Auxiliary interpretation method for quickly diagnosing acute abdominal disease through medical image in combination with deep learning

Abstract

The invention relates to the technical field of medical image artificial intelligence, in particular to an auxiliary interpretation method for medical image rapid diagnosis of acute abdominal disease in combination with deep learning, and the method comprises the steps: 1, organ targeted segmentation: obtaining CT sequences of an arterial phase, a venous phase and a delay phase, and carrying out the dynamic calculation of an interpolation interval according to the layer thickness parameter of scanning equipment, and carrying out the spatial standardization; synchronously generating an intestinal canal mask, a blood vessel mask and a peritoneum mask by adopting an acute abdominal disease directional segmentation network, wherein the network introduces an anatomical size adaptive multi-scale cavity convolution group into a deep layer of an encoder; 2, quantifying dynamic pathological signs; step 3, performing clinical-image gating fusion; and 4, multi-task cooperative diagnosis: inputting the fusion features into a pre-trained multi-task network, and outputting an acute abdominal disease cause classification probability and an operation urgency evaluation value in parallel. The multi-dimensional analysis capability of acute abdominal disease diagnosis is improved through multi-task cooperative diagnosis, and accurate classification of disease causes and evaluation of operation urgency can be carried out at the same time.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology in medical imaging, and in particular to an auxiliary interpretation method for rapid diagnosis of acute abdominal pain in medical images that combines deep learning. Background Technology

[0002] In the field of medical imaging-assisted diagnosis, the interpretation of CT images for acute abdomen faces unique challenges. Acute abdomen encompasses a variety of critical conditions, including appendicitis, intestinal obstruction, and visceral perforation. Its pathological characteristics include multi-organ interaction signs (such as the coexistence of intestinal dilatation and free gas in the abdominal cavity) and dynamic evolutionary features (such as changes in blood supply from early intestinal ischemia to necrosis). Current deep learning-based assisted diagnostic methods have the following inherent limitations:

[0003] Existing technologies directly transfer natural image recognition models (such as ResNet and VGG) to process CT images, without designing dedicated feature extraction mechanisms for the key anatomical structures in acute abdomen.

[0004] Missed organ interaction signs: The universal convolutional nucleus receptive field cannot simultaneously capture the spatial correlation between intestinal dilation (local slender structure) and free gas (diffuse distribution), resulting in insufficient recognition rate of perforation signs;

[0005] Lack of quantitative assessment of signs: The diagnostic results are directly output by the classification network, which lacks the ability to analyze quantitative indicators such as intestinal diameter and gas volume, and cannot meet the clinical needs for quantitative assessment of pathological progress.

[0006] Furthermore, the development of acute abdomen is significantly time-dependent, while current protocols rely solely on single-phase CT analysis during the venous phase.

[0007] Failure to capture evolutionary features: Pathological changes in the early stage (decreased enhancement of the intestinal wall in the arterial phase) and the late stage (no enhancement of the intestinal wall in the venous phase) of intestinal ischemia are confused in a single phase;

[0008] Clinical validation limitations: Studies have shown that the single-phase model has a diagnostic specificity of only 68% for intestinal ischemia, which is far below the clinically acceptable threshold.

[0009] Therefore, there is an urgent need for an auxiliary interpretation method that combines deep learning to quickly diagnose acute abdominal pain from medical images, in order to solve the above problems. Summary of the Invention

[0010] To achieve the above objectives, this invention provides an auxiliary interpretation method for rapid diagnosis of acute abdominal pain using medical images, incorporating deep learning, comprising:

[0011] Step 1: Organ-targeted segmentation: Acquire arterial, venous, and delayed phase CT sequences, and perform spatial standardization by dynamically calculating the interpolation interval based on the slice thickness parameters of the scanning equipment; use an acute abdomen-oriented segmentation network to simultaneously generate intestinal, vascular, and peritoneal masks, and introduce multi-scale dilated convolution groups with anatomical size adaptation in the deep layer of the encoder.

[0012] Step 2: Dynamic Pathological Sign Quantification: Performed based on the mask.

[0013] Intestinal expansion calculation: Generate cross sections at preset intervals along the centerline of the intestinal mask and measure the diameter;

[0014] Calculation of free gas volume change rate: Identify low CT value connected regions within the peritoneal mask and calculate the interphase volume ratio;

[0015] Detection of abnormal enhancement values ​​in organ walls: Measurement of interphase CT value difference in the overlapping area of ​​blood vessel and intestinal tract masks;

[0016] The above-mentioned intestinal dilatation characteristics, free gas volume change rate characteristics, and organ wall enhancement anomalies are used to construct a dynamic feature matrix in time sequence, and the evolution features are extracted by a pre-trained temporal convolutional layer with adjustable convolution width.

[0017] Step 3: Clinical-Image Gated Fusion: The patient's clinical parameters are input into a pre-trained weighted network to generate a weight vector, which is then concatenated with the dynamic features of the image and output as a fused feature through a pre-trained gated attention unit.

[0018] Step 4: Multi-task collaborative diagnosis: Input the fused features into the pre-trained multi-task network, and output the probability of acute abdominal disease etiology classification and the assessment value of surgical urgency in parallel.

[0019] Preferably, the implementation of the anatomically adaptive multi-scale dilated convolution group in step 1 includes:

[0020] For the intestinal segmentation task, the first set of cavity rates is determined based on the range of intestinal anatomical diameters, so that the maximum receptive field covers the maximum physiological expansion diameter of the intestinal segment. The specific process is as follows: the 90th percentile value of the transverse diameter of the intestinal segment in the training dataset is statistically analyzed, and the minimum cavity rate combination required to cover this distance is calculated based on twice this value.

[0021] For peritoneal segmentation, the second set of void ratios is determined based on the peritoneal curvature distance. The specific process is as follows: calculate the maximum radius of curvature of the peritoneal surface on a three-dimensional mask, and use this radius value as a reference to generate a void ratio sequence covering the curved structure.

[0022] The void ratio combination is automatically optimized through backpropagation during the model training phase.

[0023] Preferably, the calculation of the free gas volume change rate in step 2 includes:

[0024] Dual-energy CT material decomposition technology was used to separate gas components: 80kVp and 140kVp dual-phase CT data were acquired, and a gas matrix map was generated by the matrix decomposition algorithm to eliminate the interference of CT values ​​from adipose tissue.

[0025] Identifying gas connected domains within the peritoneal mask: Threshold segmentation is performed on the gas-based material map. The threshold is dynamically set according to the noise level of the scanning equipment. Specifically, the standard deviation of the CT value of the liver parenchyma region is calculated, and three times the standard deviation is taken as the lower limit of the floating threshold.

[0026] Volume change rate calculation: The number of gas connected domain voxels in the arterial and venous phases are counted separately, and then multiplied by the physical volume of a single pixel to calculate the venous phase / arterial phase volume ratio.

[0027] Preferably, the width adjustment rules for the temporal convolutional layer in step 2 include:

[0028] Read the scan timestamps from the DICOM header file and calculate the time intervals between consecutive periods;

[0029] When the time interval is less than or equal to the critical response duration of acute abdominal pathological development, a width is set to cover two consecutive phases; the critical response duration is determined according to the time window of irreversible intestinal ischemia injury in clinical guidelines.

