Diagnostic system for evaluating ultrasonic image of abdominal effusion
Through the abdominal effusion diagnosis system combined with the Laplace sharpening algorithm and the dual-branch structure of U-Net and ResNet, the problem of low accuracy and efficiency in traditional ultrasound image analysis is solved, and the automated detection and accurate classification of abdominal effusion is achieved.
Patent Information
- Application Number
- CN202510590370.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-08
- Publication Date
- 2025-08-15
AI Technical Summary
Traditional ultrasound image analysis has problems of low accuracy and low efficiency in abdominal effusion detection, especially in primary medical institutions, doctors' technical levels are uneven, resulting in misdiagnosis or missed diagnosis, and subjective interpretation of ultrasound images is difficult to achieve accurate, reliable and rapid diagnosis.
The image edges are enhanced by using Laplace sharpening algorithm, combined with U-Net and ResNet dual-branch structures to segment and severity classification of abdominal effusion areas, and diagnose them through feature fusion and multi-layer perceptrons, and data augmentation techniques such as rotation, translation, scaling and mixing operations are introduced to enrich the training samples.
It significantly improves the accuracy and efficiency of automated detection, quantitative segmentation and severity classification of abdominal effusion, and combines local characteristics and global background information to achieve more reliable diagnostic results.
Smart Images
Figure CN120495240A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of deep learning computer vision and medical image processing, and in particular relates to a diagnostic system for evaluating ultrasonic images of abdominal effusion. Background Art
[0002] Ultrasound imaging has become an important diagnostic tool in modern medicine due to its non-invasive, real-time, and easy-to-use nature. It plays a key role in the detection and severity assessment of abdominal effusions. However, traditional ultrasound image analysis relies heavily on the physician's personal experience, which limits the objectivity of diagnosis to a certain extent. This is especially true in primary healthcare institutions, where physicians' technical skills vary widely, leading to misdiagnosis or missed diagnoses. Furthermore, the subjective interpretation of ultrasound images is particularly evident in the assessment of complex abdominal effusions, making it difficult to achieve an accurate, reliable, and rapid diagnosis.
[0003] In recent years, deep learning models have demonstrated great potential in medical image analysis. By training on large-scale annotated data, they can achieve automated segmentation and quantitative analysis of abdominal effusion areas. However, existing AI-based ultrasound image analysis techniques lack robustness when dealing with complex lesion characteristics and background noise. In particular, there is still room for improvement in accuracy and efficiency in the segmentation and severity classification of abdominal effusions. Summary of the Invention
[0004] The purpose of the present invention is to provide a diagnostic system for evaluating abdominal effusion ultrasound images to solve the problems of low accuracy and low efficiency in the segmentation and severity classification of abdominal effusion.
[0005] The present invention adopts the following technical solution: a diagnostic system for evaluating ultrasonic images of abdominal effusion, comprising:
[0006] Step 1: Acquire ultrasound images of abdominal effusion using probes of the same frequency using multiple ultrasound devices, and acquire ultrasound images of abdominal effusion using probes of multiple frequencies using the same ultrasound device, thereby obtaining multiple ultrasound images of abdominal effusion;
[0007] Step 2: Use the Laplace sharpening algorithm to enhance the image edges of each abdominal effusion ultrasound image to highlight the contour of the effusion area;
[0008] Step 3: Normalize each image obtained in step 2;
[0009] Step 4: Use the trained U-Net to segment the effusion area in each image obtained in step 3;
[0010] Step 5: Use ResNet to extract the image structure and texture features of each image obtained in step 3;
[0011] Step 6: The segmented image obtained in step 4 and the extracted features obtained in step 5 are fused and spliced to obtain a result image, and then the result image is used to perform a graded diagnosis of the severity of the effusion through MLP.
[0012] Furthermore, after step 3 and before step 4, each image obtained after normalization in step 3 is rotated by ±15°, translated by ±10%, randomly scaled by 0.8–1.2 times, and mirrored and flipped to enrich the training images.
[0013] Furthermore, after step 3 and before step 4, each image obtained after normalization in step 3 is randomly mixed with two images using CutMix to generate new training samples, so as to enrich the training images.
