An Automated Segmentation Method for Pneumoperitoneum on Abdominal CT Based on U-Net with Multi-Window and Wide-Window Levels

Abstract

This invention belongs to the field of medical imaging technology and discloses an automated segmentation method for pneumoperitoneum on abdominal CT scans based on U-Net with multiple window width levels. The method includes the following steps: Step S1, collecting patient CT image data and preprocessing it; Step S2, establishing an automated segmentation model based on a U-Net network segmentation algorithm and training the model; Step S3, quantitatively optimizing the window width and window level of the trained automated segmentation model using a grid search algorithm. This invention employs the above-mentioned automated segmentation method for pneumoperitoneum on abdominal CT scans based on U-Net with multiple window width levels. By dynamically adjusting the combination of window width and window level parameters, it achieves automatic detection and accurate segmentation of free gas in the peritoneal cavity, improving the efficiency and accuracy of pneumoperitoneum diagnosis, reducing the risk of missed diagnoses, and providing a reliable basis for clinical decision-making.

Description

Technical Field

[0001] This invention relates to the field of medical imaging technology, and in particular to an automated segmentation method for pneumoperitoneum in abdominal CT scans based on U-Net with multiple windows and wide window positions. Background Technology

[0002] In the field of medical imaging, the detection of free gas in the peritoneal cavity is crucial for diagnosing acute abdominal conditions such as gastrointestinal perforation. Currently, CT imaging is widely used in clinical practice as the gold standard for detecting pneumoperitoneum. However, existing pneumoperitoneum detection techniques have many shortcomings, and there is an urgent need for a more efficient and accurate automated segmentation method to improve diagnostic efficiency and accuracy.

[0003] Traditional manual identification methods rely on the experience of radiologists, which is not only time-consuming and labor-intensive but also susceptible to subjective factors, potentially leading to missed diagnoses. For example, when processing large numbers of CT images, manual identification may overlook subtle pneumoperitoneum features due to physician fatigue or experience differences, thus delaying diagnosis. While traditional segmentation algorithms can assist physicians in image analysis to some extent, their segmentation results are often less than ideal for targets with blurred boundaries and irregular shapes, such as pneumoperitoneum. These algorithms are typically based on fixed image processing rules, making it difficult to accurately extract the target region when faced with complex pneumoperitoneum images.

[0004] In recent years, deep learning has demonstrated tremendous potential in the field of medical image segmentation. The U-Net architecture, in particular, has become a popular choice for medical image segmentation due to its excellent performance in scenarios with small sample sizes and high accuracy requirements. However, current research largely focuses on detecting lesions in the chest and lung regions, lacking in-depth modeling capabilities for the morphological characteristics and spatial distribution patterns of free gas in the abdominal cavity. Furthermore, while window width and level parameters in CT imaging significantly affect image contrast and texture, research on optimizing these parameters for pneumoperitoneum in medical imaging is relatively scarce. Existing segmentation algorithms typically employ fixed window width and level settings, failing to fully consider the impact of different parameter combinations on the segmentation accuracy of pneumoperitoneum, thus limiting model performance.

[0005] Regarding datasets, publicly available high-quality pneumoperitoneum datasets are scarce, leading to model training often relying on limited samples. This limits the generalization ability of models when faced with diverse clinical data, making it difficult to meet the needs of actual clinical applications. Furthermore, the lack of dedicated optimization strategies for pneumoperitoneum prevents segmentation models from fully utilizing their performance when processing pneumoperitoneum images, impacting segmentation accuracy and efficiency.

[0006] Existing technologies for automated segmentation of free gas in the peritoneal cavity suffer from several drawbacks, including low efficiency in manual identification, poor adaptability of traditional segmentation algorithms to complex targets, insufficient research on the impact of window width and window level parameters, and limitations in datasets. These issues not only affect the accuracy and efficiency of pneumoperitoneum detection but also restrict the widespread application and promotion of this technology in clinical practice. Therefore, there is an urgent need for an innovative method that can overcome the shortcomings of existing technologies to achieve automated detection and accurate segmentation of free gas in the peritoneal cavity, improve the level of pneumoperitoneum diagnosis, and provide reliable support for clinical decision-making. Summary of the Invention

[0007] The purpose of this invention is to provide an automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multiple window width and window level parameters. By dynamically adjusting the combination of window width and window level parameters, the method can achieve automatic detection and accurate segmentation of free gas in the peritoneal cavity, improve the efficiency and accuracy of pneumoperitoneum diagnosis, reduce the risk of missed diagnosis, and provide a reliable basis for clinical decision-making.

