Two-stage endoscopic ultrasonic tumor identification method based on context awareness

Abstract

The invention discloses a two-stage endoscopic ultrasonic tumor identification method based on context awareness, and relates to the technical field of medical image processing, and the key points of the technical scheme are as follows: collecting endoscopic ultrasonic images of various types of gastrointestinal tumor cases; the collected images are screened by professional medical image experts, samples which do not meet tumor classification requirements are removed, and labeling information is added; an endoscope ultrasonic tumor recognition model is constructed, and a self-supervised learning method and an ROI detection classification method are adopted for training; carrying out parameter optimization on the trained model by adopting an optimization algorithm; and performing performance verification on the model by adopting a clinical application data set, and outputting accuracy and robustness evaluation results of gastrointestinal tumor recognition. According to the method, the self-supervised learning and context coding technologies are introduced, potential semantic information in the EUS image is fully mined under the condition that a large amount of labeled data is not needed, and the accuracy of tumor detection and the generalization ability of the model are effectively improved in combination with a two-stage training strategy.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically, to a context-aware two-stage endoscopic ultrasound tumor identification method. Background Technology

[0002] Endoscopic ultrasound (EUS) is a commonly used medical imaging technique in clinical practice. By placing a high-frequency ultrasound probe inside the human body, it can provide high-resolution imaging of the digestive tract and surrounding tissues. EUS is widely used in the diagnosis and localization of gastrointestinal tumors, especially in the detection of benign and malignant tumors such as gastrointestinal stromal tumors (GIST) and leiomyomas. However, due to the complexity and variability in the quality of EUS images, accurate detection and classification of tumors remain challenging in clinical applications.

[0003] Traditional tumor detection relies on the doctor's experience, which is subjective and easily affected by factors such as image quality and the complexity of the lesion site, leading to misdiagnosis or missed diagnosis. In addition, the complex structure of the gastrointestinal tract, with its overlapping anatomical structures and diverse tumor morphologies, makes it more difficult to identify and distinguish lesion areas in images. These factors limit the widespread application of EUS in tumor detection and diagnosis, especially when doctors lack experience or the condition is complex, the difficulty of interpreting EUS images increases, further affecting the accuracy of diagnosis.

[0004] In recent years, artificial intelligence (AI)-based image processing and recognition technologies have been increasingly applied in the field of medical imaging, assisting doctors in image analysis through automation. However, existing AI methods still face many challenges when processing EUS images, mainly including the following aspects:

[0005] (1) Limited quality and quantity of image data: EUS images have relatively low resolution, and the gastrointestinal structures of different patients vary greatly, resulting in insufficient diversity of image data. At the same time, since labeling EUS images requires a highly specialized knowledge background, manual labeling is costly, resulting in a shortage of high-quality datasets that can be used for AI model training;

[0006] (2) Complex background and interference information: EUS images contain a large amount of background information, such as the gastrointestinal wall and mucosal folds. This complex background can interfere with the identification of tumor areas. Traditional image recognition methods have difficulty accurately extracting tumor areas from complex backgrounds;

[0007] (3) Lack of sufficient contextual information understanding: Existing AI methods, when processing EUS images, usually rely on simple feature extraction and classification, making it difficult to effectively utilize the contextual information in the image for accurate tumor localization and classification.

[0008] Therefore, the present invention aims to provide a context-aware two-stage endoscopic ultrasound tumor identification method to solve the above problems. Summary of the Invention

[0009] The purpose of this invention is to provide a context-aware two-stage endoscopic ultrasound tumor identification method. This invention employs a self-supervised learning method to extract useful contextual semantic information from a large number of unlabeled EUS images, avoiding reliance on a large amount of manually labeled data. Simultaneously, it adopts a two-stage model training strategy: first, self-supervised feature learning is performed; then, tumor regions are detected and classified through target region expansion. This not only improves the accuracy of tumor detection but also effectively solves the overfitting problem, enhancing the model's generalization ability. This makes it widely applicable in the early screening, clinical diagnosis, and treatment evaluation of gastrointestinal tumors, helping doctors more accurately identify tumor regions and classify diseases, greatly improving diagnostic efficiency and accuracy.

