Thyroid cancer central lymph node classification method

Abstract

The invention relates to the technical field of image segmentation and image classification, in particular to a thyroid cancer central lymph node classification method, which comprises the following steps: S1, shooting an ultrasonic image of a thyroid cancer central lymph node; s2, establishing a thyroid nodule segmentation network; S3, extracting radiomics features by using the ultrasonic image of the central lymph node of the thyroid cancer and the segmentation mask, performing feature screening in the radiomics features by using a self-adjusting Lasso algorithm, and selecting an optimal feature set; and S4, performing deep learning feature extraction on a feature map after fusion of the thyroid cancer central lymph node ultrasound image and the region-of-interest image through a multi-source information fusion classification network, and performing feature fusion on the image omics feature optimal feature set, clinical features and deep learning features by using a maximum covariance fusion algorithm. According to the technical scheme, the accuracy of thyroid cancer central lymph node classification is effectively improved by combining radiomics features, clinical features and deep learning.

Description

Technical Field

[0001] The present invention relates to the technical field of image segmentation and image classification, and in particular to a method for classifying central lymph nodes of thyroid cancer. Background Art

[0002] According to cancer statistics in 2020, thyroid cancer ranks ninth in incidence among all cancers, with 586,000 new cases reported. The incidence rate in women is three times higher than in men. Although thyroid cancer is an indolent tumor, it is characterized by early metastasis, primarily to central and lateral cervical lymph nodes. Lymph node metastasis is an important indicator of surgical scope, approach, and prognosis. Central lymph node metastasis is considered a negative factor in predicting metastasis and recurrence. Furthermore, the presence or absence of central lymph node metastasis may affect surgical approach and postoperative staging. In China, the latest guidelines recommend routine prophylactic central lymph node dissection, at least on the ipsilateral side of the lesion. However, the necessity of routine prophylactic central lymph node dissection remains controversial. While routine prophylactic central lymph node dissection significantly improves prognosis, it increases the risk of postoperative complications and may also lead to widespread overtreatment. Therefore, noninvasive and accurate preoperative assessment of central lymph node metastasis is crucial, as it not only helps physicians formulate appropriate surgical plans but also guides the treatment process.

[0003] Ultrasound, as a noninvasive and reproducible imaging method, is considered the preferred method for evaluating thyroid tumors and cervical lymph node metastasis. However, the accuracy of preoperative ultrasound examinations depends to some extent on the radiologist's expertise and clinical experience. Furthermore, due to the anatomical characteristics of the central neck region, the accuracy and sensitivity of ultrasound in identifying central lymph node metastasis remain unsatisfactory. It has been reported that preoperative ultrasound examinations can only detect 20%-31% of central lymph node metastases and can only change the surgical approach in 20% of patients. Therefore, a more accurate, efficient, and non-invasive method is urgently needed to predict the risk of central lymph node metastasis in patients with thyroid cancer.

[0004] To address these issues, numerous studies have explored the status of central lymph node metastasis from various perspectives. Early studies used statistical methods to predict central lymph node metastasis based on clinical factors. Roh et al. used multivariate analysis to demonstrate that tumor size and extrathyroidal extension were two significant factors for central lymph node metastasis (p < 0.05). In recent years, radiomics has garnered significant attention in medical diagnosis, extracting intensity, boundary, texture, and wavelet features from ultrasound images for preoperative prediction of central lymph node metastasis. Park VY et al. generated 730 omics features from a single ultrasound image, including energy, entropy, kurtosis, features from the GLCM and GLRLM texture matrices, and features from a single-level discrete 2D wavelet transform. Using LASSO regression, they selected 14 features associated with central lymph node metastasis, achieving an AUC of 0.621 on the test set. In these studies, whether based on clinical factors or radiomics, the completeness of the extracted features was difficult to ensure, resulting in suboptimal prediction performance for central lymph node metastasis. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology. In response to the above problems, the present invention proposes a method for classifying central lymph nodes of thyroid cancer, which effectively improves the accuracy of central lymph node metastasis classification.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A method for classifying central lymph nodes of thyroid cancer comprises the following steps:

[0008] S1: Take ultrasound images of the central lymph nodes of thyroid cancer and classify the central lymph nodes of thyroid cancer under ultrasound images into those with metastasis or those without metastasis;

[0009] S2: establishing a thyroid nodule segmentation network, wherein the thyroid nodule segmentation network is used to obtain a segmentation mask of the thyroid nodule, and obtaining a region of interest image using an ultrasound image of a central lymph node of thyroid cancer and the segmentation mask;

[0010] S3: Extract radiomic features using ultrasound images and segmentation masks of central lymph nodes of thyroid cancer, and use the self-adjusting Lasso algorithm to screen features and select the optimal feature set among the radiomic features;

[0011] S4: Deep learning feature extraction was performed on the feature map of the fusion of the central lymph node ultrasound image and the region of interest image of thyroid cancer through a multi-source information fusion classification network. The maximum covariance fusion algorithm was used to fuse the optimal feature set of imaging genomics features, clinical features and deep learning features, and neural network classification was used.

[0012] Furthermore, in S2, the thyroid nodule segmentation network includes:

[0013] A backbone feature extraction module, wherein the backbone feature extraction module uses a lightweight network MobileNetV3 to extract a feature map from an ultrasound image of a central lymph node of thyroid cancer;

[0014] A saliency feature extraction module, the saliency feature extraction module is used to capture and transmit global saliency information and local saliency information of the ultrasound image of the central lymph node of thyroid cancer;

[0015] The attention guidance module is used to suppress the interference of background on segmentation and capture the edge structure characteristics of thyroid nodules.

[0016] Furthermore, the salient feature extraction module adopts an encoder-decoder architecture, which includes an encoder and a decoder. The encoder-decoder architecture is used to set a skip connection in the middle of the framework to reduce the amount of computation. The encoder extracts local saliency information from the feature map extracted by the backbone feature extraction module, and the decoder adds global saliency information while retaining the local saliency information. Each layer of the encoder includes a convolution layer, a batch normalization layer, and a maximum pooling layer. Each layer of the decoder includes a convolution layer, a batch normalization layer, a ReLU activation function, and bilinear interpolation upsampling.

[0017] For a given feature map extracted by the backbone feature extraction module X∈F H×W×3 , the specific calculation formula for the output of each layer of encoder is as follows:

[0018]

[0019] Where C = 2048, which is halved by the maximum pooling layer; the convolution feature G n Generate deep features G through the encoder n D , where n = 1, 2, 3, 4, 5, then G n D is input to the decoder to generate G n U , where n = 2, 4, 5. , finally, G n D and G n U is simultaneously input to the decoder to obtain a mixture of global features and edge features G n M , where n=2,4.

[0020] Furthermore, the attention guidance module uses the improved ECA module, adds a global maximum pooling branch to the ECA module, multiplies the weight vector obtained by global average pooling with the input feature, and adds the result of multiplying the weight vector obtained by global maximum pooling with the input feature to obtain the final result. The specific calculation formula is as follows:

[0021]

[0022] Where X(a,b) represents the value of the element at (a,b) in the input feature map X, H and W represent the height and width of the feature map, and F A represents the weight vector of the global average pooling output, Max(·) represents the global maximum pooling operation, and F M Represents the weight vector of the global maximum pooling output, e represents the multiplication of the corresponding position elements, C1D n represents a one-dimensional convolution operation with a convolution kernel size of n, σ(·) represents the Sigmoid activation function operation, and A A Represents the result of multiplying the input feature map with the weight vector output by the global average pooling, A M Represents the result of multiplying the input feature map with the weight vector output by the global maximum pooling, A result Represents the final output result of the attention mechanism module.

