Auxiliary method for early diagnosis of digestive tract tumor

Abstract

The present application relates to the technical field of medical informatics, and in particular to an auxiliary method for early clinical diagnosis of digestive tract tumors, which comprises the following steps: collecting fecal metabolomics data, serum exosome miRNA expression profiles and white light endoscopy image data of a subject, respectively performing standardization preprocessing and feature extraction; using a self-supervised contrast learning algorithm to pre-train unlabelled data of each modality to obtain general feature representation; constructing a cross-modal Transformer fusion network based on a multi-head attention mechanism to dynamically learn the contribution weight of different modality features in different tumor types; and finally outputting a tumor risk score through a classifier and generating a visual and interpretable report containing key feature contribution. The present application effectively solves the problems of low diagnostic accuracy, high dependence on labeled data and poor model interpretability in the prior art, and significantly improves the sensitivity and specificity of early diagnosis of digestive tract tumors.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, specifically to an auxiliary method for the early clinical diagnosis of gastrointestinal tumors. Background Technology

[0002] Gastrointestinal cancers are among the most common and deadliest malignant tumors worldwide, primarily including stomach cancer, colorectal cancer, esophageal cancer, and pancreatic cancer. According to the 2020 Global Cancer Burden data released by the International Agency for Research on Cancer (IARC) of the World Health Organization, there are approximately 5 million new cases of gastrointestinal cancers globally, and approximately 4 million deaths, accounting for 26% and 35% of all new cancer cases and deaths, respectively. In my country, the incidence and mortality rates of gastrointestinal cancers remain high, with over 400,000 new cases of stomach and colorectal cancer each year. Most patients are diagnosed at an advanced stage, with a 5-year survival rate of less than 30%. However, if detected and treated promptly in the early stages, the 5-year survival rate for patients with gastrointestinal cancers can significantly increase to over 90%. Therefore, improving the early diagnosis rate of gastrointestinal cancers is crucial to reducing their mortality rate.

[0003] Currently, early diagnosis of gastrointestinal tumors mainly includes endoscopy, imaging examinations, and tumor marker testing. Endoscopy is the standard for diagnosing gastrointestinal tumors, allowing direct observation of morphological changes in the digestive tract mucosa and enabling biopsies and pathological examinations of suspicious lesions. However, endoscopy is an invasive procedure with poor patient compliance, and the results are highly dependent on the endoscopist's experience and skill, resulting in a certain rate of missed diagnoses and misdiagnoses. This is especially true for some small lesions and flat lesions, which can be easily missed even by experienced endoscopists. Imaging examinations such as CT, MRI, and PET-CT, while providing structural information about the digestive tract wall and surrounding tissues, have low sensitivity for early-stage gastrointestinal tumors and involve issues such as radiation exposure and high costs. Tumor marker tests, such as carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA19-9), and alpha-fetoprotein (AFP), have advantages such as being non-invasive, convenient, and reproducible. However, these markers generally have low sensitivity and specificity, and often do not show significant increases in patients with early-stage gastrointestinal tumors, which can easily lead to false negative results. False positive results may also occur in some patients with benign diseases. Therefore, they can only be used as auxiliary diagnostic tools.

[0004] In recent years, with the rapid development of artificial intelligence technology, computer-aided diagnostic (CAD) methods based on machine learning and deep learning have been widely used in the early diagnosis of gastrointestinal tumors. These methods automatically identify tumor characteristics and make diagnostic predictions by analyzing endoscopic images, medical imaging, or biomarker data, effectively improving diagnostic efficiency and accuracy while reducing the influence of human factors. However, most existing CAD methods are based on single-modality data for diagnosis, such as using only endoscopic images or only tumor marker data, failing to fully utilize the complementary information between different modalities, resulting in insufficient diagnostic accuracy and sensitivity. For example, endoscopic image-based CAD methods can effectively identify morphological abnormalities but cannot reflect the molecular biological characteristics of tumors; while biomarker-based CAD methods can reflect molecular-level changes in tumors but cannot provide information on the location and morphology of lesions.

