A lung nodule classification method based on a cascaded convolutional neural network model

Abstract

The application provides a lung nodule classification method based on a cascade convolutional neural network model, and relates to the technical field of deep learning medical image processing. The method comprises the following steps: acquiring a lung CT image and performing pretreatment, generating a sample data set according to the pretreated image; labeling the sample data in the sample data set, dividing the sample data set into multiple sub-sample data sets according to the labels, performing data augmentation on part of the sub-sample data sets, and then dividing each sub-sample data set into a training set, a verification set and a test set; constructing a binary classification model based on a ResNet residual neural network, training and verifying the binary classification model by using multiple sub-sample data sets, and correspondingly obtaining multiple binary classification models; constructing a cascade network by using the obtained multiple binary classification models, inputting the test set in the multiple sub-sample data sets into the constructed cascade network for classification, and obtaining a classification result. Through the above method, the lung nodule is classified, and the classification accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning medical image processing, and particularly relates to a lung nodule classification method based on a cascade convolutional neural network model. BACKGROUND

[0002] The early clinical manifestation of lung cancer is solitary pulmonary nodule. In medicine, pulmonary nodules are divided into six categories: ground glass benign, ground glass malignant, ground glass component-based, ground glass malignant solid component-based, good prognosis, poor prognosis, solid benign, and solid malignant. Screening of pulmonary nodules requires layer-by-layer screening of CT images to find cases of pulmonary nodules, and such a huge workload may lead to mis-screening. Therefore, it is necessary to establish a lung nodule classification model with good performance, which classifies the read-in CT images through a lung nodule classification algorithm, and once similar lesions are found, the patient needs to be regularly tracked and observed to prevent the patient's condition from deteriorating or even canceration.

[0003] By using computer classification algorithms to provide auxiliary information, specifically by using codes to pre-process CT images such as labeling, enhancement, cropping, segmentation, etc., and then classifying the nodules in the CT images. Artificial intelligence plays an important role in image classification, which can significantly improve the classification degree and accuracy, and is applied in many fields such as medical imaging. Machine learning is an important branch of artificial intelligence, and deep learning is the most important algorithm in machine learning, which aims to extract high-level abstract features of data through multiple nonlinear transformations, learn the potential distribution of data, and thus acquire the ability to make reasonable judgments or predictions on new data. Deep learning has begun to stand out in various fields due to its strong fitting ability, especially in the field of image classification, which can also benefit from deep learning methods.

[0004] Chinese patent application CN115578307A entitled "A lung nodule benign and malignant classification method and related products" discloses a lung nodule classification method as follows: first, physiological characteristics of a lung nodule patient and marker characteristics of the lung nodule under imaging are obtained; then, based on the physiological characteristics and the marker characteristics, a first evaluation result of the lung nodule is obtained; next, a detection result of a methylation marker of circulating tumor DNA of the lung nodule patient is obtained; finally, based on the detection result, a second evaluation result of the lung nodule is obtained; and according to the first evaluation result and the second evaluation result, a comprehensive evaluation result is obtained. This method only classifies lung nodules as benign or malignant, without more detailed classification.

[0005] The Chinese patent application CN112215799A entitled "A ground glass lung nodule automatic classification method and system" discloses a lung nodule automatic classification method, which is to process and analyze chest CT images through image processing methods, feature engineering and machine learning technology, effectively distinguish ground glass lung nodules representing lung micro-infiltrating adenocarcinoma and infiltrating adenocarcinoma, and quantitatively evaluate the classification effect. Although this method can realize the automatic classification of ground glass lung nodules representing lung micro-infiltrating adenocarcinoma and infiltrating adenocarcinoma, it does not effectively classify solid nodules. SUMMARY

[0006] In view of the above problems of the prior art, the present application provides a lung nodule classification method based on a cascade convolutional neural network model, which can comprehensively and meticulously classify lung nodules through hierarchical classification, making the lung nodule classification more comprehensive and detailed, thereby improving the accuracy of lung nodule classification.

