Crohn disease intestinal stenosis type identification method based on radiomics and deep learning

Abstract

The invention relates to the field of medical image processing, and discloses a Crohn disease intestinal stenosis type identification method based on radiomics and deep learning, which comprises the following steps: marking a region of interest of an intestinal stenosis section through standardized ultrasonic gray-scale image acquisition, and extracting radiomics features and deep learning features; constructing a classification model based on image omics by using a random forest algorithm, and extracting deep features and constructing a classification model by using a ResNet50 deep learning network; in order to further improve the classification performance, in combination with the results of the radiomics model and the deep learning model, multi-modal classification is realized through weighted fusion; the classification result is used for distinguishing inflammatory stenosis and fibrous stenosis of the intestinal tract of the Crohn's disease. According to the method, the dependence of traditional ultrasonic diagnosis on experience of operators is overcome, the diagnosis consistency and accuracy are remarkably improved, and reliable support is provided for non-invasive accurate diagnosis of the Crohn disease intestinal stenosis type.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and specifically to a method for identifying the intestinal stenosis types of Crohn's disease based on radiomics and deep learning. Background Art

[0002] Crohn's Disease (CD) is a chronic inflammatory bowel disease with the characteristics of protraction and recurrence. Its main feature is the transmural inflammation involving the entire intestinal wall. During the disease progression, more than half of the patients will develop intestinal stenosis, and the stenosis types mainly include inflammatory stenosis and fibrotic stenosis. The accurate distinction between these two stenosis types is crucial because their treatment strategies are completely different. Inflammatory stenosis can usually be relieved by drug treatment, while fibrotic stenosis mostly requires non-drug intervention means, such as endoscopic balloon dilation, strictureplasty or surgical resection. Therefore, accurately distinguishing the stenosis types is of great significance for optimizing the treatment plan and improving the prognosis of patients.

[0003] However, the current diagnostic methods still have significant limitations in identifying the intestinal stenosis types of Crohn's disease. First, since the formation of fibrotic stenosis mainly involves intestinal wall fibrosis, and the fibrosis process cannot be directly detected by conventional endoscopy or tissue biopsy, clinically, cross-sectional imaging techniques are mainly relied on for evaluation. For example, magnetic resonance imaging (MRI), computed tomography enterography (CTE) and intestinal ultrasound examination (IUS) have potential application value in evaluating the fibrosis degree of intestinal stenosis. Among them, intestinal ultrasound is recommended as the preferred method for the diagnosis and monitoring of Crohn's disease by multiple guidelines because of its non-invasive, simple and highly repeatable characteristics. However, although ultrasound has certain capabilities in detecting intestinal stenosis and indicating the characteristics of intestinal wall tissue, its diagnostic accuracy is still limited by many aspects.

[0004] At present, the main challenges in differentiating stenosis types by intestinal ultrasound include the following aspects: First, the results of ultrasound examination highly depend on the experience level of the operator, and there are significant differences in the results among different operators. Second, the evaluation criteria for ultrasound images have not been fully standardized, lacking consistency and repeatability, making it difficult to accurately identify the stenosis types. In addition, traditional ultrasound diagnostic methods usually infer the stenosis types through intuitive image observation, with problems of strong subjectivity and insufficient capture of detailed information. These deficiencies significantly affect the application effect of ultrasound in the diagnosis of Crohn's disease stenosis types.

[0005] In recent years, radiomics and deep learning techniques have shown great potential in medical image analysis. These techniques can extract deep information that is difficult to observe intuitively from medical images through high-throughput feature extraction and automated pattern recognition. However, the application of these techniques in differentiating the types of intestinal strictures in Crohn's disease is still relatively limited. Radiomics techniques can extract statistical, texture, and shape features of images, but in traditional implementations, it is still necessary to manually annotate the regions of interest (ROIs), which is a complex process and has a certain degree of subjectivity. In contrast, deep learning techniques can capture more complex image patterns through automated feature extraction, but their application requires a large amount of high-quality image data support, and the interpretability of the results is relatively weak.

