Thyroid cancer lateral neck lymph node metastasis prediction method based on multi-modal feature decoupling

Abstract

The application provides a thyroid cancer lateral neck lymph node metastasis prediction method based on multi-modal feature decoupling and fusion. The method comprises the following steps: collecting ultrasound images, text reports and clinical structured data of thyroid cancer patients, and preprocessing to improve the usability of the data; then, according to the distribution characteristics of different modal data, a model suitable for different data feature extraction is designed to obtain initial features of different modal data; further, by combining contrast learning and cross-modal feature reconstruction tasks, the multi-modal features are decoupled, the common representation and the specific representation of each modality are separated, so as to reduce information redundancy and noise interference; finally, based on the cross-attention mechanism, the common representation and the specific representation of each modality are fused, the deep correlation between multi-modal data is fully mined, and the fused representation is input into a classifier to realize the prediction of lateral neck lymph node metastasis.

Description

Technical Field

[0001] This invention relates to the field of digital medical technology, and more specifically to a method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling. Background Technology

[0002] Thyroid cancer, a highly prevalent malignant tumor worldwide, has a biological behavior closely related to cervical lymph node metastasis, which is a core risk factor for clinical prognosis. Cervical lymph node metastasis is divided into central and lateral cervical lymph node metastasis, with an overall incidence rate of 20%-50%. Among them, lateral cervical lymph node metastasis (LLNM) is more aggressive than central metastasis, increasing patient mortality by approximately 46%, and significantly increasing the risk of recurrence and shortening disease-free survival.

[0003] Preoperative lymph node biopsy is the standard method for clinically assessing the status of lateral cervical lymph node metastasis in patients with thyroid cancer. Currently, the most commonly used technique in clinical practice is ultrasound-guided fine-needle aspiration biopsy. Although it is a minimally invasive procedure and widely used, it suffers from problems such as target localization errors and missed suspicious positive lymph nodes. Some patients also exhibit characteristic skip metastasis, meaning that while central lymph node metastasis is negative, positive lymph node metastasis has already occurred in the lateral cervical region. This phenomenon further reduces the accuracy of this technique in assessing LLNM, making it difficult to meet the needs of precise clinical diagnosis. Therefore, there is an urgent need in clinical practice to develop efficient and accurate techniques for predicting LLNM, assisting clinicians in developing individualized optimal treatment plans, effectively reducing the risk of tumor progression caused by missed lymph node anatomy, and providing more targeted clinical support and comprehensive nursing interventions. Current prediction methods based on multimodal fusion mostly directly fuse heterogeneous features from multiple sources such as images, text, and clinical structured data. They do not fully consider the expression differences and information redundancy between different modalities. They may associate redundant features within a single modality with other modalities, thereby introducing noise and limiting the model's fusion efficiency and generalization ability. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide a method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling. This method extracts features from clinical structured data, ultrasound images, and corresponding text reports. Through multimodal feature decoupling, the original features of each modality are decomposed into common and specific representations. The effectiveness of feature decoupling is verified by combining a cross-modal feature reconstruction task, avoiding interference from redundant information between modalities. Furthermore, the common and specific representations are fused through a cross-attention mechanism to improve the accuracy of LLNM prediction in thyroid cancer.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling, characterized by comprising the following steps: Step 1: Collect ultrasound images, text reports, and clinical structured data from thyroid cancer patients and perform preprocessing. Step 2: Design a dedicated network architecture to extract initial features of different modal data based on their distribution characteristics; Step 3: Combine contrastive learning and cross-modal feature reconstruction tasks to achieve multimodal feature decoupling and obtain common and specific representations of different modalities; Step four: Based on cross-attention, the common representation and the modality-specific features are fused, and the fused representation is input into the classifier to obtain the prediction result.

