Pelvic fat disease auxiliary diagnosis method fusing domain knowledge

Abstract

The invention belongs to the field of computer-aided medical treatment, and relates to a pelvic fat diagnosis method fusing clinical diagnosis knowledge. Comprising the following steps: firstly, performing multi-organ segmentation on a 3D abdominal cavity computed tomography image; secondly, in combination with a reference standard of clinical diagnosis and treatment of a doctor, extracting and representing the content of fat in the pelvic cavity and morphological change characteristics of peripheral organs on the basis of a segmentation result; then, taking the semantic radiomics characteristics as important prior knowledge in a diagnosis process, and fusing the important prior knowledge with observation evidence obtained by the 3D evidence vision converter network to obtain evidence prior distribution; and finally, training a 3D evidence visual converter network by using the corresponding evidence prior loss function, thereby realizing accurate diagnosis of the pelvic fat disease. The method not only can identify typical image features of pelvic fat, but also can capture fine but key morphological changes, and provides more accurate and reliable results; and the interpretability of the model is obviously enhanced.

Description

Technical Field

[0001] The present application relates to the field of computer-assisted medical treatment, and in particular to a method for assisting the diagnosis of pelvic floor fat disease based on probability distribution of domain knowledge and deep learning of evidence. Background Art

[0002] Pelvic fat disease is a rare benign proliferative disease characterized by abnormal proliferation of adipose tissue in the pelvis. Because this hyperplastic fat lacks specific imaging features and is difficult to distinguish from normal fat by location or grayscale intensity alone, diagnosis requires the extensive experience of urologists and radiologists. In addition, the pathological definition of pelvic fat disease is still unclear, and conventional biochemical indicators (such as blood or hormone levels) are usually normal, which makes it difficult for traditional diagnostic methods to achieve high-precision diagnosis. Given the powerful feature extraction capabilities of deep neural networks, they have been widely used in computer-aided diagnosis (CAD) of disease subtypes. Compared with the inefficiency and poor repeatability of manual measurement of imaging parameters, disease diagnosis technology based on deep learning has significant advantages in automation, efficiency and accuracy.

[0003] However, existing deep learning-based CAD methods usually rely on large-scale datasets for training, while the number of pelvic fat disease cases is very limited. In addition, the imaging differences between normal and abnormal adipose tissue are very weak. It is difficult to obtain ideal diagnostic results by training deep learning models based solely on raw image data. In addition, in current research on the diagnosis of pelvic fat disease, omics features are mostly measured manually by radiologists. This method is not only time-consuming and inefficient, but also easily affected by subjective judgment. During the diagnosis process, doctors usually comprehensively evaluate the factors such as compression, deformation, and dislocation caused by abnormal fat proliferation on adjacent organs, while combining clinical semantic information of inflammatory response. However, existing deep learning models have limited utilization of such high-order semantic priors and are usually only able to integrate single domain knowledge. They are difficult to directly apply to diseases such as pelvic fat disease that require comprehensive judgment of multiple types of domain knowledge with a high degree of semantic abstraction.

[0004] Despite significant progress in medical image analysis using deep learning technology, many challenges remain in the diagnosis of pelvic fat syndrome. First, the number of pelvic fat syndrome cases is small, and existing deep learning methods typically rely on large-scale data for training, resulting in poor diagnostic accuracy when using small sample sizes. Second, pelvic fat syndrome is characterized by minimal differences in imaging, making it difficult for existing methods to effectively capture the subtle but critical morphological changes caused by the compression, deformation, and dislocation of adjacent organs by abnormally proliferating fat. Furthermore, doctors' diagnostic processes often rely on a variety of complex semantic information (such as organ changes and inflammatory responses assessed clinically), and existing deep learning methods still have technical limitations when integrating clinical semantic priors. For example, the tumor detection and classification method proposed by Kibriya et al. relies on a simple concatenation strategy of gray-level co-occurrence matrix texture features and pre-trained VGG-16 network features. While this concatenated feature is discriminated using shallow classifiers such as support vector machines and K-nearest neighbor classifiers, it fails to establish a deep coupling mechanism between medical prior knowledge and data features. This results in ineffective exploration of pathological semantic associations between multi-source features. Furthermore, it lacks the ability to model clinical diagnostic logic in the analysis of complex anatomical structures, making it difficult to directly apply to diseases such as pelvic lipomatosis, which require comprehensive discrimination based on multiple types of domain knowledge with a high degree of semantic abstraction. Finally, current methods are often "black box" models that lack a transparent representation of the physician's diagnostic logic, limiting their interpretability and clinical applicability. Therefore, how to effectively integrate multiple high-level semantic priors to improve diagnostic accuracy while maintaining model interpretability remains a key challenge in the intelligent diagnosis of pelvic lipomatosis. Summary of the Invention

