Prediction method of lymph node metastasis status of esophageal cancer based on multiple instance learning model

Abstract

The present invention relates to the field of imaging artificial intelligence technology, and discloses a method for predicting the state of esophageal cancer lymph node metastasis based on a multiple instance learning model, comprising the steps of: collecting CT images of esophageal cancer patients; in the CT images of esophageal cancer patients, marking all visible lymph nodes; for each marked lymph node, extracting its radiomics features and deep learning features; based on a multi-attention mechanism, constructing a multiple instance learning model MILRad through radiomics features, and constructing a multiple instance learning model MILDL through deep learning features; combining the prediction results of the multiple instance learning model MILRad and the multiple instance learning model MILDL by weighted averaging, to form a lymph node metastasis state prediction. The purpose of the present invention is to utilize radiomics features and deep learning features, introduce an attention mechanism, and construct a multiple instance learning model to predict the probability of esophageal cancer lymph node metastasis, thereby improving the prediction accuracy and interpretability.

Description

Technical Field

[0001] The present invention relates to the field of imaging artificial intelligence technology, and in particular to a method for predicting esophageal cancer lymph node metastasis status based on a multi-instance learning model. Background Art

[0002] Accurately predicting whether patients with esophageal cancer are at risk of lymph node metastasis is extremely important for the formulation of esophageal cancer treatment plans and prognosis assessment. Currently, clinicians mainly determine whether there is lymph node metastasis by measuring the short diameter size of lymph nodes on images of patients with esophageal cancer. However, the accuracy of this method is limited and it is difficult to meet the clinical needs for high-precision prediction. This method fails to fully utilize the rich information of all visible lymph nodes in CT images, limiting the accuracy and interpretability of the prediction. Summary of the invention

[0003] The purpose of the present invention is to utilize imaging genomics features and deep learning features, introduce an attention mechanism, and construct a multi-instance learning model to predict the probability of esophageal cancer lymph node metastasis, thereby providing an esophageal cancer lymph node metastasis prediction method based on a multi-instance learning model, which improves the prediction accuracy and interpretability.

[0004] In order to achieve the above-mentioned object of the invention, the embodiment of the present invention provides the following technical solutions:

[0005] The method for predicting the status of esophageal cancer lymph node metastasis based on a multi-instance learning model includes the following steps:

[0006] Step 1, image data collection and preprocessing: collecting CT images of patients with esophageal cancer and preprocessing the collected CT images;

[0007] Step 2, lymph node annotation: annotate all visible lymph nodes in the CT images of patients with esophageal cancer;

[0008] Step 3, feature extraction: for each labeled lymph node, extract its radiomics features and deep learning features;

[0009] Step 4: Construct a multi-instance learning model: Based on the multi-attention mechanism, construct a multi-instance learning model MILRad through radiomics features, and construct a multi-instance learning model MILDL through deep learning features;

[0010] Step 5, comprehensive prediction: The prediction results of the multi-instance learning model MILRad and the multi-instance learning model MILDL are combined by weighted averaging to form the final lymph node metastasis status prediction.

[0011] Compared with the prior art, the present invention has the following beneficial effects:

[0012] (1) The present invention fully combines radiomics features and deep learning features. Radiomics features include but are not limited to the morphological features, texture features, and edge features of lymph nodes, revealing the visualization and quantitative representation of lymph nodes; while deep learning features extract high-dimensional abstract representations of lymph node images through a pre-trained Resnet18 neural network model, capturing patterns that are difficult to intuitively identify. This dual feature extraction strategy greatly enhances the ability to capture lymph node image features.

[0013] (2) The present invention designs a multi-instance learning algorithm, which analyzes all visible lymph nodes on the patient's CT images as a whole (package) instead of limiting prediction to a single lesion, greatly improving the full utilization of the patient's imaging information. At the same time, by integrating the attention mechanism, the multi-instance learning algorithm can more easily focus on the lymph node instances that are critical to the prediction, thereby improving the accuracy and interpretability of the prediction.