[0006] As a further improvement to the present invention, the specific steps for acquiring ultrasound images, text reports, and clinical structured data of thyroid cancer patients and performing preprocessing in step one are as follows: Step 1: Local contrast of ultrasound images is enhanced by limiting contrast adaptive histogram equalization. This is achieved by dividing the image into multiple small regions and performing local histogram equalization on each region to adapt to the non-uniformity of the image at different depths and locations. At the same time, the contrast enhancement amplitude is limited. By cropping the histogram peaks and redistributing pixels, the tendency to over-enhance noise as "details" in uniform regions is suppressed, while key features such as microcalcifications and low echoes are preserved. Steps one and two: Simultaneously collect the formatted imaging diagnostic report text corresponding to the ultrasound images, describe multiple qualitative features of the nodule in standardized language, including but not limited to: location, shape, echo, boundary, calcification and blood flow signal, convert all English characters in the report text to a uniform lowercase form, and remove special characters, redundant punctuation marks, numbers and meaningless format codes from the report, while retaining the core descriptive text; Step 13 involves collecting easily quantifiable basic clinical information about the patient that is closely related to the diagnosis. This includes, but is not limited to, age, sex, BMI, duration of underlying thyroid disease, family history, and clinical laboratory indicators such as the seven thyroid function tests. Considering the dimensional differences between different data points, the data are standardized using a Z-score. Subsequently, statistical tests are used to screen for significant differences in whether patients have lateral cervical lymph node metastasis. For discrete characteristics such as sex and family history, a chi-square test is used; for continuous characteristics such as age and BMI, a t-test is used.

[0007] As a further improvement of the present invention, the specific steps in step two, which combine contrastive learning and cross-modal feature reconstruction tasks to achieve multimodal feature decoupling and obtain common and specific representations of different modalities, are as follows: Step 21: An architecture combining multi-scale convolutional neural networks and feature pyramid networks is adopted. Multiple layers of CNNs are used to progressively extract features at different levels of the image, generating multi-level feature maps. For a given level of feature map... Its multi-scale features Obtained through the following formula: in, Indicates the kernel size as The convolution operation, and Designed for extracting point-like, region-like, and global features respectively. For channel splicing operations; Step 22 introduces an FPN structure to achieve top-down feature enhancement. This involves upsampling deep, semantically strong but low-resolution feature maps and adding them element-wise with shallow, high-resolution feature maps, thereby passing high-level semantic information to the lower layers. This process can be represented as: in, For the enhanced feature map, This is an upsampling operation; Steps two and three: The fused multi-scale semantic features extracted from the FPN output layer are denoted as ultrasound image features. It simultaneously encodes the microscopic calcification features and macroscopic morphological features of the nodules.

[0008] As a further improvement of the present invention, the different modal data in step two include ultrasound text report data and clinical structured data; For ultrasound report data, which includes sonographic features of nodules and preliminary BI-RADS classification, containing rich qualitative diagnostic knowledge, a pre-trained BERT model is used as the feature extractor for a given ultrasound report text sequence. After word segmentation and embedding, the text is input into the BERT model. Through stacked Transformer encoder blocks, the context-dependent representation of each word is calculated. Finally, the hidden state corresponding to the CLS token is taken as the aggregate representation of the entire text sequence, and it is mapped to the feature space through an additional projection layer. in, This is the extracted text report feature vector, which condenses the semantic information of key diagnostic descriptors in the report and can be used for complementary verification with image features. For structured clinical data, high-dimensional nonlinear mapping is performed using deep neural networks. First, categorical variables are one-hot encoded and then concatenated with numerical variables to form the original input vector. Then, the vector is fed into the multilayer perceptron: in, and For trainable parameters, Using the activation function, this network is able to learn complex interaction effects between clinical variables and map low-dimensional inputs to high-dimensional feature representations. .