[0005] To address the limitations of single-modal diagnosis, some researchers have begun exploring the fusion of multimodal data for the diagnosis of gastrointestinal tumors. However, most existing multimodal fusion methods employ simple feature concatenation or weighted averaging strategies, failing to consider the complex nonlinear relationships and interactions between different modalities and thus failing to fully exploit the potential value of multimodal data. Furthermore, existing methods largely rely on supervised learning using large amounts of labeled medical data. Labeling medical data requires significant time and effort from medical professionals, resulting in high costs. Moreover, the quantity and quality of labeled data often fall short of the training requirements of deep learning models, leading to overfitting and poor generalization ability. Additionally, most existing AI diagnostic models are black-box models, unable to explain the basis of their diagnostic decisions. Clinicians struggle to understand and trust the model's output, significantly limiting the widespread application of AI-assisted diagnostic technology in clinical practice.

[0006] Therefore, there is an urgent need to develop an auxiliary method for the early clinical diagnosis of gastrointestinal tumors that can make full use of complementary information from multimodal data, reduce dependence on labeled data, and has good interpretability, so as to improve the early diagnosis rate of gastrointestinal tumors and reduce mortality. Summary of the Invention

[0007] To address the problems in existing technologies, this invention provides an auxiliary method for the early clinical diagnosis of gastrointestinal tumors. This method effectively solves the problems of low accuracy of single-modal diagnosis, high dependence on labeled data, and poor model interpretability in existing technologies by using techniques such as multimodal data fusion, self-supervised comparative learning, and interpretability analysis.

[0008] The technical solution adopted by this invention to solve its technical problem is: an auxiliary method for early clinical diagnosis of gastrointestinal tumors, comprising the following steps: S1: Data acquisition, collecting fecal metabolomics data, serum exosome miRNA expression profile data, and white light endoscopy image data of subjects to be diagnosed; S2: Data preprocessing, which involves standardizing, normalizing and denoising the data of the three modalities mentioned above, and extracting the initial feature vectors of each modality; S3: Feature pre-training. A self-supervised contrastive learning algorithm is used to pre-train the feature extractors of each modality using large-scale unlabeled multimodal medical data to obtain a general feature representation that can capture the inherent patterns of the data. S4: Cross-modal feature fusion. The pre-trained features of each modality are input into a cross-modal Transformer fusion network based on a multi-head attention mechanism. The correlation between different features is learned through inter-modal attention interaction, the contribution weight of each modality feature is dynamically allocated, and a fused feature vector is generated. S5: Diagnostic prediction and interpretation. It integrates feature vectors into the classifier, outputs a gastrointestinal tumor risk score, and generates a visual diagnostic report through the interpretability module. The report includes the top N key features that contribute the most to the diagnostic results and their clinical significance.

[0009] Specifically, in step S2, the preprocessing of fecal metabolomics data includes missing value imputation, logarithmic transformation, and Z-score normalization; the preprocessing of serum exosome miRNA expression profile data includes low expression filtering, median normalization, and batch effect correction; and the preprocessing of white light endoscopy image data includes image cropping, grayscale normalization, histogram equalization, and deblurring.

[0010] Specifically, in step S3, the self-supervised contrastive learning algorithm adopts the SimCLR framework, which generates positive sample pairs by performing different data augmentations on the same input sample and generates negative sample pairs for different input samples, thereby maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, and thus training the feature extractor.

[0011] Specifically, in step S4, the cross-modal Transformer fusion network includes three encoder layers and one cross-modal attention layer. Each encoder layer consists of a multi-head self-attention sub-layer and a feedforward neural network sub-layer. The cross-modal attention layer is used to calculate the attention weights between different modal features.

[0012] Specifically, the multi-head attention mechanism has 8 heads, each attention head has a dimension of 64, the hidden layer dimension of the feedforward neural network is 512, and the activation function is GELU.

[0013] Specifically, in step S5, the interpretability module uses the Gradient Weighted Class Activation Mapping (Grad-CAM) algorithm to calculate the contribution of each feature to the final diagnostic result, and highlights the top 10 features with the highest contribution and their corresponding original data fragments in the visualization report.

[0014] Specifically, in step S5, the classifier uses a logistic regression model, and the output tumor risk score ranges from 0 to 1. When the risk score is greater than or equal to 0.5, it indicates that the subject has a high risk of gastrointestinal tumors and it is recommended to conduct further pathological examination.