[0014] Furthermore, after step 3 and before step 4, each image obtained after normalization in step 3 is linearly interpolated with two images and their labels using MixUp to generate new training samples, so as to enrich the training images.
[0015] Furthermore, in step 4, the U-Net is trained using the images obtained in step 3, and a trained U-Net is obtained.
[0016] The beneficial effects of the present invention are:
[0017] The present invention combines local effusion characteristics with global background information, and realizes automated detection, quantitative segmentation and severity classification of abdominal effusion through a dual-branch structure, significantly improving diagnostic efficiency and accuracy.
[0018] The present invention adopts a dual-branch structure. The U-Net branch focuses on the precise segmentation of the abdominal effusion area and extracts local features. The ResNet branch captures global background information to provide scale and context information support for effusion classification. The U-Net branch extracts local detail features and the ResNet branch captures global features, and the two types of information are effectively integrated during feature fusion. At the same time, the multi-layer perceptron performs classification and prediction, which can not only accurately determine the type of lesion, but also assess its severity.
[0019] The present invention introduces Laplace sharpening and percentile-based normalization technology to improve image quality and ensure the consistency and reliability of input data. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 2 is an architectural diagram of the diagnostic system of the present invention. DETAILED DESCRIPTION
[0021] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
[0022] The present invention discloses a diagnostic system for evaluating ultrasound images of abdominal effusion, such as Figure 1 Shown, including:
[0023] Step 1: Acquire ultrasound images of abdominal effusion using probes of the same frequency of multiple ultrasound devices, and acquire ultrasound images of abdominal effusion using probes of multiple frequencies of the same ultrasound device, thereby obtaining multiple ultrasound images of abdominal effusion.
[0024] Step 2: The Laplace sharpening algorithm is used to enhance the image edges of each abdominal effusion ultrasound image to highlight the contours of the effusion area. Based on image processing techniques, the acquired abdominal effusion ultrasound images are subjected to denoising, edge enhancement, and contrast enhancement to ensure image quality and consistency. The Laplace sharpening algorithm enhances the edge information of the effusion in the image, improving the clarity and segmentation accuracy of the effusion area.
[0025] Step 3: Normalize the images obtained in step 2; standardize the image grayscale values to reduce brightness differences caused by different devices and enhance the contrast of the effusion area.
[0026] Step 4: Use the trained U-Net to segment the effusion regions in each image obtained in Step 3. The U-Net branch (local feature extraction module) focuses on extracting detailed local features of the abdominal effusion region, including its edges, morphology, and location. This precise segmentation is achieved through a classic encoder-decoder architecture. The encoder extracts deep features through layer-by-layer downsampling, while the decoder restores the spatial information of the feature map through layer-by-layer upsampling, ultimately achieving precise localization and segmentation of the lesion region.
[0027] Step 5: Use ResNet to extract the image structure and texture features of each image obtained in step 3; the ResNet branch (global feature extraction module) is responsible for extracting the global background information and scale characteristics of the abdominal effusion ultrasound image, capturing the overall distribution of the effusion and the complex background relationship, and providing classification guidance for the segmented effusion area.
[0028] Step 6: The segmented image obtained in step 4 and the extracted features obtained in step 5 are fused and spliced to obtain a result image, and then the result image is used to perform a graded diagnosis of the severity of the effusion through MLP.
[0029] The local features extracted by the UNet branch are fused with the global features extracted by the ResNet branch, and fusion features with comprehensive expression capabilities are generated through splicing or weighted operations, providing more reliable information support for the classification of the severity of effusion.
[0030] Based on the result graph, a multi-layer perceptron (MLP) is used for classification prediction to grade the severity of the effusion, providing classification results including mild, moderate, and severe.
[0031] Preferably, after step 3 and before step 4, each image obtained after normalization in step 3 is rotated by ±15°, translated by ±10%, randomly scaled by 0.8–1.2 times, and mirror-flipped to enrich the training images.
[0032] Preferably, after step 3 and before step 4, each image obtained after normalization in step 3 is subjected to random region mixing of two images using CutMix to generate new training samples, so as to enrich the training images.
[0033] Preferably, after step 3 and before step 4, each image obtained after normalization in step 3 is linearly interpolated with two images and their labels using MixUp to generate new training samples, so as to enrich the training images.