[0009] As a further improvement of the present invention, the specific steps in step three, which combine contrastive learning and cross-modal feature reconstruction tasks to achieve multimodal feature decoupling and obtain common and specific representations of different modalities, are as follows: Step 31: For the multimodal initial features obtained in Step 2 , , In order to extract common representations from data of different modalities With specific characterization A lightweight feature separation head is designed for each modality, specifically: Regarding image features: Regarding text features: For structured table features: in, It is a learnable weight matrix. As a bias term, the feature separation head enables the model to adaptively separate common representations and modality-specific representations from the same modality; Step 32 introduces a two-way contrastive learning mechanism, including contrastive learning of common representations and contrastive learning of specific representations; Step 33: Introduce a cross-modal feature reconstruction task as a self-supervised signal, using shared common representations. Specific characterization of a particular mode The original features of this modality are reconstructed, prompting the encoder to retain the fine structural information in the ultrasound image, the key semantic content in the text description, and the quantitative index patterns in the table, and to enhance the integrity and reconstructability of the decoupled features. The corresponding calculation method is as follows: in The reconstruction decoder, representing the corresponding modality, consists of fully connected layers, while , , The images, text, and tables represent the reconstructed features, respectively, and the corresponding reconstruction loss is calculated as follows: As a further improvement of the present invention, a bidirectional contrastive learning mechanism is introduced in step three-two, including the specific methods of common representation contrastive learning and specific representation contrastive learning as follows: Common representation contrastive learning specifically refers to: for N patients, each with K modal features, then there are common features... Common characteristics , where i represents the patient and k represents the modality. In the commonality representation contrastive learning, positive sample pairs represent common features of different modalities within the same patient, reflecting their cross-modal description of the same pathological state; negative sample pairs represent common features of different patients, used to enhance the model's ability to distinguish between different patients, and their corresponding contrastive loss function is: in, For cosine similarity, For temperature hyperparameters, This indicates different modalities under the same patient i. and The similarity of common representations between them, here, Traversal All modalities other than those mentioned above ensure that the common characteristics of all modalities of the same patient are brought closer together; This indicates different patient j, modality The similarity between common representations and anchor representations can be used to obtain more discriminative and modality-invariant common feature representations through comparative learning of common representations. Specific characterization contrastive learning specifically involves: positive sample pairs used for learning represent specific features of different patients within the same modality, while negative sample pairs represent specific features of different modalities. The corresponding contrastive loss function is: In summary, the total loss function of the final decoupling module is: in To balance the weights of the various loss terms and the weights of the bidirectional contrastive learning and reconstruction tasks, the model can maintain semantic alignment, modality discrimination, and information integrity during the decoupling process.

[0010] Finally, after decoupling, the aggregated representation of common characteristics across modalities is extracted as common features, specifically: .

[0011] As a further improvement of the present invention, the specific method for fusing common representations and modality-specific features based on cross-attention in step four, and inputting the fused representations into the classifier to obtain the prediction results, is as follows: The common representation of each modality is obtained through the multimodal feature decoupling module. And the specific characteristics of each mode: , , To facilitate the interaction between common features and modality-specific features, a cross-attention mechanism is used to separately address the following: Integrate with each specific feature; For ultrasonic modes, As a query As keys and values, their fused features are computed via scaled dot product attention. : in, , , For a learnable projection matrix, The dimension of the key vector. This process allows the model to focus on specific morphological patterns in ultrasound images most relevant to lymph node metastasis, such as microcalcifications or areas of abnormal blood flow, from a shared feature perspective; Performing cross-attention operations on the text modality and the table modality respectively yields: To integrate the information from the three fused features, learnable weights are used for weighted ensemble, resulting in the final joint feature representation used for classification: in, The learnable weight parameters allow the model to dynamically adjust the contribution of each modality to the current diagnostic task, thereby adapting to the differences in modal information quality in different cases; finally, the final joint representation is... The data is fed into a classifier, which passes through a fully connected layer to ultimately obtain the probability of lateral cervical lymph node metastasis. in, It is the predicted probability of lymph node metastasis. The higher the value, the higher the risk of metastasis. Pathological confirmation or intervention should be carried out as early as possible to help develop an individualized treatment plan.