[0008] To achieve the above objectives, this invention provides an automated segmentation method for pneumoperitoneum on abdominal CT scans based on U-Net with multiple windows and wide windows, comprising the following steps:

[0009] Step S1: Collect patient CT image data and perform preprocessing;

[0010] Step S2: Based on the U-Net network segmentation algorithm, establish an automated segmentation model and train the model;

[0011] Step S3: Quantitatively optimize the window width and window level of the trained automated segmentation model using a grid search algorithm.

[0012] Preferably, in step S1, patient CT image data is collected, and the specific process is as follows:

[0013] First, collect patient CT images and exclude patient images with imaging artifacts and motion artifacts to ensure the quality of the dataset;

[0014] Then, the pneumoperitoneum is labeled using the diagnostic results in the radiological report to generate a gold standard label;

[0015] Finally, data were segmented based on individual patients to avoid duplicate data.

[0016] Based on a list of case IDs with pneumoperitoneum annotations, stratified random sampling was used to divide 70% of the cases into the training set and 30% into the test set, ensuring that the two sets of data did not overlap.

[0017] Preferably, in step S1, the collected CT image data is preprocessed, and the specific process is as follows:

[0018] Step S121: Generate 2D slices along the axis of the DICOM sequence for each case, and simultaneously apply three standardized window levels—abdominal window, bone window, and lung window—to each slice.

[0019] Among them, the abdominal window highlights the contrast between abdominal gas and soft tissue, the bone window enhances skeletal details, and the lung window enhances the visibility of the lungs and pneumoperitoneum area.

[0020] Step S122: Stack the single-channel images of the three windows into a multi-channel input by linear gray-level truncation and normalization;

[0021] Step S123: Perform data augmentation on the training set by randomly rotating, translating, scaling, and perturbing the brightness and contrast of the multi-channel images to simulate anatomical locations and imaging differences.

[0022] Step S124: Calculate the mean and standard deviation of the three channels based on the training set to achieve global normalization. Perform Z-score standardization on all images to eliminate grayscale differences between devices.

[0023] During the CT image data preprocessing process, the training set and the test set are strictly isolated to ensure that the preprocessing parameters are derived only from the training data.

[0024] Preferably, in step S2, an automated segmentation model is established based on the U-Net network segmentation algorithm, and the specific process is as follows:

[0025] First, a U-Net network is built, which consists of two parts: an encoder and a decoder. The encoder is used for downsampling, and the decoder is used for upsampling.

[0026] Secondly, an automated segmentation model is established using a segmentation algorithm based on the U-Net network to achieve automated segmentation of two-dimensional CT images.

[0027] Preferably, the encoder consists of several repeated convolutional blocks, each of which contains two 3×3 convolutions, a ReLU activation function, and a 2×2 max pooling operation to achieve spatial downsampling and semantic extraction.

[0028] The first encoder-based The layer outputs a feature map, as shown below:

[0029] ;

[0030] in, Indicates the convolution operation; Represents the ReLU activation function; , For the first Convolutional kernels of layers; , For bias terms; For encoder Layer output; For encoder The output of the layer.

[0031] Preferably, the decoder restores the low-resolution feature map to the original image size through deconvolution or upsampling operations, and stitches the encoder feature map of the corresponding layer at each layer to achieve high-precision recovery;

[0032] Based on the decoder The layer outputs a feature map, as shown below:

[0033] ;

[0034] in, Indicates the decoder's first Layer upsampling results and encoder 1 Layer feature map stitching; Indicates the transposed convolution kernel; This is the bias term for the transposed convolution; For the decoder's first Layer output.

[0035] Preferably, the decoder and encoder are connected via skip connections. An attention mechanism is introduced into the skip connections to retain only the feature information relevant to the current segmentation target, reducing redundancy, as shown below:

[0036] ;

[0037] ;

[0038] in, Encoder features; Encoder features The linear transformation weight matrix; Features of the decoder; Decoder features The linear transformation weight matrix; This is an attention weight map; For bias terms; This is pixel-by-pixel multiplication; These are decoder features filtered by the attention mechanism;

[0039] Finally, a 1×1 convolution is used to map the multi-channel feature map into a binary classification probability map, and the final probability is output using the Sigmoid function.