[0030] When the time interval is greater than the critical response duration, the width is set to cover three consecutive phases;

[0031] When the convolution kernel slides along the time dimension, it adopts a causal convolution mode, allowing only historical phase features to influence the current output.

[0032] Preferably, the training method for the weighted network in step 3 includes:

[0033] The network structure is configured with two fully connected layers. The number of neurons in the first layer is equal to the dimension of the clinical parameter features, and the output dimension of the second layer is the same as the dimension of the clinical parameter features.

[0034] During training, all trainable parameters of the image feature extraction network are frozen.

[0035] The weighted network parameters are updated using the gradient backpropagation signal from the acute abdomen etiology classification task. The specific process is as follows:

[0036] Forward propagation calculates the etiology classification loss value, which is calculated using the cross-entropy loss function. The backpropagation process only unfreezes the parameter gradients of the weighted network and updates the weight matrix and bias vector of the weighted network based on the gradient descent algorithm.

[0037] The training termination condition is set to the validation set accuracy improving by less than a set threshold for three consecutive iterations.

[0038] Preferably, the detection of abnormal organ wall enhancement values ​​in step 2 includes:

[0039] Generate an intestinal wall ring detection area: Extend a preset distance inward and outward along the normal direction of the boundary of the intestinal tract mask. The distance is set according to the average thickness of the intestinal wall. Specifically, the median value of the intestinal wall thickness in the training set is statistically analyzed, and 1.5 times this value is used to generate a symmetrical ring area.

[0040] Calculation of abnormal enhancement values ​​of organ walls: The average CT values ​​of the arterial and venous phases are statistically analyzed within the annular region, and the difference between the venous and arterial phases is calculated as the abnormal enhancement value of the organ wall.

[0041] Standardize abnormal enhancement values ​​of organ walls: Calculate the relative enhancement ratio based on the enhancement value of the aorta at the same level.

[0042] Preferably, the operation of the gating attention unit in step 3 includes:

[0043] Gating value generation: The stitched image features and clinical features are input into the fully connected layer, and the output scalar is mapped to the [0,1] interval by the Sigmoid activation function;

[0044] Compensation value calculation: Perform an arithmetic operation of subtracting the gating value from 1;

[0045] Feature fusion: The image feature vector is multiplied element-wise with the gate value, and the clinical feature vector is multiplied element-wise with the compensation value. The two result vectors are then added together and output.

[0046] Preferably, the construction of the surgical urgency assessment value in step 4 includes:

[0047] Tag generation: Patients are classified according to the time interval between imaging examination and surgery. The time window is divided based on the following criteria:

[0048] An interval of ≤2 hours corresponds to label 1.0;

[0049] A label of 0.7 corresponds to an interval of 2 hours to 6 hours.

[0050] An interval > 6 hours corresponds to a label of 0.3;

[0051] The loss function uses Huber loss, and its threshold parameter is dynamically adjusted according to the quantiles of the label distribution.

[0052] Preferably, the loss function optimization method for the pre-trained multi-task network in step 4 includes:

[0053] The focus loss function is used as the basic loss function, and its adjustment factor γ is dynamically set according to the proportion of each class of training data. The setting rule is: the smaller the sample size of the class, the larger the γ value.

[0054] Establish a dynamic weight adjustment mechanism: After each training cycle, calculate the validation set confusion matrix, count the misdiagnosis rate of each category, calculate the average misdiagnosis rate, and increase the loss weight value of categories whose misdiagnosis rate exceeds the average misdiagnosis rate by a multiple.

[0055] The weight adjustment process includes mathematical constraints: the sum of the loss weights for all categories remains constant, and weight normalization is achieved through scaling.

[0056] After the loss function is updated, the model training stability verification condition must be met: the accuracy of the validation set under the new loss function is not lower than the accuracy of the original loss function.

[0057] Preferably, spatial standardization in step 1 includes:

[0058] Interlayer interpolation spacing calculation: Parse the SliceThickness parameter in the DICOM header file and take the minimum layer thickness among the three phases as the reference layer thickness;

[0059] Interpolation algorithm execution: For phases with a layer thickness greater than the reference layer thickness, cubic spline interpolation is performed along the Z-axis direction, and the interpolation point spacing is set to the reference layer thickness;

[0060] Spatial alignment: Using the bifurcation of the portal vein in the liver as an anatomical landmark, three-phase images were registered using affine transformation.

[0061] The beneficial effects of this invention are:

[0062] 1. This invention optimizes the features of key anatomical structures in acute abdomen by designing a dedicated feature extraction mechanism and employing multi-scale convolutional kernels and joint extraction of local and global information from deep learning models. By finely adjusting the receptive field of the convolutional kernels, this method can accurately capture the spatial relationship between slender intestinal loops and diffusely distributed free gas, thereby improving the recognition rate of perforation signs and avoiding missed detections of organ interaction signs.

[0063] 2. This invention, by combining regression analysis methods from deep learning, designs a quantitative model that can not only classify the diagnosis of acute abdomen but also quantitatively assess pathological features, such as intestinal diameter and gas volume. This innovative feature enables the auxiliary diagnostic method to provide detailed assessments of pathological progression, meeting the clinical need for quantitative analysis of acute abdomen and improving its clinical applicability.

[0064] 3. This invention employs multi-phase CT image analysis, combined with a deep learning model of multi-phase CT images including arterial and venous phases, to capture the evolution of intestinal ischemia at different time points. Through joint analysis of multi-phase CT images, the model can identify enhancement changes in the intestinal wall during the arterial and venous phases, thereby accurately diagnosing the pathological stage of intestinal ischemia. This technology effectively solves the problem of failing to capture the evolutionary characteristics of acute abdomen, improving diagnostic accuracy.

[0065] 4. Through comprehensive analysis of multi-phase CT images and training of the optimized model, this invention significantly improves the specificity of acute abdominal diagnosis, especially the diagnostic accuracy of intestinal ischemia. By combining arterial and venous phase CT image data, the model can more accurately distinguish between early and late changes in intestinal ischemia, thereby improving the diagnostic specificity of intestinal ischemia and reaching or exceeding clinically acceptable diagnostic criteria. This innovation effectively solves the clinical validation deficiencies of single-phase models, giving the diagnostic method higher clinical value in practical applications. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0068] Figure 2 This is a flowchart of the steps for calculating the free gas volume change rate in step 2 of the method of the present invention;

[0069] Figure 3 This is a flowchart of the steps for detecting abnormal organ wall enhancement values ​​in step 2 of the method of the present invention. Detailed Implementation

[0070] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0071] Please see Figures 1-3This invention provides an auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning. In step 1, CT sequence images of the arterial, venous, and delayed phases are first acquired. The spatial resolution of these images is affected by the slice thickness of the scanning device. Therefore, spatial standardization is performed by dynamically calculating the interpolation interval based on the slice thickness parameters of the device to ensure image consistency. Subsequently, an acute abdomen-oriented segmentation network is used, introducing multi-scale dilated convolutional groups with anatomically adaptive dimensions into the deep layers of the encoder to better extract complex anatomical structural features, especially in key areas such as the intestines, blood vessels, and peritoneum. This network can simultaneously generate intestinal masks, vascular masks, and peritoneal masks, providing accurate region segmentation for subsequent analysis.