[0012] The beneficial effects of this invention are as follows: By introducing a multimodal feature decoupling mechanism, this invention separates common representations from specific representations, solves the problems of redundancy and noise interference between modalities, and extracts purer and more discriminative fusion features; further, it adopts a cross-attention fusion mechanism to achieve the fusion of common and specific representations, and enhances the model's generalization ability by adaptively learning the weights of different modal fusion features, thereby achieving accurate prediction of preoperative LLNM and reducing the risk of cancer progression due to missed metastases. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to the present invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the embodiments shown in the accompanying drawings.

[0015] Example 1: As Figure 1 As shown, a method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling is described, and the method steps are as follows: Step 1: Collect ultrasound images, text reports, and clinical structured data from thyroid cancer patients and perform preprocessing; Ultrasound imaging is the preferred tool for screening and evaluating thyroid nodules, providing a direct visual representation of key features such as nodule size, shape, internal structure, echo characteristics, and blood flow. However, ultrasound images are formed by the echoes of sound waves due to differences in acoustic impedance at different tissue interfaces. Many important tissue structures have only slight differences in acoustic impedance compared to surrounding normal tissues, resulting in very similar echo signal intensities, which manifest as low contrast and blurred boundaries in the image. Therefore, Limit Contrast Adaptive Histogram Equalization (CLAHE) is employed to enhance the local contrast of ultrasound images. This method divides the image into multiple small regions and performs local histogram equalization on each region to adapt to the non-uniformity at different depths and locations within the image. Simultaneously, it sets a contrast limit enhancement amplitude, effectively suppressing the tendency to over-enhance noise as "detail" in uniform regions by cropping histogram peaks and redistributing pixels, while preserving key features such as microcalcifications and low echogenicity.

[0016] Simultaneously, formatted image diagnostic report text corresponding to ultrasound images is collected, using standardized language to describe multiple qualitative features of the nodules, including but not limited to: location, shape, echogenicity, boundary, calcification, and blood flow signals. The standardized and semantically refined text reports, containing the expert knowledge and diagnostic logic of clinicians, help the model learn high-level semantic features consistent with medical practice. To ensure vocabulary consistency, all English characters in the report text are converted to a uniform lowercase form, while special characters, redundant punctuation marks, numbers, and meaningless formatting codes are removed, retaining only core descriptive text. Furthermore, common stop words that contribute little to semantic analysis are filtered out, allowing the model to focus more on key terms relevant to medical diagnosis.

[0017] Collect easily quantifiable basic clinical information of patients closely related to diagnosis. This includes, but is not limited to, age, gender, BMI, duration of underlying thyroid disease, family history, and clinical laboratory indicators such as thyroid function tests (7 items). Considering the dimensional differences between different data, the data were standardized using Z-scores. Subsequently, statistical tests were used to screen for significant differences in whether patients had lateral cervical lymph node metastasis. Chi-square tests were used for discrete characteristics such as gender and family history, while t-tests were used for continuous characteristics such as age and BMI.

[0018] Step 2: Design a dedicated network architecture to extract initial features of different modal data based on the distribution characteristics of different modal data; To comprehensively capture the multi-scale features of thyroid ultrasound nodules, this module employs an architecture combining a multi-scale convolutional neural network (CNN) and a feature pyramid network (FPN). The multi-scale CNN consists of three CNNs with kernels of different sizes, and it progressively extracts features at different levels of the image through multiple CNN layers, generating multi-level feature maps. Specifically, for a given level of feature map... Its multi-scale features It can be obtained through the following formula: in, Indicates the kernel size as The convolution operation, and Designed for extracting point-like, region-like, and global features respectively. This is for channel splicing operations.

[0019] Subsequently, an FPN structure is introduced to achieve top-down feature enhancement. Deep feature maps with strong semantics but low resolution are upsampled and then element-wise added to shallow, high-resolution feature maps, thereby passing high-level semantic information to the lower levels. This process can be represented as: in, For the enhanced feature map, This is an upsampling operation. Finally, the fused multi-scale semantic features extracted from the FPN output layer are denoted as ultrasound image features. It simultaneously encodes the microscopic calcification features and macroscopic morphological features of the nodules.