[0010] The above-mentioned technical objective of the present invention is achieved through the following technical solution: a context-aware two-stage endoscopic ultrasound tumor identification method, comprising the following steps:

[0011] S1. Data collection: Collect endoscopic ultrasound images of various types of gastrointestinal tumor cases, which cover different pathological diagnostic reports and diverse pathological types;

[0012] S2. Data cleaning and annotation: Professional medical imaging experts screen the collected endoscopic ultrasound images, remove samples that do not meet the tumor classification requirements, and then add annotation information to the screened endoscopic ultrasound images.

[0013] S3. Model Development and Training: Construct an endoscopic ultrasound tumor recognition model and train it using self-supervised learning and ROI detection and classification methods;

[0014] S4. Model parameter optimization: The stochastic gradient descent optimization algorithm is used to fine-tune the parameters of the endoscopic ultrasound tumor recognition model after training.

[0015] S5. Clinical Validation: The performance of the endoscopic ultrasound tumor recognition model after parameter tuning is validated using a clinical application dataset, and the accuracy and robustness evaluation results of gastrointestinal tumor recognition are output.

[0016] The present invention is further configured such that the annotation information added in step S2 includes tumor feature information, tumor type information and tumor location information.

[0017] The present invention is further configured such that step S3, which employs a self-supervised learning method, includes the following steps:

[0018] S311. Adjust the endoscopic ultrasound images after adding annotation information to the preset size;

[0019] S312, Construct a feature encoder G e (·), Feature decoder G d (·) and discriminator G D The self-supervised learning framework (·) adopts image restoration task-driven training. When the resized endoscopic ultrasound image is input into the self-supervised learning framework, a local region of the endoscopic ultrasound image is randomly removed as a mask region, which serves as the true label for the image restoration task.

[0020] S313, via feature encoder G e (·) Convert the resized endoscopic ultrasound image into latent semantic features, and then use the feature decoder G d (·) Reconstruct the latent semantic features into an endoscopic ultrasound image to repair the masked region, using discriminator G. D (·) Determine the similarity between the repaired endoscopic ultrasound image and the original endoscopic ultrasound image, and iteratively train the self-supervised learning framework for the endoscopic ultrasound tumor recognition model to learn the gastrointestinal structure and contextual semantic information of the endoscopic ultrasound image.

[0021] The present invention is further configured such that the preset size uniformly adjusted in step S311 is 256×256 pixels.

[0022] The present invention is further configured such that the ROI detection and classification method in step S3 includes the following steps:

[0023] S321. Use a target detection model to detect tumor regions in the endoscopic ultrasound image after adding annotation information, obtain the tumor region bounding box, and then expand the tumor region bounding box by a preset factor to include sufficient background information.

[0024] S322, Construct the feature encoder F e (·) and classifier, the feature encoder being compared with the feature encoder G in the self-supervised learning method. e (·) Share parameters and simultaneously input the expanded tumor region into the feature encoder F e (·), extract potential tumor features;

[0025] S323. Input the extracted potential tumor features into the classifier, and output the tumor type classification result corresponding to the endoscopic ultrasound image through the classifier.

[0026] The present invention is further configured such that: the target detection model in step S321 is the OSFormer model, and the tumor region bounding box is expanded by a preset factor of 1.5 times to ensure that the expanded tumor region contains sufficient contextual information.

[0027] The present invention is further configured such that: the classifier in step S322 is a three-layer fully connected layer structure, wherein the first layer and the second layer each contain 256 neurons for processing features and extracting deep information, and the third layer contains 2 neurons for outputting the differentiation results between gastrointestinal stromal tumors and leiomyomas.

[0028] The present invention is further configured such that, when the parameters of the endoscopic ultrasound tumor recognition model are optimized in step S4, the specific hyperparameters are set as follows: learning rate of 0.00025, momentum of 0.9, weight decay of 0.0001, batch size of 8, and number of iterations of 20000.

[0029] The present invention is further configured such that: the training of the self-supervised learning method and the ROI detection and classification method in step S3 also includes preprocessing of the endoscopic ultrasound image. The preprocessing operation specifically includes random flipping, random rotation and random cropping of the endoscopic ultrasound image to improve the generalization ability of the model.