[0007] The technical scheme of the present application is as follows:

[0008] A lung nodule classification method based on a cascade convolutional neural network model, the method comprising the following steps:

[0009] Step 1: Obtain a lung CT image, pre-process the obtained CT image, and generate a sample data set according to the pre-processed image;

[0010] Step 2: Label the sample data in the sample data set with a label, and divide the sample data set into multiple sub-sample data sets according to the label, compare the difference in the number of samples in any two sub-sample data sets with a set threshold, and according to the comparison result, data augmentation is performed on the sub-sample data set with fewer samples, and each sub-sample data set is divided into a training set, a validation set and a test set;

[0011] Step 3: Construct a binary classification model based on a ResNet residual neural network, input the training sets in the multiple sub-sample data sets into the binary classification model in turn for training, and correspondingly obtain multiple trained binary classification models, and use the validation sets in the multiple sub-sample data sets to verify the trained binary classification models, and correspondingly obtain multiple final binary classification models;

[0012] Step 4: Construct a cascade network using the multiple binary classification models obtained in step 3, input the test sets in the multiple sub-sample data sets into the constructed cascade network for classification, and obtain the classification results;

[0013] Further, the preprocessing method is: screening CT images with lung nodules from the acquired lung CT images, marking the positions of the nodules in the screened CT images, and cropping the current CT image according to the positions of the nodules, calculating the length and width of the nodules in the CT image according to the marked positions of the lung nodules, and adjusting the nodules to the middle position of the image according to the calculated results to realize the centered display of the nodules;

[0014] Further, the step 2 comprises:

[0015] Step 2.1: labeling the sample data in the sample data set with labels;

[0016] The labels include: ground glass nodule, solid nodule, ground glass benign nodule, ground glass malignant nodule, solid benign nodule, solid malignant nodule, ground glass malignant ground glass component dominant nodule, ground glass malignant solid component dominant nodule, ground glass malignant solid component dominant nodule with good prognosis, ground glass malignant solid component dominant nodule with poor prognosis;

[0017] Step 2.2: dividing the sample data set into a plurality of sub-sample data sets according to the labels of the sample data, wherein each sub-sample data set has two labels;

[0018] The sub-sample data set is 5, including: data set 1: ground glass nodule, solid nodule; data set 2: ground glass benign nodule, ground glass malignant nodule; data set 3: solid benign nodule, solid malignant nodule; data set 4: ground glass malignant ground glass component dominant nodule, ground glass malignant solid component dominant nodule; data set 5: ground glass malignant solid component dominant nodule with good prognosis, ground glass malignant solid component dominant nodule with poor prognosis;

[0019] Step 2.3: comparing the difference in the number of samples in any two sub-sample data sets with a set threshold value, and when the difference in the number of samples between any two sub-sample data sets reaches the set threshold value, data augmentation is performed on the sub-sample data set with fewer samples;

[0020] The data augmentation method comprises: using one or more of image rotation, image mirroring and image brightening to perform data augmentation;

[0021] Step 2.4: dividing each sub-sample data set into a training set, a validation set and a test set according to a set proportion;

[0022] Further, the trained binary classification model in step 3 is 5, denoted as Model 1, Model 2, Model 3, Model 4, and Model 5, wherein Model 1 is used for classifying the input training set into ground glass nodule and solid nodule; Model 2 is used for classifying the input training set into ground glass benign nodule and ground glass malignant nodule; Model 3 is used for classifying the input training set into solid benign nodule and solid malignant nodule; Model 4 is used for classifying the input training set into ground glass malignant ground glass component dominant nodule and ground glass malignant solid component dominant nodule; and Model 5 is used for classifying the input training set into ground glass malignant solid component dominant nodule with good prognosis and ground glass malignant solid component dominant nodule with poor prognosis.