[0014] (3) The present invention comprehensively considers the prediction results of the multi-instance learning model MILRad based on radiomics features and the multi-instance learning model MILDL based on deep learning features through a weighted average method, further improving the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 is a flow chart of the method of the present invention;

[0017] Figure 2 It is a schematic diagram of feature extraction of the present invention;

[0018] Figure 3 Schematic diagram of the multi-instance learning model structure of the present invention. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0020] It should be noted that similar numbers and letters represent similar items in the following figures, so once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. At the same time, in the description of the present invention, the terms "first", "second", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance, or implying any such actual relationship or order between these entities or operations. In addition, the terms "connected", "connected", etc. can be directly connected between elements, or indirectly connected via other elements.

[0021] Embodiment 1:

[0022] The present invention is achieved through the following technical solutions: Figure 1 As shown, a method for predicting the status of esophageal cancer lymph node metastasis based on a multi-instance learning model is proposed, which includes the following steps:

[0023] Step 1: Image data collection and preprocessing: Collect CT images of esophageal cancer patients and preprocess the collected CT images.

[0024] The inclusion criteria for CT imaging data of patients with esophageal cancer include: only patients with esophageal cancer confirmed by postoperative pathology; patients must have enhanced CT imaging data within 2 weeks before surgery; patients have not received any form of adjuvant treatment before surgery; patients must have thorough lymph node dissection during surgery, and the number of lymph nodes dissected is not less than 15; patients must have complete clinical pathological data. In addition, if the enhanced CT image is not good, it will not be included in the collection criteria.

[0025] The preprocessing of CT images includes but is not limited to: format conversion, voxel normalization, and privacy removal.

[0026] Step 2, lymph node annotation: Annotate all visible lymph nodes in the CT images of patients with esophageal cancer.

[0027] In this step, professional physicians mark all visible lymph nodes in the CT images of esophageal cancer patients in the mediastinal window. The edges of the lymph nodes are outlined during marking, and the marked lymph node area information is saved for subsequent feature extraction.

[0028] Step 3, feature extraction: For each labeled lymph node, extract its radiomics features and deep learning features.

[0029] For each lymph node annotated by a patient with esophageal cancer, the radiomics features are extracted using the Python package Pyradiomics. The extracted radiomics features include shape features, texture features, first-order statistical features, and high-order statistical features. For each lymph node, m radiomics features can be extracted. This example is only used as an example. Assuming that the jth patient has x lymph nodes, the radiomics feature matrix Rad of all lymph nodes extracted from the jth patient is j for:

[0030]

[0031] The above imaging genomics feature matrix Rad j In, Rad jik represents the kth radiomics feature of the i-th lymph node of the j-th patient; i=1,2,...,x, x represents the total number of lymph nodes of the j-th patient; k=1,2,...,m, m represents the total number of radiomics features of each lymph node.

[0032] For each lymph node annotated by the esophageal cancer patient, the largest layer of the lymph node and the adjacent upper and lower layers are combined into a three-channel color image, and then the deep learning features are extracted through the pre-trained Resnet18 neural network model. For each lymph node, n deep learning features can be extracted. The deep learning feature matrix DL of all lymph nodes of the jth patient is extracted. j for:

[0033]

[0034] The above deep learning feature matrix DL j In, DL jil represents the lth deep learning feature of the i-th lymph node of the j-th patient; i=1,2,...,x, x represents the total number of lymph nodes of the j-th patient; l=1,2,...,n, n represents the total number of deep learning features of each lymph node.

[0035] The image feature matrix of x lymph nodes of the jth patient is: Rad j =[Rad j1 ,...,Rad jx] and DL j =[DL j1 ,...,DL jx ], where Rad ji represents the radiomics features of the i-th lymph node of the j-th patient, DL ji represents the deep learning features of the i-th lymph node of the j-th patient. The radiomics features and deep learning features of the i-th lymph node are taken as a package vector i, such as Figure 2 shown.