[0072] By employing efficient multi-scale dilated convolution and an anatomical size adaptive mechanism, this step can accurately extract anatomical structures related to acute abdomen. Compared to traditional fixed-size convolution, this method is better able to adapt to the differences in anatomical structures among different patients, improving the segmentation accuracy of organs such as intestines, blood vessels, and peritoneum, and solving the problem that traditional segmentation methods cannot accurately capture local structures.

[0073] In step 2, the intestinal tract mask, vascular mask and peritoneal mask obtained in step 1 are used for quantitative analysis of pathological features.

[0074] Intestinal expansion calculation: Cross sections are generated at preset intervals along the centerline of the intestinal mask, and the diameter of each cross section is measured to quantify the expansion of the intestinal tract.

[0075] Free gas volume change rate calculation: Identify low CT value connected regions within the peritoneal mask, calculate the volume ratio between different phases, and quantify the change in free gas.

[0076] Detection of abnormal enhancement values ​​in organ walls: By measuring the difference in CT values ​​between phases in the overlapping area of ​​blood vessel-intestinal tract masks, abnormal enhancement changes in organ walls can be identified.

[0077] These quantified features are then used to construct a dynamic feature matrix in time series, and evolutionary features are extracted through pre-trained temporal convolutional layers with adjustable convolution width to capture the dynamic pathological changes of acute abdomen.

[0078] This step quantifies key pathological features such as intestinal dilatation, changes in free gas volume, and abnormal organ wall enhancement, providing more accurate data on pathological progression. Traditional methods often rely on qualitative analysis, while this invention, through quantitative analysis, provides more clinically valuable dynamic indicators, helping physicians monitor disease progression in real time.

[0079] In step 3, the patient's clinical parameters (such as age, medical history, and clinical symptoms) are input into a pre-trained weighted network to generate a weight vector. Then, the clinical weight vector is concatenated with the image dynamic feature matrix, and the fused features are output through a pre-trained gating attention unit. The gating mechanism can automatically adjust the weights of each input feature based on the correlation between image and clinical data, further improving the model's focus on important information.

[0080] This step, by integrating clinical and imaging data, enables a more comprehensive diagnosis of acute abdomen. The gating attention mechanism automatically adjusts the importance of input data, ensuring the proper integration of imaging data and clinical parameters, thereby enhancing the model's accuracy and robustness and making the diagnostic results more consistent with clinical realities.

[0081] In step 4, the fused features are input into a pre-trained multi-task network for analysis. This network outputs in parallel the probabilities of etiological classification for acute abdominal pain (such as appendicitis, intestinal obstruction, visceral perforation, etc.) and the assessment value of surgical urgency (such as whether emergency surgery is required). This multi-task network performs classification and regression tasks simultaneously in the same model, improving the efficiency and accuracy of diagnosis.

[0082] Multi-task collaborative diagnosis enhances the multi-dimensional analytical capabilities for diagnosing acute abdomen, enabling simultaneous and precise classification of etiologies and assessment of surgical urgency. Compared to traditional single-task models, it provides clinicians with more comprehensive and specific diagnostic information, facilitating rapid decision-making and treatment.

[0083] This invention, combining deep learning technology, effectively addresses key issues in the diagnosis of acute abdomen through steps such as organ-targeted segmentation, dynamic pathological sign quantification, clinical-image gating fusion, and multi-task collaborative diagnosis. These technical features not only improve diagnostic accuracy and stability but also provide clinicians with more quantitative and dynamic pathological data, thereby enhancing the rapid diagnostic capability for acute abdomen and demonstrating high clinical practical value.

[0084] In one possible implementation, in the intestinal segmentation task, the cavity rate is first determined based on the range of intestinal anatomical diameters in the training dataset. Specifically, the 90th percentile of the transverse diameter of the intestinal segments in the training dataset is statistically analyzed, and twice this value is used as a benchmark to calculate the minimum combination of cavity rates required to cover the maximum physiological expansion diameter of the intestinal segments. The cavity rate defines the proportion of holes (i.e., non-information regions in the receptive field) in the convolutional kernel. Increasing the cavity rate expands the receptive field of the convolutional kernel, allowing the convolutional layer to capture features over a wider range.

[0085] This method dynamically adjusts the cavity rate based on the maximum anatomical diameter of the intestine, enabling the convolutional neural network to fully cover the physiological expansion range of the intestine while maintaining computational efficiency. This helps to accurately extract the boundary information of the intestine, especially when the intestine is significantly dilated or changes in size, providing more accurate segmentation results and reducing the segmentation inaccuracies caused by anatomical differences in traditional methods.

[0086] For peritoneal segmentation, the void ratio is determined based on the curvature of the peritoneum. On a 3D mask, the maximum radius of curvature of the peritoneal surface is first calculated, reflecting the degree of curvature. Using this radius of curvature as a benchmark, a void ratio sequence is generated that covers this curved structure. This step ensures that the convolutional kernel can adapt to the curved surface structure of the peritoneum, thus effectively extracting features even in the curved regions of the peritoneum.

[0087] The curved surface structure of the peritoneum poses a significant challenge to image segmentation, especially in the junction region between the intestinal lining and the peritoneum. By adaptively adjusting the cavity rate based on the peritoneal curvature, a convolutional neural network can precisely adapt to different anatomical structures, ensuring accurate capture of curved regions during segmentation. This method avoids the problem of traditional methods struggling to handle complex structures due to fixed cavity rate settings, thus improving the segmentation accuracy of the peritoneum.

[0088] During model training, the combination of dilatation rates is automatically optimized through backpropagation. By optimizing the loss function, the network can progressively adjust the combination of dilatation rates, allowing the convolutional kernels to adapt to the structural features of different organs. During training, the dilatation rates are automatically adjusted to maximize segmentation results, thereby improving the performance of each organ segmentation task.

[0089] By introducing a backpropagation mechanism to automatically optimize the hole rate during training, the receptive field of each convolutional layer can be autonomously adjusted to better adapt to the needs of different anatomical structures. This adaptive optimization enables the network to achieve good segmentation results on image data from different patients, improving the model's generalization ability, especially when processing images of patients with different physiological characteristics.

[0090] By adaptively adjusting the cavity rate in the segmentation tasks of the intestine and peritoneum, the convolutional neural network can be optimized for the characteristics of different anatomical structures, thereby achieving higher accuracy in image segmentation. By dynamically setting the cavity rate combination based on anatomical diameter and radius of curvature, and combining it with automatic optimization via backpropagation, the application effect of the segmentation model in the diagnosis of acute abdomen is greatly improved. This adaptive mechanism enables the deep learning model to better process image data from different patients, providing more accurate medical image analysis results.

[0091] In one possible implementation, during dual-energy CT scanning, gas components are separated using a matrix decomposition algorithm based on dual-phase CT data acquisition at 80 kVp and 140 kVp. This algorithm effectively distinguishes different substances (such as gases, fats, and soft tissues) based on the differences in energy dependence of CT images. Matrix decomposition generates a gas matrix map, thereby accurately extracting gas components and eliminating interference from adipose tissue on CT values. Since adipose tissue has a low density in CT images and may affect gas resolution, this separation step is crucial.

[0092] Using dual-energy CT material decomposition technology can improve the resolution and accuracy of gas components, avoid interference from CT values ​​in adipose tissue, and ensure precise gas separation and detection. This provides clearer images for subsequent gas connectivity identification and volume change rate calculation, thereby improving the diagnostic accuracy of acute abdomen.