[0023] Furthermore, in step S2, the steps of screening omics features by the self-adjusting Lasso algorithm are specifically as follows:

[0024] Step S21: Input the radiomics feature set to be screened where z i,j For the sample, L i is the set of category numbers, l is the number of categories, and N is the total number of samples; and the feature set is standardized;

[0025] Step S22: Initialize the range of ω, for example, 0<ω<1, select 10 candidate values ​​ω with equal distance from the defined range and input them into the Lasso algorithm to obtain the feature set selected for each ω. The specific calculation formula is as follows:

[0026]

[0027] Where arg α min represents the value of parameter α when the function reaches the minimum value, L i represents the category of the i-th sample, z i,j represents the jth feature of the i-th sample, N represents the number of samples, M represents the number of features, ω represents a constant, α j represents the regression coefficient.

[0028] Step S23: According to the selected feature set of each ω, each feature set is input into SVM, 10-fold cross validation is used to implement the SVM algorithm, and the AUC corresponding to each feature set is calculated. The candidate ω is found according to the calculated AUC. h and best ω b , the new ω range will be identified as [ω b -|ω b -ω h |,ω b +|ω b -ω h |], these iterations are repeated when there is any improvement in the performance and number of features of the new ω, and finally, the ω with the best performance and the least number of features is selected.

[0029] Furthermore, in step S4, the multi-source information fusion classification network includes:

[0030] A channel fusion module is used to fuse the ultrasound image of the central lymph node of thyroid cancer with the image of the region of interest by splicing them in the channel dimension before the convolutional neural network performs feature extraction, and use the fused feature map as the input feature map of the multi-source information fusion classification network;

[0031] The maximum covariance fusion module is used to use the maximum covariance fusion algorithm to fuse the optimal feature set of imaging genomics features, clinical features and deep learning features to obtain a stronger feature vector and use a neural network for classification.

[0032] Furthermore, the maximum covariance fusion algorithm calculates the two eigenvectors M1 and M2 by the maximum covariance, and the calculation steps are as follows:

[0033] Step A1: Assume that the dimensions of M1 and M2 are n×m and n×q, respectively, where n represents the number of samples, m represents the length of the deep learning feature vector, and q represents the length of the feature vector after the optimal set of omics features and clinical features are concatenated. M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are concatenated. To make the vectors of equal length, find the vector with the largest length and perform mean padding.

[0034] Step A2: Calculate the time series projection. The calculation formula is as follows:

[0035]

[0036] In the formula, λ1, λ2 represent the projection of M1, M2 on the time series, μ1 T and μ2 TRepresents the transpose of the feature vector. M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are spliced ​​together;

[0037] Step A3: By calculating the covariance between μ1 and μ2 Maximize the covariance between λ1 and λ2 To find the optimal solution λ1 and λ2, the calculation formula is as follows:

[0038]

[0039] Where Cov represents the covariance operation, represents the covariance between λ1 and λ2, represents the covariance between μ1 and μ2, n represents the number of samples, λ1, λ2 represent the projection of M1, M2 on the time series, μ1 T and μ2 T Represents the transpose of the feature vector, M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are spliced ​​together; the feature pair with the largest covariance is selected, and the corresponding features are saved in the final fusion vector.

[0040] Beneficial effects of the present invention:

[0041] Based on ultrasound images, this method introduces a salient feature extraction module and an attention guidance module to establish a pixel-by-pixel relationship model between thyroid nodules and background pixels, optimizing the ability to distinguish thyroid cancer boundaries. A self-regulating Lasso algorithm is used to screen and select the optimal feature set from imaging omics features. Maximum covariance calculation is performed on deep learning features, optimal omics features, and clinical features to establish a central lymph node classification model, effectively improving the accuracy of central lymph node classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the overall flow chart of the present invention;

[0043] Figure 2 Schematic diagram of the segmented network in the present invention;

[0044] Figure 3 Schematic diagram of the multi-source information fusion classification network in the present invention;

[0045] Figure 4 This is a visualization diagram of the segmentation network in the present invention;

[0046] Figure 5 This is the ROC curve diagram of the classification effect in the present invention. DETAILED DESCRIPTION

[0047] The present invention will be described in further detail below with reference to the accompanying drawings.