[0036] The image feature matrix F of all lymph nodes of the jth patient j for:

[0037]

[0038] The image feature set of all lymph nodes of all patients, Patients, can be expressed as:

[0039]

[0040] Where j=1,2,...,p, and p represents the total number of patients.

[0041] Step 4: Build a multi-instance learning model: Based on the multi-attention mechanism, build a multi-instance learning model MILRad through radiomics features, and build a multi-instance learning model MILDL through deep learning features.

[0042] Taking the construction of the multi-instance learning model MILRad as an example, a detailed explanation is given. The multiple lymph nodes of an esophageal cancer patient are regarded as a set (package). The multi-instance learning model MILRad is trained based on the radiomics features of the lymph nodes. The network structure of the multi-instance learning model MILRad includes an embedding layer, a gated attention layer, a weighted sum layer, and a classification layer. The specific implementation process is as follows:

[0043] (1) Embedding layer: This layer is implemented using a sequence network that includes a fully connected layer, a ReLU activation function layer, and a Dropout layer. The input radiomics features are processed for dimensionality reduction and feature weighting to obtain a refined embedding feature representation of the lymph nodes. This embedding layer is designed to efficiently extract the core features of each instance (lymph node) and prepare a detailed feature representation foundation for the attention mechanism of the multi-instance learning model MILRad.

[0044] Assuming that each lymph node of a patient contains 512 radiomic features (k=512), the input radiomic feature vector is ,h i Equivalent to the radiomics feature matrix Rad jThe radiomics features of the i-th row in , that is, the i-th lymph node contains k radiomics features. The radiomics feature vector h of lymph node i i Dimensionality reduction is performed, first to 128 dimensions, and then further to 32 dimensions. At the same time, feature weighting is performed to obtain the embedded feature representation e of lymph node i i :

[0045]

[0046] e i is a 32-dimensional feature vector. and Both represent weight matrices; and All represent bias terms; ReLU represents the activation function operation; Dropout represents the Dropout layer operation.

[0047] (2) Gated attention layer: Through an attention mechanism network including the Tanh activation function, the 32-dimensional embedding feature representation e of lymph node i is transformed into i Simplified to 1 dimension, calculate the original attention score a for each instance i In addition, this layer also incorporates a gating mechanism, which generates a gating signal g corresponding to each instance through the Sigmoid activation function. i . The gating signal g of several lymph node instances of the patient i Compared with the original attention score a i Perform element-by-element multiplication and apply the Softmax function to the result to obtain the normalized gated attention weights The gated attention layer enhances the model’s ability to capture nonlinear relationships between instances and responds to different input conditions by dynamically adjusting the importance of each instance. This approach makes the assignment of instance weights more precise and adaptable, improving the model’s attention and processing capabilities for key information. This enables the multi-instance learning model MILRad to further identify and focus on lymph node instances that are critical to the prediction process.

[0048] Calculate the original attention score a of lymph node i i :

[0049]

[0050] in, and Both represent weight matrices; and Both represent bias terms; Tanh represents the activation function operation.

[0051] Generate a gating signal for each instance through the Sigmoid activation function :

[0052]

[0053] in, and Both represent weight matrices; and Both represent bias terms; Sigmoid represents the activation function operation.

[0054] The gating signal of each instance With the original attention score a i Perform multiplication and apply the Softmax function to the result to obtain the normalized gated attention weights :

[0055]

[0056] Among them, softmax represents the normalization operation.

[0057] (3) Weighted summation layer: The multi-instance learning model MILRad combines the normalized gated attention weights of several lymph node instances with the corresponding embedded feature representations and performs weighted summation to generate a global representation that integrates the information of all lymph node instances, focusing on those lymph node instances that have greater value for prediction.

