Application of liver fibrosis detection grading model construction method to schistosomiasis

Abstract

The application discloses a liver fibrosis detection grading model construction method applied to schistosomiasis, relates to the technical field of medical detection, collects multi-modal liver sample data for standardized processing to generate standard multi-modal data, adopts corresponding processing methods for feature processing on different standard multi-modal data, obtains liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis, constructs a multi-modal joint grading diagnosis model based on the features to detect liver fibrosis of patients with schistosomiasis, verifies the detection result of the liver fibrosis detection, dynamically optimizes the multi-modal joint grading diagnosis model based on the detection result, evaluates the model indication of the multi-modal joint grading diagnosis model after the dynamic optimization is completed, and determines whether to adjust the optimization parameter of the dynamic optimization of the multi-modal joint grading diagnosis model based on the model indication.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of medical detection, in particular to a liver fibrosis detection grading model construction method applied to blood fluke disease. BACKGROUND

[0002] Blood fluke disease is a parasitic disease that seriously endangers human health, is widely prevalent in many regions of the world, especially in regions with relatively poor sanitary conditions and water sources that are easy to be polluted, and after long-term infection with blood flukes, the liver is extremely vulnerable to damage, causing liver fibrosis, and the treatment strategy and prognosis are completely different for different degrees of liver fibrosis.

[0003] In the past, the detection means for liver fibrosis caused by blood fluke disease is single and scattered, mainly including traditional liver imaging examination such as ultrasound and CT, serum biochemical marker detection such as detection of specific proteins and enzyme indicators to reflect liver damage, and pathological section examination, different detection means are isolated from each other, lack effective integration, and each modality data is not fully mined and cooperatively used, which cannot provide comprehensive, accurate and dynamic liver fibrosis grading information for the clinic, resulting in low diagnosis efficiency and poor accuracy of blood fluke disease liver fibrosis, delaying the best treatment opportunity of patients, and increasing medical costs and patient pain. SUMMARY

[0004] In order to solve the above problems, the purpose of the present application is to provide a liver fibrosis detection grading model construction method applied to blood fluke disease.

[0005] The purpose of the present application can be achieved by the following technical solutions: a liver fibrosis detection grading model construction method applied to blood fluke disease, comprising the following steps:

[0006] Step S1: collecting multi-modal liver sample data related to blood fluke disease, and standardizing the multi-modal liver sample data to generate respective standard multi-modal data;

[0007] Step S2: adopting corresponding processing methods for feature processing on different standard multi-modal data to obtain liver imaging features, serum biochemical marker features and pathological section features representing the influence of blood fluke disease on liver fibrosis;

[0008] Step S3: constructing a multi-modal joint grading diagnosis model based on the liver imaging features, serum biochemical marker features and pathological section features, and detecting liver fibrosis of blood fluke disease related patients through the multi-modal joint grading diagnosis model;

[0009] Step S4: verifying the detection result of the liver fibrosis detection, dynamically optimizing the multi-modal joint grading diagnosis model based on the detection result, and evaluating the model indication of the multi-modal joint grading diagnosis model after the dynamic optimization is completed;

[0010] Step S5: determining whether to adjust the optimization parameter of the dynamic optimization of the multi-modal joint grading diagnosis model based on the model indication;

[0011] Collecting multi-modal liver sample data related to schistosomiasis, and standardizing the multi-modal liver sample data to generate respective corresponding standard multi-modal data includes:

[0012] Selecting a number of diagnosed patients and a number of healthy control groups in the schistosomiasis epidemic area, and setting the selection ratio of diagnosed patients and healthy control groups;

[0013] Configure the ultrasonic diagnostic instrument, set the gain value of the TGC curve corresponding to the near field, the middle field and the far field respectively, collect the liver multi-slice dynamic image, and perform two-dimensional wavelet transform filtering, adaptive histogram equalization and cubic spline interpolation processing on the liver multi-slice dynamic image;

[0014] Scan the liver multi-slice dynamic image of the diagnosed patient with schistosomiasis by spiral CT equipment, and perform image gray normalization and rigid registration processing after scanning, and perform MRI scanning according to the set parameters by using the superconducting magnetic resonance imaging system, and perform gray normalization and rigid registration after scanning;

[0015] Further, the standardized ultrasound image corresponding to the diagnosed patient is obtained, and the standardized ultrasound image includes MRI image and CT image;

[0016] Collect a few milliliters of venous blood of the diagnosed patient, centrifuge the serum, detect specific biochemical markers and perform standardization conversion, obtain corresponding serum biochemical marker data, and perform non-local mean filtering, color calibration and brightness contrast adjustment under the guidance of ultrasound.