[0034] Random transformation operations, including rotation, flipping, and scaling, are performed on the abdominal effusion ultrasound images and the corresponding mask labels to increase data diversity and improve the generalization ability of the model.
[0035] Preferably, in step 4, the U-Net is trained using the images obtained in step 3 to obtain a trained U-Net. The training data uses the preprocessed ultrasound images of abdominal effusion and the corresponding mask labels as model input, and the data is divided into a training set and a test set in a ratio of 8:2.
[0036] Example 1
[0037] The specific process of this embodiment is as follows:
[0038] Step 1: Acquire ultrasound images of abdominal effusion using probes of the same frequency of multiple ultrasound devices, and acquire ultrasound images of abdominal effusion using probes of multiple frequencies of the same ultrasound device, thereby obtaining multiple ultrasound images of abdominal effusion.
[0039] Through cooperation with multiple medical institutions, a large number of abdominal ultrasound images of abdominal effusion and lesion area annotation files were obtained, thus achieving the acquisition of target data.
[0040] Step 2: Use the Laplace sharpening algorithm to enhance the image edges of each abdominal effusion ultrasound image to highlight the contour of the effusion area.
[0041] To ensure the quality of input data, image processing technology is used to perform denoising, edge enhancement, and contrast enhancement on the original image data; the median filtering method is used to denoise the image to improve the quality of ultrasound images; Laplace sharpening technology is used to enhance image edge details to make the lesion area clearer; and percentile-based normalization technology is used to standardize the grayscale value of the image to ensure the consistency of image brightness and contrast and reduce the interference of external factors on diagnosis.
[0042] Median filtering denoising: For each pixel I(x,y) in the image, a 3×3 window centered at (x,y) is selected and all pixel values within the window are collected to form a neighborhood set N(x,y).
[0043]
[0044] Sort all pixel values in the neighborhood set N(x,y), and then take the median m as the output pixel value.
[0045] Laplace sharpening: Use the Laplace convolution kernel to calculate the Laplace operator value for each pixel in the image, thereby enhancing the edge information of the image. The Laplace convolution kernel used is:
[0046]
[0047] For each pixel I(x,y) in the image, the Laplacian value L(x,y) is calculated using the following formula:
[0048]
[0049] Where K(i,j) is the value of the Laplacian convolution kernel, and I(x+i,y+j) is the grayscale value of the corresponding pixel in the image. Next, the original image and the Laplacian image after absolute value are weighted and summed to enhance the edges. This process can be expressed as follows:
[0050] S(x,y)=α·I(x,y))+β·∣L(x,y)∣+γ
[0051] Wherein, α and β are weight parameters that adjust the contribution ratio of the original image and the Laplacian image, and γ is an offset (usually set to 0). In this embodiment, the parameters are set as: α = 1.5, β = -0.5, γ = 0.
[0052] Step 3: Normalize the images obtained in step 2.
[0053] Percentile-based normalization: This method removes outliers by clipping the extreme values based on the selected percentile to reduce the influence of noise and extreme values in the data. The clipped data is then scaled to the standard range [0, 1] to improve the overall image contrast.
[0054] Mask label generation and data augmentation: In this example, to enable effective model training and validation, a mask label image is generated based on the acquired annotation file. Specifically, the coordinates of the polygonal regions are extracted from the annotation file and used to draw a mask image. This marks the lesion area as the foreground, while other areas remain as the background, providing accurate label data for model training.
[0055] At the same time, random image transformation is used to enhance data diversity. The rotation function, flip function, and scaling function in the Python image processing library are used to perform random rotation, flip, and scaling operations on the image to enrich the data and improve the robustness of the model.
[0056] Step 4: Use the trained U-Net to segment the effusion area in each image obtained in step 3.
[0057] The U-Net branch extracts local features. Using an encoder-decoder architecture, it gradually extracts detailed local features from the image, such as the morphology and boundaries of the lesion. The encoder downsamples the image to extract deeper local features. The decoder upsamples the image to restore its spatial information, ensuring accurate location of the lesion.
[0058] Specifically:
[0059] Encoder: For the input image I, features are gradually extracted through a series of convolution operations and downsampling.