[0040] Preferably, in step S2, the automated segmentation model is trained, and the specific process is as follows:

[0041] During training, an enhanced training set is used, and the RMSprop optimizer is employed, combined with the cross-entropy loss function. This is used to measure the difference between the predicted probability distribution and the true distribution, as shown below:

[0042] ;

[0043] Where N is the number of categories; For the first The loss for the nth sample represents the model's loss on the nth sample. The difference between the predicted results and the true labels on each sample; This is a category index used to iterate through all possible categories for each sample; Total number of categories; This is an indicator function that indicates which category the element belongs to; To predict probabilities, the observed sample belongs to a category. The predicted probability needs to be estimated and calculated in advance.

[0044] Preferably, in step S3, the trained automated segmentation model is quantitatively optimized using a grid search algorithm, as follows:

[0045] Step S31, Preliminary Exploration Stage: Under the lung window, bone window, and abdominal window respectively, compare the model's performance in extracting image features using different window width and window level combinations, and select the window with the best performance in extracting image features.

[0046] Step S32, Window width optimization stage: Fix the window position of the optimal window, adjust the window width every 100 units, train the model and test it to determine the window width that makes the performance of each indicator optimal.

[0047] Step S33, Window Position Optimization Stage: Fix the selected window width parameters, adjust the window position every 100 units, test and analyze the results, and determine the optimal window position configuration.

[0048] Preferably, a series of performance evaluation indicators are used to measure the accuracy of the present invention in detecting free air in the peritoneal cavity in CT images, including pixel accuracy, class accuracy, class average pixel accuracy, crossover ratio (CRO), average CRO, Dess coefficient, and recall; the specific calculation methods are as follows:

[0049] Pixel accuracy: ;

[0050] Category accuracy: ;

[0051] Category average pixel accuracy: ;

[0052] Intersection over Union (IOU) = ;

[0053] Average crossover ratio: ;

[0054] Dess coefficient: ;

[0055] Recall rate ;

[0056] in, The number of true positive pixels in the test set, i.e., the pixels that are marked as pneumoperitoneum and correctly predicted as pneumoperitoneum; The number of true negative pixels in the test set, i.e., the pixels that are marked as non-pneumoperitoneous and correctly predicted as non-pneumoperitoneous. The number of false positive pixels in the test set, i.e., pixels that are marked as non-pneumoperitoneum but predicted as pneumoperitoneum; The number of false negative pixels in the test set, i.e., pixels that are labeled as pneumoperitoneum but predicted as non-pneumoperitoneum images; For pixels of the first category; ; The intersection-union ratio of the first category; The crossover ratio for the second category;

[0057] Pixel accuracy and category pixel accuracy reflect the accuracy on the whole and on specific categories; intersection-union ratio and Dess coefficient are key indicators for evaluating segmentation accuracy, measuring the degree of overlap between the predicted region and the real region and the difference between the predicted region and the real region, respectively; in pneumoperitoneum segmentation, recall reflects the sensitivity in detecting free air in the peritoneal cavity, i.e. the ability to avoid missed detections.

[0058] Therefore, the present invention adopts the above-mentioned automated segmentation method for pneumoperitoneum based on U-Net multi-window wide-window level abdominal CT. By dynamically adjusting the combination of window width and window level parameters, it realizes automatic detection and accurate segmentation of free gas in the peritoneal cavity, improves the efficiency and accuracy of pneumoperitoneum diagnosis, reduces the risk of missed diagnosis, and provides a reliable basis for clinical decision-making.

[0059] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0060] Figure 1 This is the technical solution flow of an automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net multi-window wide-window level.

[0061] Figure 2 These are comparison images of different window settings for the same body part; (a) is the abdominal window; (b) is the bone window; and (c) is the lung window. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0063] like Figure 1 As shown, the present invention provides an automated segmentation method for pneumoperitoneum on abdominal CT scans based on U-Net with multiple windows and wide windows, comprising the following steps:

[0064] Step S1: Collect patient CT image data and perform preprocessing;

[0065] Step S2: Based on the U-Net network segmentation algorithm, establish an automated segmentation model and train the model;

[0066] Step S3: Quantitatively optimize the window width and window level of the trained automated segmentation model using a grid search algorithm.

[0067] Example 1

[0068] Step S1: Collect patient CT image data and perform preprocessing.

[0069] Step S11: Collect patient CT image data.