[0093] Within the peritoneal mask, the generated gas-based material map is segmented using a thresholding method. The threshold is dynamically adjusted based on the noise level of the scanning equipment. Specifically, the standard deviation of the CT values ​​for the liver parenchyma is first calculated, and a floating lower threshold, which is three times the standard deviation, is then calculated. This dynamic threshold setting effectively addresses noise variations under different patient and equipment conditions, ensuring that only gas-related connected regions are preserved during segmentation, while excluding other noise influences.

[0094] By setting dynamic thresholds and calculating floating thresholds, this method can automatically adapt to different scanning devices and patient image quality, thereby accurately identifying gas connectivity regions within the peritoneum. This is crucial for accurately assessing gas accumulation and distribution changes in patients with acute abdomen, effectively improving diagnostic accuracy.

[0095] The calculation of the volume change rate is divided into two parts: the arterial phase and the venous phase. First, in the CT images of the arterial and venous phases, the number of voxels in the gas connected domain is counted and multiplied by the physical volume of a single pixel to calculate the volume of the gas connected domain in the arterial and venous phases. Then, the ratio of the venous phase volume to the arterial phase volume is calculated to obtain the free gas volume change rate. This ratio reflects the accumulation and distribution of gas, thus providing a basis for the diagnosis of acute abdomen.

[0096] By calculating the rate of change in volume, it is possible to dynamically assess changes in gas at different time points (such as the arterial and venous phases), thereby determining gas flow or retention. For the diagnosis of acute abdomen, especially in assessing intestinal perforation or abdominal infection, the rate of change in free gas is a crucial indicator. This method can accurately measure gas changes, providing clinicians with objective and quantitative reference data.

[0097] By combining innovative methods such as dual-energy CT material decomposition, dynamic threshold segmentation, and volume change rate calculation, changes in free gas in images of patients with acute abdomen can be identified and analyzed efficiently and accurately. These techniques not only improve image processing accuracy but also effectively eliminate noise and interference factors, thus providing more reliable data support for the diagnosis of acute abdomen. The application of this method is of great significance for improving the early diagnosis capability of acute abdomen, reducing the misdiagnosis rate, and improving treatment outcomes.

[0098] In one possible implementation, the scan timestamps are first read from the DICOM format medical images. The DICOM header file contains detailed information about the image acquisition, including the specific time points of the scan. By calculating the time intervals between consecutive phases (such as the arterial phase, venous phase, etc.), the temporal relationships between image data can be accurately assessed. This information is crucial for subsequent temporal convolution operations, ensuring that the temporal dependencies between different phase data are utilized appropriately.

[0099] By accurately acquiring time interval information, it is possible to ensure that the temporal differences between image data are fully considered, thereby enhancing the ability of subsequent temporal convolutional layers to understand image sequences at different stages. This is particularly important for dynamically monitoring the development of acute abdominal lesions, providing rich temporal features for deep learning models and improving diagnostic efficiency.

[0100] The convolution width is set based on the calculated time interval. Specifically, when the time interval is less than or equal to the critical response duration for the development of acute abdominal pathology, the convolution width is set to cover two consecutive phases; when the time interval is greater than the critical response duration, the convolution width is set to cover three consecutive phases. The critical response duration here is determined based on the time window for irreversible intestinal ischemia-related injury in clinical guidelines, and is usually a fixed time range (e.g., 3 hours). The selection of this time window ensures a timely response in cases of rapid pathological progression in acute abdominal conditions.

[0101] By adjusting the convolution width according to the actual time interval, it is possible to more flexibly adapt to imaging changes at different pathological stages. When the time interval is short, the lesion may progress rapidly, so a smaller convolution width is used to cover two phases to avoid over-reliance on earlier imaging information and ensure sensitive interpretation of rapidly progressing acute abdomen. When the time interval is long, pathological changes may be slower, and in this case, by expanding the convolution width to cover three phases, information from multiple phases can be fully utilized for accurate diagnosis.

[0102] When the convolutional kernel slides along the time dimension, a causal convolution mode is employed. In causal convolution, the output at each time step is only affected by information from historical phases (i.e., previous phases) and not by future phases. This ensures the causality of the model, meaning that the prediction at the current moment depends only on past observations and not on future phase data, thus better aligning with the actual diagnostic logic of medical imaging.

[0103] Causal convolution can effectively ensure the temporal consistency and causal relationships of the model, thereby preventing the "leakage" of future information from affecting current predictions. In the interpretation of acute abdominal pain, the pathological development is often time-dependent. Therefore, by using causal convolution, the disease development process can be simulated more accurately, improving the model's sensitivity to the temporal features of acute abdominal pain and its predictive accuracy.

[0104] By combining time interval information, dynamic convolution width adjustment rules, and causal convolution patterns, efficient temporal analysis of acute abdominal images was achieved. The dynamic adjustment of the convolution width allows the model to flexibly adapt to the characteristics of different phases based on the rate of pathological progression, while the causal convolution pattern ensures the accuracy of the temporal sequence. These technical features significantly improve the accuracy of auxiliary interpretation of medical images for the rapid diagnosis of acute abdominal conditions, providing clinicians with more timely and accurate diagnostic support, especially in determining the progression stage of acute abdominal conditions, providing crucial evidence for early intervention and treatment.

[0105] In one possible implementation, the weighted network is configured with two fully connected layers. The first layer has the same number of neurons as the clinical parameter features, and the second layer's output dimension also matches the clinical parameter features. This configuration ensures that the network can process information matching the clinical parameter features, thus maintaining consistency between input and output dimensions. The fully connected layers allow the network to effectively learn complex relationships between different features, especially the nonlinear connections between imaging features and clinical data.

[0106] This network configuration helps the network maintain efficient and intuitive output when capturing the correlation between imaging features and clinical parameter features. By accurately modeling the input features, the network can enhance its comprehensive feature analysis capabilities in the interpretation of acute abdomen, thereby improving the overall performance of the model.

[0107] During training, all trainable parameters of the image feature extraction network are frozen. This means that the image feature extraction network does not participate in gradient updates during training, but only provides information for the subsequent weighted network through its output image features. This strategy effectively shifts the training focus to the weighted network, avoiding model overfitting caused by over-adjusting the parameters of the image feature extraction network.

[0108] Freezing the parameters of the image feature extraction network effectively reduces computational resource consumption and allows for focused training of the weighted network, enabling it to concentrate on combining image features and clinical parameters for the diagnosis of acute abdomen. This strategy helps improve training speed and stability, especially in deep learning models, preventing potential parameter redundancy and overfitting issues.

[0109] The parameters of the weighted network are updated using the backpropagation signal of the gradient from the acute abdomen etiology classification task. During forward propagation, the loss value for etiology classification is calculated, and the cross-entropy loss function is used to measure the difference between the predicted result and the true label. During backpropagation, only the gradients of the weighted network parameters are unfrozen, while the parameters of the image feature extraction network remain frozen. Then, the weight matrix and bias vector of the weighted network are updated based on the gradient descent algorithm.

[0110] This gradient-based backpropagation training method ensures that the weighted network can extract effective etiological classification information from image features and clinical parameters. The cross-entropy loss function performs well in classification tasks, effectively guiding the network to learn how to distinguish between cases of acute abdomen with different etiologies, thereby improving the accuracy and reliability of diagnosis.