[0020] For ultrasound report data, which includes sonographic features of nodules and preliminary BI-RADS classification, containing rich qualitative diagnostic knowledge, a pre-trained BERT model is used as the feature extractor. For a given sequence of ultrasound report texts... After word segmentation and embedding layers, the text is input into the BERT model. Context-dependent representations of each word are computed using stacked Transformer encoder blocks. Finally, the hidden states corresponding to the CLS tokens are taken as the aggregate representation of the entire text sequence, and mapped to the feature space through an additional projection layer. in, This refers to the extracted text report feature vector, which condenses the semantic information of key diagnostic descriptors in the report and can be used for complementary verification with image features.

[0021] For structured clinical data, a high-dimensional nonlinear mapping is performed using a deep neural network. First, categorical variables are one-hot encoded and concatenated with numerical variables to form the original input vector. The vector is then fed into the multilayer perceptron: in, and For trainable parameters, The activation function is used. This network is able to learn complex interaction effects between clinical variables and map low-dimensional inputs to high-dimensional feature representations. .

[0022] Step 3: Combine contrastive learning and cross-modal feature reconstruction tasks to achieve multimodal feature decoupling and obtain common and specific representations of different modalities; Data from different modalities, such as ultrasound images, text reports, and clinical forms, can provide complementary pathological information at the morphological, semantic, and quantitative levels, but significant intermodal differences also exist. Traditional multimodal fusion methods often directly concatenate or weightedly fuse features from different modalities, ignoring differences in feature distribution, semantic hierarchy, and noise patterns. This leads to redundant features, impure representations, and low information coupling, thus affecting prediction accuracy. To address these issues, this paper proposes a method that combines contrastive learning and cross-modal feature reconstruction tasks to decouple multimodal features. The original features of each modality are explicitly decomposed into shared pathological characteristics and modality-specific characteristics, thereby preserving cross-modal consistency information during fusion.

[0023] For the multimodal initial features obtained in step two , , In order to extract common representations from data of different modalities With specific characterization A lightweight feature separation head is designed for each modality, specifically: Regarding image features: Regarding text features: For structured table features: in It is a learnable weight matrix. This is a bias term. The feature separation head allows the model to adaptively separate common representations from mode-specific representations within the same mode.

[0024] To enhance decoupling and ensure high consistency of pathological semantics in common representations across modalities, a bidirectional contrastive learning mechanism is introduced, including contrastive learning of common representations and contrastive learning of specific representations. The goal of contrastive learning of common representations is to make the common representations of different modalities of the same patient as close as possible, and the common representations of different patients as far apart as possible, thereby learning the shared disease essence information across modalities. For N patients, each with K modal features (here, K=3), the total number of features is... Common characteristics Where i represents the patient and k represents the modality. In the commonality representation contrastive learning, positive sample pairs represent common features of different modalities within the same patient, reflecting their cross-modal description of the same pathological state; negative sample pairs represent common features of different patients, used to enhance the model's ability to distinguish between different patients. Their corresponding contrastive loss function is: in For cosine similarity, This refers to temperature hyperparameters. This indicates different modalities under the same patient i. and The similarity of common representations between them, here, Traversal All modalities other than those in the same patient ensure that common representations across all modalities are brought closer together. This indicates different patient j-modes The similarity between common representations and anchor representations. By learning through contrastive learning of common representations, more discriminative and mode-invariant common feature representations can be obtained.