[0006] Therefore, aiming at the deficiencies of the existing technologies, how to use radiomics and deep learning techniques to extract features and perform classification analysis on intestinal ultrasound images, reduce operator dependence, improve diagnostic consistency and accuracy, and at the same time achieve accurate differentiation of the types of strictures is an important technical problem that urgently needs to be solved at present. Summary of the Invention

[0007] Aiming at the deficiencies of the existing technologies, the present invention provides a method for identifying the types of intestinal strictures in Crohn's disease based on radiomics and deep learning, which solves the problems in the existing technologies that the ultrasonic diagnosis is highly dependent on the operator's experience and lacks standardization, and cannot accurately distinguish inflammatory strictures and fibrotic strictures of the intestine in Crohn's disease.

[0008] To achieve the above objectives, the present invention is realized through the following technical solutions: A method for identifying the types of intestinal strictures in Crohn's disease based on radiomics and deep learning, comprising the following steps: Obtain the ultrasonic gray-scale images of the intestines of Crohn's disease patients, use standard cross-sections to image the strictured segments of the intestine, and annotate the regions of interest in the images; Extract features based on the gray-scale images, including quantitative features extracted by radiomics methods and features automatically extracted by deep learning networks; Respectively use the features extracted by radiomics methods and the features extracted by deep learning networks to construct a classification model for differentiating inflammatory strictures and fibrotic strictures; Identify the types of intestinal strictures based on the output results of the classification model.

[0009] Preferably, the acquisition of the ultrasonic gray-scale images includes the following steps: Use a convex array probe and a linear array probe in combination to comprehensively scan the intestines of Crohn's disease patients; Manually adjust the gain of the ultrasonic device and save the typical longitudinal section images of the intestinal stricture areas.

[0010] Preferably, the annotation of the region of interest is completed by an image processing tool. The region of interest includes the intestinal wall part of the narrow region and excludes the irrelevant regions in the image.

[0011] Preferably, the quantitative features extracted by the radiomics method include: Extract first-order statistical features through the PyRadiomics library, including mean, variance, and skewness; Extract texture features, including gray-level co-occurrence matrix, gray-level run length matrix, and gray-level size zone matrix features; Extract shape features, including area and aspect ratio.

[0012] Preferably, the selection of the quantitative features extracted by the radiomics method is completed through Pearson correlation coefficient analysis, and only the features significantly related to the classification task are retained, with the significance level set at p < 0.05.

[0013] Preferably, the deep learning network is a convolutional neural network, including: Convolutional layer, used to extract low-level features from ultrasonic gray-scale images; Residual block, used to reduce the problem of gradient vanishing through skip connections; Global average pooling layer, used to reduce the feature dimension and reduce overfitting; Fully connected layer, used to generate classification outputs.

[0014] Preferably, during the training process of the deep learning network, data augmentation is performed on the input ultrasonic gray-scale images to improve the generalization ability of the model.

[0015] Preferably, the performance evaluation of the classification model includes the following indicators: accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and area under the receiver operating characteristic curve.

[0016] Preferably, the features extracted by the radiomics method and the features extracted by the deep learning network are integrated through a fusion model, and the fusion model assigns different weights to the two types of features to optimize the classification performance.

[0017] Preferably, the output of the classification model is used to label the intestinal stenosis type as inflammatory stenosis or fibrotic stenosis.

[0018] The present invention provides a method for identifying intestinal stenosis types of Crohn's disease based on radiomics and deep learning. It has the following beneficial effects: 1. The present invention solves the problems of inconsistent image quality and sections caused by strong operator dependence in traditional ultrasonic diagnosis by establishing a standardized ultrasonic gray-scale image acquisition process. By adopting the technology of standard sections, the imaging specifications of the intestinal stenosis segments are clarified and unified, making the extracted image data more consistent and ensuring the accuracy of subsequent feature extraction and classification models.