[0015] Specifically, the digestive tract tumors include gastric cancer, colorectal cancer, esophageal cancer, and pancreatic cancer.

[0016] Specifically, the method also includes differential diagnosis of gastrointestinal tumors at different stages. By adding tumor stage labels during the classifier training process, the model can simultaneously output tumor type and stage prediction results.

[0017] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned auxiliary method for early clinical diagnosis of gastrointestinal tumors.

[0018] The beneficial effects of this invention are: This invention integrates three modalities of data: fecal metabolomics, serum exosome miRNA, and white light endoscopy images. By fully utilizing the complementary information between different modalities, it significantly improves the accuracy and sensitivity of early diagnosis of gastrointestinal tumors.

[0019] This invention employs a self-supervised contrastive learning algorithm to pre-train the feature extractor, effectively reducing the model's dependence on labeled data, lowering data labeling costs, and simultaneously improving the model's generalization ability.

[0020] This invention constructs a cross-modal Transformer fusion network based on a multi-head attention mechanism, which can dynamically learn the contribution weights of different modal features in different tumor types, thus achieving more effective multimodal feature fusion.

[0021] This invention introduces an interpretability module that can generate visualized diagnostic reports, solving the black box problem of AI models and improving clinicians' acceptance and trust in the models.

[0022] The method of this invention is non-invasive, convenient, and highly reproducible, making it suitable for large-scale population screening. It can effectively improve the early diagnosis rate of gastrointestinal tumors and reduce mortality. Attached Figure Description

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Figure 1 A flowchart of an auxiliary method for early clinical diagnosis of gastrointestinal tumors provided by the present invention; Figure 2 A diagram of the self-supervised contrastive learning pre-training framework in the auxiliary method for early clinical diagnosis of gastrointestinal tumors provided by this invention; Figure 3 A schematic diagram of the cross-modal Transformer fusion network in the auxiliary method for early clinical diagnosis of gastrointestinal tumors provided by the present invention; Figure 4 The flowchart illustrates the process of generating an interpretable diagnostic report in the auxiliary method for early clinical diagnosis of gastrointestinal tumors provided by this invention. Detailed Implementation

[0025] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0026] like Figures 1-4 As shown, the auxiliary method for early clinical diagnosis of gastrointestinal tumors according to the present invention includes the following steps: S1: Data Acquisition. Fecal metabolomics data, serum exosome miRNA expression profiles, and white light endoscopy images were collected from subjects seeking diagnosis. Fecal metabolomics data were obtained using liquid chromatography-mass spectrometry (LC-MS), reflecting overall changes in gut microbiota metabolism and host metabolism. Serum exosome miRNA expression profiles were obtained using high-throughput sequencing; exosome miRNAs exhibit high stability and tumor specificity, making them ideal early tumor markers. White light endoscopy images were acquired using an electronic endoscopy system, clearly displaying the morphological characteristics of the digestive tract mucosa.

[0027] S2: Data Preprocessing. Standardization, normalization, and denoising were performed on the three modalities to extract initial feature vectors for each modality. Specifically, for fecal metabolomics data, the K-nearest neighbor algorithm was first used for missing value imputation, followed by logarithmic transformation to eliminate skewed distribution, and finally Z-score standardization to ensure each feature has a mean of 0 and a variance of 1. For serum exosome miRNA expression profile data, miRNAs with expression levels below a threshold were first filtered out, followed by median normalization to eliminate technical differences between samples, and finally, the ComBat algorithm was used for batch effect correction. For white light endoscopy image data, the images were first cropped to a uniform size, followed by grayscale normalization and histogram equalization to enhance image contrast, and finally, Gaussian filtering was used to remove image noise.

[0028] S3: Feature Pre-training. A self-supervised contrastive learning algorithm is employed to pre-train the feature extractors for each modality using large-scale unlabeled multimodal medical data, obtaining general feature representations capable of capturing the inherent patterns in the data. This invention uses the SimCLR framework for self-supervised contrastive learning. This framework generates positive sample pairs by performing different data augmentation operations on the same input sample and negative sample pairs on different input samples. The feature extractor is then trained by maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs. For fecal metabolomics data and serum exosome miRNA expression profile data, data augmentation methods such as random masking and adding Gaussian noise are used; for white light endoscopy image data, data augmentation methods such as random cropping, rotation, flipping, and color jittering are used. Through self-supervised pre-training, the feature extractor can learn the essential features of the data without requiring a large amount of labeled data, thereby improving the model's generalization ability.