[0111] The training termination condition is set to the validation set accuracy improving by less than a set threshold for three consecutive iterations. This strategy allows training to automatically stop when model performance stabilizes, avoiding overfitting and improving training efficiency.

[0112] By setting training termination conditions, ineffective overtraining can be avoided during model training, ensuring that the model stops in a timely manner when it reaches optimal performance, thereby saving computational resources and preventing overfitting. This strategy helps ensure the model's generalization ability and improves its performance in real-world applications.

[0113] By carefully designing the structure and training strategy of the weighted network, image features and clinical parameters can be effectively combined to assist in the diagnosis of acute abdomen. By freezing the parameters of the image feature extraction network and updating only the parameters of the weighted network, overtraining of the image feature extraction component is avoided, ensuring the model's efficiency and stability. Simultaneously, a gradient update strategy based on the cross-entropy loss function allows the weighted network to focus on the etiological classification of acute abdomen, improving diagnostic accuracy. Setting a training termination condition further enhances training efficiency, ensuring the model stops at the optimal moment, thereby improving the model's applicability and practicality.

[0114] In one possible implementation, firstly, a predetermined distance is extended inward and outward along the normal direction from the mask boundary of the intestinal tract. This predetermined distance is not fixed but is set based on the average thickness of the intestinal wall. Specifically, the median value of the intestinal wall thickness in the training set is statistically analyzed, and then 1.5 times this median value is taken as the extension distance to generate a symmetrical ring-shaped detection area. This method, by flexibly adjusting the size of the detection area, can adapt to individual differences in intestinal wall thickness, ensuring efficient detection of abnormal enhancement areas in patients of different body types or with different lesion degrees.

[0115] The generated annular detection area ensures that the detection range matches the actual structure of the intestinal wall, effectively avoiding omissions or misdiagnoses caused by uneven intestinal wall thickness or local variations. Furthermore, by calculating the median intestinal wall thickness and setting the extension distance accordingly, the personalization and adaptability of the detection area are ensured, improving detection accuracy.

[0116] Within the aforementioned annular region, the average CT values ​​for the arterial and venous phases are calculated. By comparing the differences in CT values ​​between the venous and arterial phases, the abnormal enhancement value of the organ wall is calculated. The difference in CT values ​​between the arterial and venous phases is an important indicator reflecting changes in organ wall enhancement. Under normal circumstances, the degree of enhancement of the organ wall in the arterial and venous phases follows a certain pattern. When the difference is abnormal, it usually indicates the presence of lesions or abnormal blood flow changes in the organ wall.

[0117] This method of calculating enhancement anomalies based on the difference in CT values ​​between the arterial and venous phases can effectively reflect the blood flow and enhancement of organ walls, and is an important basis for diagnosing the cause of acute abdomen. By calculating the difference in CT values ​​between the arterial and venous phases, it can highlight abnormal phenomena that need attention in diagnosis, thereby improving the disease recognition rate and the accuracy of judgment.

[0118] To further improve the accuracy and contrast of outlier detection, the enhancement abnormalities of organ walls need to be standardized. The standardization process uses the enhancement value of the aorta at the same level as a benchmark to calculate the relative enhancement ratio. This ratio allows for comparison of organ wall enhancement abnormalities with the enhancement status of the aorta, thereby eliminating differences between patients due to factors such as body size and age, making the assessment of enhancement abnormalities more objective and consistent.

[0119] Standardization allows for more accurate comparison of organ wall enhancement abnormalities among different patients, eliminating the influence of individual differences and making the interpretation results more reliable and consistent. Standardized data also aids in the training of deep learning models, further enhancing their discriminative and generalization abilities.

[0120] By generating personalized circular detection areas of the intestinal wall and combining the differences in CT values ​​between the arterial and venous phases, abnormal enhancement values ​​are calculated and standardized to make the detection results more accurate and reliable. This technical step can effectively improve the rapid diagnosis of acute abdomen, especially in clinical image analysis, helping doctors to more accurately identify potential organ lesions and reduce the risk of misdiagnosis and missed diagnosis. Through the combination of deep learning and image features, this invention not only optimizes the diagnostic process of acute abdomen but also improves the overall performance and adaptability of image-assisted diagnostic methods.

[0121] In one possible implementation, the gating value is generated by processing the stitched image features and clinical features into a fully connected layer. The stitched feature vector includes image features (such as CT images, MRI images, etc.) and clinical features (such as medical history, age, gender, etc.), which are input together into the fully connected layer. After processing, the output scalar value is processed by a sigmoid activation function and mapped to the [0,1] interval. This scalar value represents a gating value, reflecting the importance or influence of a specific feature combination on the diagnosis.

[0122] By using the Sigmoid function, the gate value can be normalized within the [0,1] interval, allowing it to be used as a weight to adjust the influence of imaging and clinical features in further calculations. A larger gate value indicates that imaging features are more important in the current diagnosis, while a smaller gate value suggests that clinical features or other factors may be more important. This operation effectively extracts the correlation between different features and provides a basis for subsequent decision-making.

[0123] The compensation value is calculated by subtracting 1 from the gate value, i.e., compensation value = 1 - gate value. The compensation value is essentially a complement to the gate value, reflecting the weight of clinical features related to the imaging characteristics. A larger compensation value indicates a stronger influence of the clinical features in the case; conversely, a larger compensation value when the gate value is small indicates a more prominent importance of the clinical information.

[0124] The calculation of compensation values ​​ensures the complementarity between imaging and clinical features. This complementarity helps enhance the accuracy of feature fusion, avoids over-reliance on a single feature type (such as imaging or clinical data), and thus improves the reliability and robustness of diagnosis.

[0125] The feature fusion operation weights image features with a gating value and clinical features with a compensation value through element-wise multiplication. Specifically, each element in the image feature vector is multiplied element-wise with the gating value, and each element in the clinical feature vector is multiplied element-wise with the compensation value. Then, these two weighted result vectors are added together to output the final fused feature.

[0126] The purpose of feature fusion is to weight imaging and clinical features with different weights, thereby comprehensively considering the contributions of both types of features to diagnostic decisions. In this way, the model can automatically learn how to balance the weights of imaging and clinical information, improving the accuracy and efficiency of diagnosis. The integrated use of imaging and clinical features allows the model to not only extract subtle lesion information from imaging data but also fully incorporate the patient's clinical background, enhancing the model's clinical applicability.

[0127] The gated attention unit in this embodiment of the invention finely adjusts the combination of image features and clinical features through three steps: gating value generation, compensation value calculation, and feature fusion, thereby enhancing the auxiliary role of medical imaging in the diagnosis of acute abdomen. It can dynamically adjust the weights of different features, thus more accurately reflecting lesion information and the patient's clinical background. The feature fusion step ensures the complementary use of image data and clinical data, improving the comprehensiveness and reliability of diagnostic results. Through training with a deep learning model, this method can automatically optimize the feature fusion process, continuously improving diagnostic accuracy.