[0030] The present invention also provides a context-aware two-stage endoscopic ultrasound tumor identification system, which is applied to a context-aware two-stage endoscopic ultrasound tumor identification method. The system is characterized by including a data processing module, a self-supervised learning module, a tumor identification module, a model optimization module, and a clinical validation module.

[0031] The data processing module is used to collect and screen endoscopic ultrasound images of various types of gastrointestinal tumor cases, remove samples that do not meet the tumor classification requirements, and then add annotation information to the screened endoscopic ultrasound images.

[0032] The self-supervised learning module includes an image preprocessing unit and a self-supervised model training unit. The image preprocessing unit is used to uniformly adjust the endoscopic ultrasound image after adding annotation information to a preset size. The self-supervised model training unit is used to construct a self-supervised learning framework, drive training with an image restoration task, randomly remove a local region of the endoscopic ultrasound image as a mask region, extract latent semantic features of the endoscopic ultrasound image through a feature encoder, reconstruct the endoscopic ultrasound image through a feature decoder to restore the mask region, and judge the similarity between the restored endoscopic ultrasound image and the original endoscopic ultrasound image through a discriminator.

[0033] The tumor identification module includes a ROI detection unit, a feature extraction unit, and a classification unit. The ROI detection unit uses a target detection model to detect tumor regions and expands the tumor region bounding box by a preset factor. The feature extraction unit is used to extract potential features of the expanded tumor region. The classification unit uses a three-layer fully connected layer structure to classify the potential tumor features and output tumor type results.

[0034] The model optimization module is used to fine-tune the parameters of the trained endoscopic ultrasound tumor recognition model using a stochastic gradient descent optimization algorithm.

[0035] The clinical validation module is used to validate the performance of the parameter-tuned endoscopic ultrasound tumor recognition model using a clinical application dataset, and outputs the accuracy and robustness evaluation results of tumor recognition.

[0036] The present invention also provides a context-aware two-stage endoscopic ultrasound tumor identification device, comprising at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor to implement a context-aware two-stage endoscopic ultrasound tumor identification method.

[0037] The present invention also provides a computer-readable storage medium storing computer instructions for execution by a computer to implement a context-aware two-stage endoscopic ultrasound tumor identification method.

[0038] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements a context-aware two-stage endoscopic ultrasound tumor identification method.

[0039] In summary, the present invention has the following beneficial effects:

[0040] 1. In the self-supervised learning stage, this invention uses unlabeled endoscopic ultrasound image data to pre-train the model through an image restoration task, thereby learning the complex contextual information in the endoscopic ultrasound images. The self-supervised learning method enables this invention to not rely on manually labeled data, while effectively extracting the latent semantic features of the images.

[0041] 2. In the ROI detection and classification stage, this invention uses a target detection model to accurately identify and classify tumor regions in endoscopic ultrasound images. This stage utilizes the contextual information learned in the self-supervised learning stage to increase the reference of the contextual background by expanding the boundary of the tumor region, thereby further improving the accuracy and reliability of classification. Through this two-stage training strategy, this invention can significantly improve the ability to identify tumors such as gastrointestinal stromal tumors (GIST) and leiomyomas, while reducing the misdiagnosis rate.

[0042] 3. This invention not only optimizes the workflow of tumor detection, but also enhances the generalization ability of the model, enabling it to be effectively used in diverse clinical environments. Furthermore, clinical validation has demonstrated that this invention exhibits high classification accuracy and robustness in practical applications, showcasing broad application prospects, especially its potential value in early screening and precision medicine. Attached Figure Description

[0043] Figure 1 This is a data collection flowchart of a context-aware two-stage endoscopic ultrasound tumor identification method according to Embodiment 1 of the present invention;

[0044] Figure 2 This is a flowchart illustrating the overall structure of a context-aware two-stage endoscopic ultrasound tumor identification method according to Embodiment 1 of the present invention.

[0045] Figure 3 This is a schematic diagram of the self-supervised learning process in Embodiment 1 of the present invention;

[0046] Figure 4 This is a schematic diagram showing the comparison between the performance of a context-aware two-stage endoscopic ultrasound tumor identification method and the tumor classification performance of endoscopists in Embodiment 1 of the present invention. Detailed Implementation

[0047] The following is in conjunction with the appendix Figures 1-4 The present invention will be described in further detail below.

