Method for automatically identifying abnormal image signs of intestinal tract in transabdominal ultrasound image

Abstract

The invention belongs to the technical field of image recognition and classification, and particularly relates to an automatic recognition method for intestinal abnormal image signs of a transabdominal ultrasound image, which is based on an end-to-end dynamic double-branch multi-scale fusion mode, can automatically perform ultrasound region interception on an acquired transabdominal intestinal ultrasound image, and can automatically identify the intestinal abnormal image signs of the transabdominal ultrasound image. And automatically identifying the intestinal abnormality of the intercepted image through a neural network model. The algorithm provided by the invention has relatively high specificity, is simple and efficient in structure, and can be used for accurately identifying the intestinal tract abnormal image in the ultrasonic image in a full-automatic manner. According to the method, an effective computer-aided tool is provided for early recognition, detection and evaluation of intestinal tract abnormity through the ultrasonic image, and powerful support is provided for illness state monitoring and decision making in clinical practice.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and target recognition technology, specifically relating to an automatic method for recognizing abnormal intestinal imaging features in transabdominal ultrasound images. Background Technology

[0002] Intestinal abnormalities are widely seen in various gastrointestinal diseases, typically exhibiting chronic, progressive, and destructive pathological features. Their imaging manifestations are complex and diverse, and can involve multiple intestinal segments. Common abnormal imaging signs include intestinal wall thickening, abnormal intestinal wall layering, intestinal luminal narrowing, and increased echogenicity of periintestinal fat. Individuals with intestinal abnormalities may experience serious complications such as intestinal stenosis, fistulas, abscesses, or perforation in the early stages, significantly impacting quality of life and treatment strategy selection. Therefore, early identification and dynamic assessment of intestinal abnormalities are of great significance. Transabdominal ultrasound, as a non-invasive, real-time imaging, cost-effective, and patient-compliant imaging technique, has been widely used in the detection and follow-up of intestinal abnormalities. This method can visualize changes in intestinal wall structure, intestinal lumen status, and surrounding tissue responses, demonstrating high clinical applicability. However, due to the complexity of the intestinal structure and the influence of various factors on ultrasound image quality, traditional image interpretation is highly dependent on operator experience, exhibiting strong subjectivity and poor repeatability. There is an urgent need to utilize automation technology to improve analysis efficiency and stability. In recent years, deep learning methods such as Convolutional Neural Networks (CNNs) have gradually become important tools in medical image processing due to their powerful feature extraction and image understanding capabilities. In complex ultrasound images, CNNs can automatically capture potential fine-grained features and spatial structural information, providing a new solution for the identification of abnormal intestinal imaging features.

[0003] Current research has attempted to apply general deep learning models to intestinal image analysis tasks, but they generally have the following limitations: First, the models are not optimized for the structural characteristics of intestinal ultrasound images, and their recognition ability is limited when faced with real-world scenarios with large image morphological changes and low signal-to-noise ratios; Second, the image preprocessing stage relies on manual cropping of intestinal regions, which is cumbersome, has a low degree of automation, and is not conducive to efficient deployment in real-world clinical environments. Summary of the Invention

[0004] To address the problems existing in current technologies, this invention discloses an automatic identification method for abnormal intestinal imaging features in transabdominal ultrasound images. This method employs an end-to-end dynamic dual-branch multi-scale fusion approach, constructing a dual-branch feature extraction architecture that focuses on local details and global structural information respectively. A multi-scale fusion mechanism is introduced to enhance the model's ability to integrate features at different levels. Simultaneously, this method achieves fully automated processing from raw image input to recognition result output, eliminating the need for manual annotation or cropping, effectively improving the model's adaptability, accuracy, and clinical usability. The method specifically includes the following steps:

[0005] S1: Obtain the ultrasound image dataset of the transabdominal intestine, label and classify the abnormal and normal images in the ultrasound image dataset, and divide the abnormal and normal images into training set, validation set and test set respectively.

[0006] S2: Automatically identify ultrasound regions in the ultrasound image dataset and automatically detect ultrasound image regions.

[0007] S3: Establish a deep neural network model, which includes a convolutional initiation feature extraction layer, multiple dynamic convolutional multi-scale attention integration modules, and an abnormal image feature recognition module. The dynamic convolutional multi-scale attention integration module includes a local and global feature extraction module, a local and global feature fusion module, and a multi-scale feature extraction module.

[0008] S4: The convolutional initiation feature extraction layer performs preliminary feature extraction on the ultrasound image; the local and global feature extraction module obtains local and global features of the intestinal region from both local structural and global semantic levels, and performs in-depth analysis of the intestinal state in the ultrasound image; the local and global feature fusion module fuses the local and global features to generate a fused feature map, which represents the intestinal features of the suspected lesion area; the multi-scale feature extraction module performs multi-scale modeling on the fused feature map, and extracts the shape and size information of the multi-scale lesion area through different receptive fields;

[0009] S5: The deep neural network model is trained using the training set, and the parameter information of the deep neural network model is updated through forward propagation, loss function optimization, and backpropagation.