[0070] Collect CT images of adult patients presenting with trauma or abdominal pain, excluding images of patients with imaging artifacts or motion artifacts to ensure dataset quality. Annotate pneumoperitoneum using diagnostic results from radiological reports to generate a gold standard label.

[0071] During the data partitioning phase, it is essential to ensure that each case is treated as a separate patient to avoid duplication. Based on a list of case IDs labeled with pneumoperitoneum, stratified random sampling (balanced distribution according to the severity of pneumoperitoneum or other key labels) is used to allocate 70% of cases to the training set and 30% to the test set, ensuring no overlap between the two sets.

[0072] Step S12: Preprocess the collected CT image data.

[0073] Step S121: Generate 2D slices along the axial direction of the DICOM sequence for each case, and simultaneously apply three standardized windows to each slice: abdominal window (window width 400HU / window level 60HU) to highlight the contrast between abdominal gas and soft tissue, bone window (window width 1500HU / window level 300HU) to enhance bone details, and lung window (window width 1500HU / window level -400HU) to enhance the identification of the lung and pneumoperitoneum regions. Figure 2 As shown.

[0074] Step S122: By linear gray-level truncation and normalization, the single-channel images of the three windows are stacked into a multi-channel input (i.e., the same slice generates three images of the abdomen, bone and lung windows, which are then merged into an H×W×3 matrix).

[0075] Step S123: Perform data augmentation on the training set by randomly rotating (±15°), translating (≤10% range), scaling (90%~110% ratio), and perturbing brightness and contrast (±20%) the multi-channel images to simulate anatomical positions and imaging differences.

[0076] Step S124: Global normalization calculates the mean and standard deviation of the three channels based on the training set, and performs Z-score normalization on all images to eliminate grayscale differences between devices.

[0077] The entire process requires strict isolation between the training and test sets to ensure that preprocessing parameters are derived only from the training data.

[0078] Step S2: Based on the U-Net network segmentation algorithm, establish an automated segmentation model and train the model.

[0079] Step S21: Establish an automated segmentation model based on the U-Net network segmentation algorithm.

[0080] First, the U-Net network was constructed, which consists of two parts: an encoder (downsampling) and a decoder (upsampling). It was chosen because of its symmetrical "U"-shaped structure and its excellent performance in biomedical image processing.

[0081] The encoder mainly consists of several repeated convolutional blocks, each of which contains two 3×3 convolutions (padding=1, keeping the size constant), a ReLU activation function, and a 2×2 max pooling operation to achieve spatial downsampling and semantic extraction.

[0082] The first encoder-based The layer outputs a feature map, as shown below:

[0083] ;

[0084] in, Indicates the convolution operation; Represents the ReLU activation function; , For the first Convolutional kernels of layers; , For bias terms; For encoder Layer output; For encoder The output of the layer.

[0085] The decoder gradually restores the low-resolution feature map to the original image size through deconvolution (transposed convolution) or upsampling operations, and stitches the encoder feature map of the corresponding layer at each layer to achieve high-precision restoration.

[0086] Based on the decoder The layer outputs a feature map, as shown below:

[0087] ;

[0088] in, Indicates the decoder's first Layer upsampling results and encoder 1 Layer feature map stitching; Indicates the transposed convolution kernel; This is the bias term for the transposed convolution; For the decoder's first Layer output.

[0089] The decoder and encoder are connected via skip connections. An attention mechanism is introduced into the skip connections to retain only the feature information relevant to the current segmentation target, reducing redundancy, as shown below:

[0090] ;

[0091] ;

[0092] in, Encoder features; Encoder features The linear transformation weight matrix; Features of the decoder; Decoder features The linear transformation weight matrix; This is an attention weight map; For bias terms; This is pixel-by-pixel multiplication; These are decoder features filtered by the attention mechanism.

[0093] The multi-channel feature map is mapped to a binary classification probability map (pneumoperitoneum / non-pneumoperitoneum) through a 1×1 convolution, and the final probability is output using the Sigmoid function.

[0094] Step S22: Use a segmentation algorithm based on U-Net network to establish an automated segmentation model to achieve automated segmentation of two-dimensional CT images.

[0095] The automated segmentation model can process a large number of CT images in a short time, significantly reducing diagnostic time. Experimental data shows that the time required for the model to process each image is far less than the average time taken by a clinically experienced radiologist to process one CT slice for pneumoperitoneum issues, thus improving diagnostic efficiency.