[0048] Example 1: A context-aware two-stage endoscopic ultrasound tumor identification method

[0049] (1) Data collection

[0050] First, a large number of endoscopic ultrasound (EUS) images were collected, covering different types of gastrointestinal tumor cases to ensure data diversity and representativeness. In this embodiment, a model development dataset, an external testing dataset, and a clinical application dataset were constructed. The model development dataset included 1101 samples, which were used for model training and validation. The external testing dataset included 241 samples, which were mainly used to evaluate the model's performance on external data. The clinical application dataset included 59 samples, which were used for the final clinical evaluation.

[0051] In this embodiment, the datasets were all collected through actual clinical diagnostic processes, covering different pathological diagnostic reports and diverse pathological types, ensuring the richness and authenticity of the data.

[0052] (2) Data cleaning and labeling

[0053] During the data cleaning and annotation phase, professional medical imaging experts rigorously screened the collected EUS images to ensure data quality and annotation accuracy. In this embodiment, 472 samples that did not meet the tumor classification requirements were removed during the data cleaning process, leaving 911 samples for model training. The cleaned dataset contained high-quality annotation information, including imaging feature descriptions, disease categories, and tumor location annotations. In the imaging feature descriptions, experts wrote detailed imaging feature descriptions for each sample, providing characteristic information about the tumor. In the disease categories, each sample was classified into different tumor types (such as gastrointestinal stromal tumors or leiomyomas). In the tumor location annotations, the precise location of the tumor in each EUS image was marked for use by the model's region detection module.

[0054] (3) Model development and training

[0055] Model development includes a first-stage self-supervised learning and a second-stage ROI-based tumor classification. In the first stage of model training, this embodiment uses a self-supervised learning method to pre-learn the latent semantic information in EUS images without requiring a large amount of labeled data. The implementation steps are as follows: all labeled EUS images are adjusted to 256×256 pixels to unify the image size for model processing; an image inpainting task is used as the basis for self-supervised learning. Specifically, when the model is given an input image, a local region (as a "mask region") is randomly removed. This region is used to simulate the ground truth in the image inpainting task; the model learns the contextual semantic information of the image to predict the removed pixels. This process relies on the following three sub-models:

[0056] Feature encoder G e (·): Used to convert image information into latent semantic features; in this embodiment, the encoder uses the VGG16 model to ensure its ability to extract complex image features.

[0057] Feature decoder G d (·): Used to convert encoded latent features back into image information for image restoration;

[0058] Discriminator G D (·): Used to determine the similarity between the repaired image and the original image, ensuring that the decoded image is consistent with the real image.

[0059] In this embodiment, the unique feature of the first stage of self-supervised learning is that it can use EUS images that do not contain tumor information to focus on learning the structural features of the gastrointestinal tract. The first stage used 911 samples (all samples in the model development dataset) and performed a 7:3 ratio of training and validation to make full use of the existing data.

[0060] In the second stage, the focus of this embodiment is on the accurate detection and classification of tumor regions (ROIs) in EUS images; in ROI detection, the target detection model F is first used. ROI (·) Detect the region where the tumor is located. The OSFormer model is an advanced object segmentation model that can accurately locate the boundaries of the tumor. In order to capture more contextual information, the tumor region bounding box predicted by OSFormer is expanded by 1.5 times to ensure that it contains enough background information for subsequent classification.

[0061] During feature extraction, the detected tumor regions are input into the feature encoder F. e In (·), F e (·) and the first-stage feature encoder G e (·) Sharing the same parameters ensures that the semantic information learned in the early self-supervised learning is fully utilized.

[0062] In the classification task, the extracted tumor features are classified by a classifier with three fully connected layers. The first and second layers of the classifier each contain 256 neurons to process the input features and extract deeper information. The last layer contains 2 neurons and outputs results to distinguish between gastrointestinal stromal tumors (GIST) and leiomyomas.

[0063] In this embodiment, data augmentation techniques (such as image flipping, random rotation, and cropping) were used in the model training at this stage to further increase the diversity of training data, prevent model overfitting, and ensure that the final classification results have high accuracy and robustness.