[0025] Specific representation contrastive learning aims to capture and standardize the unique information expression patterns of each modality, making the specific representations of different patients within the same modality as similar as possible, and the specific representations of different modalities as dissimilar as possible, highlighting the irreplaceable information contribution of each modality in clinical diagnosis. Positive sample pairs used for learning represent the specific features of different patients within the same modality, while negative sample pairs represent the specific features of different modalities. The corresponding contrastive loss function is: Through collaborative optimization using a bidirectional contrastive learning mechanism, the model can align cross-modal common semantic information, including morphological features and pathophysiological mechanisms of lesions, while preserving information unique to each modality, such as pixel-level anatomical details of images, semantic integrity of text descriptions, and individual differences in structured data.

[0026] After decoupling, the aggregated representation of common characteristics of each modality is extracted as common features. Specifically: Furthermore, to ensure that no key information is lost during the decoupling process, a cross-modal feature reconstruction task is introduced as a self-supervised signal, using shared common representations. Specific characterization of a particular mode The original features of this modality are reconstructed, prompting the encoder to retain the fine structural information in the ultrasound image, the key semantic content in the text description, and the quantitative index patterns in the table, and to enhance the integrity and reconstructability of the decoupled features. The corresponding calculation method is as follows: in The reconstruction decoder, representing the corresponding modality, consists of fully connected layers, while , , The images, text, and tables represent the reconstructed features, respectively, and the corresponding reconstruction loss is calculated as follows: In summary, the total loss function of the final decoupling module is: in To balance the weights of the various loss terms and the weights of the bidirectional contrastive learning and reconstruction tasks, the model can maintain semantic alignment, modality discrimination, and information integrity during the decoupling process.

[0027] Step 4: Based on cross-attention, the common representation and the modality-specific features are fused, and the fused representation is input into the classifier to obtain the prediction result; The common representation of each modality is obtained through the multimodal feature decoupling module. And the specific characteristics of each mode: , , To facilitate the interaction between common features and modality-specific features, a cross-attention mechanism is used to separately address the following: It is fused with various specific features.

[0028] For ultrasonic modes, As a query As keys and values, their fused features are computed through scaled dot product attention. : in , , For a learnable projection matrix, The dimension of the key vector. This process allows the model to focus on specific morphological patterns in ultrasound images that are most relevant to lymph node metastasis, such as microcalcifications or areas of abnormal blood flow, from a shared feature perspective.

[0029] Similarly, by performing cross-attention operations on the text modality and the table modality respectively, we can obtain: Text modality fusion can guide the model to further associate shared features with key semantics in pathological descriptions, such as unclear lymph node boundaries and uneven internal recovery. Similarly, tabular modality fusion can guide the model to focus on key clinical indicators such as patient age and thyroglobulin.

[0030] To integrate the information from the three fused features, learnable weights are used for weighted ensemble, resulting in the final joint feature representation used for classification: in The learnable weight parameters allow the model to dynamically adjust the contribution of each modality to the current diagnostic task, thus adapting to the differences in modal information quality across different cases. Finally, the final joint representation... The data is fed into a classifier, which passes through a fully connected layer to ultimately obtain the probability of lateral cervical lymph node metastasis. in It is the predicted probability of lymph node metastasis. The higher the value, the higher the risk of metastasis. Pathological confirmation or intervention should be carried out as early as possible to help develop an individualized treatment plan.

[0031] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling, characterized in that: Includes the following steps: Step 1: Collect ultrasound images, text reports, and clinical structured data from thyroid cancer patients and perform preprocessing. Step 2: Design a dedicated network architecture to extract initial features of different modal data based on their distribution characteristics; Step 3: Combine contrastive learning and cross-modal feature reconstruction tasks to achieve multimodal feature decoupling and obtain common and specific representations of different modalities; Step four: Based on cross-attention, the common representation and the modality-specific features are fused, and the fused representation is input into the classifier to obtain the prediction result.

