Data processing system for auxiliary diagnosis of biliary tract diseases
By integrating biochemical indicators and ultrasound imaging data, and utilizing LASSO feature screening and multi-layer perceptron neural networks, the difficult problem of differential diagnosis between CC and CBA in early infancy was solved, high-precision auxiliary diagnosis was achieved, and timely treatment was ensured for CBA patients.
Patent Information
- Application Number
- CN202510841600.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-23
- Publication Date
- 2025-10-28
AI Technical Summary
Existing technologies make it difficult to accurately distinguish between common bile duct cysts (CC) and cystic biliary atresia (CBA) in early infancy, resulting in a high misdiagnosis rate and missed optimal treatment opportunities. Existing diagnostic methods lack sensitivity and specificity, and there is a lack of comprehensive diagnostic tools.
By integrating biochemical indicators and ultrasound imaging data, employing LASSO feature screening and multilayer perceptron neural networks, and combining synthetic minority class oversampling technology, a high-precision data processing system is constructed to provide auxiliary diagnostic support.
It achieved an accuracy rate of 95.8%, a sensitivity of 92.9% and a specificity of 96.3%, significantly improving the accuracy of early differential diagnosis, reducing misdiagnosis and ensuring that CBA patients receive timely treatment.
Smart Images
Figure CN120853873A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of disease auxiliary diagnosis technology, specifically relating to a data processing system for auxiliary diagnosis of biliary tract diseases. Background Technology
[0002] Choledochal cysts (CC) and cystic biliary atresia (CBA) are common causes of obstructive jaundice in children. CC is characterized by varying degrees of dilation of the common bile duct, which may be accompanied by dilation of intrahepatic bile ducts. According to the Japanese Clinical Practice Guidelines for Abnormal Pancreaticobiliary Connections, cyst excision and hepatic-enteric anastomosis are recommended as soon as possible for children with symptomatic CC, while asymptomatic children should be closely monitored for liver function and symptoms, and surgery should be performed electively between 3 and 6 months of age.
[0003] Biliary atresia (BA) is characterized by inflammation, progressive liver fibrosis, and extrahepatic bile duct obstruction. The prognosis of this disease is closely related to the timing of surgery; Kasai hepatic-enteric anastomosis performed within 60 days of birth has a better prognosis, while delayed surgery may lead to severe liver fibrosis due to prolonged cholestasis, affecting the surgical outcome. Cystic biliary atresia (CBA) is a subtype of BA; although the prognosis is good, early surgery is still necessary to avoid excessive liver fibrosis leading to a poor prognosis.
[0004] In early infancy, CBA and CC can be difficult to distinguish due to their similar symptoms and imaging findings. Currently, the gold standard for diagnosis is intraoperative cholangiography, which is challenging to perform preoperatively. If CBA is misdiagnosed as CC, the optimal surgical window may be missed, complicating treatment decisions.
[0005] Tang et al. investigated the role of gamma-glutamyl transferase (GGT) in the diagnosis of biliary atresia; Tanaka et al. found that patients with elevated total bile acid (TBA) levels should be suspected of having biliary atresia. When TBA > 111 μmol / L in patients with hilar cysts, congenital biliary atresia (CBA) should be considered instead of common bile duct (CC). However, the specificity and sensitivity of a single indicator are limited. Kim et al. and Lal et al. studied the ultrasound manifestations of CC with or without biliary atresia in neonates and infants; Shin et al. suggested that postpartum cysts ≤ 2.2 cm in diameter and without distal common bile duct on MRI scans can differentiate between CC and CBA. However, these methods have limited accuracy in early infancy. Yu et al. and Chen et al. conducted clinical controlled studies of CC and CBA, comparing the clinical characteristics of the two groups of patients, including biochemical markers and ultrasound measurements, but no unified and operational diagnostic criteria were established.
[0006] These existing technical solutions have the following drawbacks: 1) The diagnostic value of a single indicator is limited, leading to a high misdiagnosis rate. This is because both CC and CBA may present as cholangiocysts in the early stages, and a single parameter is insufficient to capture the complex pathophysiological characteristics of the disease; 2) Existing diagnostic methods lack sensitivity and specificity, especially in early infancy, which is the critical period for CBA treatment. Missing this window will significantly reduce the prognosis of CBA patients; 3) There is a lack of comprehensive diagnostic models that integrate multiple clinical parameters; 4) There are no practical diagnostic tools that are convenient for doctors to use.