[0128] In one possible implementation, label generation is based on the time interval between patient imaging examination and surgery. The specific time window division is as follows:

[0129] When the time interval is less than or equal to 2 hours, a label of 1.0 is assigned, indicating that the patient needs surgery as soon as possible and the urgency is extremely high;

[0130] When the time interval is greater than 2 hours and less than or equal to 6 hours, a label of 0.7 is assigned, indicating that the surgery is highly urgent, but can still be slightly delayed;

[0131] When the time interval is greater than 6 hours, a label of 0.3 is assigned, indicating that the urgency of the surgery is low and it can be delayed.

[0132] By dividing the time window, tag generation can effectively reflect the urgency of surgery for patients with acute abdomen. Different tags represent different clinical decision-making information, helping doctors make timely and appropriate judgments based on time and changes in the patient's condition.

[0133] The tags are generated based on the patient's actual condition and the timing of surgery, enabling a quantifiable assessment of the urgency of surgery for patients with acute abdomen, providing a standardized and quantifiable reference indicator. In this way, doctors can not only rely on imaging information for diagnosis, but also use these tags to determine the optimal timing of surgery, reducing subjective bias in clinical decision-making and improving treatment efficiency.

[0134] To optimize the model's learning process, the Huber loss function was adopted. The Huber loss function is more tolerant of large errors but remains sensitive to small errors, thus effectively balancing noise and accuracy during model training. This loss function has a certain degree of robustness, reducing the impact of outliers when processing medical image data, thereby improving the model's stability and generalization ability.

[0135] The Huber loss function effectively handles outlier data or erroneous labels that may exist during label generation while avoiding overfitting, thus improving the stability of model training. In medical image analysis, due to the complexity and heterogeneity of image data, using Huber loss can ensure higher model robustness, thereby providing a more accurate basis for the rapid diagnosis of acute abdominal pain.

[0136] In the Huber loss function, the threshold parameter setting is a crucial factor affecting training performance. This invention dynamically adjusts the threshold parameter based on the quantiles of the label distribution to adapt to variations in different clinical situations and patient groups. By dynamically adjusting the threshold, the loss function can better adapt to different label distributions, thereby improving the model's generalization ability across various patient groups.

[0137] By dynamically adjusting the threshold parameter of the loss function, it is possible to flexibly adjust it according to the distribution of labels and the actual situation of the data, avoiding the excessive constraints caused by a fixed threshold. This method effectively improves the model's performance in practical applications, enabling it to adapt to a wider range of patient data and enhancing the accuracy of acute abdominal diagnosis.

[0138] In this embodiment of the invention, the construction of the surgical urgency assessment value effectively quantifies the surgical urgency of patients with acute abdomen through technical features such as label generation, Huber loss function, and dynamic adjustment of threshold parameters, providing timely decision support for clinical practice. The implementation of these technical features not only improves the accuracy of rapid diagnosis of acute abdomen but also ensures the stability and generalization ability of the model, thus possessing high value and application prospects in practical clinical applications.

[0139] In one possible implementation, the focus loss function is commonly used to address class imbalance. In medical imaging, due to the limited number of samples in certain classes, traditional loss functions may lead to poor model recognition performance for these classes. The focus loss function enhances the difference between easy and difficult samples by introducing a regulation factor γ, where a larger γ value indicates a greater difference in influence between easy and difficult samples. To further enhance the handling of class imbalance, the regulation factor γ in this invention is dynamically set according to the sample size ratio of each class in the training data, with the rule being: the smaller the sample size of a class, the larger the γ value. This ensures the importance of rare classes during training, preventing them from being dominated by majority class samples and improving the recognition performance of minority classes.

[0140] By dynamically adjusting the γ value, the focus loss function can adaptively handle class imbalance in the data. Classes with larger sample sizes do not gain excessive influence during training, while classes with smaller sample sizes (such as certain specific diseases in acute abdominal pain) receive more attention, thereby improving the model's predictive ability for rare diseases and enhancing the comprehensiveness and accuracy of diagnosis.

[0141] After each training cycle, the confusion matrix on the validation set is first calculated, and the false diagnosis rate for each category is tallied. Then, the average false diagnosis rate for all categories is calculated. For categories with false diagnosis rates exceeding a certain multiple of the average false diagnosis rate, the model increases their loss weights. This adjustment mechanism, by increasing the loss weights of categories with more false diagnoses, makes the model pay more attention to these categories in subsequent training, thereby reducing their false diagnosis rates.

[0142] This dynamic adjustment mechanism ensures that the model can flexibly adapt to the learning needs of different categories during training, especially for categories that are easily misdiagnosed, allowing for more correction during training. By increasing the weights of these categories, the model can more accurately identify these "difficult-to-diagnose" conditions, ultimately improving the overall accuracy and reliability of diagnosis.

[0143] When adjusting the weights of the loss function, the sum of the loss weights for all classes must remain constant. To achieve this, the weights are normalized using a scaling method, which involves adjusting the weights of the increased classes while reducing the weights of the other classes, so that the total sum of the loss weights for all classes remains unchanged.

[0144] Mathematical constraints prevent training instability issues that might arise from over-adjusting the weights of a particular category. Maintaining a constant sum of loss weights across all categories ensures that the model pays attention to difficult-to-diagnose categories while also giving appropriate consideration to other categories, thus avoiding excessive bias towards any one category during training.

[0145] After updating the loss function, it is essential to ensure that the accuracy on the validation set under the new loss function is not lower than that under the original loss function. This validation condition ensures that the newly optimized loss function will not lead to a decrease in model performance. Training updates will only proceed if the new loss function achieves equivalent or better performance on the validation set.

[0146] This condition effectively prevents the risk of overfitting or optimization failure, ensuring the stability of model training. By setting the validation condition, it is possible to detect in real time whether the optimization method has improved the model's generalization ability, ensuring the accuracy of the final model in practical applications.

[0147] By combining a focus-based loss function with a dynamic adjustment mechanism, the loss function in multi-task networks can be effectively optimized, improving the accuracy of acute abdominal pain diagnosis. Through adaptive adjustment of class imbalance, dynamic weighting of misdiagnosis rates, and the appropriate application of mathematical constraints, this method enhances the model's ability to identify minority classes and difficult-to-diagnose diseases, thereby improving the comprehensiveness and accuracy of diagnosis. The implementation of these technical features not only improves the model's training effect but also provides more accurate and reliable auxiliary interpretation tools for clinical decision-making.

[0148] In one possible implementation, in medical image processing, the header information of the DICOM file contains the slice thickness parameter for each layer. For three-phase images, since the slice thickness may differ between phases, these slice thicknesses need to be standardized. First, the slice thickness parameter in the DICOM header file is parsed to obtain the slice thickness information for each phase in the three-phase image. Then, the minimum slice thickness among the three phases is taken as the baseline slice thickness. This minimum slice thickness value is used as the standard, and the slice thicknesses of other phase images are adjusted to this standard.

[0149] By standardizing the inter-layer interpolation interval, the problem of inconsistent layer thickness in images of different phases can be solved, ensuring that all images of different phases have the same resolution in the spatial dimension. This guarantees the consistency of input data for subsequent deep learning models, thereby improving the stability and accuracy of model training.

[0150] If the thickness of a certain phase is greater than the reference layer thickness, cubic spline interpolation is required along the Z-axis. Cubic spline interpolation is a smooth interpolation method that can effectively reduce errors caused by inconsistent layer thicknesses. The spacing between interpolation points is set to the reference layer thickness to ensure that the resolution of the interpolated image in the Z-axis direction is consistent with the reference layer thickness.