[0029] S4: Cross-modal feature fusion. Pre-trained features from each modality are input into a cross-modal Transformer fusion network based on a multi-head attention mechanism. Through inter-modal attention interactions, the network learns the relationships between different features, dynamically assigns contribution weights to each modality feature, and generates a fused feature vector. The cross-modal Transformer fusion network constructed in this invention comprises three encoder layers and one cross-modal attention layer. Each encoder layer consists of a multi-head self-attention sublayer and a feedforward neural network sublayer. The multi-head self-attention sublayer captures the dependencies between different features within the same modality, while the feedforward neural network sublayer performs nonlinear transformations on the features. The cross-modal attention layer calculates the attention weights between features from different modalities, thereby achieving information interaction and fusion between modalities. In this way, the model can automatically adjust the contribution weights of each modality feature according to the characteristics of different tumor types and samples, fully utilizing the complementary information of multimodal data.

[0030] S5: Diagnostic Prediction and Interpretation. The fused feature vector is input into a classifier, which outputs a gastrointestinal tumor risk score. A visual diagnostic report is then generated through an interpretability module. This invention uses a logistic regression model as the classifier, with the output tumor risk score ranging from 0 to 1. A higher score indicates a greater risk of gastrointestinal tumors in the subject. When the risk score is greater than or equal to 0.5, it indicates a high risk of gastrointestinal tumors in the subject, and further pathological examination is recommended. The interpretability module uses the Gradient Weighted Class Activation Mapping (Grad-CAM) algorithm to calculate the contribution of each feature to the final diagnostic result. The top 10 features with the highest contribution and their corresponding original data segments are highlighted in the visual report, along with an explanation of the clinical significance of each key feature. In this way, clinicians not only receive the model's diagnostic results but also understand the basis for the model's diagnosis, thereby increasing their trust in the model.

[0031] Example 1: Auxiliary methods for early diagnosis of gastric cancer: This embodiment provides an auxiliary method for early diagnosis of gastric cancer based on the method of the present invention, and the specific steps are as follows: Dataset Construction: This study collected data from 1200 participants from three tertiary hospitals, including 600 patients with pathologically confirmed early-stage gastric cancer and 600 healthy controls. All participants signed informed consent forms, and the study was approved by the hospital ethics committees. All data were randomly divided into training, validation, and test sets in a 7:1:2 ratio, with the training set containing 840 samples, the validation set containing 120 samples, and the test set containing 240 samples. In addition, 5000 unlabeled multimodal medical data points were collected for self-supervised pre-training.

[0032] Data preprocessing: Fecal metabolomics data were analyzed using an Agilent 1290 Infinity II liquid chromatography system and an Agilent 6545 Q-TOF mass spectrometer to obtain metabolite peak intensity data for each sample. First, missing values ​​were imputed using the K-nearest neighbor algorithm, with K set to 5. Then, a logarithmic transformation was performed to ensure the data followed a normal distribution. Finally, Z-score standardization was applied to obtain a metabolomics feature vector with a dimension of 200.

[0033] For serum exosomal miRNA expression profiling data, serum exosomes were first extracted using the ExoQuick kit, and then miRNAs were sequenced using the Illumina NovaSeq 6000 sequencing platform to obtain miRNA expression levels for each sample. First, miRNAs expressed in less than 10% of the samples were filtered out, retaining 150 miRNAs. Then, median normalization was performed to eliminate technical differences between samples. Finally, the ComBat algorithm was used to correct for batch effects caused by different sequencing batches, resulting in miRNA feature vectors with a dimension of 150.

[0034] White light endoscopic image data were acquired using an Olympus GIF-H290 electronic endoscope system, with 10-20 gastric mucosal images taken from different angles for each subject. First, all images were cropped to a uniform size of 224×224 pixels; then, grayscale normalization was performed, scaling pixel values ​​to between 0 and 1; next, histogram equalization was performed to enhance image contrast; finally, a 3×3 Gaussian filter was used to remove image noise. A pre-trained ResNet-50 network was used as the image feature extractor to extract the last layer of convolutional features for each image. The average of all image features from the same subject was then calculated to obtain a 2048-dimensional image feature vector.

