers, with few research efforts in cervical cancer screening. Mo and the like adopt Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) algorithms and combine serum FT-IR spectrum to identify cervical cancer patients and healthy control groups, and the diagnosis accuracy, specificity and sensitivity reach 98 percent. Although the experiment was very effective, diagnosis of CIN patients was ignored. Nordstrom et al identified normal tissues, CIN I, CIN II/III, using a multivariate algorithm based on the UV fluorescence spectral characteristics of the biopsy tissue site, with sensitivity and specificity for classification of CIN II/III and normal tissue of 91% and 93%, respectively, and sensitivity and specificity for classification of CIN I and normal tissue of 86% and 87%, respectively. Although the classification results are more accurate, the experiment can further increase the identification of different stages of cervical high-grade precancerous lesions. Furthermore, Yang et al propose a feature fusion-based method for classifying cervicitis, low-grade squamous intraepithelial lesions, high-grade squamous intraepithelial lesions, cervical squamous cell carcinoma and cervical adenocarcinoma, and the classification accuracy rates of KNN, ELM, ABC-SVM, CS-SVM, PSO-SVM and CNN-LTSM are 60.91%, 67.84%, 77.64%, 78.49%, 75.54% and 70.72% respectively, which is improved compared with the accuracy rate of the original spectrum classification without feature fusion. However, as in most studies using CNN as a feature extraction or classifier, they do not optimize the network structure of CNN, and limit the recognition effect of CNN to some extent.

In view of the above, the present invention provides a new model for rapidly screening cervical tumors, and further optimizes the existing early screening model.

Disclosure of Invention

The invention aims to provide a method for establishing a model for rapidly screening cervical tumors, which can improve the identification effect by optimizing a PSO-CNN algorithm.

In order to realize the purpose, the adopted technical scheme is as follows:

a method of building a model for rapid screening of cervical tumors, comprising the steps of:

(1) collecting serum samples of cervical tumor patients with different lesion degrees, and measuring and acquiring Fourier infrared spectrum data;

(2) and establishing a classification model by the Fourier infrared spectrum data through a PSO-CNN algorithm to obtain the model for rapidly screening the cervical tumors.

Further, in the step (1), the different pathological changes include cervical cancer, precancerous lesion I grade, precancerous lesion II grade, precancerous lesion III grade and hysteromyoma.

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

Still further, in the step (2), the initialization process of the particle swarm is as follows: firstly setting the range of the number of network layers, then randomly selecting a convolution layer, a maximum pooling layer, an average pooling layer and a full-connection layer as current layers, and after parameters of the current layers are prepared, performing particle swarm initialization calculation;

wherein 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 layer, the maximum pooling layer, the average pooling layer and the fully-connected layer is randomly interspersed in the middle.

Still further, in the step (2), an architecture obtained before performing the particle group initialization calculation in the process of initializing the particle group is as follows:

the first layer is a convolution layer, the kernel size is 6 multiplied by 1, and the number of output channels is 136;

the second layer is a convolution layer, the core size is 6 multiplied by 1, and the number of output channels is 90;

the third layer is a convolution layer, the size of the core is 5 multiplied by 1, and the number of output channels is 217;

the fourth layer is a convolution layer, the kernel size is 6 multiplied by 1, and the number of output channels is 141;

the fifth layer is a convolution layer, the core size is 3 multiplied by 1, and the number of output channels is 197;

the sixth layer is a full-connection layer, and the number of nuclear neurons is 82;

the seventh layer is a full-junction layer, and the number of nuclear neurons is 5.

the first layer is a convolution layer, the kernel size is 6 multiplied by 1, and the number of output channels is 51;

the second layer is in average pooling, the pooling size is 3 multiplied by 1, and the step length is 2;

the third layer is a convolution layer, the core size is 4 multiplied by 1, and the number of output channels is 97;

the fourth layer is a convolution layer, the kernel size is 5 multiplied by 1, and the number of output channels is 228;

the fifth layer is a convolution layer, the core size is 5 multiplied by 1, and the number of output channels is 228;

the sixth layer is a full-junction layer, and the number of nuclear neurons is 279;

Further, the parameters used in the PSO-CNN algorithm include: particle swarm algorithm parameters, CNN architecture initialization parameters and CNN training parameters.