[0019] 2. The present invention innovatively combines radiomics features and deep learning features. It not only extracts interpretable statistical and geometric features by radiomics methods but also captures complex deep characteristic information through deep learning networks. This multi-modal feature fusion effectively makes up for the limitations of a single method and greatly improves the ability of the classification model to distinguish between inflammatory and fibrotic stenoses.

[0020] 3. Traditional ultrasonic examinations have high requirements for the operator's experience and are prone to poor observer consistency. The present invention reduces human operation errors through automated feature extraction and classification technologies, making the diagnostic process more objective and standardized, thus improving the reliability and repeatability of diagnostic results.

[0021] 4. The non-invasive diagnostic method of the present invention based on ultrasonic images is more convenient, economical and suitable for wide application compared with other imaging technologies (such as CT or MRI). Combined with artificial intelligence analysis, the present invention can quickly and accurately identify the stenosis type, provide efficient support for the diagnostic decisions of clinicians, and avoid the invasive harm and high cost that may be brought by traditional methods.

[0022] 5. By combining radiomics and deep learning technologies, the present invention further expands the application scope of ultrasonic diagnosis, developing it from traditional manual interpretation to an automated analysis system relying on artificial intelligence assistance. This intelligent technology not only improves the diagnostic efficiency but also reduces the technical threshold for operators, creating conditions for the popularization and promotion of ultrasonic technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a schematic flow chart of the method of the present invention; Figure 2 is a schematic flow chart of feature extraction based on the radiomics method of the present invention; Figure 3 is a schematic flow chart of feature extraction based on the Resnet50 model of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0024] Next, in conjunction with the accompanying drawings of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] Please refer to the attached Figure 1 - attached Figure 3 , the present invention provides a method for identifying the types of intestinal strictures in Crohn's disease based on radiomics and deep learning, aiming to accurately distinguish inflammatory strictures and fibrotic strictures through non-invasive techniques.

[0026] As Figure 1 shown, the method for identifying the types of intestinal strictures in Crohn's disease based on radiomics and deep learning may include the following steps: S1. Obtain the ultrasonic grayscale images of the intestines of Crohn's disease patients; S2. Extract features based on the ultrasonic grayscale images; S3. Construct a classification model for distinguishing inflammatory strictures and fibrotic strictures; S4. Identify the types of intestinal strictures based on the output results of the classification model.

[0027] For step S1, in this embodiment, obtaining the ultrasonic grayscale images of the intestines of Crohn's disease patients is carried out through standardized ultrasonic scanning techniques and image acquisition methods to ensure clear and consistent image data, providing high-quality basic data for subsequent feature extraction and model training.

[0028] It should be noted that the ultrasonic grayscale images are obtained using high-resolution ultrasonic equipment, including but not limited to Philips iU22 ultrasonic instruments and SuperSonic Aixplorer ultrasonic instruments. As an option, a combination of convex array (C5-2) and linear array (L9-3) probes is used for image acquisition. Specifically, the convex array probe is used for preliminary scanning of the overall range of the patient's intestines, and the linear array probe is used for local detailed examination, especially for the intestinal wall area of the stricture segment.

[0029] In a possible implementation, the patient needs to fast for at least 8 hours before the examination to reduce the interference of intestinal gas and ensure clear images. The patient lies in the supine position for the examination and is adjusted to the left lateral position if necessary to obtain the best imaging effect of the intestinal tract. Specifically, the ultrasonic probe contacts the patient's intestinal tract area through the abdominal skin and scans the colon and small intestine segment by segment, including segments such as the ileocecal region, ascending colon, transverse colon, descending colon, and sigmoid colon.

[0030] In some embodiments, to ensure the consistency and standardization of images, standard section techniques are used in this embodiment to obtain ultrasonic grayscale images. Specifically, the area with the most obvious intestinal wall stenosis is selected, and the gain, depth, and focus of the ultrasonic device are adjusted to ensure clear images with high resolution and capable of clearly showing the intestinal wall structure.