[0035] Self-supervised pre-training: The SimCLR framework was used for self-supervised pre-training of the feature extractors for each modality. For metabolomics and miRNA data, data augmentation methods included random masking (mask ratio of 0.2) and adding Gaussian noise (mean 0, standard deviation 0.1). For image data, data augmentation methods included random cropping, random horizontal flipping, random vertical flipping, color jittering, and Gaussian blurring. During pre-training, the batch size was set to 128, the temperature coefficient to 0.5, the optimizer to Adam, the learning rate to 0.001, and the number of training epochs to 100. After pre-training, the weight parameters of each feature extractor were saved.

[0036] Training cross-modal fusion networks: The weight parameters of each modal feature extractor, pre-trained, were used as initial values ​​to construct a cross-modal Transformer fusion network. The network consists of three encoder layers and one cross-modal attention layer. The multi-head attention mechanism has eight heads, each with a dimension of 64. The hidden layer dimension of the feedforward neural network is 512, and the activation function is GELU. Modal features from the training set samples are input into the network for fine-tuning. During training, the batch size is set to 32, the optimizer is AdamW, the learning rate is 0.0001, the weight decay is 0.01, and the number of training epochs is 50. An early stopping strategy is employed to prevent overfitting; training stops when the validation set loss does not decrease for 10 consecutive epochs.

[0037] Diagnostic prediction and interpretability analysis: The test set samples are input into the trained model, which outputs a gastric cancer risk score. Accuracy, sensitivity, specificity, and AUC are used as evaluation metrics. Experimental results show that the method of this invention achieves an accuracy of 92.5%, a sensitivity of 91.7%, a specificity of 93.3%, and an AUC of 0.972 on the test set.

[0038] To verify the superiority of the method of the present invention, the following three sets of comparative examples were set up: Control Example 1: Diagnosis was performed using only white light endoscopic image data, employing the same ResNet-50 network and classifier. Experimental results showed that Control Example 1 achieved an accuracy of 81.7%, a sensitivity of 79.2%, a specificity of 84.2%, and an AUC of 0.895.

[0039] Control Example 2: A simple feature splicing method was used for multimodal fusion, which directly spliced ​​metabolomics, miRNA, and image features before inputting them into the classifier. Experimental results showed that Control Example 2 had an accuracy of 86.3%, a sensitivity of 84.2%, a specificity of 88.3%, and an AUC of 0.931.

[0040] Comparative Example 3: No self-supervised pre-training was used; the feature extractor was trained directly using random initialization. Experimental results show that the accuracy of Comparative Example 3 was 83.8%, the sensitivity was 81.7%, the specificity was 85.8%, and the AUC value was 0.914.

[0041] The results above show that the evaluation metrics of the method of this invention are significantly better than those of the three control examples, indicating that multimodal data fusion, attention-based cross-modal fusion network, and self-supervised pre-training all make important contributions to improving diagnostic performance.

[0042] Furthermore, the interpretability module of this invention can generate visual diagnostic reports. For example, for a patient with early-stage gastric cancer, the model outputs a risk score of 0.89, with the top three contributing features being: the metabolite feature "lactic acid," the miRNA feature "miR-21," and the image feature "irregular bulges in the antral mucosa." The report also provides explanations of the clinical significance of these features: elevated lactate is associated with enhanced glycolytic metabolism in tumor cells; miR-21 is a proven gastric cancer-associated miRNA that is highly expressed in gastric cancer tissue; and irregular bulges in the antral mucosa are a typical morphological feature of early-stage gastric cancer. In this way, clinicians can make more accurate diagnoses based on this information and their own experience.