[0007] Developing a high-precision data processing system for the auxiliary diagnosis of biliary tract diseases is of great significance for improving the accuracy of early differential diagnosis of biliary tract diseases (CC) and biliary tract atrophy (CBA). Summary of the Invention
[0008] The present invention aims to at least partially solve one of the technical problems existing in the prior art. To this end, the present invention provides a data processing system for auxiliary diagnosis of biliary tract diseases.
[0009] According to one aspect of the present invention, a data processing system for auxiliary diagnosis of biliary tract diseases is provided, comprising: (a) a data acquisition module configured to acquire biochemical index detection data and ultrasound image data of a patient; (b) a data preprocessing module for standardizing the detection data; (c) a feature screening module for extracting key biometric parameters; (d) a classification decision module including a multilayer perceptron neural network structure for generating biliary tract abnormality classification indicators; and (e) a result output module displaying auxiliary diagnostic reference information including classification confidence levels.
[0010] Preferably, the biochemical index detection data includes at least five of the following: total bile acid TBA, gamma-glutamyl transferase GGT, total bilirubin TBIL, direct bilirubin DBIL, aspartate aminotransferase AST, and alanine aminotransferase ALT; and the ultrasound imaging data includes the maximum longitudinal diameter LDC and the maximum transverse diameter TDC of the cyst extracted by an image segmentation algorithm.
[0011] Preferably, the feature screening module is a LASSO feature screening module, configured to perform regression analysis on standardized biochemical index data and imaging measurement data to screen out a key feature set including γ-glutamyl transferase GGT, total bile acid TBA, direct bilirubin DBIL, total bilirubin TBIL, alanine aminotransferase ALT, maximum longitudinal diameter LDC of cysts, and maximum transverse diameter TDC.
[0012] Preferably, the multilayer perceptron neural network structure has a neural network architecture with an input layer, at least three hidden layers, and an output layer. The number of nodes in the input layer corresponds to the seven parameters of the key feature set, and the output layer consists of binary classification neurons.
[0013] Preferably, the classification decision module uses the Synthetic Minority Oversampling Technique (SMOTE) to process the data.
[0014] Preferably, the data preprocessing module uses the Z-score method for standardization, with the formula: z = (x - μ) / σ, where x is the original value, μ is the mean of the training set, and σ is the standard deviation of the training set.
[0015] Preferably, the confidence level grading logic of the result output module includes: when the confidence level is <70%, activating the first prompt signal and calling the image review submodule; when the confidence level is 70% ≤ confidence level <90%, activating the second prompt signal and associating with the electronic medical record database; when the confidence level is ≥90%, activating the third prompt signal and generating a surgical plan index link.
[0016] Preferably, it also includes a model update module, configured to: receive newly labeled clinical data through the PCIe interface; execute an incremental learning algorithm to update the weights of the multilayer perceptron classifier; and trigger a full training mode when the model accuracy drops by more than 5%.
[0017] This invention aims to address the challenge of differential diagnosis between choledochal cysts (CC) and cystic biliary atresia (CBA) in early infancy. These two diseases are extremely similar in clinical presentation and imaging, but their treatment plans and prognoses are drastically different. CBA requires Kasai surgery within 60 days of birth to prevent cirrhosis, while surgery for CC can be postponed until 3-6 months. In existing technologies, relying solely on traditional biochemical indicators or ultrasound examinations is insufficient to accurately distinguish between these two diseases. Intraoperative cholangiography is the gold standard but not suitable for preoperative diagnosis, often leading to misdiagnosis of CBA as CC and missing the optimal surgical window. This invention integrates multiple clinical parameters and utilizes machine learning algorithms to establish a highly accurate differential diagnostic model, providing clinicians with preoperative diagnostic support tools to avoid misdiagnosis and delayed treatment of CBA. Attached Figure Description
[0018] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:
[0019] Figure 1 This is a bar chart with seven non-zero coefficients: GGT, TBA, DBIL, TBIL, ALT, LDC, and TDC, provided according to an embodiment of the present invention.
[0020] Figure 2 This is a user interface diagram of diagnostic software provided according to an embodiment of the present invention. Detailed Implementation
[0021] The following examples are provided to help those skilled in the art better understand the present invention. It should be noted that the following examples are not intended to limit the scope of protection claimed by the present invention, but are merely illustrative. Unless otherwise specified, the raw materials, reagents, or devices mentioned in the following examples are commercially available or obtained through known existing methods.