[0031] It should be noted that in the selection of standard sections, images of intestinal stenosis segments with typical features are preferentially retained, such as intestinal wall thickening and changes in the layered structure. At the same time, image acquisition parameters (such as gain value, depth, focus position, etc.) need to be recorded to ensure data consistency.

[0032] Exemplarily, in actual operation, the ultrasonic grayscale images of the intestinal wall stenosis area can show the thickness of the intestinal wall, echo characteristics, and the degree of intestinal lumen stenosis. These parameters are crucial for subsequent feature extraction and classification modeling. For example, by adjusting the high-frequency mode of the linear array probe, the layered condition of the intestinal wall in the stenosis area can be observed more precisely.

[0033] In a possible implementation, the obtained grayscale images are annotated with regions of interest (ROI) in image processing software (such as ImageJ). Specifically, the regions of interest include intestinal stenosis segments, while excluding possible interfering tissues, such as intestinal lumen contents and surrounding adipose tissues. The annotation process is usually manually completed by professional operators to ensure the accuracy of the annotated areas. It should be noted that the annotation of ROI is crucial for subsequent radiomics feature extraction and the training of deep learning models.

[0034] It can be understood that to further standardize the data, preprocessing steps such as denoising and contrast enhancement can be applied to the grayscale image data. For example, algorithms based on histogram equalization are used to enhance image details, or Gaussian filtering is used to remove noise, thereby improving the accuracy of subsequent feature extraction.

[0035] In this embodiment, the obtained grayscale images are finally saved in a standard format (such as DICOM or PNG format) for subsequent radiomics and deep learning analysis. These image data will be characterized by high quality and consistency, ensuring the reproducibility of subsequent analysis processes.

[0036] It should be emphasized that the key to this step is to obtain high-resolution and diagnostically significant grayscale image data through standardized ultrasonic scanning techniques and section selection, while accurately annotating the regions of interest.

[0037] For step S2, in this embodiment, radiomics methods and deep learning methods are used to extract features from the intestinal ultrasonic grayscale images of Crohn's disease patients to provide high-dimensional feature data for subsequent classification models. The steps include two parts: radiomics-based feature extraction and deep learning-based feature extraction.

[0038] Radiomics-based feature extraction In one embodiment of the present invention, one or several most representative ultrasound images of each patient are selected from the original dataset, and the processed dataset is generated after cropping and normalization preprocessing. The final dataset contains images of 87 patients, and the training set and the test set are divided in a ratio of 8:2.

[0039] Specifically, the image cropping is based on the annotation results of the target box (ROI), and only the intestinal stenosis area is retained. The SimpleITK tool is used to read and crop the images to ensure the consistency of the input data.

[0040] In the feature extraction stage, the PyRadiomics library of Python is used to extract radiomics features from the cropped grayscale images. The extracted features include but are not limited to the following categories: First-order statistical features: reflecting the basic statistical characteristics of the gray-scale distribution, including mean, standard deviation, skewness, and kurtosis.

[0041] Texture features: quantifying texture characteristics based on the spatial relationship of gray-scale pixels in the image, mainly including: Gray-level co-occurrence matrix (GLCM) features: such as contrast, energy, entropy, and uniformity, used to describe the spatial relationship between pixel gray levels.

[0042] Gray-level run length matrix (GLRLM) features: such as short run high intensity ratio and long run emphasis intensity, quantifying the characteristics of the continuous gray-scale pixel distribution.

[0043] Gray-level size zone matrix (GLSZM) features: such as regional size mean, describing the spatial distribution characteristics of similar gray-scale regions in the texture.

[0044] Shape features: such as volume, area, aspect ratio, etc., used to describe the geometric characteristics of the stenosis area.