[0005] To address the challenges of existing technologies, this paper proposes an auxiliary diagnosis method for pelvic fat disease that integrates multiple high-level semantic priors. This method reduces reliance on large-scale data while significantly improving the model's performance and clinical applicability, as well as the lack of interpretability in the diagnostic process. Specifically, by converting doctors' clinical diagnostic experience into computable semantic omics features, these semantic features are incorporated into the diagnostic model as prior evidence using an evidence-based deep learning framework, significantly improving the model's diagnostic performance, interpretability, and adaptability to small sample data.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] An auxiliary diagnosis method for pelvic fat disease integrating domain knowledge includes the following steps:

[0008] Step 1: Abdominal data acquisition and preprocessing. Perform multi-organ segmentation on 3D abdominal computed tomography imaging data to obtain accurate segmentation results of major organs in the abdominal cavity, providing a basis for subsequent extraction of advanced semantic omics features.

[0009] Step 2: Extraction and representation of semantic omics features. Based on the multi-organ segmentation results and in conjunction with clinical diagnosis and treatment reference standards, high-level semantic features of pelvic fat content and surrounding organ morphological changes are extracted and represented, including key information such as fat distribution, organ compression deformation, and misalignment.

[0010] Step 3: Fusion of prior and observational evidence.

[0011] The image semantic omics features extracted in step 2 are converted into class probability distributions and used as prior evidence describing dyslipidemia. These are then fused with observational evidence extracted using the 3D Vision Transformer (ViT) network to obtain a prior distribution of evidence reflecting both the clinician's diagnostic information and the imaging features.

[0012] Step 4: Evidence-based deep learning model training and optimization. By designing an evidence prior loss function, the 3D evidence visual transformer network is trained to effectively integrate prior and observational evidence, enabling accurate diagnosis of pelvic floor fat disease and improving the performance of the diagnostic model and its adaptability to complex cases.

[0013] By adopting the above technical solution, leveraging the mathematical properties of evidence-based deep learning and incorporating physician expertise to alter the feature distribution originally learned by the neural network based solely on imaging, changes in prediction errors can be more accurately assessed when calculating the loss function. This fusion strategy enables the model to not only identify typical imaging features of pelvic floor fat disease, but also capture subtle but critical morphological changes. This method reduces the model's dependence on data because clinical knowledge is already internalized as part of the model. Furthermore, it significantly enhances the model's interpretability, allowing physicians to more intuitively understand how the model integrates various clinical information to make diagnostic decisions.

[0014] Furthermore, in step 2, the semantic omics features include: the distance D from the bladder to the rectum, the roundness R of the rectum, the angle θ between the bladder wall and the prostate seminal vesicle, and the relative volume V of the pelvic fat.

[0015] Furthermore, the fusion of prior and observational evidence described in step 3 includes two parts: constructing prior evidence based on semantic omics features and constructing fusion of prior and observational evidence.

[0016] Prior evidence based on semantic omics features extracts and represents doctors' clinical diagnostic experience, introducing it into the model in numerical form, improving the model's ability to capture advanced semantic omics features. The process of fusing prior and observational evidence enables the model to more comprehensively understand global and local patterns in the input data by combining prior information built based on domain knowledge with fine-grained information from actual observational data. Furthermore, the prior evidence introduced during model training not only reduces reliance on large-scale annotated data but also improves the model's interpretability and generalization performance for specific tasks through the constraints of domain priors, thereby providing a more reliable solution for the intelligent analysis of complex medical images.

[0017] Furthermore, in step 4, the evidence prior loss function includes a Bayesian risk-based cross entropy loss function and a regularization loss function.

[0018] Furthermore, the Bayesian risk-based cross-entropy loss function is used to calculate the difference between the predicted value based on prior evidence and the actual value. By introducing the prior distribution of evidence for class probabilities, the Bayesian risk-based cross-entropy loss function is designed from an integral perspective to minimize the prediction error, thereby outputting more accurate prediction results. At the same time, by integrating prior evidence from doctors' clinical knowledge, the training of the 3D evidence visual transformer network is guided by domain prior knowledge, and the model parameters are optimized, reducing the model's dependence on large amounts of labeled data and improving the model's interpretability and classification performance.