[0048] First of all, it needs to be explained that the English explanations mentioned in this invention are as follows:

[0049] abbreviation

[0050] Lasso: least absolute shrinkage and selection operator;

[0051] ECA: Effective Channel Attention;

[0052] SVM: support vector machine;

[0053] AUC: Area under the ROC curve.

[0054] Reference Figure 1-5 A method for classifying central lymph nodes of thyroid cancer comprises the following steps:

[0055] S1: Take ultrasound images of the central lymph nodes of thyroid cancer and classify the central lymph nodes of thyroid cancer under ultrasound images into those with metastasis or those without metastasis;

[0056] S2: Establish a thyroid nodule segmentation network, which is used to obtain the segmentation mask of the thyroid nodule and obtain the region of interest image using the ultrasound image of the central lymph node of thyroid cancer and the segmentation mask;

[0057] S3: Extract radiomic features using ultrasound images and segmentation masks of central lymph nodes of thyroid cancer, and use the self-adjusting Lasso algorithm to screen features and select the optimal feature set among the radiomic features;

[0058] S4: Deep learning feature extraction was performed on the feature map of the fusion of the central lymph node ultrasound image and the region of interest image of thyroid cancer through a multi-source information fusion classification network. The maximum covariance fusion algorithm was used to fuse the optimal feature set of imaging genomics features, clinical features and deep learning features, and neural network classification was used.

[0059] In S2, the thyroid nodule segmentation network includes:

[0060] The backbone feature extraction module uses the lightweight network MobileNetV3 to extract feature maps from the ultrasound image of the central lymph node of thyroid cancer;

[0061] A salient feature extraction module is used to capture and transmit global and local salient information of ultrasound images of central lymph nodes of thyroid cancer;

[0062] It is worth noting that global saliency information refers to the macroscopic semantic features extracted from the entire image. These features are usually used to identify the overall structure, context, position relationship, etc.

[0063] Local saliency information refers to the features of small areas or detailed areas in an image.

[0064] Attention Guidance Module,The attention guidance module is used to suppress the interference of background on,segmentation and capture the edge structure features of thyroid nodules.

[0065] In addition, it is important to explain that background areas usually lack clear texture and semantic features, and have low average activation values ​​and maximum values, so they account for a low proportion in GAP / GMP. The attention mechanism improves the accuracy of foreground segmentation by suppressing areas with low responses (such as the background) and enhancing the channel weights of the lesion area.

[0066] The saliency feature extraction module uses an encoder-decoder architecture, which includes an encoder and a decoder. The encoder-decoder architecture is used to set up skip connections in the middle of the framework to reduce the amount of computation. The encoder extracts local saliency information from the feature maps extracted by the backbone feature extraction module, and the decoder adds global saliency information while retaining local saliency information. Each encoder layer includes a convolutional layer, a batch normalization layer, and a maximum pooling layer. Each decoder layer includes a convolutional layer, a batch normalization layer, a ReLU activation function, and bilinear interpolation upsampling.

[0067] For a given feature map extracted by the backbone feature extraction module X∈F H×W×3 , the specific calculation formula for the output of each layer of encoder is as follows:

[0068]

[0069] Where C = 2048, which is halved by the maximum pooling layer; the convolution feature G n Generate deep features G through the encoder n D , where n = 1, 2, 3, 4, 5, then G n D is input to the decoder to generate G n U , where n = 2, 4, 5. , finally, G n D and G n U is simultaneously input to the decoder to obtain a mixture of global features and edge features G n M , where n=2,4.