[0045] It should be noted that in this embodiment, more than 1100 high-dimensional radiomics features are extracted from the ROI of each image. These features comprehensively quantify the gray-scale, texture, and shape information of the image, providing rich input data for the subsequent classification model.

[0046] To further optimize the feature set, the Pearson correlation coefficient is used to screen the features. Specifically, the correlation between each feature and the target variable is calculated, and the features with low correlation or redundancy are removed, and only the features with statistical significance p < 0.05 are retained for the training of the classification model.

[0047] Such as Figure 2As shown, the process of feature extraction is presented, including feature selection and classification model construction. After feature selection, a high-dimensional feature vector containing key features is finally obtained for subsequent classification.

[0048] Feature Extraction Based on Deep Learning In this embodiment, automatic feature extraction of grayscale images is performed through deep learning methods. The deep learning model adopted is ResNet50, which can capture complex features related to stenosis types through multi-layer convolutional operations.

[0049] It should be noted that the design of the ResNet50 model uses residual blocks (Residual Blocks), and its core lies in introducing skip connections. By directly passing the input to the output, the problem of gradient disappearance caused by the increase in network depth is avoided, thereby improving the learning efficiency of the model.

[0050] Specifically, the main structure of the ResNet50 model includes: Convolutional layer: Used to extract basic features of the image (such as edges and textures).

[0051] Residual block: Effectively transmits information through skip connections, improving the training effect and convergence speed of deep networks.

[0052] Global Average Pooling layer (GAP): Averages the pixel values of each feature map to reduce the feature dimension and the risk of overfitting.

[0053] Fully connected layer: Maps the extracted high-dimensional features to classification results.

[0054] The input data is the cropped ROI image. To improve the robustness and generalization ability of the model, data augmentation is performed on the input image during the training process, including operations such as random rotation, flipping, and adding noise. The processed data is divided into a training set and a test set according to an 8:2 ratio, which is consistent with the dataset of the radiomics method.

[0055] As Figure 3 shown, the overall process of deep learning feature extraction is presented. Starting from the convolutional operation of the input data, after passing through multiple residual blocks, a fixed-length feature vector is finally generated for classification tasks.

[0056] In some embodiments, the radiomics and deep learning methods share the same dataset, which contains a total of 87 ultrasound images. The dataset is divided into a training set and a test set according to an 8:2 ratio, which are used for feature extraction, model training, and performance verification respectively.

[0057] It should be noted that the features extracted by radiomics have high interpretability and are suitable for quantifying known structural characteristics in images; while deep learning models can capture more implicit and complex feature patterns. The combination of the two provides data support for the intestinal stricture type classification task.

[0058] In this step, the comprehensive capture of multi-dimensional information of the intestinal stricture area is achieved through feature extraction, where radiomics features quantify explicit statistical information and deep learning features mine implicit pattern information. These features will be used as the input of the classification model, directly determining the accuracy and robustness of subsequent tasks.

[0059] For step S3, in this embodiment, a classification model for distinguishing the intestinal stricture types (inflammatory stricture or fibrotic stricture) of Crohn's disease patients is constructed by using the features extracted by radiomics methods and the features extracted by deep learning networks respectively.

[0060] It should be noted that radiomics features and deep learning features quantify the characteristics of ultrasound gray-scale images from different perspectives. Radiomics features pay more attention to the interpretation of statistical and geometric features, while deep learning features can capture more complex deep pattern information. The combination of the two features enables the classification model to comprehensively and accurately identify the types of intestinal strictures.

[0061] Construction of a classification model based on radiomics features In a possible implementation, after the radiomics features are extracted, the selected highly correlated features are input into a random forest classifier for model training and classification tasks.

[0062] Specifically, random forest is a machine learning algorithm based on the integration of decision trees, which completes classification by voting or weighted averaging the results of multiple decision trees. It should be noted that the design of random forest can effectively alleviate the overfitting problem and improve the generalization ability of the model.

[0063] As an option, the key parameter configurations of random forest include: Number of decision trees: Set to 100 to ensure the robustness of the model.