[0058] Using normalized gated attention weights and embedding feature representation e i Perform weighted summation to obtain the global feature representation g:

[0059]

[0060] The above global feature representation g is also a 32-dimensional feature vector.

[0061] (4) Classification layer: In the final stage of the multi-instance learning model MILRad, a classification layer consisting of a fully connected layer and a ReLU activation function is used to make a final prediction based on the global feature representation g, outputting the probability y that the patient has lymph node metastasis:

[0062]

[0063] in, and Both represent weight matrices; and Both represent bias terms; ReLU represents the activation function operation.

[0064] The above is the network structure of the multi-instance learning model MILRad. The network structure of the multi-instance learning model MILDL is similar to it. The difference lies in the different inputs of the embedding layer, and the different weight matrices and bias items therein. They can be modified according to actual conditions, so they will not be repeated here.

[0065] In the training of the multi-instance learning model MILRad, back propagation is used to update the model parameters. In order to prevent the model from overfitting and perform feature selection at the same time, an L1 regularization function is added:

[0066]

[0067] Among them, L1 reg represents the L1 regularization function; W q represents the qth weight matrix; parameters represents all trainable parameters in the model, that is, weight matrix parameters; Represents the weight matrix W q The L1 norm of the weight matrix W q The absolute value of all parameters).

[0068] The total loss function of the multi-instance learning model MILRad is the binary cross entropy loss plus the L1 regularization function:

[0069]

[0070] Among them, y j Represents the output of the multi-instance learning model MILRad for the jth patient prediction; y` j represents the true label of the jth patient; represents the regularization coefficient.

[0071] Similarly, the loss function of the multi-instance learning model MILDL is the same, so it will not be repeated here.

[0072] Step 5, comprehensive prediction: The prediction results of the multi-instance learning model MILRad and the multi-instance learning model MILDL are combined by weighted averaging to form the final lymph node metastasis status prediction.

[0073] In this step, the weighted average method is used to fuse the prediction results of the multi-instance learning model MILRad and the multi-instance learning model MILDL. By analyzing the performance of the two models on the training set and the test set, the appropriate weight distribution is determined to ensure the accuracy and robustness of the final prediction.

[0074] The final predicted probability of lymph node metastasis status P is:

[0075]

[0076] Among them, MILRad represents the prediction result of the multiple instance learning model MILRad; MILDL represents the prediction result of the multiple instance learning model MILDL; Represents the weight of the multi-instance learning model MILRad; Represents the weight of the multi-instance learning model MILDL.

[0077] In summary, the present invention extracts the imaging features and deep learning features of all visible lymph nodes on the patient's CT images, designs a multi-instance learning algorithm to comprehensively analyze and process the lymph node image information, and thus realizes accurate prediction of the lymph node metastasis status of esophageal cancer patients.

[0078] As another feasible approach, this scheme weighs the relative importance of radiomics features and deep learning features by analyzing the similarities between multiple lymph nodes of a patient, thereby assigning weights to the multi-instance learning model MILRad and the multi-instance learning model MILDL.

[0079] Specifically, based on the radiomics feature matrix Rad j , the kth radiomics feature vector of all lymph nodes of patient j is expressed as , the lth deep learning feature vector of all lymph nodes of patient j is expressed as ; Create a correlation vector , c q Indicates the qth correlation, there are Q correlations in total, which are used to characterize the correlation between the same radiomic features of any two lymph nodes. Calculate the similarity of the same radiomic features of the uth lymph node and the vth lymph node:

[0080]

[0081] in, It represents the similarity of any same radiomics feature (such as the k-th radiomics feature) of the u-th lymph node and the v-th lymph node under the q-th association relationship; is the vector representation of the kth radiomics feature of lymph node u; is the vector representation of the kth radiomics feature of lymph node v; Indicates c q ; u=1,2,...,x; v=1,2,...,x; x is the total number of lymph nodes; sig represents the sigmoid function.