2. The method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to claim 1, characterized in that: The specific steps for collecting and preprocessing ultrasound images, text reports, and clinical structured data from thyroid cancer patients in step one are as follows: Step 1: Local contrast of ultrasound images is enhanced by limiting contrast adaptive histogram equalization. This is achieved by dividing the image into multiple small regions and performing local histogram equalization on each region to adapt to the non-uniformity of the image at different depths and locations. At the same time, the contrast enhancement amplitude is limited. By cropping the histogram peaks and redistributing pixels, the tendency to over-enhance noise as "details" in uniform regions is suppressed, while key features such as microcalcifications and low echoes are preserved. Steps one and two: Simultaneously collect the formatted imaging diagnostic report text corresponding to the ultrasound images, describe multiple qualitative features of the nodule in standardized language, including but not limited to: location, shape, echo, boundary, calcification and blood flow signal, convert all English characters in the report text to a uniform lowercase form, and remove special characters, redundant punctuation marks, numbers and meaningless format codes from the report, while retaining the core descriptive text; Step 13 involves collecting easily quantifiable basic clinical information about the patient that is closely related to the diagnosis. This includes, but is not limited to, age, gender, BMI, duration of underlying thyroid disease, family history, and clinical laboratory indicators such as the seven thyroid function tests. Then, considering the dimensional differences between different data points, the data are standardized using a Z-score. Subsequently, statistical tests are used to screen for significant differences in whether patients have lateral cervical lymph node metastasis. For discrete characteristics such as gender and family history, a chi-square test is used. Continuous characteristics such as age and BMI are screened using the T-test.

3. The method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to claim 1 or 2, characterized in that: In step two, the specific steps for decoupling multimodal features by combining contrastive learning and cross-modal feature reconstruction tasks to obtain common and specific representations of different modalities are as follows: Step 21: An architecture combining multi-scale convolutional neural networks and feature pyramid networks is adopted. Multiple layers of CNNs are used to progressively extract features at different levels of the image, generating multi-level feature maps. For a given level of feature map... Its multi-scale features Obtained through the following formula: in, Indicates the kernel size as The convolution operation, and Designed for extracting point-like, region-like, and global features respectively. For channel splicing operations; Step 22 introduces an FPN structure to achieve top-down feature enhancement. This involves upsampling deep, semantically strong but low-resolution feature maps and adding them element-wise with shallow, high-resolution feature maps, thereby passing high-level semantic information to the lower layers. This process can be represented as: in, For the enhanced feature map, This is an upsampling operation; Steps two and three: The fused multi-scale semantic features extracted from the FPN output layer are denoted as ultrasound image features. It simultaneously encodes the microscopic calcification features and macroscopic morphological features of the nodules.

4. The method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to claim 3, characterized in that: The different modal data in step two include ultrasound text report data and clinical structured data; For ultrasound text report data, which includes sonographic features of nodules and preliminary BI-RADS classification, containing rich qualitative diagnostic knowledge, a pre-trained BERT model is used as the feature extractor for a given ultrasound report text sequence. After word segmentation and embedding, the text is input into the BERT model. Through stacked Transformer encoder blocks, the context-dependent representation of each word is calculated. Finally, the hidden state corresponding to the CLS token is taken as the aggregate representation of the entire text sequence, and it is mapped to the feature space through an additional projection layer. in, This is the extracted text report feature vector, which condenses the semantic information of key diagnostic descriptors in the report and can be used for complementary verification with image features. For structured clinical data, high-dimensional nonlinear mapping is performed using deep neural networks. First, categorical variables are one-hot encoded and then concatenated with numerical variables to form the original input vector. Then, the vector is fed into the multilayer perceptron: in, and For trainable parameters, Using the activation function, this network is able to learn complex interaction effects between clinical variables and map low-dimensional inputs to high-dimensional feature representations. .