[0023] Further, the method for constructing the cascade network in step 4 is: dividing the cascade network into 4 levels, wherein the first level uses Model 1 to distinguish whether the input test set data is a ground glass nodule or a solid nodule; the second level uses Model 2 to distinguish whether the ground glass nodule is a ground glass benign nodule or a ground glass malignant nodule, and uses Model 3 to distinguish whether the solid nodule is a solid benign nodule or a solid malignant nodule; the third level uses Model 4 to distinguish whether the ground glass malignant nodule is a ground glass malignant ground glass component dominant nodule or a ground glass malignant solid component dominant nodule; and the fourth level uses Model 5 to distinguish whether the ground glass malignant solid component dominant nodule is a ground glass malignant solid component dominant nodule with good prognosis or a ground glass malignant solid component dominant nodule with poor prognosis.

[0024] The technical scheme has the beneficial effects that:

[0025] The existing method simply classifies lung nodules into two categories of benign and malignant, or classifies lung nodules into two categories of solid and ground glass nodules, and lacks comprehensive and detailed classification. Compared with the existing lung nodule classification method, the method finally classifies lung nodules into 6 categories: ground glass benign nodule, ground glass malignant ground glass component dominant nodule, ground glass malignant solid component dominant nodule with good prognosis, ground glass malignant solid component dominant nodule with poor prognosis, solid benign nodule, and solid malignant nodule, so that the classification result is more comprehensive and detailed.

[0026] The method adopts one or more of image rotation, image mirroring, and image brightening to perform data expansion on the sub-sample data set, increases the number of samples in the sub-sample data set, and improves the classification accuracy of model training.

[0027] The method of the present application adds a cascaded convolutional neural network on the basis of a traditional deep learning network, solves the problem of being unable to accurately classify at one time due to the multiple types of data sets and similar features through the cascaded network, reduces the labels of the data set through the cascaded network, improves the classification accuracy, greatly reduces the current manual screening and classification operation of CT images, improves the classification accuracy, and provides a fast, accurate and reliable decision-making information basis for improving the classification efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A flowchart of the lung nodule classification method based on the cascaded convolutional neural network model in the embodiment;

[0029] Figure 2 A schematic diagram of the pre-processing of the CT image data in the embodiment;

[0030] Figure 3 A schematic diagram of the cascaded network in the embodiment;

[0031] Figure 4 A flowchart of the classification of the test set sample data by the cascaded network in the embodiment. DETAILED DESCRIPTION

[0032] In order to facilitate the understanding of the present application, the specific embodiments of the present application are described in further detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but are not used to limit the scope of the present application. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0033] For the lung nodule classification problem, the existing researches are simple lung nodule benign and malignant classification or solid and ground glass nodule classification, and lack comprehensive and detailed classification and evaluation for prognosis. According to the early investigation results, ordinary doctors classify the types of lung nodules according to experience, and the accuracy is generally 60%+ to 70%+. Even for doctors with rich experience, the classification accuracy is only about 85%. According to the early research, compared with continuously adjusting the parameters of the deep learning neural network model and small-scale modifying the network model, self-adaptive pre-processing of the image data itself can bring more significant improvement effect.

[0034] The embodiment proposes a lung nodule classification method based on a cascade convolutional neural network model, and focuses on the preprocessing of sample data, data augmentation of sub-sample data sets, and construction of a cascade network, while paying attention to the use of a pre-trained model. The core idea of the application includes: starting from the actual needs of a hospital, proposing a lung nodule cascade classification, i.e., first classifying lung nodules into solid nodules and ground glass nodules, then classifying the two kinds of nodules into benign and malignant nodules respectively, and then classifying ground glass malignant nodules into ground glass component dominant nodules and solid component dominant nodules, wherein the solid component dominant nodules are further classified into good prognosis nodules and poor prognosis nodules; the acquired lung CT image data is screened, the selected image data is appropriately selected in a targeted manner, and the approximate contour points are drawn and superimposed on the cropped image to complete the cropping of the CT image, thereby generating a sample data set; the sample data set is divided into a plurality of sub-sample data sets, and part of the sub-sample data sets are augmented, and the augmented sub-sample data sets are used for binary classification model training, and a cascade network is built according to the principle of lung nodule classification by using the trained binary classification model, thereby realizing the hierarchical classification of lung nodules.