[0019] The Bayesian risk-based cross entropy loss function is:

[0020]

[0021] Among them, i represents the i-th sample, and there are N samples in total. and They represent the labels of the i-th sample belonging to pelvic fat syndrome and the i-th sample not belonging to pelvic fat syndrome, and represents the observational evidence learned by the 3D evidence visual transformer network, corresponding to the diseased and control groups, respectively, and represents prior evidence based on multiple semantic omics features, corresponding to the diseased and control groups, respectively. ψ(·) represents the digamma function.

[0022] Furthermore, a regularization loss function is used to minimize the evidence of the wrong class while keeping the evidence of the correct class at a constant level.

[0023] The regularization loss function is:

[0024]

[0025] in, Γ(·) represents the gamma function.

[0026] Regularization loss functions prevent model overfitting and improve generalization to unseen data by penalizing incorrect classes that do not contribute to the data fit. If a sample cannot be correctly classified, a total evidence of zero is preferred, as this indicates that the model is uncertain about the classification rather than making a prediction based on unreliable evidence. In this case, the model's uncertainty about the classification can be viewed as a uniform distribution over all possible classes. Regularization loss functions help the model strike a balance between fitting the training data distribution well and avoiding overconfidence and incorrect predictions on unseen data.

[0027] Furthermore, the evidence prior loss function is in is the annealing coefficient, e represents the current training round, and the loss function is used Trained 3D Evidence Vision Transformer Network model.

[0028] Beneficial effects

[0029] Based on physicians' clinical experience, this paper extracts and represents the morphological features of pelvic organs. These semantically informed radiomics features are then integrated into a deep learning model for pelvic adiposity diagnosis, improving the model's interpretability and diagnostic accuracy. This paper proposes a method for extracting and fusing multiple semantic priors for pelvic adiposity. First, four semantic prior extraction methods are constructed based on clinicians' diagnostic experience. Second, a method is proposed to effectively fuse these four semantic priors with a deep neural network based on evidence. This innovative approach not only reduces the reliance on large datasets, as physicians' expertise is internalized as part of the model, but also significantly enhances the model's interpretability. Physicians can intuitively understand how the model integrates various clinical information to make diagnostic decisions. Through this approach, the proposed model not only identifies typical imaging features of pelvic adiposity but also captures subtle but critical morphological changes. This in-depth analytical capability enables the model to provide more accurate and reliable results for pelvic adiposity diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present invention will be further described below with reference to the accompanying drawings and examples.

[0031] Figure 1 is a process flow chart of the method of the present invention;

[0032] Figure 2 It is the module division and data flow diagram of the method of the present invention;

[0033] Figure 3 is an example diagram showing the distance from the bladder to the rectum according to an embodiment of the present invention;

[0034] Figure 4 This is a diagram showing a failure example of applying the bladder-to-rectum distance to an embodiment of the present invention;

[0035] Figure 5 is an example diagram showing the distance from the bladder to the rectum based on fat according to an embodiment of the present invention;

[0036] Figure 6 is an example diagram showing the roundness of the rectum according to an embodiment of the present invention;

[0037] Figure 7 This is an example diagram showing the angle between the bladder wall and the prostate seminal vesicle according to an embodiment of the present invention;

[0038] Figure 8 is a schematic diagram showing the angle between the bladder wall and the prostate seminal vesicle according to an embodiment of the present invention;

[0039] Figure 9 This is an example diagram showing the relative volume of fat in the pelvic cavity according to an embodiment of the present invention. DETAILED DESCRIPTION

[0040] The technical solution provided by this application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of this application will become more apparent with reference to the following description.

[0041] This proposed method leverages computer technology and deep learning to perform intelligent, assisted diagnosis of pelvic fat syndrome. This method integrates medical image processing, semantic omics feature extraction, neural network model application, and prior evidence fusion, potentially improving the accuracy and reliability of pelvic fat syndrome diagnosis in complex clinical scenarios.

[0042] like Figure 1 is a process flow chart of the method of the present invention;

[0043] like Figure 2 It is the module division and data flow diagram of the method of the present invention.

[0044] A method for assisted diagnosis of pelvic fat disease based on probability distribution of domain knowledge and deep learning of evidence, comprising the following steps: (e.g. Figure 1 、 Figure 2 )

[0045] Step 1: Abdominal cavity data acquisition and preprocessing.