5. The method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to claim 4, characterized in that: The specific steps in step three, which combine contrastive learning and cross-modal feature reconstruction tasks to decouple multimodal features and obtain common and specific representations of different modalities, are as follows: Step 31: For the multimodal initial features obtained in Step 2 , , In order to extract common characteristics from data of different modalities With specific characterization A lightweight feature separation head is designed for each modality, specifically: Regarding image features: Regarding text features: For structured table features: in, It is a learnable weight matrix. As a bias term, the feature separation head enables the model to adaptively separate common representations and modality-specific representations from the same modality; Step 32 introduces a two-way contrastive learning mechanism, including contrastive learning of common representations and contrastive learning of specific representations; Step 33: Introduce a cross-modal feature reconstruction task as a self-supervised signal, using shared common representations. Specific characterization of a particular mode The original features of this modality are reconstructed, prompting the encoder to retain the fine structural information in the ultrasound image, the key semantic content in the text description, and the quantitative index patterns in the table, and to enhance the integrity and reconstructability of the decoupled features. The corresponding calculation method is as follows: in The reconstruction decoder, representing the corresponding modality, consists of fully connected layers, while , , The images, text, and tables represent the reconstructed features, respectively, and the corresponding reconstruction loss is calculated as follows: 。 6. The method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to claim 5, characterized in that: Step 3.2 introduces a bidirectional contrastive learning mechanism, including the specific methods of common representation contrastive learning and specific representation contrastive learning as follows: Common representation contrastive learning specifically refers to: for N patients, each with K modal features, then there are common features... Common characteristics , where i represents the patient and k represents the modality. In the commonality representation contrastive learning, positive sample pairs represent common features of different modalities within the same patient, reflecting their cross-modal description of the same pathological state; negative sample pairs represent common features of different patients, used to enhance the model's ability to distinguish between different patients, and their corresponding contrastive loss function is: in, For cosine similarity, For temperature hyperparameters, This indicates different modalities under the same patient i. and The similarity of common representations between them, here, Traversal All modalities other than those mentioned above ensure that the common characteristics of all modalities of the same patient are brought closer together; This indicates different patient j, modality The similarity between common representations and anchor representations can be used to obtain more discriminative and modality-invariant common feature representations through comparative learning of common representations. Specific characterization contrastive learning specifically involves: positive sample pairs used for learning represent specific features of different patients within the same modality, while negative sample pairs represent specific features of different modalities. The corresponding contrastive loss function is: In summary, the total loss function of the final decoupling module is: in To balance the weight hyperparameters of each loss term and the weights of bidirectional contrastive learning and reconstruction tasks, the model can maintain semantic alignment, modality discrimination and information integrity during the decoupling process. Finally, after decoupling, the aggregated representation of common characteristics across modalities is extracted as common features, specifically: 。 7. The method for predicting lateral cervical lymph node metastasis in thyroid cancer based on multimodal feature decoupling according to claim 1 or 2, characterized in that: In step four, the fusion of common representations and modality-specific features based on cross-attention is achieved. The specific method for inputting the fused representations into the classifier to obtain the prediction results is as follows: The common representation of each modality is obtained through the multimodal feature decoupling module. And the specific characteristics of each mode: , , To facilitate the interaction between common features and modality-specific features, a cross-attention mechanism is used to separately address the following: Integrate with various specific features; For ultrasonic modes, As a query As keys and values, their fused features are computed via scaled dot product attention. : in, , , For a learnable projection matrix, The dimension of the key vector. This process allows the model to focus on specific morphological patterns in ultrasound images most relevant to lymph node metastasis, such as microcalcifications or areas of abnormal blood flow, from a shared feature perspective; Performing cross-attention operations on the text modality and the table modality respectively yields: To integrate the information from the three fused features, learnable weights are used for weighted ensemble, resulting in the final joint feature representation used for classification: in, The learnable weight parameters allow the model to dynamically adjust the contribution of each modality to the current diagnostic task, thereby adapting to the differences in modal information quality in different cases; finally, the final joint representation is... The data is fed into a classifier, which passes through a fully connected layer to ultimately obtain the probability of lateral cervical lymph node metastasis. in, It is the predicted probability of lymph node metastasis. The higher the value, the higher the risk of metastasis. Pathological confirmation or intervention should be carried out as early as possible to help develop an individualized treatment plan.

Images