[0070] The attention guidance module uses the improved ECA module and adds a global maximum pooling branch to the ECA module. It multiplies the weight vector obtained by global average pooling with the input feature, and adds the result of multiplying the weight vector obtained by global maximum pooling with the input feature to obtain the final result. The specific calculation formula is as follows:

[0071]

[0072] Where X(a,b) represents the value of the element at (a,b) in the input feature map X, H and W represent the height and width of the feature map, and F A represents the weight vector of the global average pooling output, Max(·) represents the global maximum pooling operation, and F M Represents the weight vector of the global maximum pooling output, e represents the multiplication of the corresponding position elements, C1D n represents a one-dimensional convolution operation with a convolution kernel size of n, σ(·) represents the Sigmoid activation function operation, and A A Represents the result of multiplying the input feature map with the weight vector output by the global average pooling, A M Represents the result of multiplying the input feature map with the weight vector output by the global maximum pooling, A result Represents the final output result of the attention mechanism module.

[0073] In step S2, the steps of the self-adjusting Lasso algorithm to screen the omics features are as follows:

[0074] Step S21: Input the radiomics feature set to be screened where z i,j For the sample, L i is the set of category numbers, l is the number of categories, and N is the total number of samples; and the feature set is standardized;

[0075] Step S22: Initialize the range of ω, for example, 0<ω<1, select 10 candidate values ​​ω with equal distance from the defined range and input them into the Lasso algorithm to obtain the feature set selected for each ω. The specific calculation formula is as follows:

[0076]

[0077] Where arg α min represents the value of parameter α when the function reaches the minimum value, L i represents the category of the i-th sample, z i,j represents the jth feature of the i-th sample, N represents the number of samples, M represents the number of features, ω represents a constant, α j represents the regression coefficient;

[0078] Step S23: According to the selected feature set of each ω, each feature set is input into SVM, 10-fold cross validation is used to implement the SVM algorithm, and the AUC corresponding to each feature set is calculated. The candidate ω is found according to the calculated AUC. h and best ω b , the new ω range will be identified as [ω b -|ω b -ω h |,ω b +|ω b -ω h |], these iterations are repeated when there is any improvement in the performance and number of features of the new ω, and finally, the ω with the best performance and the least number of features is selected.

[0079] In step S4, the multi-source information fusion classification network includes:

[0080] The channel fusion module is used to fuse the ultrasound image of the central lymph node of thyroid cancer with the image of the region of interest by splicing them in the channel dimension before the convolutional neural network performs feature extraction. The fused feature map is used as the input feature map of the multi-source information fusion classification network.

[0081] The maximum covariance fusion module is used to use the maximum covariance fusion algorithm to fuse the optimal feature set of imaging genomics features, clinical features and deep learning features to obtain a stronger feature vector and use neural networks for classification.

[0082] The maximum covariance fusion algorithm calculates the two eigenvectors M1 and M2 through the maximum covariance. The calculation steps are as follows:

[0083] Step A1: Assume that the dimensions of M1 and M2 are n×m and n×q, respectively, where n represents the number of samples, m represents the length of the deep learning feature vector, and q represents the length of the feature vector after the optimal set of omics features and clinical features are concatenated. M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are concatenated. To make the vectors of equal length, find the vector with the largest length and perform mean padding.

[0084] Step A2: Calculate the time series projection. The calculation formula is as follows:

[0085]

[0086] In the formula, λ1, λ2 represent the projection of M1, M2 on the time series, μ1 T and μ2 T Represents the transpose of the feature vector. M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are spliced ​​together;

[0087] Step A3: By calculating the covariance between μ1 and μ2 Maximize the covariance between λ1 and λ2 To find the optimal solution λ1 and λ2, the calculation formula is as follows:

[0088]

[0089] Where Cov represents the covariance operation, represents the covariance between λ1 and λ2, represents the covariance between μ1 and μ2, n represents the number of samples, λ1, λ2 represent the projection of M1, M2 on the time series, μ1 T and μ2 T Represents the transpose of the feature vector, M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are spliced ​​together; the feature pair with the largest covariance is selected, and the corresponding features are saved in the final fusion vector.