[0151] Cubic spline interpolation can smoothly insert new slice points, avoiding image quality degradation caused by inappropriate interpolation algorithms. This smooth interpolation method can ensure the spatial continuity of the image, reduce noise or distortion that may be introduced during the interpolation process, thereby improving image quality and providing more consistent and high-quality input for deep learning models.

[0152] During spatial alignment, the bifurcation point of the hepatic portal vein was used as an anatomical landmark for image spatial registration. First, the location of the hepatic portal vein bifurcation point in the image was detected as a reference point for image alignment. Then, an affine transformation method was used to register the three-phase images, aligning the hepatic portal vein bifurcation point in all images to ensure spatial consistency.

[0153] Spatial alignment using the bifurcation of the hepatic portal vein as a landmark ensures consistency in anatomical structure across images from different phases. The bifurcation of the hepatic portal vein, as a highly recognizable anatomical landmark, effectively avoids spatial location information deviations caused by different scanning angles or equipment, ensuring accurate image alignment. This is crucial for subsequent deep learning models to extract image features, effectively improving the model's ability to recognize image details.

[0154] By applying inter-slice interpolation spacing calculation, interpolation algorithm execution, and spatial alignment techniques, this invention enables standardized processing of three-phase medical images. These technical steps effectively address issues such as inconsistent slice thickness and spatial location information deviations in image data, providing unified and standardized input data for subsequent deep learning models. Through these processing steps, not only is the quality and consistency of medical image data improved, but more reliable input is also provided for accurate model diagnosis, ultimately enhancing the auxiliary diagnostic capabilities for acute abdominal conditions.

[0155] The following examples will illustrate this in detail:

[0156] This invention significantly improves the accuracy and efficiency of diagnosing acute abdominal pain through a series of algorithms, including image preprocessing, feature extraction, spatial alignment, and focus loss function optimization.

[0157] This invention is applied in emergency medicine and radiology to help doctors automatically identify and diagnose CT image data of patients with acute abdominal conditions. This embodiment uses intestinal obstruction and appendicitis in acute abdominal conditions as examples to demonstrate the application process in a real medical scenario.

[0158] This method uses a publicly available dataset of CT images of acute abdomen, containing 5,000 CT images of patients with acute abdomen.

[0159] Pixel values ​​of all CT images were normalized to the [0,1] interval to eliminate the influence of equipment differences. Gaussian filtering was used to denoise the CT images, with a standard deviation of σ = 1.0 and a filter kernel size of 3x3. The slice thickness of different CT images was standardized to a uniform 0.8 mm. If the slice thickness was inconsistent in the images, cubic spline interpolation was used for adjustment: for each image, the slice thickness was standardized to 0.8 mm.

[0160] The interpolation formula is:

[0161] I new (x)=(1-t)I0(x)+tI1(x), t∈[0,1];

[0162] Where I0(x) and I1(x) are the pixel values ​​of adjacent layers, and t is the interpolation weight.

[0163] Convolutional Neural Networks (CNNs) were chosen for feature extraction and classification. The model architecture is as follows:

[0164] Input layer: Receives images of size 224x224.

[0165] Convolutional layer 1: 32 3x3 convolutional kernels with a stride of 1 and ReLU activation function.

[0166] Pooling layer 1: max pooling, pooling window size is 2x2, stride is 2.

[0167] Convolutional layer 2: 64 3x3 convolutional kernels with a stride of 1 and ReLU activation function.

[0168] Pooling layer 2: max pooling, with a pooling window size of 2x2 and a stride of 2.

[0169] Fully connected layer: contains 128 neurons, activation function ReLU.

[0170] Output layer: The output layer contains two nodes, which are the classification probabilities of intestinal obstruction and appendicitis, respectively, and the activation function is Softmax.

[0171] The batch size is set to 32.

[0172] The learning rate was set to 0.001, and the optimizer used was Adam.

[0173] The number of training epochs is set to 50.

[0174] To address the problem of imbalanced data classes, a focus loss function is used to optimize model training, preventing samples from the larger class from dominating the model's training and thus improving the classification accuracy of the smaller class. The formula for the focus loss function is as follows:

[0175] L focal =-α(1-p t ) γ log(p t );

[0176] Where: p t α is the model's predicted probability for the current sample; α is the class weight factor, considering the impact of class imbalance, set to α = 0.25; γ is the adjustment factor, used to suppress the influence of easily classified samples, set to γ ​​= 2.

[0177] To ensure consistency between different CT images, an affine transformation is used to align all images. The mathematical form of the affine transformation is:

[0178]

[0179] Where A is a 2×2 rotation matrix and b is a translation vector.

[0180] To increase the robustness of the model, data augmentation strategies such as random rotation (±10°), flipping, and cropping are adopted.

[0181] The 5000 CT images were divided into an 80% training set (4000 images) and a 20% validation set (1000 images). During training, the loss value and accuracy of the validation set were output every 100 rounds.

[0182] Experimental comparison:

[0183] Method 1 (Traditional Method): Training is performed using the standard cross-entropy loss function.

[0184] Method 2 (the method of this invention): Training is performed using a focus loss function, combined with image interpolation, data augmentation, and spatial alignment.

[0185] Experimental results:

[0186] Method 1 (traditional method) has an accuracy rate of 85% in the diagnosis of acute abdomen.

[0187] Method 2 (the method of this invention) has an accuracy rate of 92% in the diagnosis of acute abdomen.

[0188] The method of this invention improves the recall rate from 80% to 90% in diagnosing appendicitis and increases the accuracy by 7% in diagnosing intestinal obstruction.

[0189] Evaluation indicators:

[0190] Accuracy: Method 1: 85%, Method 2: 92%;

[0191] Recall rates: 80% for Method 1 and 90% for Method 2;

[0192] F1-score: 0.85 for Method 1 and 0.91 for Method 2.

[0193] Based on the above evaluation metrics, it can be seen that the method of the present invention has significant improvements over the traditional method in all metrics, especially in recall and F1-score, indicating that the present invention effectively improves the identification ability of small class samples.

[0194] The specific application process of this invention in the automatic diagnosis of acute abdominal pain medical images includes steps such as data preprocessing, neural network model design, application of the focal loss function, spatial alignment, and data augmentation, all of which effectively improve the accuracy and reliability of diagnosis. Comparison with traditional methods verifies the outstanding advantages of this invention in improving the diagnostic accuracy of small-category samples, further supporting the technical feasibility and innovation of this invention.