[0096] An enhanced training set was used during training to prevent overfitting and improve the model's generalization ability. Model training was conducted using the PyTorch framework in a Linux environment equipped with an NVIDIA 3090 GPU, with 200 training epochs, a batch size of 2, and an initial learning rate of 0.00001.

[0097] The RMSprop optimizer is used, with a momentum of 0.9 and a weight decay of 1e-8.

[0098] Introducing the cross-entropy loss function This is used to measure the difference between the predicted probability distribution and the true distribution, as shown below:

[0099] ;

[0100] Where N is the number of categories; For the first The loss for the nth sample represents the model's loss on the nth sample. The difference between the predicted results and the true labels on each sample; This is a category index used to iterate through all possible categories for each sample; Total number of categories; This is an indicator function that indicates which category the element belongs to; To predict probabilities, the observed sample belongs to a category. The predicted probability needs to be estimated and calculated in advance.

[0101] Step S3: Quantitatively optimize the window width and window level of the trained automated segmentation model using a grid search algorithm.

[0102] Step S31, Preliminary Exploration Stage: The image feature extraction performance of different window width and level combinations was compared under lung, bone, and abdominal windows. This stage determined that the image feature extraction under the lung window was relatively better, and could more clearly display the feature information of the pneumoperitoneum region.

[0103] Step S32, Window Width Optimization Stage: Fix the lung window level (-400 HU), adjust the window width every 100 units, train the model and test it. It was found that the performance of various indicators was best when the window width was 800 HU. At this time, the model can segment the pneumoperitoneum area most accurately and effectively distinguish pneumoperitoneum from surrounding tissues.

[0104] Step S33, Window Position Optimization Stage: Fix the window width at 800 HU, adjust the window position every 100 units, test and analyze the results, and determine the optimal window position configuration. The experimental results are shown in Tables 1 and 2.

[0105] Table 1 Results of Fixed Window Level and Quantized Window Width

[0106]

[0107] Table 2. Results of Fixed Window Width and Quantized Window Level

[0108]

[0109] Through this series of optimization experiments, the optimal window width and window level combination for pneumoperitoneum segmentation was finally determined, which significantly improved the accuracy and efficiency of image segmentation, provided standardized guidance for clinical diagnosis, ensured the consistency of diagnostic results among different hospitals and doctors, and improved the overall quality of medical care.

[0110] In addition, the automated segmentation model also includes an automatic detection module that can replace manual labeling of pneumoperitoneum areas, greatly reducing the workload of radiologists and allowing them to devote more energy to other key diagnostic tasks and patient treatment.

[0111] Example 2

[0112] This embodiment uses a series of performance evaluation metrics to measure the accuracy of the algorithm in detecting free air in the peritoneal cavity in CT images, including pixel accuracy, class accuracy, class-average pixel accuracy, crossover ratio (CRO), average crossover ratio (ARO), Dyess coefficient, and recall; the specific calculation methods are as follows:

[0113] Pixel accuracy: ;

[0114] Category accuracy: ;

[0115] Category average pixel accuracy: ;

[0116] Intersection over Union (IOU) = ;

[0117] Average crossover ratio: ;

[0118] Dess coefficient: ;

[0119] Recall rate ;

[0120] in, The number of true positive pixels in the test set, i.e., the pixels that are marked as pneumoperitoneum and correctly predicted as pneumoperitoneum; The number of true negative pixels in the test set, i.e., the pixels that are marked as non-pneumoperitoneous and correctly predicted as non-pneumoperitoneous. The number of false positive pixels in the test set, i.e., pixels that are marked as non-pneumoperitoneum but predicted as pneumoperitoneum; The number of false negative pixels in the test set, i.e., pixels that are labeled as pneumoperitoneum but predicted as non-pneumoperitoneum images; For pixels of the first category; ; The intersection-union ratio of the first category; This represents the crossover ratio for the second category.

[0121] Pixel accuracy and class pixel accuracy reflect the algorithm's accuracy on an overall scale and for specific classes, respectively. Intersection over Union (IoU) and Dessell coefficient are key metrics for evaluating segmentation accuracy, measuring the overlap between the predicted and ground truth regions and the difference between them, respectively. In pneumoperitoneum segmentation, recall reflects the algorithm's sensitivity in detecting free air within the peritoneal cavity, i.e., its ability to avoid missed detections. These comprehensive metrics allow for a complete evaluation of the algorithm's performance and facilitate improvements to address its limitations.