[0090]

[0091] Table 1. Comparison of prediction results

[0092] The present invention uses evaluation indicators such as AUC, accuracy, sensitivity, specificity, NPV, PPV, and F1-score to evaluate the effect of the model. To demonstrate the effectiveness of the present invention, the present invention compares the predictive performance of the present method with that of omics models, clinical models, and deep learning models. The results are shown in Table 1.

[0093] The embodiments of the present invention are described in detail above, but the contents described are only preferred embodiments of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for classifying central lymph nodes of thyroid cancer, characterized in that: The steps include: S1: Take ultrasound images of the central lymph nodes of thyroid cancer and classify the central lymph nodes of thyroid cancer under ultrasound images into those with metastasis or those without metastasis; S2: establishing a thyroid nodule segmentation network, wherein the thyroid nodule segmentation network is used to obtain a segmentation mask of the thyroid nodule, and obtaining a region of interest image using an ultrasound image of a central lymph node of thyroid cancer and the segmentation mask; S3: Extract radiomic features using ultrasound images and segmentation masks of central lymph nodes of thyroid cancer, and use the self-adjusting Lasso algorithm to screen features and select the optimal feature set among the radiomic features; S4: Deep learning feature extraction was performed on the feature map of the fusion of the central lymph node ultrasound image and the region of interest image of thyroid cancer through a multi-source information fusion classification network. The maximum covariance fusion algorithm was used to fuse the optimal feature set of imaging genomics features, clinical features and deep learning features, and neural network classification was used.

2. A method for classifying central lymph nodes of thyroid cancer according to claim 1, characterized in that: In S2, the thyroid nodule segmentation network includes: A backbone feature extraction module, wherein the backbone feature extraction module uses a lightweight network MobileNetV3 to extract a feature map from an ultrasound image of a central lymph node of thyroid cancer; A saliency feature extraction module, the saliency feature extraction module is used to capture and transmit global saliency information and local saliency information of the ultrasound image of the central lymph node of thyroid cancer; The attention guidance module is used to suppress the interference of background on segmentation and capture the edge structure characteristics of thyroid nodules.

3. A method for classifying central lymph nodes of thyroid cancer according to claim 2, characterized in that: The salient feature extraction module adopts an encoder-decoder architecture, which includes an encoder and a decoder. The encoder-decoder architecture is used to set a skip connection in the middle of the framework to reduce the amount of computation. The encoder extracts local saliency information from the feature map extracted by the backbone feature extraction module, and the decoder adds global saliency information while retaining the local saliency information. Each layer of the encoder includes a convolution layer, a batch normalization layer, and a maximum pooling layer. Each layer of the decoder includes a convolution layer, a batch normalization layer, a ReLU activation function, and bilinear interpolation upsampling. For a given feature map extracted by the backbone feature extraction module X∈F H×W×3 , the specific calculation formula for the output of each layer of encoder is as follows: Where C = 2048, which is halved by the maximum pooling layer; the convolution feature G n Generate deep features G through the encoder n D , where n = 1, 2, 3, 4, 5, then G n D is input to the decoder to generate G n U , where n = 2, 4, 5. , finally, G n D and G n U is simultaneously input to the decoder to obtain a mixture of global features and edge features G n M , where n=2,4.