[0046] First, 3D abdominal computed tomography image data is acquired and cropped to retain the pelvic region image. Specifically, the upper boundary of the pelvic region is the location of the iliac crest on the skeleton, and the lower boundary of the pelvic region is the lower edge of the pubic symphysis.

[0047] Subsequently, 30 samples with masks of fat, bladder, prostate, seminal vesicles, and rectum manually labeled by doctors were used as training data for the segmentation network (15 patients with pelvic fat disease and 15 controls). The segmentation network used the Unet automatic segmentation model based on deep neural networks.

[0048] Finally, the trained segmentation model is used to perform multi-organ segmentation on the cropped 3D abdominal computed tomography images.

[0049] Step 2: Extraction and representation of semantic omics features.

[0050] Based on the multi-organ segmentation results obtained in step 1, four semantic radiomics features are further extracted and calculated, including: the distance from the bladder to the rectum D, the circularity of the rectum R, the angle θ between the bladder wall and the prostate seminal vesicle, and the relative volume of pelvic fat V.

[0051] The specific instructions are as follows:

[0052] (1) Distance from bladder to rectum

[0053] Reference Figure 3 In this 3D abdominal computed tomography image slice, the left image shows the distance (red line) from the bladder (yellow area) to the rectum (blue area) in patients with pelvic fat syndrome, while the right image shows the distance from the bladder to the rectum in a control group. Because patients with pelvic fat syndrome have abnormal fat growth in the pelvis, the space between the bladder and rectum is filled with fat, which pushes them apart and increases the distance between them.

[0054] Define the distance d from the bladder to the rectum pixel is the shortest distance from any point a on the bladder to any point b on the rectum, using the following formula:

[0055]

[0056] The lowest point of the bladder and the uppermost point of the rectum satisfy:

[0057]

[0058] Where A represents the set of bladder pixels, B represents the set of rectal pixels, a represents the bladder pixel, b represents the rectal pixel, and d represents the pixel spacing.

[0059] However, statistical analysis revealed that the difference between patients with pelvic fat and controls in this feature was not significant. This is because in some samples, the increased distance between the bladder and rectum was mainly due to the presence of other organs in between, rather than due to abnormal fat proliferation. Figure 4 , the slice of the 3D abdominal computed tomography image shows that the absolute distance from the bladder to the rectum is larger in the control group. Therefore, this application no longer measures the absolute distance, but focuses on measuring the distance increase caused by abnormal fat proliferation. Specifically, refer to Figure 5 In this slice of a 3D abdominal computed tomography image, the left image shows the fat-based (red) distance from the bladder (yellow area) to the rectum (blue area) in patients with pelvic adiposity (green line), and the right image shows the fat-based distance from the bladder to the rectum in the control group.

[0060] The present invention multiplies the number N of fat pixels in the absolute distance from the bladder to the rectum by the pixel spacing d to obtain the fat-based distance D from the bladder to the rectum, which is in the following form:

[0061] D=N×d,

[0062] Where N = ∑ t∈l(x) I(t), t represents the pixel point on the line l(x), l(x) represents the point passing the lowest edge of the bladder and the uppermost point of the rectum The straight line, F represents the set of fat pixels.

[0063] (2) The roundness of the rectum

[0064] Reference Figure 6 In this 3D abdominal computed tomography image slice, the upper image shows the rectal segmentation results (blue area) of a patient with pelvic adiposity, while the lower image shows the rectal segmentation results of a control group. In patients with pelvic adiposity, the rectum is deformed by the compression of abnormal fat growth in the pelvis.

[0065] Therefore, the circularity R of the rectum is defined as the ratio of the area S enclosed by the rectal contour to the square of the contour perimeter C, in the following form:

[0066]

[0067] (3) Angle between the bladder wall and the prostate seminal vesicle

[0068] Reference Figure 7In this 3D abdominal computed tomography image slice, the left image shows the angle between the bladder wall (yellow area) and the prostate seminal vesicle (green area) in a patient with pelvic fat, while the right image shows the angle between the bladder wall and the prostate seminal vesicle in a control group. Abnormal fat proliferation compresses the bladder and prostate seminal vesicle, causing the angle between them to increase.