Still further, the particle swarm algorithm parameters are as follows: the iteration number, the population size and the Cg are respectively set to be 30, 10 and 0.5;

initializing the CNN architecture: the maximum network layer number is set to be 15, the selection range of the number of convolution kernels is [3,256], the selection range of the size of the convolution kernels is [3 × 1,7 × 1], the selection range of the number of neurons of a full connection layer is [1,300], the number of neurons of an output layer is 5, the generation probability of the convolution layer is 0.6, the generation probability of a pooling layer is 0.3, and the generation probability of the full connection layer is 0.1;

the CNN training parameters are as follows: the training epoch number for particle evaluation is 1, the global optimal particle training epoch number is 200, and Dropout is 0.3.

Further, in the step (2), the particles are updated by the following formula:

wherein k is the current iteration number;

and

respectively representing the speed and the position of the dimension d of the particle i in the next iteration; cg is a threshold value; r is a number between 0 and 1 randomly generated at each position of the particle.

Compared with the prior art, the invention has the beneficial effects that;

the optimal CNN framework is selected in a self-adaptive mode through an optimization algorithm, so that early cervical cancer screening based on the FT-IR technology is realized quickly and efficiently. In the field of CNN parameter optimization, genetic algorithms and particle swarm algorithms are mostly used. In general, PSO is simpler and converges faster than GA because there are no crossover and mutation operations. Furthermore, the PSO algorithm also achieves excellent results due to the communication between particles and randomness.

Based on the characteristics of few parameters and fast convergence of the particle swarm algorithm, the particle swarm algorithm is used for automatically building CNN structures with indefinite layer number and indefinite layer type parameters, and a CNN model for early screening of cervical cancer based on the FT-IR technology is designed. The accuracy of the verification set is used as an evaluation basis in network training, and experimental results show that the CNN structure obtained by the PSO-CNN algorithm has a better detection result in the aspect of screening early-stage cervical cancer tumors than other classical CNN structures.

Drawings

FIG. 1 is a graph of the average FT-IR spectra of five serum samples; wherein the shaded area represents the standard deviation;

FIG. 2 is a schematic diagram of particle architecture update;

FIG. 3 is an 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

In order to further illustrate the model for rapidly screening cervical tumors and the method for establishing the same according to the present invention, the following embodiments, structures, features and effects thereof will be described in detail with reference to the accompanying drawings. In the following description, different "one embodiment" or "an embodiment" refers to not necessarily the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

Before describing the model for rapidly screening cervical tumors and the method for establishing the same in detail, it is necessary to further describe the relevant materials mentioned in the present invention to achieve better results.

The model for rapidly screening cervical tumors and the establishment method thereof according to the present invention will be further described in detail with reference to the following specific examples:

the invention provides a cervical cancer early tumor screening model based on one-dimensional FT-IR spectral data and PSO-CNN on the basis of automatically selecting a CNN network layer architecture and super parameters thereof by using a Particle Swarm Optimization (PSO), and provides reference for selecting a CNN framework in the field in the future.

The technical scheme of the invention is as follows:

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

Further preferably, in the step (2), the initialization process of the particle group is: setting a network layer number range, randomly selecting a convolutional layer, a maximum pooling layer, an average pooling layer and a full-link layer as current layers, and performing particle swarm initialization calculation after parameters of the current layers are prepared;

Further preferably, in the step (2), an architecture obtained before performing the particle group initialization calculation in the process of initializing the particle group is as follows:

the third layer is a convolution layer, the core size is 5 multiplied by 1, and the number of output channels is 217;

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

Further preferably, the particle swarm algorithm parameters are: the iteration number, the population size and the Cg are respectively set to be 30, 10 and 0.5;

initializing the CNN architecture: the maximum network layer number is set to be 15, the selection range of the number of convolution kernels is [3,256], the selection range of the size of the convolution kernels is [3 × 1,7 × 1], the selection range of the number of neurons of the full-connection layer is [1,300], the number of neurons of the output layer is 5, the generation probability of the convolution layer is 0.6, the generation probability of the pooling layer is 0.3, and the generation probability of the full-connection layer is 0.1;

Preferably, in the step (2), the particles are updated by the following formula:

wherein k is the current iteration number;

and

Example 1.