[0010] S6: Input the test set into the trained neural network model and evaluate the model's performance metrics; input the original transabdominal intestinal ultrasound images into the deep neural network model and output the abnormal image judgment results.

[0011] Furthermore, S11: A total of 622 transabdominal intestinal ultrasound images were collected using a Canon i800 ultrasound device, including 311 images of intestinal abnormalities and 311 images of normal intestines. The image data were divided according to their purpose of training, validation and testing, with 422 images in the training set (including 82 images in the validation set) and 200 images in the test set, which were used for model training and performance evaluation.

[0012] Further, S21: Read the input transabdominal intestinal ultrasound image file; if the image cannot be read successfully, terminate the current processing flow;

[0013] S22: Perform grayscale processing on the successfully read image: if the image is a color image, convert it to a grayscale image; if it is already a grayscale image, copy and keep it directly.

[0014] S23: Apply the Adaptive Histogram Equalization (CLAHE) algorithm to the grayscale image to enhance local contrast, highlight dark corner areas in the image, and improve the accuracy of subsequent target region segmentation;

[0015] S24: Perform gamma correction on the enhanced image to further improve overall brightness and contrast, and improve the visual quality of the image;

[0016] S25: Binarize the processed image using a fixed threshold to achieve preliminary separation of the target region from the background;

[0017] S26: Perform morphological closing and opening operations on the binary image to remove small noise regions and fill holes in the target region, thereby enhancing the integrity and connectivity of the target region.

[0018] S27: Extract all external contours from the morphologically processed image and select the contour with the largest area as the boundary of the main target region;

[0019] S28: Based on the outline of the main target region, calculate its bounding rectangle to obtain the coordinate range of the region of interest (ROI);

[0020] S29: Based on the coordinates of the bounding rectangle, crop the ROI region from the original image and adjust it to a size of 224×224, which will serve as the target region for subsequent image analysis or saving.

[0021] Furthermore, when building a deep neural network model, the working steps of this network model are as follows: The workflow of the convolutional initiation feature extraction layer is as follows.

[0022] S31: First, features are extracted using a 7×7 convolutional layer. Then, the feature map is fed into a Batch Normalization (BN) layer. The introduction of the BN layer helps reduce internal covariate shifting, accelerates the training process of the neural network, reduces gradient explosion and vanishing problems, and allows for a higher learning rate, improving the model's generalization performance. The calculation process of the BN layer is as follows: Given a set of inputs from a single neural network layer: X = [x1, x2, x3, ... x n ]

[0023] In the formula, x is the number of samples, n is the batch size, and x n Representing the nth sample, first, calculate the mean of the batch elements:

[0024]

[0025] in, It is the mean of the samples within the batch. Let be the i-th sample in the batch, and n be the batch size. Then, calculate the variance of the batch elements:

[0026]

[0027] in, This is the variance of the within-batch samples. Then, normalization is performed on each element in the mini-batch:

[0028]

[0029] in, It is normalized data. Used to prevent the denominator from being zero;

[0030] Finally, the BN layer scales and shifts the normalized output to preserve the network's representational power; this step uses two learnable parameters, γ scaling and β translation, to account for the network's nonlinear expression.

[0031]

[0032] in, It is the output of the BN layer. A learnable scaling factor. It is a learnable bias factor;

[0033] S32: Based on the output obtained through the BN layer Then proceed to the ReLU activation function and the max pooling layer;

[0034] ReLU is a commonly used nonlinear activation function that sets negative values ​​in the input to zero while keeping positive values ​​unchanged, thus introducing nonlinearity.

[0035]

[0036] Max pooling layers are used to downsample feature maps, preserving the maximum values ​​within local regions, thereby reducing the number of parameters and computational cost, and possessing a certain degree of translation invariance for the input feature map. In a window size of And the step size is In this case, the output of max pooling is:

[0037]

[0038] in, Indicates the first The area covered by a pooled window.

[0039] The workflow of the local and global feature extraction modules is as follows:

[0040] S41: Local Feature Extraction Branch: This branch is used to extract fine-grained spatial local information from the input feature map and enhance the local feature representation capability by combining the global context;

[0041] S411: Global Feature Pooling: Based on the last output in S3 Perform global average pooling to obtain global context information:

[0042]

[0043] in, This represents the initial feature map of the input. For the number of channels, For height, For width, Indicates the first Each channel is located in eigenvalues ​​at that location Indicates the first Global context information corresponding to each channel;

[0044] S412: Weight Generation Network: Input global context information into the weight generation network shown below (including 1×1 convolution, GN normalization, GELU activation, etc.), and output the weight coefficients of each convolution kernel. :

[0045]

[0046] in, Indicates the core size is Convolution operation, For group normalization methods, channels are divided into For each group, normalize the features in that group:

[0047]

[0048] in, Indicates the number within the group 1 eigenvalue, Represents the normalized eigenvalues. It is the first The mean of all features within a group. It is the first The variance of all features within a group. A minimal constant introduced to prevent the denominator from being zero;

[0049] The normalized features are linearly recalibrated after scaling and shifting operations, and are defined as follows:

[0050]

[0051] in, A learnable scaling factor. A learnable bias factor. This represents the output feature value after scaling and offset transformation.