[0035] It should be noted that the classification result of the lung nodule in the method of the application is only used to assist in predicting the nodule type of the lung nodule patient, because the characteristics of some patients' nodules are inconsistent with the actual type, and other means need to be used to comprehensively judge the nodule type of the patient. The direct purpose of the method of the application is to obtain a preliminary prediction result to assist doctors in diagnosis, not the diagnosis result.

[0036] The lung nodule classification method based on a cascade convolutional neural network model of the embodiment includes the following steps, as shown in Figure 1

[0037] Step 1: acquiring lung CT images, preprocessing the acquired CT images, and generating a sample data set according to the preprocessed images;

[0038] Further, the preprocessing method is as follows: the CT images with lung nodules are selected from the acquired lung CT images, the positions of the nodules in the selected CT images are marked, and the current CT image is cropped according to the positions of the nodules; the length and width of the nodules in the CT image are calculated according to the marked lung nodule positions, and the nodules are adjusted to the middle position of the image according to the calculated results to realize the centered display of the nodules;

[0039] In the embodiment, a plurality of lung CT images provided by a cooperative hospital are acquired. Since one CT contains several hundred images, but not every image has a nodule, the images with nodules need to be selected first. As shown in Figure 2 ​As shown, due to the less features of the nodule in the image, it is difficult to accurately extract the features of the nodule in the process of feature extraction, so the lung nodule in the image is marked according to the marking information of the lung nodule contained in the CT image, in order to better improve the effect of feature extraction, the image needs to be cropped, and only a small part of the area around the lung nodule is reserved according to the size of the lung nodule, and the length and width of the nodule also need to be calculated, and finally the nodule is centered to the middle position of the image, wherein the size of the CT image is fixed 64px*64px, when the lung nodule is located at the center position of the image, the remaining area is the small part of the reserved area, which is 10-60px.

[0040] Step 2: label the sample data in the sample data set, and divide the sample data set into a plurality of sub-sample data sets according to the label, compare the difference of the sample quantity in any two sub-sample data sets with the set threshold value, and according to the comparison result, the sub-sample data set with less sample quantity is data augmented, and each sub-sample data set is divided into a training set, a validation set and a test set;

[0041] Step 2.1: label the sample data in the sample data set;

[0042] The label includes: ground glass nodule, solid nodule, ground glass benign nodule, ground glass malignant nodule, solid benign nodule, solid malignant nodule, ground glass malignant ground glass component dominant nodule, ground glass malignant solid component dominant nodule, ground glass malignant solid component dominant good prognosis nodule, ground glass malignant solid component dominant poor prognosis nodule;

[0043] Step 2.2: divide the sample data set into a plurality of sub-sample data sets according to the label of the sample data, wherein each sub-sample data set has two labels;

[0044] The sub-sample data set is 5, including: data set 1: ground glass nodule, solid nodule; data set 2: ground glass benign nodule, ground glass malignant nodule; data set 3: solid benign nodule, solid malignant nodule; data set 4: ground glass malignant ground glass component dominant nodule, ground glass malignant solid component dominant nodule; data set 5: ground glass malignant solid component dominant good prognosis nodule, ground glass malignant solid component dominant poor prognosis nodule;

[0045] In the embodiment, the sample data in the sample data set is labeled, the CT images of the same category are put into the same label, the name of the label is named by the type of the image, and the types of the pulmonary nodules are divided into six categories: ground glass benign nodules, ground glass malignant nodules, ground glass malignant nodules with ground glass component as the main component, ground glass malignant nodules with solid component as the main component, solid benign nodules, and solid malignant nodules. In order to build a cascade network, the sample data set needs to be divided into five sub-sample data sets, each of which has two labels. Data set 1: ground glass nodules, solid nodules; data set 2: ground glass benign nodules, ground glass malignant nodules; data set 3: solid benign nodules, solid malignant nodules; data set 4: ground glass malignant nodules with ground glass component as the main component, ground glass malignant nodules with solid component as the main component; data set 5: ground glass malignant nodules with solid component as the main component with good prognosis, ground glass malignant nodules with solid component as the main component with poor prognosis.