[0022] 1. Data preparation and preprocessing
[0023] The study included 319 diagnosed children (274 with CC and 45 with CBA) aged ≤60 days and with total bilirubin >34.2 μmol / L. Biochemical markers (BM) were collected one day before surgery: total bile acid (TBA), gamma-glutamyl transpeptidase (GGT), total bilirubin (TBIL), direct bilirubin (DBIL), aspartate aminotransferase (AST), and alanine aminotransferase (ALT); and ultrasound measurements (UM) were taken: the maximum longitudinal diameter (LDC, Length of Cyst from superior to inferior point) and the maximum transverse diameter (TDC, Transverse Diameter of Cyst from left to right point) of the cyst were measured by fasting ultrasound before surgery.
[0024] To address the characteristics of the data distribution, missing values are first imputed using the number of digits in the training set (e.g., ALT is imputed with 35 U / L). Extreme outliers exceeding the mean ± 3 times the standard deviation (e.g., samples with GGT > 800 U / L) are removed. Subsequently, numerical features are standardized using Z-score, calculated as z = (x - μ) / σ, where μ and σ are taken from the mean (e.g., μ = 285.3 U / L for GGT) and standard deviation (σ = 120.5) of each feature in the training set, respectively, to eliminate the influence of dimensional differences on the model.
[0025] Based on the standardized data, a LASSO regression model is constructed, whose objective function incorporates L1 regularization constraints:
[0026]
[0027] Where yi is the class label (CC=0, CBA=1), βj is the feature coefficient, and λ is the regularization strength parameter. The coordinate descent algorithm is used for iterative solution, with the search range of λ set to 10. -5 Up to 10 1 In logarithmic space, a 5-fold stratified cross-validation method was used to evaluate the model performance. During the cross-validation process, the original ratio of CC to CBA was maintained in each fold (close to 6:1). The average classification error corresponding to each λ value was calculated, and the optimal λ value with the smallest cross-validation error (λ = 0.0010) was finally selected. At the same time, the "one standard deviation rule" was referred to ensure the generalization ability of the model.
[0028] When λ = 0.0010, LASSO regression compresses the coefficients of five non-critical features, including AST, to zero, retaining seven non-zero coefficient features: GGT, TBA, DBIL, TBIL, ALT, LDC, and TDC, forming the optimal feature subset. Among them, TBA (coefficient +0.156) and LDC (coefficient -0.212) contribute the most and are highly consistent with the pathological mechanism: elevated TBA reflects the interruption of enterohepatic circulation due to cholestasis in CBA patients, while a smaller LDC is consistent with the imaging characteristics of CBA cysts being mostly confined to the porta hepatis. To verify the necessity of the features, ablation experiments were conducted. After removing TBA or LDC, the accuracy of the model test set decreased by 13.5% and 16.7%, respectively, demonstrating their irreplaceability. Figure 1 ).
[0029] The LASSO regression equation is: label = 0.4999999999999994 + 0.067599·GGT + 0.155626·TBA + 0.134957·DBIL + 0.018696·TBIL - 0.044751·ALT - 0.212238·LDC + 0.021129·TDC.
[0030] 2. Optimize data distribution
[0031] This invention addresses the class imbalance problem in the differential diagnosis of ductal cysts (CC) and cystic biliary atresia (CBA) (274 cases of CC vs. 45 cases of CBA). It employs Synthetic Minority Oversampling Technique (SMOTE) to optimize data distribution. The specific implementation process is as follows:
[0032] Data distribution analysis and segmentation strategies
[0033] The original dataset had a CC to CBA ratio of 6.1:1. Direct training could easily lead to a bias towards the majority class (CC), significantly increasing the risk of missing CBA diagnoses. To address this issue, the 319 cases were first stratified in a 7:3 ratio into a training set (223 cases, 192 CC, 31 CBA) and a test set (96 cases, 82 CC, 14 CBA), strictly maintaining the original proportions of the two classes in the training and test sets. Data augmentation of any kind was strictly prohibited in the test set throughout the process to accurately reflect the low proportion of CBA in clinical settings and avoid evaluation bias caused by synthetic data contamination.
[0034] SMOTE Synthetic Sample Generation and Quality Control
[0035] Within the training set, the SMOTE algorithm was applied to 31 CBA samples to generate synthetic samples:
[0036] Nearest neighbor search: For each CBA sample, a Euclidean distance space is constructed based on 7 standardized features such as GGT and TBA, and the k=5 nearest neighbor samples are searched (verified by grid search, k=5 is the optimal value between generation diversity and computational efficiency).
[0037] Linear interpolation, randomly selecting nearest neighbor samples, according to formula x new =x i +δ×(x jn -x i Generate new samples, where δ is a random number uniformly distributed in the range [0,1]. For example, when δ = 0.3, the feature value of the new sample is the 30% difference between the original sample and its nearest neighbor samples.