[0082] The cumulative similarity of radiomics features is:

[0083]

[0084] in, Indicates the cumulative similarity of radiomics features. When calculating the cumulative similarity, the similarity when u=v is not calculated.

[0085] Based on the same association vector C, calculate the similarity of the same deep learning features of the u-th lymph node and the v-th lymph node:

[0086]

[0087] in, Represents the similarity of any same deep learning feature (such as the lth deep learning feature) of the uth lymph node and the vth lymph node under the qth association relationship; is the vector representation of the lth deep learning feature of lymph node u; is the vector representation of the lth deep learning feature of lymph node v.

[0088] The cumulative similarity of deep learning features is:

[0089]

[0090] in, Indicates the cumulative similarity of deep learning features. When calculating the cumulative similarity, the similarity when u=v is not calculated.

[0091] Then, if , it means that for all lymph nodes of the patient, the extracted radiomics features are more important than the deep learning features, because different lymph nodes have a higher degree of correlation under the same radiomics features, so it is necessary to assign a higher weight to the multi-instance learning model MILRad, that is . For example, to calculate the importance weight:

[0092]

[0093] like , then , .

[0094] On the contrary, if , it means that for all lymph nodes of this patient, the extracted deep learning features are more important than the imaging features. .

[0095] Embodiment 2:

[0096] This embodiment provides experimental data based on embodiment 1 and proposes a method for predicting the status of esophageal cancer lymph node metastasis based on a multi-instance learning model, including the following steps:

[0097] Step 1: Image data collection and preprocessing.

[0098] Inclusion criteria include: limited to patients with esophageal cancer confirmed by postoperative pathology; patients must have enhanced CT imaging data within 2 weeks before surgery; patients have not received any form of adjuvant therapy before surgery; ensure that the patient has undergone thorough lymph node dissection during surgery, and the number of lymph nodes dissected is not less than 15; patients must have complete clinical pathological data. In addition, if the quality of enhanced CT images is poor, they will be excluded. The collected original images are preprocessed by format conversion, voxel normalization, and privacy removal.

[0099] Step 2: Lymph node labeling.

[0100] like Figure 2 As shown in the figure, professional physicians marked all visible lymph nodes in the CT images of patients with esophageal cancer in the mediastinal window (the window width was set to 350 HU and the window level was set to 40 HU). When marking, the lymph nodes were outlined along the edges and the marked lymph node regional information was saved.

[0101] Step 3: feature extraction.

[0102] like Figure 2 As shown in the figure, the radiomic features of each lymph node of the patient were extracted by the Python package Pyradiomics, including shape features, texture features, wavelet features and other aspects of radiomic features, totaling 1379 radiomic features. To ensure the effectiveness and consistency of radiomic features in the model, all extracted radiomic features were normalized so that the feature values ​​were in the numerical range of 0 to 1. The deep learning features of each lymph node of the patient were extracted by combining the largest layer of the lymph node and the adjacent upper and lower layers into three-channel images, and processing these three-channel images using the pre-trained Resnet18 neural network model to extract deep learning features. In particular, deep learning features were extracted from the last pooling layer of the Resnet18 neural network model, and a total of 512 deep learning features were obtained.

[0103] Step 4: Build a multi-instance learning model.

[0104] like Figure 3 As shown in the figure, based on the multi-attention mechanism, the multi-instance learning model MILRad is constructed through the radiomics features, and the multi-instance learning model MILDL is constructed through the deep learning features. The multiple lymph nodes of an esophageal cancer patient are regarded as a set (package), and the multi-instance learning model MILRad is trained based on the radiomics features of the lymph nodes. The network structure of the multi-instance learning model MILRad includes an embedding layer, a gated attention layer, a weighted sum layer, and a classification layer. Finally, the classification layer outputs the probability of lymph node metastasis in the patient, as shown in Figure 3As shown in the figure, pN- indicates that there is no lymph node metastasis in the postoperative pathology, and pN+ indicates that there is lymph node metastasis. The network structure of the multi-instance learning model MILDL is similar to it, except that the input of the embedding layer, the weight matrix and the bias terms are different, which can be modified according to actual conditions, so it will not be repeated here.