[0069] Therefore, refer to Figure 8 In a two-dimensional computed tomography image, this application defines the right angle θ1 between the bladder wall and the seminal vesicle as the angle between the line between the rightmost edge point B1 and the rightmost lower edge point B2 of the bladder and the line between the rightmost edge point P1 and the rightmost upper edge point P2 of the seminal vesicle. The left angle θ2 between the bladder wall and the seminal vesicle is defined as the angle between the line between the leftmost edge point B3 and the leftmost lower edge point B4 of the bladder and the line between the leftmost edge point P3 and the leftmost upper edge point P4 of the seminal vesicle. The right angle is expressed as follows:

[0070]

[0071] The left angle is as follows:

[0072]

[0073] Calculate the average of the left and right angles to get the angle θ, which is as follows:

[0074]

[0075] (4) Relative volume of pelvic fat

[0076] Reference Figure 9 In the slice of the 3D abdominal computed tomography image, the upper figure shows the segmentation result of the sample pelvic fat (red area), and the lower figure shows the segmentation result of the sample pelvic cavity (purple area).

[0077] Define the relative volume of fat V as fat volume V fat and pelvic volume V pelvic The ratio is as follows:

[0078]

[0079] Table 1 summarizes the above four semantic omics features of pelvic fat syndrome and the control group, and it can be seen that there are significant differences in the features of the two groups.

[0080] Table 1 Semantic omics features between pelvic fat syndrome and control group

[0081]

[0082] Step 3: Fusion of prior and observational evidence.

[0083] Prior evidence:

[0084] The image semantic omics features extracted in step 2 are converted into class probability distributions and used as prior evidence to describe abnormal lipid accumulation. The details are as follows:

[0085] For the feature of the distance from bladder to rectum, the mean values ​​of the samples belonging to pelvic fat syndrome and the control group were calculated, which are and The standard deviations for pelvic fat syndrome and the control group are and

[0086] Assume that the distance from the bladder to the rectum follows a Gaussian distribution, and the probability density function of the Gaussian distribution is And the cumulative distribution function is

[0087] For each sample x i , then the probability of it belonging to pelvic fat syndrome under the characteristics of the distance from the bladder to the rectum is Confidence level of pelvic fat syndrome Among them, D i represents the fat-based distance from the bladder to the rectum of the i-th sample.

[0088] Using the same method, we calculate the features of the rectal roundness, the angle between the bladder wall and the prostate seminal vesicle, and the relative volume of pelvic fat. The i-th sample x i The probability of pelvic fat disease is and The confidence levels of pelvic fat syndrome are and

[0089] Based on the above prior evidence, sample x i The definition of prior evidence for pelvic fat syndrome is as follows:

[0090]

[0091] Similarly, sample x i The definition of prior evidence for belonging to the control group is as follows:

[0092]

[0093] in, and The probability of belonging to the control group is the distance from the bladder to the rectum based on fat, the roundness of the rectum, the angle between the bladder wall and the prostate seminal vesicle, and the relative volume of pelvic fat; and They are the distance from the bladder to the rectum based on fat, the roundness of the rectum, the angle between the bladder wall and the prostate seminal vesicle, and the confidence that the relative volume of pelvic fat belongs to the control group.

[0094] Through normalization operation, we can get

[0095] Observational evidence:

[0096] Use the evidence deep neural network to process 3D abdominal computed tomography imaging data to obtain observational evidence e + and e - , representing the observational evidence belonging to pelvic fat disease and the control group, respectively.

[0097] The Evidence Deep Neural Network (EDN) is implemented using a 3D Evidence Visual Transformer Network. Traditional deep neural networks convert continuous activations in the output layer into class probabilities. However, the softmax operator exponentially amplifies small differences between logarithms, leading to over-inflated probabilities for certain classes despite their low activation values. This type of point estimate can lead to overconfidence. In contrast, the Evidence Deep Neural Network considers the class probabilities output by the neural network to be multivariate random variables with a certain distribution, rather than fixed but unknown values. By calculating the expected value of the distribution associated with the class probabilities, more accurate and reliable classification results can be obtained while reducing the model's dependence on labeled data.

[0098] Fusion of prior evidence and observational evidence:

[0099] According to Bayes' theorem, the posterior distribution of a multivariate random variable can be expressed as:

[0100]

[0101] Where p represents the class probability output by the 3D Evidence Visual Transformer network, and f like (x|p) represents the likelihood function, f prior (p) represents the prior distribution.

[0102] Given the properties of the conjugate distribution, this method assumes that the likelihood function of the input data follows a binomial distribution. Where p1 represents the probability that the sample output by the 3D evidence visual transformer network belongs to the class of pelvic fat disease, p0 represents the probability that the sample belongs to the control group, and e + and e - They represent the observation evidence obtained by the 3D evidence visual transformer network belonging to pelvic fat disease and the control group respectively.