[0064] (4) Model parameter optimization

[0065] To ensure the efficiency and accuracy of the model, this embodiment determined the optimal hyperparameter settings through multiple experiments. During model training, the stochastic gradient descent (SGD) optimization algorithm was used, and the specific hyperparameter settings are as follows:

[0066] Learning rate: 0.00025;

[0067] Momentum: 0.9;

[0068] Weight decay: 0.0001;

[0069] Batch size: 8;

[0070] Number of iterations: 20,000.

[0071] After these parameters were optimized, the model in this embodiment performed excellently on internal and external test datasets, especially achieving high accuracy in tumor detection and classification.

[0072] (5) Clinical validation

[0073] The model was finally validated using 59 clinical samples. Validation results showed that the method of this invention can accurately distinguish between different types of gastrointestinal tumors and exhibits high robustness in detecting small or complex-shaped tumors. Clinical evaluation showed that the EUS image tumor detection and classification method of this invention has high practicality and reliability in real-world applications. Experimental results are as follows: Figure 4 As shown, this paper compares the detection and classification performance of two methods for gastrointestinal stromal tumors (GISTs) and leiomyomas. Figure 4 This paper compares the performance of the self-supervised learning and two-stage model training method proposed in this invention (blue bar) with that of a human endoscopist (light purple bar) in identifying gastrointestinal stromal tumors (GISTs) and leiomyomas. In GIST identification, the method in this embodiment outperforms the human endoscopist in all performance metrics, including accuracy (ACC), sensitivity (SE), specificity (SP), positive predictive value (PPV), negative predictive value (NPV), and AUC. In leiomyoma detection, although the performance of both methods is similar, the technique of this invention shows a slight advantage in specificity and AUC, demonstrating its potential advantage in automated tumor identification. This indicates that the method of this invention can effectively improve the accuracy and reliability of diagnosis, reduce human error, and improve the efficiency of medical diagnosis when handling these challenging medical imaging tasks.

[0074] Example 2: A context-aware two-stage endoscopic ultrasound tumor recognition system, comprising a data processing module, a self-supervised learning module, a tumor recognition module, a model optimization module, and a clinical validation module; the data processing module collects and filters endoscopic ultrasound images of various types of gastrointestinal tumor cases, removing samples that do not meet the tumor classification requirements, and then adds annotation information to the filtered endoscopic ultrasound images; the self-supervised learning module includes an image preprocessing unit and a self-supervised model training unit; the image preprocessing unit adjusts the annotated endoscopic ultrasound images to a preset size; the self-supervised model training unit constructs a self-supervised learning framework, uses an image restoration task to drive training, randomly removes a local region of the endoscopic ultrasound image as a mask region, and extracts latent semantic features of the endoscopic ultrasound image through a feature encoder. The endoscopic ultrasound image is reconstructed using a feature decoder to repair the masked region. A discriminator then judges the similarity between the repaired endoscopic ultrasound image and the original endoscopic ultrasound image. The tumor recognition module includes a ROI detection unit, a feature extraction unit, and a classification unit. The ROI detection unit uses a target detection model to detect the tumor region and expands the tumor region bounding box by a preset factor. The feature extraction unit extracts the potential features of the expanded tumor region. The classification unit uses a three-layer fully connected layer structure to classify the potential tumor features and output the tumor type result. The model optimization module uses a stochastic gradient descent optimization algorithm to fine-tune the parameters of the trained endoscopic ultrasound tumor recognition model. The clinical validation module uses a clinical application dataset to validate the performance of the parameter-tuned endoscopic ultrasound tumor recognition model and outputs the accuracy and robustness evaluation results of tumor recognition.

[0075] Example 3: A context-aware two-stage endoscopic ultrasound tumor identification device includes at least one processor and a memory communicatively connected to the at least one processor. The memory stores instructions executable by the processor to implement a context-aware two-stage endoscopic ultrasound tumor identification method. This method includes: collecting endoscopic ultrasound images of various types of gastrointestinal tumor cases, covering different pathological diagnostic reports and diverse pathological types; having professional medical imaging experts screen the collected endoscopic ultrasound images, removing samples that do not meet tumor classification requirements, and then adding annotation information to the screened endoscopic ultrasound images; constructing an endoscopic ultrasound tumor identification model and training it using a self-supervised learning method and a ROI detection and classification method; using a stochastic gradient descent optimization algorithm to fine-tune the parameters of the trained endoscopic ultrasound tumor identification model; and using a clinical application dataset to verify the performance of the parameter-tuned endoscopic ultrasound tumor identification model, outputting the accuracy and robustness evaluation results for gastrointestinal tumor identification.