[0046] Step 2.3: Compare the difference in the number of samples in any two sub-sample data sets with a set threshold value, and when the difference in the number of samples between any two sub-sample data sets reaches the set threshold value, perform data augmentation on the sub-sample data set with fewer samples.

[0047] The data augmentation method includes one or more of image rotation, image mirroring, and image brightening.

[0048] In the embodiment, since the CT images of the pulmonary nodules belong to the private data of the patient, in order to ensure that the privacy of the patient is not infringed, the consent of the patient needs to be obtained before the CT images are obtained, which will result in a small number of individual samples, such as a low incidence rate of ground glass nodules compared to solid nodules. Therefore, the number of sample data of ground glass nodules, especially ground glass malignant nodules, is much smaller than that of other types, which results in different numbers of images in the six types of nodules, thereby causing different numbers of sample data in different sub-sample data sets. When the difference in the number of samples between any two sub-sample data sets reaches the set threshold value, it will affect the accuracy of the final classification result, wherein the set threshold value is 25%-30%. According to the actual situation, it can be known that the number of sample data in data set 2, data set 4, and data set 5 in the sub-sample data set is less than that of other data sets. In order to balance the number of sample data under different labels, one or more of image rotation, image mirroring, and image brightening is used to perform data augmentation on the sub-sample data set. The data augmentation method in the embodiment is shown in Table 1.

[0049] Table 1 Data augmentation method of sub-sample data set

[0050]

[0051]

[0052] Taking Method 1 as an example, Table 1 is a data augmentation method that simultaneously rotates and mirrors the image before brightening it. The rotation angle is 120°, and the image is augmented by a factor of 3. The image is then mirrored horizontally and vertically, resulting in a augmentation factor of 2 times. Finally, the image brightness is increased by 0.5 times. Since image brightening does not change the number of images, the augmentation factor obtained by the dataset using Method 1 is (3+2), or 5 times. Taking Method 5 as another example, Table 1 is a data augmentation method that first rotates and then mirrors the image before brightening it. The rotation angle is 120°, and the image is augmented by a factor of 3. The image is then mirrored horizontally and vertically, resulting in a augmentation factor of 3 times. Finally, the image brightness is increased by 0.5 times. Since image brightening does not change the number of images, the augmentation factor obtained by the dataset using Method 5 is 3×3, or 9 times.

[0053] Since the number of data points contained in different subsample datasets varies, the augmentation methods used for data augmentation also differ. In this embodiment, different data augmentation methods were selected for datasets 2, 4, and 5 based on experimental results, as shown in Table 2. In the datasets shown in Table 2, dataset 2 includes sample data labeled as: benign ground-glass opacity (GGO), predominantly malignant GGO, predominantly malignant solid GGO with good prognosis, and predominantly malignant solid GGO with poor prognosis; dataset 4 includes sample data labeled as: predominantly malignant GGO, predominantly malignant GGO, predominantly malignant solid GGO with good prognosis, and predominantly malignant solid GGO with poor prognosis; dataset 5 includes sample data labeled as: predominantly malignant solid GGO with good prognosis and predominantly malignant solid GGO with poor prognosis.

[0054] Table 2 lists the data augmentation methods selected for datasets 2, 4, and 5.

[0055]

[0056]

[0057] Step 2.4: Divide each subsample dataset into a training set, a validation set, and a test set according to a set ratio;

[0058] In this embodiment, each subsample dataset is divided into three parts according to a certain ratio, wherein the ratio of training set, validation set and test set is set to 14:3:3.

[0059] Step 3: Construct a binary classification model based on the ResNet residual neural network. Input the training sets from multiple subsample datasets into the binary classification model for training to obtain multiple trained binary classification models. Then, use the validation sets from multiple subsample datasets to validate the trained binary classification models to obtain multiple final binary classification models.