[0103] The prior distribution based on domain knowledge follows the Beta distribution:

[0104]

[0105] where a + and a - It is based on prior evidence of semantic omics features. is the Beta function.

[0106] According to Bayes' theorem, the posterior probability distribution of class probabilities p1 and p0 can be calculated under the above assumptions in the following form:

[0107]

[0108] The observation evidence parameter e output by the 3D evidence visual transformer network is converted into + and e - With prior evidence + and a - Combine the calculations to get the posterior probability distribution, and then find the expectation of the distribution to get the final posterior expectation:

[0109]

[0110] Where S = e + +e - +2.

[0111] This method converts semantic omics features into prior evidence, and this additional information can be used to guide the model during training, thereby alleviating the problem of poor model prediction results due to insufficient data.

[0112] Step 4: Evidence deep learning model training and optimization.

[0113] The four semantic omics features extracted from the doctors’ clinical experience in step 3 are combined with the observational evidence obtained by the 3D evidence visual transformer network to obtain the posterior distribution f post (p1,p0;e + ,e - ,a + ,a - )=Beta(e + +a + ,e - +a - ) to describe the probability of pelvic fat classification.

[0114] An evidence prior loss function is designed, and the cross entropy loss function based on Bayesian risk and the regularization loss function are comprehensively considered to optimize the network parameters of the 3D evidence visual transformer.

[0115] For the given abdominal cavity dataset It includes N labeled 3D abdominal computed tomography images. Each image consists of a pair of x iand y i Indicates that x i Represents image, y i Represents the true category. i =(y i1 ,y i0 ) is the sample x i The unique heat encoding vector of the label is (1,0) or (0,1). The former represents the sample x i Belong to pelvic fat syndrome, the latter means that the sample x i Belong to the control group.

[0116] The evidence prior loss function for diagnosing pelvic fat is defined as the cross entropy loss function based on Bayesian risk and regularized loss function The form is as follows:

[0117]

[0118] in is the annealing coefficient, and e represents the current training round.

[0119] Bayesian risk-based cross-entropy loss function: The cross-entropy loss function, which incorporates domain priors, is recalculated from an integral perspective to reduce prediction error and thus improve the accuracy of prediction results. Furthermore, by incorporating clinician expertise as prior information, the 3D Evidence Visual Transformer Network is guided by domain expert knowledge during training, further optimizing model parameters. This approach not only reduces the model's need for large amounts of annotated data but also enhances the model's interpretability and classification efficiency. The Bayesian risk-based cross-entropy loss function takes the following form:

[0120]

[0121] Where i represents the i-th sample, N is the number of samples, and They represent the labels of the i-th sample belonging to pelvic fat disease and the i-th sample not belonging to pelvic fat disease, p i1 and p i0 denote the class probabilities of the i-th sample output by the 3D evidence visual transformer network belonging to pelvic fat disease and the control group, respectively. and represents the observational evidence learned by the 3D evidence visual transformer network, corresponding to the diseased and control groups, respectively, and represents prior evidence based on multiple semantic omics features, corresponding to the diseased and control groups, respectively. ψ(·) represents the digamma function.

[0122] Regularized loss function: To prevent model overfitting, limit model complexity, and improve generalization ability, additional constraints (regularization terms) are added to the loss function. The regularized loss function is as follows:

[0123]

[0124] in, Γ(·) represents the gamma function.

[0125] To verify the effectiveness of the proposed method, a pelvic fat disease test was conducted on a 3D abdominal cavity dataset. The dataset was collected from patients who underwent CT urography (CTU) in a hospital over an 8-year period and included 126 3D CT images. A five-fold cross-validation method was used, and a 3D visual transformer network was used as the backbone network. A ReLU activation layer was introduced in the final fully connected layer as an evidence extractor to extract non-negative observational evidence. Finally, a loss function based on the fusion domain evidence prior was used. Train the 3D Evidence Visual Transformer model to obtain the final prediction results.

[0126] The proposed method is compared with the existing 3D residual network (ResNet), 3D densely connected network (DenseNet), 3DViT and 3D evidential visual transformer model based on uniform prior assumption (Evi-ViT). The prediction results are shown in Table 2. The accuracy of the prediction results is quantitatively evaluated in the form of mean ± standard deviation.

[0127] Table 2 Comparison of the accuracy of the pelvic fat disease auxiliary diagnosis method and the classification model based on the integration of domain knowledge

[0128]

[0129] From the experimental results in Table 2, it can be seen that the pelvic fat disease diagnosis method proposed by this method that integrates domain knowledge achieves the best performance.