[0076] Example 4: A computer-readable storage medium storing computer instructions for execution by a computer to implement a context-aware two-stage endoscopic ultrasound tumor identification method. The method includes: collecting endoscopic ultrasound images of various types of gastrointestinal tumor cases, encompassing different pathological reports and diverse pathological types; having professional medical imaging experts screen the collected endoscopic ultrasound images, removing samples that do not meet tumor classification requirements, and then adding annotation information to the screened endoscopic ultrasound images; constructing an endoscopic ultrasound tumor identification model and training it using a self-supervised learning method and a ROI detection and classification method; using a stochastic gradient descent optimization algorithm to fine-tune the parameters of the trained endoscopic ultrasound tumor identification model; and using a clinical application dataset to verify the performance of the parameter-tuned endoscopic ultrasound tumor identification model, outputting the accuracy and robustness evaluation results for gastrointestinal tumor identification.

[0077] Example 5: A computer program product, comprising a computer program, which, when executed by a processor, implements a context-aware two-stage endoscopic ultrasound tumor identification method. The method includes: collecting endoscopic ultrasound images of various types of gastrointestinal tumor cases, with these images covering different pathological reports and diverse pathological types; having professional medical imaging experts screen the collected endoscopic ultrasound images, removing samples that do not meet tumor classification requirements, and then adding annotation information to the screened endoscopic ultrasound images; constructing an endoscopic ultrasound tumor identification model and training it using a self-supervised learning method and a ROI detection classification method; using a stochastic gradient descent optimization algorithm to fine-tune the parameters of the trained endoscopic ultrasound tumor identification model; and using a clinical application dataset to verify the performance of the parameter-tuned endoscopic ultrasound tumor identification model, outputting the accuracy and robustness evaluation results for gastrointestinal tumor identification.

[0078] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

Claims

1. A context-aware two-stage endoscopic ultrasound tumor identification method, characterized in that: Includes the following steps: S1. Data collection: Collect endoscopic ultrasound images of various types of gastrointestinal tumor cases, which cover different pathological diagnostic reports and diverse pathological types; S2. Data cleaning and annotation: Professional medical imaging experts screen the collected endoscopic ultrasound images, remove samples that do not meet the tumor classification requirements, and then add annotation information to the screened endoscopic ultrasound images. S3. Model Development and Training: Construct an endoscopic ultrasound tumor recognition model and train it using self-supervised learning and ROI detection and classification methods; S4. Model parameter optimization: The stochastic gradient descent optimization algorithm is used to fine-tune the parameters of the endoscopic ultrasound tumor recognition model after training. S5. Clinical Validation: The performance of the endoscopic ultrasound tumor recognition model after parameter tuning is validated using a clinical application dataset, and the accuracy and robustness evaluation results of gastrointestinal tumor recognition are output.

2. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 1, characterized in that: The annotation information added in step S2 includes tumor characteristic information, tumor type information, and tumor location information.

3. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 1, characterized in that: The self-supervised learning method used in step S3 includes the following steps: S311. Adjust the endoscopic ultrasound images after adding annotation information to the preset size; S312, Construct a feature encoder G e (·), Feature decoder G d (·) and discriminator G D The self-supervised learning framework (·) adopts image restoration task-driven training. When the resized endoscopic ultrasound image is input into the self-supervised learning framework, a local region of the endoscopic ultrasound image is randomly removed as a mask region, which serves as the true label for the image restoration task. S313, via feature encoder G e (·) Convert the resized endoscopic ultrasound image into latent semantic features, and then use the feature decoder G d (·) Reconstruct the latent semantic features into an endoscopic ultrasound image to repair the masked region, using discriminator G. D (·) Determine the similarity between the repaired endoscopic ultrasound image and the original endoscopic ultrasound image, and iteratively train the self-supervised learning framework for the endoscopic ultrasound tumor recognition model to learn the gastrointestinal structure and contextual semantic information of the endoscopic ultrasound image.

4. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 3, characterized in that: The preset size uniformly adjusted in step S311 is 256×256 pixels.

5. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 3, characterized in that: The ROI detection and classification method used in step S3 includes the following steps: S321. Use a target detection model to detect tumor regions in the endoscopic ultrasound image after adding annotation information, obtain the tumor region bounding box, and then expand the tumor region bounding box by a preset factor to include sufficient background information. S322, Construct the feature encoder F e (·) and classifier, the feature encoder being compared with the feature encoder G in the self-supervised learning method. e (·) Share parameters and simultaneously input the expanded tumor region into the feature encoder F e (·), extract potential tumor features; S323. Input the extracted potential tumor features into the classifier, and output the tumor type classification result corresponding to the endoscopic ultrasound image through the classifier.

6. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 5, characterized in that: In step S321, the target detection model is the OSFormer model, and the tumor region bounding box is expanded by a preset factor of 1.5 times to ensure that the expanded tumor region contains sufficient contextual information.

7. A context-aware two-stage endoscopic ultrasound tumor identification method according to claim 5, characterized in that: In step S322, the classifier is a three-layer fully connected layer structure. The first and second layers each contain 256 neurons, which are used to process features and extract deep information. The third layer contains 2 neurons, which are used to output the differentiation results between gastrointestinal stromal tumors and leiomyomas.

8. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 1, characterized in that: In step S4, when optimizing the parameters of the endoscopic ultrasound tumor recognition model, the specific hyperparameters are set as follows: learning rate of 0.00025, momentum of 0.9, weight decay of 0.0001, batch size of 8, and number of iterations of 20000.

9. The context-aware two-stage endoscopic ultrasound tumor identification method according to claim 1, characterized in that: The training process in step S3, which employs self-supervised learning and ROI detection and classification methods, also includes preprocessing the endoscopic ultrasound images. The preprocessing operations specifically include randomly flipping, rotating, and cropping the endoscopic ultrasound images to improve the model's generalization ability.

10. A context-aware two-stage endoscopic ultrasound tumor identification system, applied to the context-aware two-stage endoscopic ultrasound tumor identification method as described in any one of claims 1-9, characterized in that: It includes a data processing module, a self-supervised learning module, a tumor identification module, a model optimization module, and a clinical validation module; The data processing module is used to collect and screen endoscopic ultrasound images of various types of gastrointestinal tumor cases, remove samples that do not meet the tumor classification requirements, and then add annotation information to the screened endoscopic ultrasound images. The self-supervised learning module includes an image preprocessing unit and a self-supervised model training unit. The image preprocessing unit is used to uniformly adjust the endoscopic ultrasound image after adding annotation information to a preset size. The self-supervised model training unit is used to construct a self-supervised learning framework, drive training with an image restoration task, randomly remove a local region of the endoscopic ultrasound image as a mask region, extract latent semantic features of the endoscopic ultrasound image through a feature encoder, reconstruct the endoscopic ultrasound image through a feature decoder to restore the mask region, and judge the similarity between the restored endoscopic ultrasound image and the original endoscopic ultrasound image through a discriminator. The tumor identification module includes a ROI detection unit, a feature extraction unit, and a classification unit. The ROI detection unit uses a target detection model to detect tumor regions and expands the tumor region bounding box by a preset factor. The feature extraction unit is used to extract potential features of the expanded tumor region. The classification unit uses a three-layer fully connected layer structure to classify the potential tumor features and output tumor type results. The model optimization module is used to fine-tune the parameters of the trained endoscopic ultrasound tumor recognition model using a stochastic gradient descent optimization algorithm. The clinical validation module is used to validate the performance of the parameter-tuned endoscopic ultrasound tumor recognition model using a clinical application dataset, and outputs the accuracy and robustness evaluation results of tumor recognition.

11. A context-aware two-stage endoscopic ultrasound tumor identification device, characterized in that: It includes at least one processor; and a memory communicatively connected to at least one of the processors; wherein the memory stores instructions executable by the processor to implement a context-aware two-stage endoscopic ultrasound tumor identification method as described in any one of claims 1-9.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions that are executed by the computer to implement a context-aware two-stage endoscopic ultrasound tumor identification method as described in any one of claims 1-9.

13. A computer program product, characterized in that: Includes a computer program, which, when executed by a processor, implements a context-aware two-stage endoscopic ultrasound tumor identification method as described in any one of claims 1-9.

Images