[0064] Maximum depth: Set to 10 to limit the complexity of a single decision tree and prevent overfitting.

[0065] Partitioning criterion: Use the Gini impurity as the evaluation criterion for splitting nodes.

[0066] Exemplarily, after the radiomics features are extracted, they are divided into a training set and a test set, where the training set is used for parameter learning of the random forest model, and the test set is used to verify the classification performance of the model.

[0067] It is understandable that the random forest achieves accurate classification of intestinal stricture types by means of multiple random samplings and integrated voting while maintaining the stability of the model.

[0068] Construction of a Classification Model Based on Deep Learning Features In another possible implementation, the deep learning features extracted by the ResNet50 model are used as the input to the fully connected layer to complete the classification task. It should be noted that the feature extraction part (convolutional layer and residual block) of the ResNet50 model has been clarified in step S2, and the focus of this step is on the feature classification part.

[0069] Specifically, the core of the deep learning classifier is the Softmax activation function, whose role is to map the features output by the model to the class probability distribution.

[0070] The structural design of the fully connected layer includes the following key parameters: Input dimension: Consistent with the output dimension of the last pooling layer of the ResNet50 model.

[0071] Output dimension: The number of classes in the classification task, that is, 2 (inflammatory stricture and fibrotic stricture).

[0072] Activation function: Use Softmax to convert the output of the fully connected layer into class probabilities.

[0073] During the training process, the cross-entropy loss function is used as the optimization objective.

[0074] It should be noted that the network parameters are updated through an optimization algorithm (such as the Adam optimizer), thereby gradually reducing the training error and improving the classification accuracy.

[0075] Fusion of Multimodal Classification Models In some embodiments, in order to further improve the classification performance, the results of the radiomics model and the deep learning model are fused in this embodiment. The fusion model achieves a comprehensive judgment by weighted averaging the classification results of the two types of features.

[0076] Specifically, the output of the radiomics model is the class probability vector P Radiomics , and the output of the deep learning model is the class probability vector P DL . The final output of the fusion model is: P Fusion = α·P Radiomics + β·P DL where α and β are fusion weights, satisfying α + β = 1. As an option, the fusion weights can be obtained through validation set optimization.

[0077] It is understandable that the multimodal fusion model combines the interpretability of radiomics features and the complexity of deep learning features, and can effectively improve the accuracy and robustness of classification tasks.

[0078] Performance Evaluation of Classification Model In the validation stage of the classification model, the test set is input into the constructed classification model, and the model performance is evaluated through the following metrics: Accuracy: The proportion of correctly classified samples in the total samples.

[0079] Sensitivity: The ability to identify inflammatory strictures.

[0080] Specificity: The ability to identify fibrotic strictures.

[0081] Area Under the Receiver Operating Characteristic Curve (AUC): An index that comprehensively reflects the performance of the classification model.

[0082] Confusion Matrix: Statistics the distribution of the classification results of the model, which is used to analyze the specific types of classification errors.

[0083] It should be noted that through the comprehensive evaluation of these metrics, the effectiveness and reliability of the model for classifying intestinal stricture types can be objectively reflected.

[0084] In this step, a classification model based on radiomics and deep learning is constructed, and the classification performance is further optimized through multimodal fusion. The constructed model can accurately distinguish between inflammatory and fibrotic stricture types, providing support for the clinical diagnosis and subsequent treatment of Crohn's disease patients.

[0085] For step S4, in this embodiment, based on the constructed classification model, the output result of the classification model is used to identify the intestinal stricture type of Crohn's disease patients. The identification process includes the inference operation of the classification model, the interpretation of the output result, and the evaluation of the classification performance.

[0086] It should be noted that the output result of the classification model includes the prediction probabilities for two stricture types (inflammatory stricture and fibrotic stricture), and these probabilities are used to judge the stricture type of the input image and further perform performance verification.