[0130] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical spirit of this invention. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A pelvic fat disease auxiliary diagnosis method integrating domain knowledge, characterized in that: The following steps are involved: Step 1, abdominal cavity data acquisition and preprocessing; Perform multi-organ segmentation on 3D abdominal computed tomography imaging data to obtain accurate segmentation results of major organs in the abdominal cavity, providing a basis for subsequent extraction of advanced semantic omics features; Step 2: Extraction and representation of semantic omics features; Integrating with doctors' clinical diagnosis and treatment reference standards, based on the multi-organ segmentation results, we extract and represent high-level semantic features of pelvic fat content and morphological changes of surrounding organs, including key information such as fat distribution characteristics, organ compression deformation, and misalignment. Step 3: Fusion of prior and observational evidence; The image semantic omics features extracted in step 2 are converted into class probability distributions and used as prior evidence describing dyslipidemia. These features are then fused with the observational evidence extracted by the 3D Evidence Visual Transformer Network to obtain a prior distribution of evidence reflecting the clinician's diagnostic information and imaging features. Step 4: Evidence deep learning model training and optimization; By designing an evidence prior loss function and training a 3D evidence visual transformer network, it is possible to effectively integrate prior and observational evidence, thereby achieving accurate diagnosis of pelvic floor fat disease and improving the performance of the diagnostic model and its adaptability to complex cases.

2. The pelvic fat disease auxiliary diagnosis method integrating domain knowledge according to claim 1 is characterized in that: Step 1 is as follows: First, 3D abdominal computed tomography image data is acquired, and the original image data is cropped to retain the image of the pelvic region; specifically, the upper boundary of the pelvic region is the location of the iliac crest on the skeleton, and the lower boundary of the pelvic region is the lower edge of the pubic symphysis; Subsequently, the samples with masks of fat, bladder, prostate, seminal vesicles, and rectum manually labeled by doctors were used as training data for the segmentation network; Finally, the trained segmentation model is used to perform multi-organ segmentation on the cropped 3D abdominal computed tomography images.

3. The pelvic fat disease auxiliary diagnosis method integrating domain knowledge according to claim 1 is characterized in that: In step 2, the semantic omics features include: the distance D from the bladder to the rectum, the roundness R of the rectum, the angle θ between the bladder wall and the prostate seminal vesicle, and the relative volume V of the pelvic fat.

4. The pelvic fat disease auxiliary diagnosis method integrating domain knowledge according to claim 3 is characterized in that: The semantic omics features are defined as follows: (1) Distance from bladder to rectum Define the distance d from the bladder to the rectum pixel is the shortest distance from any point a on the bladder to any point b on the rectum, using the following formula: The lowest point of the bladder and the uppermost point of the rectum satisfy: Where A represents the set of bladder pixels, B represents the set of rectal pixels, a represents bladder pixels, b represents rectal pixels, and d represents the pixel spacing; The absolute distance from the bladder to the rectum that passes through fat pixels N is multiplied by the pixel spacing d to obtain the fat-based distance D from the bladder to the rectum, which is in the following form: D=N×d, Where N = ∑ t∈l(x) I(t), t represents the pixel point on the line l(x), l(x) represents the point passing the lowest edge of the bladder and the uppermost point of the rectum The straight line, F represents the set of fat pixels; (2) The roundness of the rectum The circularity R of the rectum is defined as the ratio of the area S enclosed by the rectal contour to the square of the contour perimeter C, as follows: (3) Angle between the bladder wall and the prostate seminal vesicle The right angle θ1 between the bladder wall and the seminal vesicle is defined as the angle between the line between the rightmost edge point B1 and the rightmost lower edge point B2 of the bladder and the line between the rightmost edge point P1 and the rightmost upper edge point P2 of the seminal vesicle. The left angle θ2 between the bladder wall and the seminal vesicle is defined as the angle between the line between the leftmost edge point B3 and the leftmost lower edge point B4 of the bladder and the line between the leftmost edge point P3 and the leftmost upper edge point P4 of the seminal vesicle. The right angle is as follows: The left angle is as follows: Calculate the average of the left and right angles to get the angle θ, which is as follows: (4) Relative volume of pelvic fat Define the relative volume of fat V as fat volume V fat and pelvic volume V pelvic The ratio is as follows:

5. The pelvic fat disease auxiliary diagnosis method integrating domain knowledge according to claim 1 is characterized in that: In step 3, The method for calculating the prior evidence is as follows: For the feature of the distance from bladder to rectum, the mean values ​​of the samples belonging to pelvic fat syndrome and the control group were calculated, which are and The standard deviations for pelvic fat syndrome and the control group are and Assume that the distance from the bladder to the rectum follows a Gaussian distribution, and the probability density function of the Gaussian distribution is And the cumulative distribution function is For each sample x i , then the probability of it belonging to pelvic fat syndrome under the characteristics of the distance from the bladder to the rectum is Confidence level of pelvic fat syndrome Among them, D i represents the fat-based distance from the bladder to the rectum of the i-th sample; Using the same method, we calculate the features of the rectal roundness, the angle between the bladder wall and the prostate seminal vesicle, and the relative volume of pelvic fat. The i-th sample x i The probability of pelvic fat disease is and The confidence levels of pelvic fat syndrome are and Sample x i The definition of prior evidence for pelvic fat syndrome is as follows: Sample x i The definition of prior evidence for belonging to the control group is as follows: in, and The probability of belonging to the control group is the distance from the bladder to the rectum based on fat, the roundness of the rectum, the angle between the bladder wall and the prostate seminal vesicle, and the relative volume of pelvic fat; and The confidence level that the patient belongs to the control group is determined by the distance from the bladder to the rectum based on fat, the roundness of the rectum, the angle between the bladder wall and the prostate seminal vesicle, and the relative volume of pelvic fat. Use the evidence deep neural network to process 3D abdominal computed tomography imaging data to obtain observational evidence e + and e - , representing the observational evidence belonging to pelvic fat disease and the control group, respectively.

6. The pelvic fat disease auxiliary diagnosis method integrating domain knowledge according to claim 1 is characterized in that: In step 3, the prior evidence is fused with the observational evidence as follows: Assume that the likelihood function of the input data follows a binomial distribution Where p1 represents the probability that the sample output by the 3D evidence visual transformer network belongs to the class of pelvic fat disease, p0 represents the probability that the sample belongs to the control group, and e + and e - They represent the observation evidence obtained by the 3D evidence visual transformer network belonging to pelvic fat disease and the control group respectively; The prior distribution follows the Beta distribution: where a + and a - It is based on prior evidence of semantic omics features. is the Beta function; According to Bayes' theorem, the posterior probability distribution of class probabilities p1 and p0 is as follows: The observation evidence parameter e output by the 3D evidence visual transformer network is converted into + and e - With prior evidence + and a - Combine the calculations to get the posterior probability distribution, and then find the expectation of the distribution to get the final posterior expectation: and Where S = e + +e - +2.

7. The pelvic fat disease auxiliary diagnosis method integrating domain knowledge according to claim 1 is characterized in that: In step 4, the evidence prior loss function is designed, and the cross entropy loss function based on Bayesian risk and the regularization loss function are comprehensively considered to optimize the network parameters of the 3D evidence visual transformer. For a given abdominal cavity dataset It includes N labeled 3D abdominal computed tomography images; each image consists of a pair of x i and y i Indicates that x i Represents image, y i Represents the true category. i =(y i1 ,y i0 ) is the sample x i The unique heat encoding vector of the label is (1,0) or (0,1). The former represents the sample x i Belong to pelvic fat syndrome, the latter means that the sample x i belonged to the control group; The evidence prior loss function for diagnosing pelvic fat is defined as the cross entropy loss function based on Bayesian risk and regularized loss function as follows: in is the annealing coefficient, and e represents the current training round.

8. The method for assisted diagnosis of pelvic fat disease integrating domain knowledge according to claim 7, characterized in that: The Bayesian risk-based cross entropy loss function is as follows: Where i represents the i-th sample, N is the number of samples, and They represent the labels of the i-th sample belonging to pelvic fat disease and the i-th sample not belonging to pelvic fat disease, p i1 and p i0 denote the class probabilities of the i-th sample output by the 3D evidence visual transformer network belonging to pelvic fat disease and the control group, respectively. and represents the observational evidence learned by the 3D evidence visual transformer network, corresponding to the diseased and control groups, respectively, and represents prior evidence based on multiple semantic omics features, corresponding to the diseased and control groups, respectively. ψ(·) represents the digamma function.

9. The method for assisted diagnosis of pelvic fat disease integrating domain knowledge according to claim 7, characterized in that: The regularization loss function is as follows: in, Γ(·) represents the gamma function.

Images