[0052] GELU is a non-linear activation function that combines the properties of linear units with the probabilistic characteristics of a Gaussian distribution. It uses a probabilistic mechanism to determine whether to retain an input value, making the neural network more stable and flexible during training.

[0053]

[0054] in, Indicates the input feature value. Represents the hyperbolic tangent function;

[0055] This represents the normalization function, used to map the output to a probability distribution of weighting coefficients. The value range of each component in the output is within... Between these points, the sum of all outputs is 1 for the input vector Z = [z1, z2, z3, ... z2]. n ]:

[0056]

[0057] in, Represents the first input vector. One element, Represents the dimension of a vector. This indicates exponentiation. Indicates the first The weight values ​​of each element after normalization.

[0058] S413: Dynamic kernel weighted fusion: To implement the dynamic convolution mechanism, convolution kernels are randomly initialized. ,in, Indicates the first The static convolution kernel is defined by C1 (number of input channels), C2 (number of output channels), K1 (kernel height), and K2 (kernel width). This is combined with the generated weight coefficients. (Calculated from the aforementioned weight generation network), calculate the weighted and fused dynamic convolutional kernel. Its definition is:

[0059]

[0060] in, Indicates the number generated in S412 The weight coefficients corresponding to the group convolution kernel, It is the number of convolutional kernel groups. Represents the fused dynamic convolutional kernel

[0061] S414: Dynamic Convolution Operation: The dynamic convolution kernel obtained from the weighted fusion in S413 is used to perform a convolution operation on the input preliminary feature map to obtain the output feature map. :

[0062]

[0063] in, Represents the initial feature map of the input. Represents two-dimensional convolution operation, This refers to the edge padding operation performed during convolution to preserve the feature map size. The kernel size;

[0064] S42: Global Feature Extraction Branch: This branch is used to capture the dependencies between distant pixels in the image from the input feature map, helping the network understand high-level information of the entire input;

[0065] S421: Weight generation for query (Q): Preliminary features of the final output from input S3 The query weight matrix is ​​generated through 1×1 convolution:

[0066]

[0067] in, express Convolution operation is used to generate the query weight matrix. ;

[0068] S422: Spatial downsampling: If a sampling ratio is set... ( ), then for Performing spatial downsampling will improve the computational efficiency of attention:

[0069]

[0070] in, express Convolution operation and These represent batch normalization operations, express Convolution operation The activation function of the Gaussian error linear unit. This represents the feature map after downsampling;

[0071] If the sampling ratio is set Then it's not correct. Perform spatial downsampling:

[0072]

[0073] S423: Local Feature Enhancement: Downsampling features obtained from S422 Depthwise separable convolutions are used to enhance the expressive power of local features during global feature extraction. This approach is similar to residual connections, which can optimize feature propagation and improve the network's representational capabilities.

[0074]

[0075] in, express Depthwise separable convolution;

[0076] S424: Weight generation of key values ​​K and V: The updated feature map The weight parameters of the key and value matrices are generated using a 1×1 convolutional layer:

[0077]

[0078] in, This indicates the method used to simultaneously generate key and value weight matrices. Convolution operation; and These are the convolution weights for the keys and values, respectively.

[0079] S425: Attention weight calculation: utilizing the learnable linear projection layer in S421 , and Get the query matrix Key matrix Sum matrix :

[0080]

[0081] Based on the three matrices mentioned above, the attention score is calculated to obtain the global feature map. :

[0082]

[0083] in, The dimension of the key vector is used to scale the inner product result and avoid excessively large gradients. This represents a relative position encoding item, used to introduce spatial position information;

[0084] The workflow of the local and global feature fusion module is as follows:

[0085] S51: Feature Concatenation: Concatenate the local branch feature maps obtained from S41. and the global branch feature map obtained from S42 The input features are concatenated along the channel dimension to form a fused input feature:

[0086]

[0087] in, This represents a feature concatenation operation along the channel dimension;

[0088] S52: Adaptive Weight Generation: To achieve adaptive adjustment of the importance of each branch feature, the concatenated features are... Global average pooling is performed, then the input is fed into the weight generation network, and the output is a fused weight tensor. :

[0089]

[0090] in, , For batch size, For the number of channels, It is the Sigmoid activation function. For average pooling layer, This represents the modified linear unit activation function;

[0091] An activation function is used to map the input to a range of (0,1).