(1) Serum sample acquisition and spectral collection

The study cases were 38 cases of cervical cancer, 21 cases of precancerous lesion I, 27 cases of precancerous lesion II, 29 cases of precancerous lesion III, and 30 cases of uterine fibroids. Serum samples from all patients were stored in a freezer at-20 ℃ prior to testing. During the test, the serum samples were thawed naturally in a constant environment at room temperature of 22 ℃. Subsequently, 50. mu.L of each serum sample was placed on ZnSe crystals and dried to form a uniform film, followed by infrared measurement. The study has been approved by the regional ethics committee.

For the measurement of the spectra, a VERTEX70 infrared spectrometer from Bruker, Germany and an Attenuated Total Reflectance (ATR) sample measuring accessory from SPECAC were used. The sample cell of the ATR was ZnSe crystal with an incidence angle of 45 °, triple reflection. The beam splitter is KBr. Before each recording of spectral data, background data was measured using the OPUS 65 software on the windows xp system, and then each sample data was measured. After the test is completed, a fourier transform will be automatically performed in the software to obtain fourier infrared spectral data. Wherein the scanning range is 600-4000cm^-1Resolution of 8cm^-1The number of scans was 32. All patient serum samples were measured 3 times, giving 435 data total. Selection of CO₂The compensation is used as a compensation parameter.

(2) Data processing and sample partitioning

The experiment adopts a mode of randomly dividing data sets, and each type of data set is divided into 6: 2: the scale of 2 divides the training set, the validation set, and the test set. In order to ensure the reliability of the experimental results, the spectrum of a single sample is only contained in the training set, the verification set or the test set.

(3) Spectral analysis

FIG. 1 shows the mean spectra of five types of serum sample before modeling. In which the spectral standard deviation is represented by the shaded portion. Since different spectral absorption peaks correspond to different molecular structures, the spectral differences may reflect changes in the substances in the serum. The results of the prior FT-IR spectroscopy were combined to list different spectral characteristic peaks and assignments corresponding to the infrared absorption bands, as shown in Table 1.

As carcinogenesis occurs, cancer cells undergo rapid division, which is in contrast to CINI, CINII, CIN III, cervical cancer and hysteromyoma spectra at 1056cm^-1The nucleic acid band at (A) corresponds to an increased absorption, similarly located at 1400cm^-1The absorption peak of (1). In addition, there were studies showing significant differences in phosphatidylcholine, phosphatidylethanolamine, diglyceride and free fatty acid in healthy controls, CINI, CINII/III and cervical cancer patients, indicating a progression from pre-cervical lesions to cervical cancer with lipid changes. 2932cm corresponding to the spectral absorption peak associated with lipids^-1、2962cm^-1。1539cm^-1And 1647cm^-1The strong absorption peaks correspond to amide II and amide I bands respectively, the spectrum absorption peak intensity is increased along with the reduction of the precancerous lesion degree due to the vibration of amide groups in protein, the serum spectrum intensity of a cervical cancer patient is lower than that of the precancerous lesion patient, and the research result is consistent with the research result of Jusman and the like. In conclusion, the five samples mainly have differences in nucleic acid, lipid and protein bands and correspond to changes of spectral substance content, so that feasibility and biological basis are provided for identifying the five samples by combining FT-IR spectrum with a machine learning algorithm.

TABLE 1 Peak position and attribution of the major Infrared bands of human serum

(4) Optimization convolutional neural network model building based on particle swarm optimization

CNN is a multi-layer neural network that can not only analyze intrinsic information of a signal sequence, but also overcome signal nonlinearity and complexity, and has recently achieved significant success in the field of computer vision. Among them, CNN is widely used as a feature extraction and classifier in research on application of molecular spectroscopy to medical diagnosis. However, the structure of CNN is very complex, and the performance of the algorithm depends on the structure of CNN to a large extent, so how to select a suitable CNN network structure becomes an urgent problem for scholars in this field to solve. Guo et al use the PSO algorithm to optimize the hyper-parameters of CNN, such as convolution kernel parameters, pooling type, activation function, learning rate, etc., on the premise of fixing the number and arrangement of network layers. Although they further increase the number of optimized superparameters, they still do not fall under the premise of fixed network layer architecture.