[0017] Further, the corresponding processing method is adopted for feature processing of different standard multi-modal data, and the process of obtaining liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis includes:

[0018] Construct an improved convolutional neural network, divide the standardized ultrasound image into a training set and a test set in a ratio of 7:3, and use an Adam optimizer and a data enhancement technique to extract ultrasound image features in the standardized ultrasound image;

[0019] The CT liver region in the standardized ultrasound image is extracted by using Otsu threshold segmentation and region growing algorithm, the CT value, morphological and blood vessel related features are calculated, the MRI image is divided into training set and test set in the ratio of 6:4 by using 3D-U-Net model, the image is trained and normalized by using Adagrad optimizer to extract features;

[0020] The multi-modal feature vector corresponding to the serum biochemical marker data is constructed, and the fibrous tissue, hepatocytes and inflammatory cells are identified in the Masson and HE staining sections by using pathological image analysis software, and the fiber ratio, cell morphology parameters and inflammatory cell number are calculated;

[0021] The liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis are summarized.

[0022] Further, the process of constructing a multi-modal joint grading diagnosis model based on liver image features, serum biochemical marker features and pathological section features includes:

[0023] The liver image features, serum biochemical marker features and pathological section feature vectors are spliced in the ratio of 1:1:1 to form a comprehensive feature vector, and a support vector machine algorithm and a radial basis kernel function are selected to construct an initial multi-modal joint grading diagnosis model;

[0024] Obtain sample data, and divide the sample data into training set and test set in the ratio of 8:2;

[0025] Set the penalty parameter and kernel function parameter of the initial multi-modal joint grading diagnosis model;

[0026] Use grid search algorithm combined with 5-fold cross-validation to value the penalty parameter and kernel function parameter, optimize the parameters in the value range until the training set accuracy is above 95% and the test set accuracy is stable above 90%, and apply the initial multi-modal joint grading diagnosis model to detect liver fibrosis of patients with schistosomiasis and output the grading results.

[0027] Further, the process of detecting liver fibrosis of patients with schistosomiasis by the multi-modal joint grading diagnosis model includes:

[0028] Collect the relevant data of different modalities of patients with schistosomiasis who need to be detected for liver fibrosis, and standardize all the relevant data to obtain different multi-modal standard data of the patients, and the liver image features, serum biochemical marker features and pathological section features obtained after feature processing of different multi-modal standard data;

[0029] Extract the liver image feature, serum biochemical marker feature and pathological section feature corresponding to the respective multi-modal feature vector, and input all the multi-modal feature vectors into the multi-modal joint grading diagnosis model which has been trained, output the grading result of related liver fibrosis detection, and the grading result includes F0-F4 grade.

[0030] Further, the detection result of liver fibrosis detection is verified, the multi-modal joint grading diagnosis model is dynamically optimized based on the detection result, and the process of evaluating the model indication of the multi-modal joint grading diagnosis model after dynamic optimization comprises:

[0031] The liver fibrosis grading results output by the multi-modal joint grading diagnosis model are summarized, and the liver fibrosis grading results are compared and analyzed with the patient's clinical symptoms, signs and other examination results, and then the detection result is verified, the multi-modal joint grading diagnosis model is dynamically optimized by using the sliding window method, and then the model accuracy and sensitivity of the multi-modal joint grading diagnosis model are monitored, the mean and standard deviation of the last 5 results are calculated, and the model indication of the multi-modal joint grading diagnosis model is evaluated.

[0032] Further, the process of deciding whether to adjust the optimization parameters of the dynamic optimization of the multi-modal joint grading diagnosis model comprises:

[0033] The model accuracy and sensitivity of the multi-modal joint grading diagnosis model are monitored and recorded, the indication threshold is set, the current model indication is compared with the indication threshold, if the accuracy is continuously improved by at least 2% for 10 times, the accuracy difference between the training set and the test set is within 10%, the sensitivity is more than 85% and the specificity is more than 90%, it is decided to stop adjusting the optimization parameters, if the threshold is not met, such as the index improvement is slow, overfitting or underfitting, the optimization parameters are continuously adjusted.