[0195] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0196] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assisting in the rapid diagnosis of acute abdomen using medical images combined with deep learning, characterized in that, include: Step 1: Organ-targeted segmentation: Acquire arterial, venous, and delayed phase CT sequences, and perform spatial standardization by dynamically calculating the interpolation interval based on the slice thickness parameters of the scanning equipment; use an acute abdomen-oriented segmentation network to simultaneously generate intestinal, vascular, and peritoneal masks, and introduce multi-scale dilated convolution groups with anatomical size adaptation in the deep layer of the encoder. Step 2: Dynamic Pathological Sign Quantification: Performed based on the mask. Intestinal expansion calculation: Generate cross sections at preset intervals along the centerline of the intestinal mask and measure the diameter; Calculation of free gas volume change rate: Identify low CT value connected regions within the peritoneal mask and calculate the interphase volume ratio; Detection of abnormal enhancement values ​​in organ walls: Measurement of interphase CT value difference in the overlapping area of ​​blood vessel and intestinal tract masks; The above-mentioned intestinal dilatation characteristics, free gas volume change rate characteristics, and organ wall enhancement anomalies are used to construct a dynamic feature matrix in time sequence, and the evolution features are extracted by a pre-trained temporal convolutional layer with adjustable convolution width. Step 3: Clinical-Image Gated Fusion: The patient's clinical parameters are input into a pre-trained weighted network to generate a weight vector, which is then concatenated with the dynamic features of the image and output as a fused feature through a pre-trained gated attention unit. Step 4: Multi-task collaborative diagnosis: Input the fused features into the pre-trained multi-task network, and output the probability of acute abdominal disease etiology classification and the assessment value of surgical urgency in parallel.

2. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning, as described in claim 1, is characterized in that... The implementation of the anatomically adaptive multi-scale dilated convolution group in step 1 includes: For the intestinal segmentation task, the first set of cavity rates is determined based on the range of intestinal anatomical diameters, so that the maximum receptive field covers the maximum physiological expansion diameter of the intestinal segment. The specific process is as follows: the 90th percentile value of the transverse diameter of the intestinal segment in the training dataset is statistically analyzed, and the minimum cavity rate combination required to cover this distance is calculated based on twice this value. For peritoneal segmentation, the second set of void ratios is determined based on the peritoneal curvature distance. The specific process is as follows: calculate the maximum radius of curvature of the peritoneal surface on a three-dimensional mask, and use this radius value as a reference to generate a void ratio sequence covering the curved structure. The void ratio combination is automatically optimized through backpropagation during the model training phase.

3. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning, as described in claim 1, is characterized in that... The calculation of the free gas volume change rate in step 2 includes: Dual-energy CT material decomposition technology was used to separate gas components: 80kVp and 140kVp dual-phase CT data were acquired, and a gas matrix map was generated by the matrix decomposition algorithm to eliminate the interference of CT values ​​from adipose tissue. Identifying gas connected domains within the peritoneal mask: Threshold segmentation is performed on the gas-based material map. The threshold is dynamically set according to the noise level of the scanning equipment. Specifically, the standard deviation of the CT value of the liver parenchyma region is calculated, and three times the standard deviation is taken as the lower limit of the floating threshold. Volume change rate calculation: The number of gas connected domain voxels in the arterial and venous phases are counted separately, and then multiplied by the physical volume of a single pixel to calculate the venous phase / arterial phase volume ratio.

4. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning, as described in claim 1, is characterized in that... The width adjustment rules for the temporal convolutional layer in step 2 include: Read the scan timestamps from the DICOM header file and calculate the time intervals between consecutive periods; When the time interval is less than or equal to the critical response duration of acute abdominal pathological development, a width is set to cover two consecutive phases; the critical response duration is determined according to the time window of irreversible intestinal ischemia injury in clinical guidelines. When the time interval is greater than the critical response duration, the width is set to cover three consecutive phases. When the convolution kernel slides along the time dimension, it adopts a causal convolution mode, allowing only historical phase features to influence the current output.

5. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning according to claim 1, characterized in that, The training method for the weighted network in step 3 includes: The network structure is configured with two fully connected layers. The number of neurons in the first layer is equal to the dimension of the clinical parameter features, and the output dimension of the second layer is the same as the dimension of the clinical parameter features. During training, all trainable parameters of the image feature extraction network are frozen. The weighted network parameters are updated using the gradient backpropagation signal from the acute abdomen etiology classification task. The specific process is as follows: Forward propagation calculates the etiology classification loss value, which is calculated using the cross-entropy loss function. The backpropagation process only unfreezes the parameter gradients of the weighted network and updates the weight matrix and bias vector of the weighted network based on the gradient descent algorithm. The training termination condition is set to the validation set accuracy improving by less than a set threshold for three consecutive iterations.

6. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning, as described in claim 1, is characterized in that... The detection of abnormal organ wall enhancement values ​​in step 2 includes: Generate an intestinal wall ring detection area: Extend a preset distance inward and outward along the normal direction of the boundary of the intestinal tract mask. The distance is set according to the average thickness of the intestinal wall. Specifically, the median value of the intestinal wall thickness in the training set is statistically analyzed, and 1.5 times this value is used to generate a symmetrical ring area. Calculation of abnormal enhancement values ​​of organ walls: The average CT values ​​of the arterial and venous phases are statistically analyzed within the annular region, and the difference between the venous and arterial phases is calculated as the abnormal enhancement value of the organ wall. Standardize abnormal enhancement values ​​of organ walls: Calculate the relative enhancement ratio based on the enhancement value of the aorta at the same level.

7. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning, as described in claim 1, is characterized in that... The operations of the gated attention unit in step 3 include: Gating value generation: The stitched image features and clinical features are input into the fully connected layer, and the output scalar is mapped to the [0,1] interval by the Sigmoid activation function; Compensation value calculation: Perform an arithmetic operation of subtracting the gating value from 1; Feature fusion: The image feature vector is multiplied element-wise with the gate value, and the clinical feature vector is multiplied element-wise with the compensation value. The two result vectors are then added together and output.

8. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning according to claim 1, characterized in that, The construction of the surgical urgency assessment value in step 4 includes: Tag generation: Patients are classified according to the time interval between imaging examination and surgery. The time window is divided based on the following criteria: An interval of ≤2 hours corresponds to label 1.0; A label of 0.7 corresponds to an interval of 2 hours to 6 hours. An interval > 6 hours corresponds to a label of 0.3; The loss function uses Huber loss, and its threshold parameter is dynamically adjusted according to the quantiles of the label distribution.

9. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning, as described in claim 5, is characterized in that... The loss function optimization method for the pre-trained multi-task network in step 4 includes: The focus loss function is used as the basic loss function, and its adjustment factor γ is dynamically set according to the proportion of each class of training data. The setting rule is: the smaller the sample size of the class, the larger the γ value. Establish a dynamic weight adjustment mechanism: After each training cycle, calculate the validation set confusion matrix, count the misdiagnosis rate of each category, calculate the average misdiagnosis rate, and increase the loss weight value of categories whose misdiagnosis rate exceeds the average misdiagnosis rate by a multiple. The weight adjustment process includes mathematical constraints: the sum of the loss weights for all categories remains constant, and weight normalization is achieved through scaling. After the loss function is updated, the model training stability verification condition must be met: the accuracy of the validation set under the new loss function is not lower than the accuracy of the original loss function.

10. The auxiliary interpretation method for rapid diagnosis of acute abdomen using medical images combined with deep learning according to claim 1, characterized in that, The spatial standardization in step 1 includes: Interlayer interpolation spacing calculation: Parse the SliceThickness parameter in the DICOM header file and take the minimum layer thickness among the three phases as the reference layer thickness; Interpolation algorithm execution: For phases with a layer thickness greater than the reference layer thickness, cubic spline interpolation is performed along the Z-axis direction, and the interpolation point spacing is set to the reference layer thickness; Spatial alignment: Using the bifurcation of the portal vein in the liver as an anatomical landmark, three-phase images were registered using affine transformation.

Images