[0092]

[0093] Average pooling is used to divide the input tensor into non-overlapping regions, and the average value of the elements in each region is calculated:

[0094]

[0095] in, It is the input feature map. It is the size of the pooling window. Indicates the position in the window The value;

[0096] S53: Feature Weight Separation: To distinguish the feature weights of different branches, the fusion weight tensor is... The weights are divided along the channel dimension to obtain the weights corresponding to the local and global branches, respectively. and :

[0097]

[0098] in, , For batch size, For the number of channels, These represent the channel weight coefficients for local branches and global branches, respectively.

[0099] S54: Feature weighting: For each feature weight: Multiply by the corresponding weight:

[0100]

[0101]

[0102] By using feature weighting, deep neural network models adaptively adjust the importance of features in each branch based on the input features;

[0103] S55: Feature Fusion Output: The weighted features are concatenated and then... Convolutional processing performs fusion and dimensionality reduction, outputting the final adaptive fused feature map. :

[0104]

[0105] By fusing local branch features and global branch features, the output features combine the advantages of convolution and attention mechanisms, thereby significantly improving the overall feature representation capability.

[0106] The workflow of the multi-scale feature extraction module is as follows:

[0107] S61: Multi-scale dynamic convolution branch initialization: The multi-scale feature extraction module is used to enhance the scale adaptability and structural expressiveness of the fused features. The module initializes four dynamic convolution branches with different kernel sizes, namely 7×7, 5×5, 3×3 and 1×1. Each branch adopts the dynamic convolution structure of the local feature extraction branch in S41.

[0108] S62: Multi-scale feature extraction: Adaptive fusion of feature maps obtained from S5 The input is fed into four dynamic convolutional branches, which generate four sets of feature representations respectively:

[0109]

[0110]

[0111]

[0112]

[0113] in, Indicates having Dynamic convolution operations of convolution kernels can adaptively adjust convolution weights according to input features to improve the model's responsiveness to targets of different scales.

[0114] S63: Feature Concatenation: The four sets of features obtained in S62 are concatenated along the channel dimension to obtain a multi-scale concatenated feature map.

[0115]

[0116] in, For batch size, For the number of channels, , These represent the height and width of the feature map, respectively.

[0117] Feature maps of different scales are concatenated along the channel dimension to generate a feature map that integrates multi-scale information. The concatenated feature map. It retains local and global information from different convolutional kernel sizes, providing richer feature representation capabilities.

[0118] S64: Feature Fusion: The multi-scale spliced ​​feature map from S63 is then processed. Convolution, normalization, and activation operations are performed to fuse multi-scale information, and dimensionality reduction is then applied to obtain a multi-scale fused feature map. :

[0119]

[0120] S65: Residual Connectivity: The multi-scale fused feature map obtained in S5 With adaptive fusion feature maps adjusted by convolution and normalization Perform residual connections to obtain residual feature maps. :

[0121]

[0122] This residual structure enhances the stability of gradient propagation while preserving the original feature information, helping to prevent feature degradation.

[0123] S66: Activate Output: Perform GELU activation on the residual results and output the final features:

[0124]

[0125] The workflow of the abnormal image feature recognition module is as follows:

[0126] S81: Global Average Pooling and Feature Flattening: This applies to the final features of the network. Global average pooling is performed to obtain a global feature vector. Then, the pooled features are flattened into a one-dimensional vector, which serves as the input for subsequent classification.

[0127]

[0128]

[0129] S82: Fully Connected Layer Classification: Flattened Features Input the fully connected layer to calculate the predicted value of the category. This value can be transformed into a probability distribution using the Softmax function, ultimately enabling automatic identification of abnormal intestinal imaging features.

[0130]

[0131]

[0132] in, and These represent the weights and biases of the fully connected layer, It is the predicted category The score, It is the total number of categories. It is a category The predicted probability.

[0133] Furthermore, the loss function of the neural network model is defined as follows: The parameters of the neural network model for automatic recognition of intestinal abnormality imaging features are set, and the cross-entropy loss function is used to calculate the cross-entropy loss, which measures the difference between the predicted labels and the actual results of intestinal abnormality imaging feature recognition. The formula for calculating the cross-entropy loss is as follows:

[0134]

[0135] in, and These are real labels, corresponding to abnormal and normal gut conditions, respectively. and These are the probabilities predicted by the model, representing the probability values ​​for predicting intestinal abnormality and intestinal normality, respectively.

[0136] S102: Network Model Training and Validation: In training the automatic recognition model for abnormal intestinal imaging features, the batch size was set to 32, and the AdamW optimizer and cosine annealing learning rate scheduler were used; the training cycle was 200 times, the initial learning rate was 0.001, the weight decay coefficient was 0.01, and the learning rate was reduced from 0.001 to 1×10 during the warm-up phase. -6 The gradient is calculated and the model parameters are updated using the cross-entropy loss function. After each cycle, the model is validated and the weights of the model with the highest validation accuracy are saved to obtain the best recognition effect.