[0087] Inference Operation of Classification Model In a possible implementation, the inference process of the classification model is divided into two cases, respectively for the radiomics model and the deep learning model.

[0088] Specifically, the input of the model based on the radiomics method is a high-dimensional feature vector formed after preprocessing and feature extraction. By passing the input feature vector to the random forest model, the classification result is output. The random forest calculates the probability that the input belongs to each category by integrating decision trees.

[0089] In the model based on the deep learning method, after passing through the feature extraction module of the ResNet50 network, the input data is mapped into a high-dimensional feature vector of a fixed length. The feature vector is then input into the fully connected layer, and the prediction probability of each type of stenosis is calculated through the Softmax activation function.

[0090] As an option, if a multimodal fusion model is adopted, the final output probability of the fusion model is combined in a weighted manner from the output probabilities of the radiomics model and the deep learning model.

[0091] Judgment of the classification result Specifically, after model inference, the stenosis type of the input image is judged according to the output probability value. Exemplarily, when the prediction probability of inflammatory stenosis is greater than the prediction probability of fibrous stenosis, the stenosis type of the input image is judged to be inflammatory stenosis; otherwise, it is judged to be fibrous stenosis.

[0092] It should be noted that the judgment of the classification result can be achieved by setting an adjustable threshold to meet the requirements of different classification scenarios. For example, the default decision threshold of the classification model is 0.5, that is, when the prediction probability of a certain category is greater than 0.5, the image is classified into that category. If it is necessary to improve the sensitivity of a certain category, the classification threshold of that category can be appropriately reduced.

[0093] Evaluation of classification performance In one possible implementation, by inputting the test set into the classification model, comparing the prediction result with the actual category, and calculating the performance metrics of the classification model. The performance metrics include but are not limited to: Accuracy: Defined as the proportion of correctly classified samples in the total samples, and the formula is: Among them, TP is the number of samples that are truly positive and predicted to be positive, TN is the number of samples that are truly negative and predicted to be negative, and FP and FN are the numbers of false positive and false negative samples respectively.

[0094] Sensitivity: Also known as recall rate, used to measure the recognition ability of the model for the positive class (inflammatory stenosis), and the formula is: Specificity: Used to measure the recognition ability of the model for the negative class (fibrous stenosis), and the formula is: Area Under the Receiver Operating Characteristic Curve (AUC): An indicator for comprehensively evaluating the performance of a model, reflecting the classification ability of the model at different classification thresholds.

[0095] Confusion Matrix: Statistically represents the relationship between the prediction results and the true classes of a classification model in matrix form. The matrix contains the distribution of the number of correctly classified and misclassified samples, and can be used to analyze the performance differences of the model in different classes.

[0096] It should be noted that through the evaluation of the above indicators, the performance of the classification model in distinguishing inflammatory stenosis from fibrous stenosis can be comprehensively reflected, and at the same time, it provides a basis for model optimization.

[0097] In some embodiments, the classification results and performance indicators will be stored in a standardized data format (such as CSV or JSON format) for subsequent analysis or display. Specifically, the classification results of each input image include: image name, true class, predicted class, and their corresponding probabilities. The calculation results of the performance indicators include statistical indicators such as accuracy, sensitivity, specificity, and AUC.

[0098] It can be understood that these stored results can be used to generate a performance report of the classification model, providing support for further optimization and application.

[0099] This step realizes the accurate identification of intestinal stenosis types through the inference operation of the classification model and the interpretation of the output results. At the same time, through the comprehensive evaluation of multiple performance indicators, the effectiveness and reliability of the classification model are ensured, laying a foundation for subsequent clinical applications.