[0060] Five trained binary classification models were established, denoted as Model 1, Model 2, Model 3, Model 4, and Model 5. Model 1 was used to classify ground-glass opacities (GGOs) and solid nodules from the input training set; Model 2 was used to classify benign and malignant GGOs from the input training set; Model 3 was used to classify benign and malignant solid nodules from the input training set; Model 4 was used to classify malignant GGOs with predominantly ground-glass opacities and malignant GGOs with predominantly solid components from the input training set; and Model 5 was used to classify malignant GGOs with predominantly solid components and good prognosis nodules from the input training set.

[0061] In this embodiment, ResNet (Residual Neural Network) is a deep convolutional neural network architecture widely used in image recognition and computer vision tasks. The design of ResNet mainly solves the gradient vanishing and gradient exploding problems in the training process of deep neural networks. The core idea of ​​ResNet is to introduce residual connections, which skip some layers by directly connecting across layers, making it easier for the network to learn residual mappings. This design allows the network to be deeper and more complex without causing performance degradation.

[0062] This implementation utilizes the existing ResNet-34 to construct a binary classification model, including: an input layer for receiving input images or feature maps; convolutional and pooling layers, i.e., a series of stacked convolutional and pooling layers for extracting image features; a residual block, the core component of ResNet, consisting of two or more convolutional layers, where each residual block has skip connections that directly add the input to the block's output to propagate gradients without loss of information; a global average pooling layer for average pooling the feature map of the last residual block, converting it into a fixed-length vector; and fully connected layers that connect the output of the global average pooling layer to one or more fully connected layers for final classification or regression prediction. The ResNet model structure can be adjusted according to the complexity and needs of the task. The existing ResNet-34 has 34 convolutional layers, including 18 residual blocks, each containing several convolutional layers.

[0063] In this embodiment, the training sets from the five subsample datasets are sequentially input into the binary classification model for model training. Although different training sets are used in these five training sessions, the training parameters are exactly the same. In this embodiment, the epoch is set to 100, the batch size to 16, and the learning rate to 0.0001 for model training. After training, five different trained binary classification models are obtained. Then, the validation sets from the corresponding five subsample datasets are input into the five trained binary classification models to calculate the accuracy of the binary classification model. When the accuracy of each binary classification model reaches more than 75%, the construction of the binary classification model is completed, and the final five binary classification models are obtained, which are denoted as Model 1, Model 2, Model 3, Model 4, and Model 5, respectively.

[0064] Step 4: Construct a cascaded network using the multiple binary classification models obtained in Step 3. Input the test sets from the multiple subsample datasets into the constructed cascaded network for classification to obtain the classification results.

[0065] The method for constructing the cascaded network is as follows: the cascaded network is divided into 4 levels. The first level uses Model 1 to distinguish whether the input test set data is ground-glass nodules or solid nodules. The second level uses Model 2 to distinguish whether ground-glass nodules are benign or malignant, and Model 3 to distinguish whether solid nodules are benign or malignant. The third level uses Model 4 to distinguish whether ground-glass malignant nodules are predominantly ground-glass malignant nodules or predominantly solid nodules. The fourth level uses Model 5 to distinguish whether ground-glass malignant nodules are predominantly solid nodules with good prognosis or predominantly solid nodules with poor prognosis.

[0066] In this embodiment, such as Figure 3 As shown, the cascaded network consists of four levels. The test sets from all subsample datasets are input into the cascaded network. First, Model 1 is used to distinguish whether the input data is a ground-glass nodule or a solid nodule. If it is a ground-glass nodule, Model 2 is used to distinguish whether the nodule is a benign or malignant ground-glass nodule. If it is a malignant ground-glass nodule, Model 4 is used to distinguish whether the nodule is predominantly ground-glass or predominantly solid. If it is predominantly solid, Model 5 is used to distinguish whether the nodule has a good prognosis or a poor prognosis. If it is a solid nodule, Model 3 is used to distinguish whether the nodule is a benign or malignant solid nodule.