[0137] S103: Network Model Testing: After training, the automatic recognition model for abnormal intestinal imaging features is tested. First, the model weights with the highest accuracy during the validation process are loaded. Then, transabdominal ultrasound images from the test set are input into the model for prediction. For each test sample, the model outputs a corresponding predicted label, indicating whether the sample is abnormal or normal. The model's performance is quantitatively evaluated using metrics such as accuracy, precision, recall, F1 score, and specificity by calculating the difference between the predicted and true labels. Based on the evaluation results, the effectiveness and stability of the model in practical applications are verified.

[0138] Finally, an intestinal image was acquired using a Canon i800 ultrasound device. The acquired images were automatically cropped into ultrasound regions, the image size was adjusted to 224×224 and standardized, and the pre-trained optimal network model weights were loaded. The pre-processed ultrasound images were then input into an automatic intestinal abnormality image feature recognition network model for prediction. The network model analyzed the images and output the intestinal recognition results.

[0139] This invention discloses an automatic method for identifying intestinal abnormalities in transabdominal ultrasound images. Based on an end-to-end dynamic bi-branch multi-scale fusion approach, this method automatically extracts ultrasound regions from acquired transabdominal intestinal ultrasound images and uses a neural network model to automatically identify intestinal abnormalities within the extracted images. The proposed algorithm exhibits high specificity, a simple and efficient structure, and can automatically and accurately identify intestinal abnormalities in ultrasound images. This method provides an effective computer-aided tool for the early identification, detection, and assessment of intestinal abnormalities using ultrasound images, offering strong support for disease monitoring and decision-making in clinical practice. Attached Figure Description

[0140] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0141] Figure 1 This is a flowchart illustrating the implementation and application of the method of the present invention;

[0142] Figure 2 This is an example of automatically extracting the ultrasound region from the original image in this invention.

[0143] Figure 3 This is a diagram of the branch structure for local feature extraction in this invention;

[0144] Figure 4 This is a diagram of the global feature extraction branch structure in this invention;

[0145] Figure 5 This is a structural diagram of the local and global feature fusion module in this invention;

[0146] Figure 6 This is a structural diagram of the multi-scale feature extraction module in this invention;

[0147] Figure 7 This is a structural diagram of the network model for automatic recognition of abnormal intestinal imaging features in this invention;

[0148] Figure 8 This is a diagram showing the confusion matrix results of the automatic identification model for abnormal intestinal imaging features in this invention.

[0149] Figure 9 This is a graph showing the index results of the automatic identification model for abnormal intestinal imaging features in this invention. Detailed Implementation

[0150] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0151] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0152] like Figure 1 The method for automatically identifying abnormal intestinal imaging features in transabdominal ultrasound images specifically includes the following steps:

[0153] First, the acquired transabdominal intestinal ultrasound images were preprocessed, and the preprocessed image data was divided into training, validation, and test sets. Then, based on the characteristics of the images, functional modules of an automatic intestinal abnormality image feature recognition network model were designed and integrated to construct an end-to-end deep learning network model. Next, the network model was trained using a preset training strategy, and the corresponding weight parameters were saved when the model achieved optimal performance, thus obtaining the best recognition model. Then, the trained model was tested, and its performance was comprehensively evaluated using the recognition results and multiple evaluation indicators, including its generalization ability and robustness. Finally, the optimal model was applied to clinical practice to process and analyze newly acquired transabdominal intestinal ultrasound images, achieving automatic recognition of intestinal abnormalities.

[0154] like Figure 2The image shown illustrates the automatic cropping effect of the ultrasound region. For successfully retrieved images, grayscale processing is performed first: if the image is color, it is converted to grayscale; if it is already grayscale, it is directly copied and retained. Then, an adaptive histogram equalization (CLAHE) algorithm is applied to the grayscale image to enhance local contrast and highlight dark areas, thereby improving the accuracy of subsequent target region segmentation. Next, gamma correction is performed on the enhanced image to further improve overall brightness and contrast, enhancing the visual quality of the image. Finally, a fixed threshold is used to binarize the processed image, achieving separation of the target region from the background.

[0155] like Figure 3 The local feature extraction branch structure diagram shown is divided into two sub-branches after receiving the input feature map: a dynamic convolution generation branch and a dynamic feature extraction branch. First, in the dynamic convolution generation branch, the input feature map undergoes global feature pooling to extract global contextual information.

[0156]

[0157] Subsequently, the global context information is input into the weight generation network, which includes modules such as 1×1 convolutional layers, group normalization (GN) layers, and GELU activation functions, used to generate the weight coefficients for each group of convolutional kernels. :

[0158]

[0159] The obtained weighting coefficients A weighted computation is performed with a pre-randomized static convolution kernel to generate a dynamic convolution kernel. Simultaneously, a dynamic feature extraction branch receives the initial feature map and performs a convolution operation using the generated dynamic convolution kernel to extract features, thereby enhancing the representation of fine structural information in the intestinal region.