[0105] Step 5: Comprehensive prediction.

[0106] The prediction results of the multi-instance learning model MILRad based on radiomics features and the multi-instance learning model MILDL based on deep learning features are combined by weighted averaging to form the final prediction probability of lymph node metastasis status. By assigning a weight of 0.78 to the MILDL model result and 0.22 to the MILRad model result, the prediction performance can be optimized, that is, the final prediction probability P is:

[0107]

[0108] Among them, MILRad represents the prediction result of the multiple instance learning model MILRad; MILDL represents the prediction result of the multiple instance learning model MILDL.

[0109] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A method for predicting esophageal cancer lymph node metastasis status based on a multi-instance learning model, characterized by: The following steps are involved: Step 1, image data collection and preprocessing: collecting CT images of patients with esophageal cancer and preprocessing the collected CT images; Step 2, lymph node annotation: annotate all visible lymph nodes in the CT images of patients with esophageal cancer; Step 3, feature extraction: for each labeled lymph node, extract its radiomics features and deep learning features; In step 3, for each labeled lymph node, the step of extracting its radiomics features includes: For each lymph node of an esophageal cancer patient, m radiomic features are extracted. Assuming that the jth patient has x lymph nodes, the radiomic feature matrix Rad of all lymph nodes of the jth patient is extracted. j for: The above imaging genomics feature matrix Rad j In, Rad jik represents the kth radiomics feature of the i-th lymph node of the j-th patient; i=1,2,...,x, x represents the total number of lymph nodes of the j-th patient; k=1,2,...,m, m represents the total number of radiomics features of each lymph node; In step 3, for each labeled lymph node, the step of extracting its deep learning features includes: For each lymph node annotated by the esophageal cancer patient, n deep learning features are extracted. The deep learning feature matrix DL of all lymph nodes of the jth patient is extracted. j for: The above deep learning feature matrix DL j In, DL jil represents the lth deep learning feature of the i-th lymph node of the j-th patient; i=1,2,...,x, x represents the total number of lymph nodes of the j-th patient; l=1,2,...,n, n represents the total number of deep learning features of each lymph node; Step 4: Construct a multi-instance learning model: Based on the multi-attention mechanism, construct a multi-instance learning model MILRad through radiomics features, and construct a multi-instance learning model MILDL through deep learning features; Step 5, comprehensive prediction: The prediction results of the multi-instance learning model MILRad and the multi-instance learning model MILDL are combined by weighted average to form the final lymph node metastasis status prediction; The step 5 specifically comprises the following steps: Where P represents the predicted probability of lymph node metastasis status; MILRad represents the prediction result of the multi-instance learning model MILRad; MILDL represents the prediction result of the multi-instance learning model MILDL; Represents the weight of the multi-instance learning model MILRad; Represents the weight of the multi-instance learning model MILDL; Based on the radiomics feature matrix Rad j , the kth radiomics feature vector of all lymph nodes of patient j is expressed as , the lth deep learning feature vector of all lymph nodes of patient j is expressed as ; Create a correlation vector , c q Indicates the qth correlation relationship, there are Q correlation relationships in total, which is used to characterize the correlation relationship between the same radiomic features of any two lymph nodes; calculate the similarity of the same radiomic features of the uth lymph node and the vth lymph node: in, represents the similarity of any same radiomics feature of the u-th lymph node and the v-th lymph node under the q-th association relationship; is the vector representation of the kth radiomics feature of lymph node u; is the vector representation of the kth radiomics feature of lymph node v; Indicates c q Matrix transpose; u=1,2,...,x; v=1,2,...,x; x is the total number of lymph nodes; sig represents the sigmoid function; The cumulative similarity of radiomics features is: in, represents the cumulative similarity of radiomics features; Based on the same association vector C, calculate the similarity of the same deep learning features of the u-th lymph node and the v-th lymph node: in, Represents the similarity of any same deep learning feature of the u-th lymph node and the v-th lymph node under the q-th association relationship; is the vector representation of the lth deep learning feature of lymph node u; is the vector representation of the lth deep learning feature of lymph node v; The cumulative similarity of deep learning features is: in, Represents the cumulative similarity of deep learning features; Then, if , then On the contrary, if , then ; Calculate importance weights: like , then , .