[0060]
[0061] in is the feature map of the lth level, and They represent the convolution kernel and bias of the lth layer respectively, σ is the activation function ReLU, and L is the number of layers of the encoder.
[0062] Decoder: Restores image spatial information through upsampling operations.
[0063]
[0064] It is the feature map of the previous layer, and the symbol [] represents feature concatenation, and the final feature representation F is obtained. u ,in
[0065] Step 5: Use ResNet to extract the image structure and texture features of each image obtained in step 3.
[0066] The ResNet branch extracts global features. This branch uses a residual network structure and is mainly used to extract global features. The key is to learn the scale characteristics and complex background information in ultrasound images through the residual module. The basic formula of residual learning is:
[0067] F r =F(x)+x
[0068] Where F(x) represents the features obtained after several convolutional layers, and x is the input feature. In the residual module, convolutional layers and batch normalization are used to enhance learning ability:
[0069]
[0070] Step 6: The segmented image obtained in step 4 and the extracted features obtained in step 5 are fused and spliced to obtain a result image, and then the result image is used to perform a graded diagnosis of the severity of the effusion through MLP.
[0071] The features extracted by the UNet and ResNet branches are fused to achieve more accurate lesion area identification. The fusion operation can be completed by the splicing operation:
[0072] F f =[F u ,F r ]
[0073] The fused features are more comprehensive, including local lesion detail information, combined with global scale features and complex background information, making it easier to classify the lesion area more accurately.
[0074] Based on the result graph F f A multi-layer perceptron (MLP) is used for classification prediction to determine the type and severity of the lesion area. The MLP consists of multiple fully connected layers, each of which uses the activation function ReLU for nonlinear transformation.
[0075] H (l) =σ(W (l) H (l-1) +b (l) )(l=1,2,…,L)
[0076] The output of the final layer is normalized by the Softmax function to obtain the predicted probability of each type and severity:
[0077] P=Softmax(H (L) )
[0078] During the model training and validation phases, preprocessed, labeled ultrasound images were divided into training and test sets in an 8:2 ratio. The ultrasound image lesion recognition model was trained on the training set. During training, the Adam optimizer and cross-validation techniques were used to dynamically adjust model hyperparameters, including the learning rate and batch size, to accelerate model convergence and improve diagnostic accuracy.
[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A diagnostic system for evaluating an ultrasound image of abdominal effusion, characterized in that: include: Step 1: Acquire ultrasound images of abdominal effusion using probes of the same frequency using multiple ultrasound devices, and acquire ultrasound images of abdominal effusion using probes of multiple frequencies using the same ultrasound device, thereby obtaining multiple ultrasound images of abdominal effusion; Step 2: Use the Laplace sharpening algorithm to enhance the image edges of each abdominal effusion ultrasound image to highlight the contour of the effusion area; Step 3: Normalize each image obtained in step 2; Step 4: Use the trained U-Net to segment the effusion area in each image obtained in step 3; Step 5: Use ResNet to extract the image structure and texture features of each image obtained in step 3; Step 6: The segmented image obtained in step 4 and the extracted features obtained in step 5 are fused and spliced to obtain a result image, and then the result image is used to perform a graded diagnosis of the severity of the effusion through MLP.
2. A diagnostic system for evaluating abdominal effusion ultrasound images according to claim 1, characterized in that: After step 3 and before step 4, each image obtained after normalization in step 3 was rotated by ±15°, translated by ±10%, randomly scaled by 0.8–1.2 times, and mirrored to enrich the training images.
3. A diagnostic system for evaluating abdominal effusion ultrasound images according to claim 1, characterized in that: After step 3 and before step 4, each image obtained after normalization in step 3 is randomly mixed with the two images using CutMix to generate new training samples to enrich the training images.
4. A diagnostic system for evaluating abdominal effusion ultrasound images according to claim 1, characterized in that: After step 3 and before step 4, each image obtained after normalization in step 3 is linearly interpolated with two images and their labels using MixUp to generate new training samples to enrich the training images.
5. The diagnostic system for evaluating abdominal effusion ultrasound images according to claim 1, characterized in that: In step 4, the U-Net is trained using the images obtained in step 3, and a trained U-Net is obtained.