In view of this, the PSO-CNN algorithm is adopted in the technical solution of the present invention, and the particle swarm optimization algorithm can search for the optimal architecture of the convolutional neural network by using variable-length particles without being limited by size, and the experimental result shows that the optimal model found by the PSO-CNN can obtain an experimental result comparable to the complex convolutional neural network architecture without using any data enhancement technique. Based on the above, the invention further provides a model for rapidly screening the cervical tumor by combining the PSO-CNN algorithm idea, and the model can be used for assisting the early diagnosis of the cervical cancer.

Algorithm idea

The layers in CNN are stacked together so that the output of any given layer will become the input of the next layer, and the correspondence between the input and output can be expressed as equation (1). Wherein: x is input data; o is_iRepresents the output of the ith layer; f. of_i() represents the activation function of the ith layer; g_i() represents the weighted operation of the ith layer; z_iIs the output of the weighting operation at layer i before the activation function; w_iB represents the bias of the ith layer as a weight of the layer. And CNN is mainly composed of a convolutional layerThe pooling layer and the full connection layer, and for different network layers, the output corresponds to different weighting operations, as shown in formula (2). Wherein: the output of the convolutional layer is a convolution operation of its input and weight; what the pooling layer does is simplify the information in the convolutional layer output, including maximum pooling and average pooling; the fully-connected layer is similar to a conventional artificial neural network in that the output is a function of the weight multiplied by the input.

The basic idea of the particle swarm optimization algorithm is to find the optimal solution through cooperation and information sharing among individuals in a group. Each particle has an adaptation value determined by the objective function and knows the best position (pbest) found so far and the current position x_i. This can be seen as the flight experience of the particles themselves. In addition to this, each particle also knows the best position (gbest) found by all particles in the entire population so far (gbest is the optimal value in pbest). This can be seen as the experience of particle companions. The particle tracks the two extreme values pbest and gbest by updating its own speed and position:

wherein k is the current iteration number,

and

respectively representing the speed and the position of the dimension d of the particle i in the next iteration; ω is the inertial weight, therefore

The embodiment is the capability of particle inheriting the last iteration speed; c. C₁And c₂As a learning factor, c₁Expressing the degree of dependence of the particle on self-memory, c₂Determining the influence of other particles in the particle group on the particles, wherein the other particles respectively make each particle close to the positions of pbest and gbest; and rand () represents a random number obeying uniform distribution between 0 and 1, which is used to simulate slight perturbations in group behavior in nature.

The PSO algorithm converges quickly and is simple to implement, so each particle is considered as a possible solution to the optimal structure of CNN. Each particle is represented by a discrete set of network layer functions, such as: conv | conv | max _ pool | conv | fc | which meet the preset arrangement rule of the CNN network layer. And setting a fitness function in the process of optimizing the CNN network structure by the PSO as the classification accuracy of the verification set. The construction of the CNN framework is realized by using a keras framework. Because the keras framework has high encapsulation, the construction of the convolutional neural network by using the keras framework in the particle initialization and particle updating processes is simple and easy.

Algorithm flow

The algorithm consists of a particle swarm optimization algorithm framework, and comprises the following five steps for searching an optimal CNN structure besides effective CNN representation: initialization of a particle swarm, fitness evaluation of a single particle, measurement of the difference between two particles, velocity calculation and particle update.