2. The method for predicting esophageal cancer lymph node metastasis status based on a multi-instance learning model according to claim 1, characterized in that: In step 4, the step of constructing a multi-instance learning model MILRad based on the radiomics features based on the multi-attention mechanism includes: All lymph nodes of an esophageal cancer patient are considered as a set, and the multi-instance learning model MILRad is trained based on the radiomics features of the lymph nodes. The network structure of the multi-instance learning model MILRad includes an embedding layer, a gated attention layer, a weighted sum layer, and a classification layer; the imaging genomics features of all lymph nodes of an esophageal cancer patient are input into the embedding layer for dimensionality reduction and feature weighting, thereby obtaining an embedded feature representation of the lymph nodes; the gated attention layer reduces the embedded feature representation to one dimension, calculates the original attention score of each lymph node, and generates a gating signal corresponding to the original attention score, and then multiplies the original attention score with the corresponding gating signal to obtain a gated attention weight; the weighted sum layer performs a weighted summation of the gated attention weights and the embedded feature representation of all lymph nodes to obtain a global feature representation; the classification layer makes a final prediction based on the global feature representation, and outputs the probability of lymph node metastasis in the patient.

3. The method for predicting esophageal cancer lymph node metastasis status based on a multi-instance learning model according to claim 2, characterized in that: The step of inputting the radiomics features of all lymph nodes of an esophageal cancer patient into the embedding layer for dimensionality reduction processing and feature weighting, and then obtaining the embedded feature representation of the lymph nodes, is specifically as follows: Among them, e i represents the embedded feature representation of the i-th lymph node; h i Represents the radiomics feature matrix Rad j The radiomics features of the i-th row in ; and Both represent weight matrices; and All represent bias terms; ReLU represents the activation function operation; Dropout represents the Dropout layer operation.

4. The method for predicting esophageal cancer lymph node metastasis status based on a multi-instance learning model according to claim 3, characterized in that: The gated attention layer reduces the embedded feature representation to 1 dimension, calculates the original attention score of each lymph node, generates a gating signal corresponding to the original attention score, and then multiplies the original attention score with the corresponding gating signal to obtain the gated attention weight. Specifically, the steps are: Calculate the original attention score a of lymph node i i : in, and Both represent weight matrices; and Both represent bias terms; Tanh represents the activation function operation; The gating signal of lymph node i is generated by Sigmoid activation function : in, and Both represent weight matrices; and All represent bias terms; Sigmoid represents the activation function operation; The gated signal of lymph node i And the corresponding original attention score a i Perform multiplication and apply the Softmax function to the result to obtain the normalized gated attention weights : Among them, softmax represents the normalization operation; x represents the total number of lymph nodes; exp represents the natural exponent.

5. The method for predicting esophageal cancer lymph node metastasis status based on a multiple instance learning model according to claim 4, characterized in that: The weighted sum layer performs weighted summation of the gated attention weights and embedded feature representations of all lymph nodes to obtain the steps of global feature representation, specifically: Among them, g is the global feature representation.

6. The method for predicting esophageal cancer lymph node metastasis status based on a multiple instance learning model according to claim 5, characterized in that: The classification layer performs a final prediction based on the global feature representation and outputs the probability of lymph node metastasis in the patient, specifically: Among them, y represents the probability of lymph node metastasis in patients output by the multi-instance learning model MILRad; and Both represent weight matrices; and Both represent bias terms; ReLU represents the activation function operation.

Images