[0034] Further, for the multi-modal joint grading diagnosis model trained by gradient descent algorithm, the learning rate is increased when the training is slow, the cross-validation fold is adjusted when the oscillation is reduced and the generalization ability of cross-validation is decreased, and the multi-modal joint grading diagnosis model is retrained and verified.

[0035] Compared with the prior art, the beneficial effects of the present application are:

[0036] 1、The present application comprehensively collects liver image, serum biochemical marker and pathological section multi-modal liver sample data, and organically integrates different dimensional information, wherein the liver image directly shows the macroscopic morphological structure change of the liver, the serum biochemical marker reflects the microscopic change of the liver cell metabolism, and the pathological section provides accurate histological basis, the three are complementary, can capture the details of the influence of schistosomiasis on liver fibrosis in all directions, overcome the limitation of single modal detection, and significantly improve the diagnosis accuracy.

[0037] 2. After the construction of the multi-modal joint grading diagnosis model, the detection results are verified, the model is dynamically optimized continuously, different patient groups, regional schistosome strain differences and disease changes are continuously adapted, the optimization parameters are flexibly adjusted and optimized according to the model indication, and the model is ensured to be in the best diagnosis efficiency state at all times, and reliable liver fibrosis grading is provided for the clinic for a long time, and precise medical treatment is assisted. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The flowchart of the present application. DETAILED DESCRIPTION

[0039] As Figure 1 shown, the liver fibrosis detection grading model construction method applied to schistosomiasis includes the following steps:

[0040] Step S1: Collecting multi-modal liver sample data related to schistosomiasis, and standardizing the multi-modal liver sample data to generate respective corresponding standard multi-modal data;

[0041] Step S2: Taking corresponding processing methods for different standard multi-modal data for feature processing, and then obtaining liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis;

[0042] Step S3: Constructing a multi-modal joint grading diagnosis model based on the liver image features, serum biochemical marker features and pathological section features, and detecting liver fibrosis of patients with schistosomiasis through the multi-modal joint grading diagnosis model;

[0043] Step S4: Verify the detection results of the liver fibrosis detection, dynamically optimize the multi-modal joint grading diagnosis model based on the detection results, and evaluate the model indication of the multi-modal joint grading diagnosis model after completing the dynamic optimization;

[0044] Step S5: Based on the model indication, decide whether to adjust the optimization parameters of the dynamic optimization of the multi-modal joint grading diagnosis model.

[0045] It needs to be further explained that in the specific implementation process, the process of collecting multi-modal liver sample data related to schistosomiasis and standardizing the multi-modal liver sample data to generate respective corresponding standard multi-modal data includes:

[0046] First, sample selection is performed, 500 patients diagnosed with schistosomiasis are selected to constitute a patient group in a schistosomiasis epidemic area, and the selected patients need to cover different infection stages, different age levels and different genders, so as to ensure the diversity of the samples;

[0047] Meanwhile, 100 cases of healthy control population confirmed by examination were selected, and the selected healthy control population was matched with the patient group in terms of age level and gender distribution, so as to reduce the interference of individual differences on subsequent detection and model construction;

[0048] An ultrasonic diagnostic instrument equipped with a 5MHz probe was selected to obtain the internal structure image of the liver of the patient, the gain was set to 50dB, the time gain compensation curve was set to 20dB, 30dB and 40dB in the near field, the middle field and the far field respectively, through this setting, the signal attenuation caused by the increase of ultrasonic propagation distance can be effectively compensated, and the image quality of different depth regions of the liver is consistent;

[0049] When performing liver ultrasonic examination on each patient, dynamic images of the left lobe sagittal section, right lobe oblique section and subcostal transverse section of the liver were collected, and the collection time of each section was set to 10 seconds to ensure that sufficient liver structure information was obtained, covering the morphological changes of the liver at different angles and positions;

[0050] After the collection is completed, the ultrasonic image is filtered by using a two-dimensional wavelet transform filtering algorithm, the two-dimensional wavelet transform can decompose the image in different scales and directions, effectively suppress the noise and retain the edge and texture information of the image, avoiding the interference of noise affecting the subsequent feature extraction;