The particle swarm needs to set a network layer number range when initializing the network structure, and the minimum network layer number and the maximum network layer number are set to be 3 and 20 respectively in this embodiment. After the number of network layers is determined, a convolutional layer, a maximum pooling layer, an average pooling layer and a full-link layer are randomly selected as current layers, wherein the current layers need to meet the set network arrangement rule: the first layer of each particle must be a convolutional layer, the last layer must be a fully-connected layer, and the convolutional layer, the max pooling layer, the average pooling layer, and the fully-connected layer are randomly interspersed among them. Furthermore, once the fully-connected layer is selected as the current layer, only the fully-connected layer follows. The probability of randomly selecting the convolutional layer, the pooling layer and the fully-connected layer is 0.6, 0.3 and 0.1 respectively. The particle swarm initialization algorithm implementation steps are shown in table 2. The number of convolution kernels arranged in the range of [3,256] is randomly selected while selecting convolution layers, and the convolution kernel size is generated in the range of [3 × 1,7 × 1 ]. When the pooling layer is selected, the size of the pooling filter is set to 3 × 1, and the sampling step size is 2. The maximum value of the neuron number of the full connection layer is set to be 300, and the neuron number of the last layer of the output layer is 5.

TABLE 2 PSO-CNN Algorithm initialization particle swarm

In addition, a function needs to be defined to evaluate the fitness of the corresponding architecture of the particle, i.e. the accuracy of the verification set is used as the evaluation basis. When a particle is directly compiled into a complete CNN architecture, the embodiment sets the activation functions of the convolutional layer and the fully-connected layer as relu, and sets the activation function of the last output layer as softmax. In order to enable the neural network to output more stable values in the middle of each layer, batch normalization is performed on the convolutional layer and the full connection layer. The ratio of dropout between fully connected layers is set to 0.5.

In order to find the position of the global optimal solution, the particle needs to perform position update under the condition that pbest, gbest and its current position are known, as shown in formula (3) and formula (4), and the first solution to be solved is the calculation of particle difference, namely pbest-x and gbest-x. Therefore we need to specify the particle difference calculation rule. In order to avoid the occurrence of fully connected layers between the convolutional and pooling layers, the fully connected layers are separated separately when solving for particle differences, and then the layer types between the particles are compared. If the particle layer types are the same, it is 0, and if 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 greater than that of the second particle, the layer type with the first particle being more is added at the corresponding position, whereas 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.

The calculation of the velocity can be done after the calculation rules for the particle differences are defined to update the position. The CNN architecture needs to be optimized and the particles are updated without using real-valued coding. Therefore, improvement of equation (3) is required. In the embodiment, an inertia weight part is discarded, a selection threshold Cg is set, a number r between 0 and 1 is randomly generated at each position of the particle, when r is less than or equal to Cg, pbest-x is taken as the updated speed, and otherwise, the speed is updated to be gbest-x. Equation (3) can be rewritten as (5). We update the particles according to the formula (5) and the formula (4), and it can be known from the formula (4) that the new particles are composed of the original particles and the velocity added together. When the speed is P, C or F, performing replacement operation on the function block of the original particle; when the speed is 0, the functional blocks of the original particles are reserved; and when the speed is R, removing the functional block corresponding to the original particle. The particle architecture update process is shown in FIG. 2.

(5) Experimental modeling and model comparison

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

The particle swarm algorithm parameters mainly comprise iteration times, a swarm size and a selection threshold set during updating speed. Wherein a larger number of iterations increases the probability of finding a globally optimal solution. The population scale also has influence on the performance of the algorithm, the influence is not a simple linear relation, and when the population scale reaches a certain degree, the improvement of the performance of the algorithm is limited by increasing the population scale, and the calculation amount is increased; however, the group scale cannot be too small, and the intelligence of the group intelligent optimization algorithm cannot be embodied by the too small group scale, so that the performance of the algorithm is seriously damaged. The number of iterations and the population size are set to 30 and 10, respectively, taking the algorithm performance and the calculation amount into consideration.

The CNN architecture initialization parameter controls the size of the particle search. To prevent the generation of a complex particle architecture, the maximum number of network layers is set to 15 before particle initialization. The number of convolution output channels, the size of the convolution kernel and the number of neurons of the full connection layer are initialized within a given range, and the convolution layer, the pooling layer and the full connection layer are generated with a certain probability. The last type of parameters includes the epoch number of training in particle evaluation and the epoch number of global optimal particle training, and in order to reduce the overall training time, the epoch number of training in particle evaluation is set to 1. The neurons in the fully-connected layer were randomly discarded using Dropout, each neuron being discarded with a probability of 0.3 to prevent overfitting. All parameter settings in the PSO-CNN algorithm are shown in Table 3.