[0067] In this embodiment, such as Figure 4 As shown, the cascaded network is used to classify the input test set. In the diagram, `result` represents the prediction result, `predict()` represents the model call interface, and M1, M2, M3, M4, and M5 represent five binary classification models, i.e., binary classifiers. The cascaded network first reads the test data, records the label of each sample, and then predicts each sample. First, `predict(M1)` is called to determine whether it is a ground-glass nodule or a solid nodule. If it is a ground-glass nodule, `predict(M2)` is called to output the prediction result as either a benign or malignant ground-glass nodule. If it is a malignant ground-glass nodule, `predict(M4)` is called to output the prediction result as either a nodule predominantly composed of ground-glass components or a nodule predominantly composed of solid components. If it is a nodule predominantly composed of solid components, `predict(M5)` is called to output the prediction result as either a nodule with good or bad prognosis. If it is a solid nodule, `predict(M3)` is called to determine whether it is a benign or malignant solid nodule.

[0068] The test set from the subsample dataset is input into the cascaded network to obtain the classification result. The classification result is compared with the label information corresponding to the image data in the test set, and the accuracy of the cascaded network is calculated. Similarly, the accuracy of the traditional classification network is calculated, and the accuracy of the traditional classification network is compared with that of the cascaded network, as shown in Table 3. The table shows the accuracy of five binary classification models (i.e., binary classifiers) and one six-classification model, the classifier type, the classification output results, and the dataset used for training. From the data in Table 3, it can be seen that the accuracy of the five binary classification models used by the cascaded network (also called the cascaded classifier) ​​is higher than that of the six-classification model (also called the multi-classifier) ​​used by the traditional classification network.

[0069] Table 3. Accuracy of Traditional Classification Networks vs. Accuracy of Cascaded Networks

[0070]

[0071] In this embodiment, Test Accuracy, Precision, Recall, and F1 Score are used as evaluation metrics. The accuracy of the cascaded network is tested based on the classification results, and the obtained evaluation metric results are compared with the evaluation metric results of traditional classification networks to verify the classification effect of the cascaded network.

[0072] Test accuracy: the percentage of correctly predicted samples out of all predicted samples, where a positive prediction is for positive samples and a negative prediction is for negative samples. The calculation formula is as follows:

[0073]

[0074] Among them, TP is a true positive, with both the actual value and the predicted value being positive, i.e., a true positive example; FP is a false positive, with both the actual value and the predicted value being positive, i.e., a false positive example; TN is a true negative, with both the actual value and the predicted value being negative, i.e., a true negative example; and FN is a false positive, with both the actual value and the predicted value being positive, i.e., a false negative example.

[0075] Precision: the number of correctly predicted positive samples divided by the total number of predicted positive samples (for the predicted positive samples);

[0076]

[0077] Recall is the number of correctly predicted positive samples divided by the total number of positive samples (for positive samples only).

[0078]

[0079] The F1 score, the harmonic mean of the model's precision and recall, is a statistical metric used to measure the accuracy of a binary classification model.

[0080]

[0081] In this embodiment, the test results of the cascaded network and the traditional classification network are compared, as shown in Table 4.

[0082] Table 4. Test results of multi-classifiers from traditional classification networks and cascade classifiers from cascaded networks.

[0083]

[0084] Comparative testing accuracy revealed that the traditional multi-classification model achieved an accuracy of 72.96%, while the cascaded network achieved an accuracy of 80.04%. Calculations of Precision, Recall, and F1 score showed that the cascaded network outperformed the traditional multi-classification model in all evaluation metrics. Therefore, the cascaded network studied in this invention demonstrates superior classification performance compared to traditional multi-classification models, exhibiting significant advantages. This proves the feasibility of the cascaded network and its ability to meet clinical needs. It can assist in classifying risk factors such as pulmonary nodules, enabling early risk prevention, regular patient follow-up and observation, and improving the efficiency of pulmonary nodule classification.