[0122] The optimal parameter combination of a window width of 800 HU and a window level of -400 HU was quantitatively validated using a grid search algorithm. On a self-built high-quality clinical dataset, a segmentation accuracy of 92.3% and a recall of 88.6% were achieved, representing an average accuracy improvement of 23.6% compared to traditional methods. The first intelligent diagnostic benchmark for pneumoperitoneum was established, with a processing speed of 3.2 seconds per CT scan, which is clinically applicable. This significantly reduces the risk of missed diagnoses and provides reliable decision support for the diagnosis and treatment of acute abdominal conditions.

[0123] Example 3

[0124] In the emergency department of a hospital, a patient presenting with abdominal pain was received. The patient's CT images were processed using the method of this invention. First, the patient's CT images were collected and preprocessed. Then, they were input into a U-Net-based segmentation model. The model automatically segmented the pneumoperitoneum region under optimal window width and level configuration (window width 800 HU, window level -400 HU). Ultimately, the model quickly and accurately detected free gas in the patient's peritoneal cavity, providing clinicians with reliable diagnostic information and assisting them in developing timely treatment plans.

[0125] Therefore, the present invention adopts the above-mentioned automated segmentation method for pneumoperitoneum based on U-Net multi-window wide-window level abdominal CT. By dynamically adjusting the combination of window width and window level parameters, it realizes automatic detection and accurate segmentation of free gas in the peritoneal cavity, improves the efficiency and accuracy of pneumoperitoneum diagnosis, reduces the risk of missed diagnosis, and provides a reliable basis for clinical decision-making.

[0126] 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An automated segmentation method for pneumoperitoneum on abdominal CT scans based on U-Net with multi-window wide-window positioning, characterized in that, Includes the following steps: Step S1: Collect patient CT image data and perform preprocessing; In step S1, the collected CT image data is preprocessed, and the specific process is as follows: Step S121: Generate 2D slices along the axis of the DICOM sequence for each case, and simultaneously apply three standardized window levels—abdominal window, bone window, and lung window—to each slice. Among them, the abdominal window highlights the contrast between abdominal gas and soft tissue, the bone window enhances skeletal details, and the lung window enhances the visibility of the lungs and pneumoperitoneum area. Step S122: Stack the single-channel images of the three windows into a multi-channel input by linear gray-level truncation and normalization; Step S123: Perform data augmentation on the training set by randomly rotating, translating, scaling, and perturbing the brightness and contrast of the multi-channel images to simulate anatomical locations and imaging differences. Step S124: Calculate the mean and standard deviation of the three channels based on the training set to achieve global normalization. Perform Z-score standardization on all images to eliminate grayscale differences between devices. During the CT image data preprocessing process, the training set and the test set are strictly isolated to ensure that the preprocessing parameters are derived only from the training data. Step S2: Based on the U-Net network segmentation algorithm, establish an automated segmentation model and train the model; An automated segmentation model is established based on the U-Net network segmentation algorithm. The specific process is as follows: First, a U-Net network is built, which consists of two parts: an encoder and a decoder. The encoder is used for downsampling, and the decoder is used for upsampling. Secondly, an automated segmentation model is established using a segmentation algorithm based on the U-Net network to achieve automated segmentation of two-dimensional CT images; Step S3: For the trained automated segmentation model, the window width and window level are quantitatively optimized using a grid search algorithm. The specific process is as follows: Step S31, Preliminary Exploration Stage: Under the lung window, bone window, and abdominal window respectively, compare the model's performance in extracting image features using different window width and window level combinations, and select the window with the best performance in extracting image features. Step S32, Window width optimization stage: Fix the window position of the optimal window, adjust the window width every 100 units, train the model and test it to determine the window width that makes the performance of each indicator optimal. Step S33, Window Position Optimization Stage: Fix the selected window width parameters, adjust the window position every 100 units, test and analyze the results, and determine the optimal window position configuration; A series of performance evaluation metrics are used to measure the accuracy of this invention in detecting free air in the peritoneal cavity in CT images, including pixel accuracy, class accuracy, class average pixel accuracy, crossover ratio, average crossover ratio, Dess coefficient, and recall.

2. The automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multi-window wide-window positioning as described in claim 1, characterized in that, In step S1, patient CT image data is collected, and the specific process is as follows: First, collect patient CT images and exclude patient images with imaging artifacts and motion artifacts to ensure the quality of the dataset; Then, the pneumoperitoneum is labeled using the diagnostic results in the radiological report to generate a gold standard label; Finally, data were segmented based on individual patients to avoid duplicate data. Based on a list of case IDs with pneumoperitoneum annotations, stratified random sampling was used to divide 70% of the cases into the training set and 30% into the test set, ensuring that the two sets of data did not overlap.