[0160] like Figure 4 The global feature extraction branch structure diagram shown is divided into two sub-branches after receiving the initial feature map. One branch generates the weight matrix of the query (Q) through a 1×1 convolution:

[0161]

[0162] The other branch is used for spatial downsampling, passing through a 5×5 convolutional layer, a batch normalization (BN) layer, a GELU activation function, a 1×1 convolutional layer, and a second BN layer in sequence to obtain the downsampled feature map:

[0163]

[0164] When setting the sampling ratio , indicates that spatial downsampling is not performed: at this time .

[0165] To further enhance the expressive power of local feature extraction, a residual connection mechanism was designed, which extracts local features through 3×3 convolution and then adds them to the original features:

[0166]

[0167] Subsequently, a weight matrix for key values ​​K and V is generated through a 1×1 convolutional layer:

[0168]

[0169] Finally, attention weights and attention scores are calculated.

[0170]

[0171]

[0172] This branch effectively improves the model's ability to model and express features of complex structures in the intestinal region by fusing global contextual information with local details.

[0173] like Figure 5 The diagram shows the structure of the local and global feature fusion module. The input to this module comes from the feature map of the local feature extraction branch. Feature maps of the global feature extraction branch First, the two feature maps are concatenated along the channel dimension to form a fused input feature map:

[0174]

[0175] Then, the fused feature map is sequentially passed through an average pooling layer, a 1×1 convolutional layer, a ReLU activation function, a 1×1 convolutional layer, and a Sigmoid activation function to generate a fused weight tensor. :

[0176]

[0177] To distinguish the feature weights of different branches, the fusion weight tensor will be used. The weights are divided along the channel dimension to obtain the weights corresponding to the local and global branches, respectively. and :

[0178]

[0179] in, , , , Indicates the spliced Attention weights are generated for each channel. Through channel splitting, attention weights corresponding to local and global branches are obtained. , ,in ,

[0180] Then, the branch feature map from the local feature extraction is processed. Feature maps of the global feature extraction branch Perform feature weighting: for each Multiply by the corresponding weight:

[0181]

[0182]

[0183] Through the above weighting operations, the network can adaptively adjust the importance of local and global branch features according to the characteristics of the input image, thereby improving the fusion effect;

[0184] Finally, the weighted features and The features are then concatenated again, and feature fusion and dimensionality reduction are performed using 1×1 convolutions to output the final adaptive fused feature map. :

[0185]

[0186] By fusing local branch features and global branch features, the output features combine the advantages of convolution and attention mechanisms, thereby significantly improving the overall feature representation capability.

[0187] like Figure 6 The diagram shows the structure of the multi-scale feature extraction module, which adaptively fuses feature maps. As input, the overall structure consists of two branches: one for multi-scale dynamic learning and the other for residual connection.

[0188] In the multi-scale dynamic learning branch, the final adaptive fusion of feature maps is achieved. The data are fed into four dynamic convolutional branches, each using convolutional kernels of different sizes for feature extraction, thereby generating four sets of multi-scale feature representations:

[0189]

[0190]

[0191]

[0192]

[0193] Subsequently, the four sets of feature maps were concatenated along the channel dimension to obtain a multi-scale concatenated feature map that integrates information from different receptive fields:

[0194]

[0195] Feature maps of different scales are concatenated along the channel dimension to generate a feature map that integrates multi-scale information. The concatenated feature map. It retains local and global information from different convolutional kernel sizes, providing richer feature representation capabilities.

[0196] then, Multi-scale fused feature maps are obtained by using 1×1 convolution, GN normalization, GELU activation, and another 1×1 convolutional layer to reduce dimensionality and fuse the data. :

[0197]

[0198] Another residual connection adaptively fuses feature maps After performing 1×1 convolution and normalization adjustment, the feature map is fused with multi-scale features. The residual feature map is obtained by adding the features together to enhance feature representation and promote information flow. :

[0199]

[0200] Finally, the residual feature map The final feature is output by performing a nonlinear transformation using the GULE activation function.

[0201]

[0202] This module achieves multi-scale feature extraction through dynamic convolution and effectively improves feature representation and information flow efficiency by combining residual connection mechanism, thereby enhancing the model's ability to perceive and understand complex structures.

[0203] like Figure 7 The diagram shows the structure of a network model for automatic recognition of abnormal intestinal imaging features. This is an end-to-end neural network model for the automatic recognition of abnormal intestinal imaging features. The overall structure of the model consists of four cascaded stages and a recognition module.

[0204] Each stage consists of multiple dynamically convolutional multi-scale attention ensemble modules, and each module integrates four key functional modules: a local feature extraction branch, a global feature extraction branch, a local and global feature fusion module, and a multi-scale feature extraction module. Through this modular design, the network can deeply understand and analyze the features of transabdominal ultrasound images at multiple levels, accurately modeling the pathological structure and changes of the intestine.