[0085] The method used in this embodiment has low time complexity. The hardware and software environment used is as follows: hardware environment includes CPU (Intel(R) Core(TM) i7-10700 CPU @ 2.90GHz), GPU (NVIDIA GeForce RTX2060Ti), 16.00GB of memory, and 1TB of hard disk space. Software environment includes Windows 10 operating system and Python programming language. Under the above environment, this method takes an average of 1 to 3 minutes, which meets the preoperative planning requirements.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A lung nodule classification method based on a cascaded convolutional neural network model, characterized in that, The method includes the following steps: Step 1: Acquire lung CT images, preprocess the acquired CT images, and generate a sample dataset based on the preprocessed images; Step 2: Label the sample data in the sample dataset, and divide the sample dataset into multiple sub-sample datasets according to the labels. Compare the difference in the number of samples in any two sub-sample datasets with a set threshold, and augment the sub-sample dataset with the smaller number of samples according to the comparison results. Then divide each sub-sample dataset into a training set, a validation set, and a test set. The labels include: ground-glass nodule, solid nodule, benign ground-glass nodule, malignant ground-glass nodule, benign solid nodule, malignant solid nodule, malignant ground-glass nodule with predominantly ground-glass component, malignant ground-glass nodule with predominantly solid component, malignant ground-glass nodule with predominantly solid component and good prognosis, and malignant ground-glass nodule with predominantly solid component and poor prognosis. Step 3: Construct a binary classification model based on the ResNet residual neural network. Input the training sets from multiple subsample datasets into the binary classification model for training to obtain multiple trained binary classification models. Then, use the validation sets from multiple subsample datasets to validate the trained binary classification models to obtain multiple final binary classification models. The trained binary classification models consist of 5 models, denoted as Model 1, Model 2, Model 3, Model 4, and Model 5. Step 4: Construct a cascaded network using the multiple binary classification models obtained in Step 3. Input the test sets from the multiple subsample datasets into the constructed cascaded network for classification to obtain the classification results. The method for constructing the cascaded network is as follows: the cascaded network is divided into 4 levels. The first level uses Model 1 to distinguish whether the input test set data is ground-glass nodules or solid nodules. The second level uses Model 2 to distinguish whether ground-glass nodules are benign or malignant, and Model 3 to distinguish whether solid nodules are benign or malignant. The third level uses Model 4 to distinguish whether ground-glass malignant nodules are predominantly ground-glass nodules or predominantly solid nodules. The fourth level uses Model 5 to distinguish whether ground-glass malignant nodules are predominantly solid nodules with good prognosis or predominantly solid nodules with poor prognosis.

2. The lung nodule classification method based on a cascaded convolutional neural network model according to claim 1, characterized in that, The preprocessing method is as follows: CT images containing lung nodules are selected from the acquired lung CT images; the positions of the nodules in the selected CT images are marked; the current CT image is cropped according to the positions of the nodules; the length and width of the nodules in the CT image are calculated according to the marked lung nodule positions; and the nodules are adjusted to the center position of the image according to the calculated results to achieve centered display of the nodules.

3. The lung nodule classification method based on a cascaded convolutional neural network model according to claim 1, characterized in that, Each of the multiple subsample datasets has two labels.

4. The lung nodule classification method based on a cascaded convolutional neural network model according to claim 3, characterized in that, The subsample datasets consist of five groups: Dataset 1: Ground-glass nodules and solid nodules; Dataset 2: Ground-glass benign nodules and ground-glass malignant nodules; Dataset 3: Solid benign nodules and solid malignant nodules; Dataset 4: Ground-glass malignant nodules with predominantly ground-glass components and ground-glass malignant nodules with predominantly solid components; Dataset 5: Ground-glass malignant nodules with predominantly solid components and good prognosis and ground-glass malignant nodules with predominantly solid components and poor prognosis.

5. The lung nodule classification method based on a cascaded convolutional neural network model according to claim 1, characterized in that, The data augmentation method includes: performing data augmentation using one or more of the following operations: image rotation, image mirroring, and image brightening.

Images