3. The automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multi-window wide-window positioning as described in claim 1, characterized in that, The encoder consists of several repeated convolutional blocks, each containing two 3×3 convolutions, a ReLU activation function, and a 2×2 max pooling operation to achieve spatial downsampling and semantic extraction. The first encoder-based The layer outputs a feature map, as shown below: ； in, Indicates the convolution operation; Represents the ReLU activation function; , For the first Convolutional kernels of layers; , For bias terms; For encoder Layer output; For encoder Layer output; This indicates a max pooling operation.

4. The automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multi-window wide-window positioning as described in claim 1, characterized in that, The decoder restores the low-resolution feature map to the original image size through deconvolution or upsampling operations, and stitches the encoder feature map of the corresponding layer at each layer to achieve high-precision recovery. Based on the decoder The layer outputs a feature map, as shown below: ； in, Indicates the decoder's first Layer upsampling results and encoder 1 Layer feature map stitching; Indicates the transposed convolution kernel; This is the bias term for the transposed convolution; For the decoder's first Layer output; This represents the ReLU activation function.

5. The automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multi-window wide-window positioning as described in claim 4, characterized in that, The decoder and encoder are connected via skip connections. An attention mechanism is introduced into the skip connections to retain only the feature information relevant to the current segmentation target, reducing redundancy, as shown below: ； ； in, Encoder features; Encoder features The linear transformation weight matrix; Features of the decoder; Decoder features The linear transformation weight matrix; This is an attention weight map; For bias terms; This is pixel-by-pixel multiplication; These are decoder features filtered by the attention mechanism; Represents the ReLU activation function; Finally, a 1×1 convolution is used to map the multi-channel feature map into a binary classification probability map, and the final probability is output using the Sigmoid function.

6. The automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multi-window wide-window positioning as described in claim 1, characterized in that, In step S2, the automated segmentation model is trained, and the specific process is as follows: During training, an enhanced training set is used, and the RMSprop optimizer is employed, combined with the cross-entropy loss function. This is used to measure the difference between the predicted probability distribution and the true distribution, as shown below: ； Where N is the number of categories; For the first The loss for the nth sample represents the model's loss on the nth sample. The difference between the predicted results and the true labels on each sample; This is a category index used to iterate through all possible categories for each sample; Total number of categories; This is an indicator function that indicates which category an element belongs to; To predict probabilities, the observed sample belongs to a category. The predicted probability needs to be estimated and calculated in advance.

7. The automated segmentation method for pneumoperitoneum on abdominal CT based on U-Net with multi-window wide-window positioning as described in claim 1, characterized in that, A series of performance evaluation metrics are used to measure the accuracy of this invention in detecting free air in the peritoneal cavity in CT images, including pixel accuracy, class accuracy, class-average pixel accuracy, crossover ratio (CRO), average CRO, Dess coefficient, and recall; the specific calculation methods are as follows: Pixel accuracy: ; Category accuracy: ; Category average pixel accuracy: ; Intersection over Union (IOU) = ; Average crossover ratio: ; Dess coefficient: ; Recall rate ; in, The number of true positive pixels in the test set, i.e., the pixels that are marked as pneumoperitoneum and correctly predicted as pneumoperitoneum; The number of true negative pixels in the test set, i.e., the pixels that are marked as non-pneumoperitoneum and correctly predicted as non-pneumoperitoneum; The number of false positive pixels in the test set, i.e., pixels that are marked as non-pneumoperitoneum but predicted as pneumoperitoneum; The number of false negative pixels in the test set, i.e., pixels that are labeled as pneumoperitoneum but predicted as non-pneumoperitoneum images; For pixels of the first category; ; The intersection-union ratio of the first category; The crossover ratio for the second category; Pixel accuracy and category pixel accuracy reflect the accuracy on the whole and on specific categories; intersection-union ratio and Dess coefficient are key indicators for evaluating segmentation accuracy, measuring the degree of overlap between the predicted region and the real region and the difference between the predicted region and the real region, respectively; in pneumoperitoneum segmentation, recall reflects the sensitivity in detecting free air in the peritoneal cavity, i.e. the ability to avoid missed detections.

Images