[0043] Example 2: Auxiliary methods for early diagnosis of colorectal cancer: This embodiment provides an auxiliary method for early diagnosis of colorectal cancer based on the method of the present invention, and the specific steps are as follows: Dataset Construction: This study collected data from 1500 participants from four tertiary hospitals, including 750 patients with pathologically confirmed early-stage colorectal cancer and 750 healthy controls. All participants signed informed consent forms, and the study was approved by the hospital ethics committees. All data were randomly divided into training, validation, and test sets in a 7:1:2 ratio, with the training set containing 1050 samples, the validation set containing 150 samples, and the test set containing 300 samples. In addition, 6000 unlabeled multimodal medical data points were collected for self-supervised pre-training.

[0044] Data preprocessing: The data preprocessing steps are basically the same as in Example 1, except that: Fecal metabolomics data analysis yielded peak intensity data for 250 metabolites, which were then preprocessed to obtain a metabolomics feature vector with a dimension of 250.

[0045] After filtering, 180 miRNAs were retained from the serum exosomal miRNA expression profile data. After preprocessing, a miRNA feature vector with a dimension of 180 was obtained.

[0046] White light endoscopy image data are colorectal mucosal images. 15-25 images from different angles were collected for each subject. After preprocessing and feature extraction, image feature vectors with a dimension of 2048 were obtained.

[0047] Self-supervised pre-training: The self-supervised pre-training steps are basically the same as in Example 1, except that the number of training rounds is set to 120.

[0048] Training cross-modal fusion networks: The training steps for the cross-modal fusion network are basically the same as in Example 1, except that the number of training rounds is set to 60.

[0049] Diagnostic prediction and interpretability analysis: The test set samples are input into the trained model, which outputs a colorectal cancer risk score. Experimental results show that the method of this invention has an accuracy of 93.7%, a sensitivity of 92.8%, a specificity of 94.5%, and an AUC of 0.981 on the test set.

[0050] To further verify the clinical value of the method of this invention, a control example (Example 4) was set up for comparison with the detection of CEA, a commonly used clinical tumor marker for colorectal cancer. CEA detection was performed using electrochemiluminescence immunoassay, with a cutoff value of 5 ng / mL. The experimental results showed that the accuracy of CEA detection was 65.3%, the sensitivity was 58.7%, the specificity was 72.0%, and the AUC value was 0.702.

[0051] The results above demonstrate that the diagnostic performance of the method described in this invention is significantly superior to that of traditional CEA testing, particularly in terms of sensitivity, effectively reducing the missed diagnosis rate of early-stage colorectal cancer. Furthermore, the method can generate interpretable reports, providing clinicians with diagnostic information. For example, for a patient with early-stage colorectal cancer, the model outputs a risk score of 0.92, with the top three contributing features being: the metabolite feature "bile acids," the miRNA feature "miR-92a," and the image feature "sigmoid colon polypoid lesion." The report also provides explanations of the clinical significance of these features, helping clinicians better understand the model's diagnostic results.

[0052] Example 3: Auxiliary methods for staging and differentiating gastrointestinal tumors: This embodiment provides an auxiliary method for staging and identifying gastrointestinal tumors based on the method of the present invention, which can simultaneously predict the type and stage of gastrointestinal tumors. This embodiment uses gastric cancer as an example, and the specific steps are as follows: Dataset Construction: This embodiment collected data from 1000 gastric cancer patients from three tertiary hospitals, including 300 patients with stage I gastric cancer, 300 patients with stage II gastric cancer, 200 patients with stage III gastric cancer, and 200 patients with stage IV gastric cancer. All patients were pathologically diagnosed and staged using TNM staging. All data were randomly divided into training, validation, and test sets in a 7:1:2 ratio, with the training set containing 700 samples, the validation set containing 100 samples, and the test set containing 200 samples. In addition, 5000 unlabeled multimodal medical data were collected for self-supervised pre-training.

[0053] Data preprocessing: The data preprocessing steps are the same as in Example 1.

[0054] Self-supervised pre-training: The self-supervised pre-training steps are the same as in Example 1.

[0055] Training cross-modal fusion networks: A multi-classifier is added to the output layer of the cross-modal fusion network to simultaneously predict tumor type and stage. During training, a cross-entropy loss function is used to optimize the losses for both tumor type and stage classification. Other training parameters are the same as in Example 1.

[0056] Staged Identification Results and Analysis: The test set samples are input into the trained model, and the model outputs gastric cancer staging prediction results. Overall accuracy and the accuracy of each stage are used as evaluation metrics. Experimental results show that the overall staging accuracy of the method in this invention on the test set is 87.5%, with an accuracy of 85.0% for stage I gastric cancer, 88.3% for stage II, 87.5% for stage III, and 90.0% for stage IV.