[0038] The constraints are that the biochemical index values of the synthetic sample are limited to the mean ± 3 standard deviations of the original CBA sample (e.g., the upper limit of TBA is 150.2 + 3 × 80.7 = 392.3 μmol / L) to ensure that they meet the reasonable range for clinical testing; the imaging parameters LDC and TDC are limited to the anatomical variation range of the original CBA (LDC 0.8–2.5 cm).
[0039] Through the above steps, the number of CBA samples increased from 31 to 192, achieving a 1:1 balance with the number of CC samples. t-SNE dimensionality reduction visualization showed that the synthetic samples expanded along the distribution manifold of the original CBA samples without intruding into the CC sample region, indicating that the interpolation process did not destroy the classification boundaries of the original data.
[0040] 3. MLP Model Architecture Design
[0041] As mentioned earlier, the number of neurons in the input layer strictly corresponds to the seven key diagnostic features selected (GGT, TBA, DBIL, TBIL, ALT, LDC, TDC). Each neuron receives data standardized by Z-score. The standardized data has a mean of 0 and a standard deviation of 1, eliminating the influence of dimensional differences on the model weights. A fully connected approach is used between the input layer and the hidden layers, and the weights are initialized using a He normal distribution (suitable for the ReLU activation function) to accelerate training convergence.
[0042] Input Layer (IL): 7 neurons, corresponding to GGT, TBA, DBIL, TBIL, ALT, LDC, and TDC. Hidden Layers (HL): 3 layers, processing the input data and learning non-linear relationships through weighted connections. Output Layer (OL): 1 neuron, providing a binary classification output (CC or CBA).
[0043] The hidden layer employs a three-layer descending fully connected architecture to progressively extract higher-order pathological features:
[0044] The first hidden layer (128 neurons) receives 7-dimensional input and maps it to a high-dimensional space through dense connections, capturing complex interactions between features (such as the product effect of TBA and LDC).
[0045] The second hidden layer (64 neurons) compresses redundant information and focuses on key discrimination patterns (such as the synergistic signal of GGT>300U / L and TBA>100μmol / L).
[0046] The third hidden layer (32 neurons): further abstracts pathological features and identifies cross-modal associations (such as the CBA-specific pattern of mild ALT elevation accompanied by small cysts).
[0047] The training process involves forward propagation of information from the input layer through the hidden layers to the output layer, followed by backpropagation (BP) to adjust weights and iterative optimization to minimize prediction error. The optimization algorithm used in the training configuration is SGD (Stochastic Gradient Descent), with an initial learning rate of 0.001, a constant learning rate policy, a maximum of 300 iterations, and an actual number of 300 iterations. The regularization parameter (α) is 0.0001. This model is implemented using the Python scikit-learn library. All hyperparameters are optimized using either grid search or random search procedures. Setting a fixed random seed ensures the model's reproducibility.
[0048] The MLP model performed as follows on the test set: Accuracy (ACC) was 95.8%; Area Under the Curve (AUC) was 0.990 (95% confidence interval: 0.9743–1.0000); Sensitivity (SEN) was 92.9% (the ability to correctly identify CBA, which is crucial for avoiding missed diagnoses); and Specificity (SPE) was 96.3% (the ability to correctly identify CC, which reduces false positives).
[0049] Other evaluation metrics included Receiver Operating Characteristic Curve (ROC), Confusion Matrix, and Decision Curve Analysis (DCA), confirming the clinical applicability of the MLP model and showing that its net benefit was higher than the "all-inclusive" or "no-treatment" strategies. The confusion matrix showed that the CBA sensitivity in the test subset reached 100%, emphasizing its reliability.
[0050] Table 1 Comparison of diagnostic performance metrics of different machine learning models
[0051]
[0052]
[0053] 4. User software
[0054] User-friendly diagnostic software developed based on the MLP model allows clinicians to input seven routine clinical parameters, and the software immediately provides diagnostic predictions (CC or CBA) and confidence percentages, improving decision-making efficiency and accuracy.
[0055] The system uses Python 3.10, with the GUI based on the Tkinter library and the joblib library used for model loading. It employs seven features (GGT, TBA, DBIL, TBIL, ALT, LDC, TDC), each labeled with both Chinese and English descriptions as input fields.
[0056] Buttons: "Diagnose" and "ClearInputs".
[0057] Log area: Displays summary of input data, standardized values, prediction results, and medical advice.
[0058] Workflow: After the user inputs data and clicks "Diagnostic Analysis," the software performs the following steps:
[0059] Data standardization: Standardize the input data using the mean and standard deviation of the training set.