[0205] The four stages progressively pass and refine features, effectively improving the model's ability to perceive and distinguish complex lesion areas in images. Finally, the network transmits the fused and extracted high-level semantic features to the recognition module to complete the automatic recognition of abnormal intestinal imaging features.

[0206] like Figure 8 The confusion matrix result diagram of the automatic identification model for abnormal intestinal imaging features shown presents the identification results of the model of this invention on 200 transabdominal intestinal ultrasound images. In the samples where the intestines were actually normal, 91 were correctly predicted as normal, and only 9 were incorrectly predicted as abnormal; while in the samples where the intestines were actually abnormal, 94 were correctly predicted as abnormal, and only 6 were incorrectly predicted as normal. This clearly demonstrates that the model has good identification ability for both normal and abnormal cases in the automatic identification of abnormal intestinal imaging features, with few mispredictions, further verifying the reliability and effectiveness of the model.

[0207] like Figure 9 The graph showing the performance results of the automatic intestinal abnormality imaging model clearly illustrates the model's performance in key dimensions such as precision, accuracy, specificity, F1 score, and recall. Specifically, the recall rate reached 94.00%, the precision rate was 91.26%, the accuracy rate was 92.50%, the F1 score was 92.61%, and the specificity was 91.00%. All indicators are at high levels and evenly distributed, fully demonstrating the excellent overall performance of the model of this invention.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for automatic identification of abnormal intestinal imaging features in transabdominal ultrasound, characterized in that, Includes the following steps: A dataset of transabdominal intestinal ultrasound images was obtained. Abnormal and normal images in the ultrasound image dataset were labeled and classified. Abnormal and normal images were divided into training set, validation set and test set, respectively. Automatic ultrasound region identification and detection is performed on images in an ultrasound image dataset. A deep neural network model is established, which includes a convolutional initiation feature extraction layer, multiple dynamic convolutional multi-scale attention integration modules, and an abnormal image feature recognition module. The dynamic convolutional multi-scale attention integration module includes a local and global feature extraction module, a local and global feature fusion module, and a multi-scale feature extraction module. The convolutional initiation feature extraction layer performs preliminary feature extraction on the ultrasound image; the local and global feature extraction module obtains local and global features of the intestinal region from both local structural and global semantic levels, and performs in-depth analysis of the intestinal state in the ultrasound image; the local and global feature fusion module fuses the local and global features to generate a fused feature map, which represents the intestinal features of the suspected lesion area; the multi-scale feature extraction module performs multi-scale modeling on the fused feature map, and extracts the shape and size information of the multi-scale lesion area through different receptive fields; The deep neural network model is trained using a training set, and the parameter information of the deep neural network model is updated through forward propagation, loss function optimization, and back propagation. The test set is input into the trained neural network model to evaluate its performance metrics; the original transabdominal ultrasound images are input into the deep neural network model to output abnormal image judgment results.

2. The method for automatic identification of abnormal intestinal imaging features in transabdominal ultrasound images according to claim 1, characterized in that, The local and global feature extraction module includes a local feature extraction branch and a global feature extraction branch. The local feature extraction branch extracts fine-grained spatial local information from the input feature map and combines it with global context information to enhance the local feature representation capability. The global feature extraction branch is used to capture the dependencies between pixels in the image from the input feature map and analyze the high-level information of the entire input image.

3. The method for automatic identification of abnormal intestinal imaging features in transabdominal ultrasound images according to claim 2, characterized in that, The local feature extraction branch performs global average pooling on the preliminary features to obtain global context information for each channel: ， in, This represents the initial feature map of the input. For the number of channels, For height, For width, Indicates the first Each channel is located in eigenvalues ​​at that location Indicates the first Global context information corresponding to each channel; Input the global context information into the weight generation network shown below, and output the weight coefficients of each convolutional kernel. : ， in, Indicates the core size is Convolution operation, For group normalization methods, channels are divided into For each group, normalize the features in that group: ， in, Indicates the number within the group 1 eigenvalue, Represents the normalized eigenvalues. It is the first The mean of all features within a group. It is the first The variance of all features within a group. A minimal constant introduced to prevent the denominator from being zero; The normalized features are linearly recalibrated after scaling and offset operations, defined as follows: ， in, A learnable scaling factor. A learnable bias factor. This represents the output feature value after scaling and offset transformation; Let GELU be a non-linear activation function. Combining the properties of linear units with the probabilistic characteristics of Gaussian distribution, a probabilistic mechanism is used to determine whether to retain the input value, making the deep neural network model stable and flexible during training. ， in, Indicates the input feature value. Represents the hyperbolic tangent function; This represents the normalization function, used to map the output to a probability distribution of weight coefficients. The value range of each output component is within... Between these points, the sum of all outputs is 1 for the input vector Z = [z1, z2, z3, ... z2]. n The calculation formula is as follows: ， in, Represents the first input vector. One element, Represents the dimension of a vector. This indicates exponentiation. Indicates the first The normalized weight values ​​of each element; Randomly initialize multiple convolutional kernel parameter sets ,in, Indicates the first The set of static convolutional kernels, C1 (number of input channels), C2 (number of output channels), K1 (kernel height), and K2 (kernel width), are based on the generated weight coefficients. Calculate the dynamic convolution kernel after weighted fusion Its definition is: ， in, Indicates the first The weight coefficients corresponding to the group convolution kernel, It is the number of convolutional kernel groups. This represents the fused dynamic convolutional kernel; The weighted fusion-derived dynamic convolution kernel is used to perform a convolution operation on the input preliminary feature map to obtain the output feature map. : ， in, Represents the initial feature map of the input. Represents two-dimensional convolution operation, This refers to the edge padding operation performed during convolution to preserve the feature map size. is the kernel size.