4. A method for classifying central lymph nodes of thyroid cancer according to claim 3, characterized in that: The attention guidance module uses the improved ECA module, adds a global maximum pooling branch to the ECA module, multiplies the weight vector obtained by global average pooling with the input feature, and adds the result of multiplying the weight vector obtained by global maximum pooling with the input feature to obtain the final result. The specific calculation formula is as follows: Where X(a,b) represents the value of the element at (a,b) in the input feature map X, H and W represent the height and width of the feature map, and F A represents the weight vector of the global average pooling output, Max(·) represents the global maximum pooling operation, and F M Represents the weight vector of the global maximum pooling output, e represents the multiplication of the corresponding position elements, C1D n represents a one-dimensional convolution operation with a convolution kernel size of n, σ(·) represents the Sigmoid activation function operation, and A A Represents the result of multiplying the input feature map with the weight vector output by the global average pooling, A M Represents the result of multiplying the input feature map with the weight vector output by the global maximum pooling, A result Represents the final output result of the attention mechanism module.

5. The method for classifying central lymph nodes of thyroid cancer according to claim 1, wherein: In step S2, the steps of screening omics features by the self-adjusting Lasso algorithm are as follows: Step S21: Input the radiomics feature set to be screened where z i,j For the sample, L i is the set of category numbers, l is the number of categories, and N is the total number of samples; and the feature set is standardized; Step S22: Initialize the range of ω, for example, 0<ω<1, select 10 candidate values ​​ω with equal distance from the defined range and input them into the Lasso algorithm to obtain the feature set selected for each ω. The specific calculation formula is as follows: Where arg α min represents the value of parameter α when the function reaches the minimum value, L i represents the category of the i-th sample, z i,j represents the jth feature of the i-th sample, N represents the number of samples, M represents the number of features, ω represents a constant, α j represents the regression coefficient; Step S23: According to the selected feature set of each ω, each feature set is input into SVM, 10-fold cross validation is used to implement the SVM algorithm, and the AUC corresponding to each feature set is calculated. The candidate ω is found according to the calculated AUC. h and best ω b , the new ω range will be identified as [ω b -||ω b -ω h ||,ω b +||ω b -ω h ||], these iterations are repeated whenever there is any improvement in the performance and number of features of the new ω, and finally, the ω with the best performance and the least number of features is selected.

6. A method for classifying central lymph nodes of thyroid cancer according to claim 1, characterized in that: In step S4, the multi-source information fusion classification network includes: A channel fusion module is used to fuse the ultrasound image of the central lymph node of thyroid cancer with the image of the region of interest by splicing them in the channel dimension before the convolutional neural network performs feature extraction, and use the fused feature map as the input feature map of the multi-source information fusion classification network; The maximum covariance fusion module is used to use the maximum covariance fusion algorithm to fuse the optimal feature set of imaging genomics features, clinical features and deep learning features to obtain a stronger feature vector and use a neural network for classification.

7. A method for classifying central lymph nodes of thyroid cancer according to claim 6, characterized in that: The maximum covariance fusion algorithm calculates the two eigenvectors M1 and M2 through the maximum covariance. The calculation steps are as follows: Step A1: Assume that the dimensions of M1 and M2 are n×m and n×q, respectively, where n represents the number of samples, m represents the length of the deep learning feature vector, and q represents the length of the feature vector after the optimal set of omics features and clinical features are concatenated. M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are concatenated. To make the vectors of equal length, find the vector with the largest length and perform mean padding. Step A2: Calculate the time series projection. The calculation formula is as follows: In the formula, λ1, λ2 represent the projection of M1, M2 on the time series, μ1 T and μ2 T Represents the transpose of the feature vector. M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are spliced ​​together; Step A3: By calculating the covariance between μ1 and μ2 Maximize the covariance between λ1 and λ2 To find the optimal solution λ1 and λ2, the calculation formula is as follows: Where Cov represents the covariance operation, represents the covariance between λ1 and λ2, represents the covariance between μ1 and μ2, n represents the number of samples, λ1, λ2 represent the projection of M1, M2 on the time series, μ1 T and μ2 T Represents the transpose of the feature vector, M1 represents the deep learning feature vector, and M2 represents the feature vector after the optimal set of omics features and clinical features are spliced ​​together; the feature pair with the largest covariance is selected, and the corresponding features are saved in the final fusion vector.

Images