TABLE 3 PSO-CNN Algorithm parameter settings

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 sets of the CNN frames obtained in the second and fifth times is higher, the average accuracy rates of the five classifications are 87.2% and 82.6%, respectively, and the optimal frames obtained by the two times of optimization are shown in fig. 3.

The structure in the PSO-CNN algorithm used in the present invention is not discovered without any domain knowledge about the problem. Although the algorithm is tested on a small data set, the experimental results of the embodiment show the feasibility of the CNN network architecture found by the PSO-CNN algorithm in early diagnosis of cervical cancer patients on the serum FT-IR spectrum data set.

TABLE 4 PSO-CNN Algorithm five modeling results

In this embodiment, the network architectures obtained by the PSO-CNN algorithm for the second time and the fifth time are compared with the network structures of the classical Lenet, AlexNet, VGG16, and google Lenet, and the experimental results are shown in table 5. From table 5, it can be seen that the accuracy of the test set of the PSO-CNN algorithm is higher than that of the other four classical network architectures, and the superiority of the network structure found by the algorithm is demonstrated. Furthermore, the PSO-CNN used should achieve better results than here. From table 3, it can be seen that the fully-connected layer generated by the PSO-CNN algorithm has only 300 neurons at the maximum, which is much smaller than VGG16, which uses 4096 neurons in the fully-connected layer. This is mainly due to the fact that the available hardware does not allow searching for more complex networks. In addition, algorithms with large numbers of particles cannot be run to make more iterations. More particles will help the algorithm explore more CNN architectures. Therefore, this also greatly limits the performance of the PSO-CNN algorithm. But at the same time, this illustrates that the solution of the invention can obtain results with less data.

TABLE 5 PSO-CNN model in comparison with other classical models

Based on the characteristics of few parameters and fast convergence of the particle swarm algorithm, the particle swarm algorithm is used for automatically building CNN structures with indefinite layer number and indefinite layer type parameters, and a CNN model for early screening of cervical cancer based on the FT-IR technology is designed. The accuracy of the verification set is used as an evaluation basis in network training, and experimental results show that the CNN structure obtained by the PSO-CNN algorithm has a better detection result in the aspect of screening early-stage cervical cancer tumors than other classical CNN structures. The effectiveness of the PSO-CNN structure optimizing method in the aspect of FT-IR technology is shown, and the method can provide reference value for selecting the CNN structure for subsequent cancer diagnosis.

The above description is only a preferred embodiment of the present invention, and is not intended to limit the present invention in any way, and any simple modification, equivalent change and modification made on the above embodiments according to the technical spirit of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A method for building a model for rapid screening of cervical tumors, comprising the steps of:

2. The method of establishing according to claim 1,

in the step (1), the different pathological changes comprise cervical cancer, precancerous lesion I grade, precancerous lesion II grade, precancerous lesion III grade and hysteromyoma.

3. The method of establishing according to claim 1,

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

4. The method of establishing according to claim 3,

in the step (2), the initialization process of the particle swarm is as follows: setting a network layer number range, randomly selecting a convolutional layer, a maximum pooling layer, an average pooling layer and a full-link layer as current layers, and performing particle swarm initialization calculation after parameters of the current layers are prepared;

5. The method of establishing according to claim 4,

in the step (2), the architecture obtained before performing particle swarm initialization calculation in the process of initializing the particle swarm is as follows:

6. The method of establishing according to claim 4,

7. The method of establishing according to claim 1,

the parameters used in the PSO-CNN algorithm comprise: particle swarm algorithm parameters, CNN architecture initialization parameters and CNN training parameters.

8. The method of establishing according to claim 7,

the particle swarm algorithm parameters are as follows: the iteration number, the population size and the Cg are respectively set to be 30, 10 and 0.5;

9. The method of establishing according to claim 1,

in the step (2), the particles are updated by the following formula:

wherein k is the current iteration number;

and

10. A model for rapid screening of cervical tumors, obtained by the method of construction according to any one of claims 1 to 9.