4. The method for automatic identification of abnormal intestinal imaging features in transabdominal ultrasound images according to claim 2, characterized in that, The global feature extraction branch is based on preliminary features. Generate the query weight matrix, which is calculated through convolution operations: ， in, express Convolution operation is used to generate the query weight matrix. ; If the sampling ratio is set ( ), then for input features Spatial downsampling is performed to improve the efficiency of attention computation. ， in, express Convolution operation and These represent batch normalization operations, express Convolution operation The activation function of the Gaussian error linear unit. This represents the feature map after downsampling; If the sampling ratio is set If spatial downsampling is not performed, then: ， Downsampling features Depthwise separable convolutions are performed, and residual connections are used to optimize feature propagation, resulting in an updated feature map representation. : ， in, express Depthwise separable convolution; Weight generation for key values ​​K and V: This involves updating the feature map. The convolutional layer generates keys K and values ​​V, where the weight parameters of the matrices for keys K and values ​​V are: ， in, This represents a convolution operation used to simultaneously generate key and value weight matrices; and These are the convolution weights for the keys and values, respectively. The query matrix is ​​obtained by mapping the input features through learnable linear projection layers. Key matrix Sum matrix : ， Based on the three matrices mentioned above, the attention score is calculated to obtain the global feature map. : ， in, The dimension of the key vector is used to scale the inner product result and avoid excessively large gradients. This represents a relative position encoding item, used to introduce spatial position information.

5. The method for automatic identification of abnormal intestinal imaging features in transabdominal ultrasound images according to claim 2, characterized in that: The local and global feature fusion module is used to adaptively weight and fuse local branch features and global branch features; and outputs the feature map from the local branch. With global branch output feature map By concatenating the data along the channel dimension, the fused input features are obtained: ， in, This represents a feature concatenation operation along the channel dimension; For fused input features Perform global average pooling, then input it into the weight generation network, and output the fused weight tensor. : ， in, , For batch size, For the number of channels, It is the Sigmoid activation function. For average pooling layer, This represents the modified linear unit activation function; An activation function is used to map the input to the interval (0,1). ， Average pooling is used to divide the input tensor into non-overlapping regions, and the average value of the elements in each region is calculated: ， in, It is the input feature map. It is the size of the pooling window. Indicates the position in the window The value; fused weight tensor The weights are divided along the channel dimension to obtain the weights corresponding to the local and global branches, respectively. and : ， in, , For batch size, For the number of channels, These represent the channel weight coefficients for local branches and global branches, respectively. To each Multiply by the corresponding weights to achieve adaptive weighting: ， ， By using feature weighting, deep neural network models adaptively adjust the importance of features in each branch based on the input features; The weighted features are concatenated and then fused and dimensionality-reduced using convolution to output a fused feature map. : 。 6. The method for automatic identification of abnormal intestinal imaging features in transabdominal ultrasound images according to claim 5, characterized in that: The multi-scale feature extraction module is used to enhance the scale adaptability and structural expressiveness of the fused features. The module initializes four dynamic convolutional branches with different kernel sizes, namely 7×7, 5×5, 3×3 and 1×1. Fuse feature maps The input is fed into four dynamic convolutional branches, which generate four sets of feature representations respectively: ， ， ， ， in, Indicates having Dynamic convolution operation of convolution kernels adaptively adjusts convolution weights based on input features; The four sets of multi-scale features are concatenated along the channel dimension to obtain a multi-scale concatenated feature map: ， in, For batch size, For the number of channels, , These represent the height and width of the feature map, respectively. Multi-scale stitched feature maps Convolution, normalization, and activation operations are performed to fuse multi-scale information, followed by dimensionality reduction to obtain a multi-scale fused feature map. : ， Multi-scale fusion feature maps With adaptive fusion feature maps adjusted by convolution and normalization Perform residual connections to obtain residual feature maps. : ， For residual characteristics Perform GELU activation and output the final features. : 。

Images