[0057] To verify the superiority of the method of the present invention, a control example (5) was set up for comparison with the traditional pathological staging method. The traditional pathological staging method determines the tumor stage by performing pathological examination on the surgically removed specimen, and is currently the standard for gastric cancer staging. However, pathological staging requires surgery, and cannot be applied to some advanced patients who are not candidates for surgery. The method of the present invention is a non-invasive staging identification method that can provide important reference information for clinicians before surgery.

[0058] Furthermore, the method of this invention can also generate interpretable reports that indicate which features are closely related to tumor staging. For example, studies have found that the expression levels of metabolite features “glucose” and “glutamine” and miRNA features “miR-155” and “miR-210” are positively correlated with gastric cancer staging, while image features “ulcer size” and “depth of invasion” are also closely related to gastric cancer staging.

[0059] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of protection claimed by the present invention. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An auxiliary method for early clinical diagnosis of gastrointestinal tumors, characterized in that, Includes the following steps: S1: Data acquisition, collecting fecal metabolomics data, serum exosome miRNA expression profile data, and white light endoscopy image data of subjects to be diagnosed; S2: Data preprocessing, which involves standardizing, normalizing and denoising the data of the three modalities mentioned above, and extracting the initial feature vectors of each modality; S3: Feature pre-training. A self-supervised contrastive learning algorithm is used to pre-train the feature extractors of each modality using large-scale unlabeled multimodal medical data to obtain a general feature representation that can capture the inherent patterns of the data. S4: Cross-modal feature fusion. The pre-trained features of each modality are input into a cross-modal Transformer fusion network based on a multi-head attention mechanism. The correlation between different features is learned through inter-modal attention interaction, the contribution weight of each modality feature is dynamically allocated, and a fused feature vector is generated. S5: Diagnostic prediction and interpretation. It integrates feature vectors into the classifier, outputs a gastrointestinal tumor risk score, and generates a visual diagnostic report through the interpretability module. The report includes the top N key features that contribute the most to the diagnostic results and their clinical significance.

2. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: In step S2, the preprocessing of fecal metabolomics data includes missing value imputation, logarithmic transformation, and Z-score normalization; Preprocessing of serum exosomal miRNA expression profile data included low expression filtering, median normalization, and batch effect correction; preprocessing of white light endoscopy image data included image cropping, grayscale normalization, histogram equalization, and deblurring.

3. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: In step S3, the self-supervised contrastive learning algorithm adopts the SimCLR framework. It generates positive sample pairs by performing different data augmentations on the same input sample and generates negative sample pairs for different input samples, thereby maximizing the similarity between positive sample pairs and minimizing the similarity between negative sample pairs, thus training the feature extractor.

4. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: In step S4, the cross-modal Transformer fusion network includes three encoder layers and one cross-modal attention layer. Each encoder layer consists of a multi-head self-attention sub-layer and a feedforward neural network sub-layer. The cross-modal attention layer is used to calculate the attention weights between different modal features.

5. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 4, characterized in that: The multi-head attention mechanism has 8 heads, each attention head has a dimension of 64, the hidden layer dimension of the feedforward neural network is 512, and the activation function is GELU.

6. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: In step S5, the interpretability module uses a gradient-weighted class activation mapping algorithm to calculate the contribution of each feature to the final diagnostic result, and highlights the top 10 features with the highest contribution and their corresponding original data fragments in the visualization report.

7. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: In step S5, the classifier uses a logistic regression model, and the output tumor risk score ranges from 0 to 1. When the risk score is greater than or equal to 0.5, it indicates that the subject has a high risk of gastrointestinal tumors.

8. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: The digestive tract tumors include gastric cancer, colorectal cancer, esophageal cancer, and pancreatic cancer.

9. The auxiliary method for early clinical diagnosis of gastrointestinal tumors according to claim 1, characterized in that: The method also includes differential diagnosis of gastrointestinal tumors at different stages. By adding tumor stage labels during the classifier training process, the model can simultaneously output tumor type and stage prediction results.

Images