[0100] Generally speaking, the present invention extracts key radiomics features and deep learning features of the intestinal stenosis area through standardized ultrasonic gray-scale image acquisition and processing, and constructs classification models using random forest and ResNet50 deep learning networks respectively. At the same time, the present invention combines the interpretability of radiomics features with the complex pattern recognition ability of deep learning features through multi-modal fusion technology, significantly improving the classification accuracy. Through comprehensive performance indicator evaluation, the effectiveness of the model is verified. The present invention overcomes the dependence of traditional ultrasonic diagnosis on the operator's experience, provides reliable support for the non-invasive accurate diagnosis of intestinal stenosis types in Crohn's disease, and provides an important reference for the formulation of personalized treatment plans.

[0101] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying the types of intestinal strictures in Crohn's disease based on radiomics and deep learning, characterized in that, It includes the following steps: Obtain the ultrasonic gray-scale images of the intestines of Crohn's disease patients, image the stenotic segments of the intestinal tract using standard sections, and label the regions of interest in the images; Extract features based on the gray-scale images, including quantitative features extracted by radiomics methods and features automatically extracted by deep learning networks; Respectively use the features extracted by radiomics methods and the features extracted by deep learning networks to construct a classification model for distinguishing inflammatory stenosis and fibrotic stenosis; Identify the type of intestinal stenosis based on the output results of the classification model.

2. The method for identifying the intestinal stricture type of Crohn's disease based on radiomics and deep learning according to claim 1, wherein The acquisition of the ultrasonic gray-scale images includes the following steps: Use a convex array probe and a linear array probe in combination to comprehensively scan the intestines of Crohn's disease patients; Manually adjust the gain of the ultrasonic device and save the typical longitudinal section images of the intestinal stenosis area.

3. The method for identifying the intestinal stenosis type of Crohn's disease based on radiomics and deep learning according to claim 1, wherein The annotation of the regions of interest is completed by an image processing tool. The regions of interest include the intestinal wall parts of the stenotic regions and exclude the irrelevant regions in the images.

4. The method for identifying the intestinal stenosis type of Crohn's disease based on radiomics and deep learning according to claim 1, wherein The quantitative features extracted by the radiomics methods include: Extract first-order statistical features through the PyRadiomics library, including mean, variance, and skewness; Extract texture features, including gray-level co-occurrence matrix, gray-level run length matrix, and gray-level size zone matrix features; Extract shape features, including area and aspect ratio.

5. The method for identifying the intestinal stenosis type of Crohn's disease based on radiomics and deep learning according to claim 1, wherein The selection of the quantitative features extracted by the radiomics methods is completed through Pearson correlation coefficient analysis, and only the features significantly related to the classification task are retained, with the significance level set at p < 0.

05.

6. The method for identifying the intestinal stricture type of Crohn's disease based on radiomics and deep learning according to claim 1, wherein The deep learning network is a convolutional neural network, including: Convolutional layers, used to extract low-level features from ultrasonic gray-scale images; Residual blocks, used to reduce the problem of gradient disappearance through skip connections; Global average pooling layers, used to reduce the feature dimension and reduce overfitting; Fully connected layers, used to generate classification outputs.

7. The method for identifying the intestinal stricture type of Crohn's disease based on radiomics and deep learning according to claim 6, characterized in that During the training process of the deep learning network, data augmentation is performed on the input ultrasonic gray-scale images to improve the generalization ability of the model.

8. The method for identifying the intestinal stricture type of Crohn's disease based on radiomics and deep learning according to claim 1, characterized in that The performance evaluation of the classification model includes the following metrics: accuracy, sensitivity, specificity, positive predictive value, negative predictive value, F1 score, and area under the receiver operating characteristic curve.

9. The method for identifying the intestinal stenosis type of Crohn's disease based on radiomics and deep learning according to claim 1, wherein The features extracted by the radiomics methods and the features extracted by the deep learning network are integrated through a fusion model, and the fusion model assigns different weights to the two types of features to optimize the classification performance.

10. The method for identifying the intestinal stenosis type of Crohn's disease based on radiomics and deep learning according to claim 1, characterized in that The output of the classification model is used to label the type of intestinal stenosis as inflammatory stenosis or fibrotic stenosis.

Images