[0060] Prediction: Input standardized data into the MLP model and output the predicted category and probability.
[0061] The results show that the log area displays the diagnostic result (CC or CBA), the confidence level (e.g., 70%), and provides medical advice based on the confidence level.
[0062] Confidence level <70%: It is recommended to combine clinical judgment and other auxiliary examinations may be required.
[0063] 70% ≤ confidence level < 90%: moderate confidence level, other clinical manifestations should be considered.
[0064] Confidence level ≥ 90%: High confidence level, but it is still recommended to combine it with other indicators.
[0065] Security and Privacy: The software runs locally and does not transmit data, ensuring the confidentiality of patient information and complying with data protection regulations. Figure 2 )
[0066] This invention, for the first time, integrates seven key clinical indicators—biochemical markers (GGT, TBA, DBIL, TBIL, ALT) and ultrasound imaging parameters (LDC, TDC)—into a unified diagnostic model. This overcomes the limitations of traditional single-indicator methods (such as the TBA threshold method) or isolated image analysis, significantly improving diagnostic efficacy through cross-modal data synergy. A deep neural network with three hidden layers (128-64-32 neurons) constructed based on the aforementioned data achieved 95.8% accuracy, 92.9% sensitivity, 96.3% specificity, and an AUC of 0.990 in a clinical trial of 319 cases, demonstrating significant advantages over traditional logistic regression (AUC 0.902) and random forest (AUC 0.948). Unlike black-box deep learning systems, this invention's model provides a clear ranking of feature importance, aligning with clinical understanding and enhancing physicians' confidence in the diagnostic results. By helping CBA patients receive timely surgical treatment, it can reduce liver damage and potential liver transplant needs caused by delayed diagnosis, lower medical costs, and improve the efficiency of medical resource utilization.
[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A data processing system for auxiliary diagnosis of biliary tract diseases, characterized in that, include: (a) Data acquisition module, configured to acquire patients’ biochemical index test data and ultrasound image data; (b) A data preprocessing module, used to standardize and transform the detection data; (c) Feature filtering module, used to extract key biometric parameters; (d) Classification decision module, which includes a multilayer perceptron neural network structure, is used to generate classification indicators for biliary tract abnormalities; (e) Results output module, which displays auxiliary diagnostic reference information including classification confidence.
2. The data processing system according to claim 1, characterized in that, The biochemical index detection data includes at least five of the following: total bile acid TBA, gamma-glutamyl transferase GGT, total bilirubin TBIL, direct bilirubin DBIL, aspartate aminotransferase AST, and alanine aminotransferase ALT. The ultrasound imaging data includes the maximum longitudinal diameter LDC and the maximum transverse diameter TDC of the cyst extracted by an image segmentation algorithm.
3. The data processing system according to claim 1, characterized in that, The feature screening module is the LASSO feature screening module, which is configured to perform regression analysis on standardized biochemical index data and imaging measurement data to screen out a key feature set including γ-glutamyl transferase GGT, total bile acid TBA, direct bilirubin DBIL, total bilirubin TBIL, alanine aminotransferase ALT, maximum longitudinal diameter LDC of cysts, and maximum transverse diameter TDC.
4. The data processing system according to claim 1, characterized in that, The multilayer perceptron neural network structure has a neural network architecture consisting of an input layer, at least three hidden layers, and an output layer. The number of nodes in the input layer corresponds to the seven parameters of the key feature set, and the output layer consists of binary classification neurons.
5. The data processing system according to claim 1, characterized in that, The classification decision module uses the Synthetic Minority Oversampling Technique (SMOTE) to process the data.
6. The data processing system according to claim 1, characterized in that, The data preprocessing module uses the Z-score method for standardization, with the formula: z = (x - μ) / σ, where x is the original value, μ is the mean of the training set, and σ is the standard deviation of the training set.
7. The data processing system according to claim 1, characterized in that, The confidence level grading logic of the result output module includes: When the confidence level is less than 70%, the first prompt signal is activated and the image verification submodule is invoked; When the confidence level is 70% ≤ confidence level < 90%, activate the second prompt signal and link it to the electronic medical record database; When the confidence level is ≥90%, the third prompt signal is activated and a surgical plan index link is generated.
8. The data processing system according to claim 1, characterized in that, It also includes a model update module, configured as follows: Receive newly labeled clinical data via the PCIe interface; Execute the incremental learning algorithm to update the weights of the multilayer perceptron classifier; Full training mode is triggered when the model accuracy drops by more than 5%.