[0051] The contrast is enhanced by using an adaptive histogram equalization method, which adjusts the histogram adaptively according to the gray distribution of the local region of the image, so that the details of the liver tissue are clearer, such as the running of the blood vessels in the liver and the texture of the liver parenchyma, which is convenient for subsequent analysis;

[0052] The image resolution is uniformly resampled to 512x512 pixels by using a cubic spline interpolation method, which ensures the consistency of the image in subsequent processing and provides a standardized image data basis for feature extraction;

[0053] The 128-slice spiral CT device is used for scanning, before scanning, the patient needs to fast for more than 6 hours to reduce the interference of gastrointestinal contents on liver imaging, ensure the clarity of liver image, and take appropriate amount of contrast agent, which can enhance the contrast between liver tissue and surrounding tissue, making the boundary and internal structure of the liver more clear and identifiable;

[0054] After scanning, the CT image is subjected to gray scale normalization processing, the CT value is mapped to the [0, 1] interval, and the differences caused by different devices and scanning parameters are eliminated, so that the CT images of different patients are comparable, a rigid registration algorithm based on mutual information is used for spatial standardization registration, and the best spatial transformation parameter is found, so that all images have the same position and direction in the spatial coordinate system.

[0055] It needs to be further explained that in the specific implementation process, different standard multi-modal data are processed by corresponding processing methods for feature processing, and then the liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis are obtained.

[0056] The improved convolutional neural network model based on ResNet-50 is constructed, the first four convolutional modules of the original ResNet-50 are retained, the global average pooling layer and the full connection layer are removed, two convolutional layers are added after the fourth convolutional module, one of which is set to 3x3, the other is set to 1x1, the step is 1, ReLU is selected as the activation function, the pooling layer is set to average pooling, the pooling kernel size is 2x2, and the step is 2;

[0057] The standardized ultrasound image data is divided into training set and test set according to the proportion of 7:3, the Adam optimizer is adopted, the initial learning rate of the Adam optimizer is set to 0.0001, and the data enhancement technology is used for random rotation, horizontal flip and vertical flip operation of the ultrasound image within the range of-15° to 15°;

[0058] Through the above operation, the convolutional neural network model learns the texture features of the patient's liver, including the uniformity, fineness and shape features of the echo and blood vessels, etc.

[0059] Feature mining is performed through the 3D-U-Net deep learning model, the 3D-U-Net model includes an encoder, a decoder and a skip connection structure, the encoder part is composed of multiple 3D convolutional layers and 3D pooling layers, which is used to extract deep features of the image.

[0060] The decoder part is composed of multiple 3D deconvolutional layers and 3D convolutional layers, which is used to restore the spatial resolution of the image, and the skip connection structure fuses the feature maps of different levels in the encoder with the feature maps of corresponding levels in the decoder, improving the feature utilization rate.

[0061] The MRI image data is divided into training set and test set according to the proportion of 6:4, the Adagrad optimizer is used, and the learning rate of the Adagrad optimizer is set to 0.001. The MRI image is normalized, the image intensity value is normalized to the interval [0, 1], and the signal intensity difference and tissue contrast of the liver tissue under different sequences are learned.

[0062] The dimension reduction algorithm of principal component analysis combined with factor analysis is adopted to standardize the biochemical marker data of hyaluronic acid, laminin, type III procollagen and type IV collagen in serum, so that the data mean is 0 and the standard deviation is 1.

[0063] The covariance matrix of the data after standardization is calculated, the eigenvalues and eigenvectors are obtained by eigenvalue decomposition, the principal components with a cumulative contribution rate of more than 90% are selected, the initial eigenvectors are constructed, the principal components are further explained and reconstructed by factor analysis, the factors with actual biological significance are extracted, and finally the comprehensive serum biochemical marker eigenvector is constructed. The serum biochemical marker eigenvector is used to represent the serum biochemical marker characteristics;

[0064] The pathological sections of the liver tissue are analyzed by means of pathological image analysis software. In the stained sections, the color threshold segmentation algorithm is used to identify the blue fibrous tissue, the morphological operation is used to remove noise and small fragments, and the percentage of fibrous tissue in the whole liver tissue area is calculated.

[0065] By setting the color characteristics and morphological parameters of the cell nucleus, the morphological parameters including circularity, area and perimeter, the hepatocytes are identified by using the cell detection algorithm, 10 high-power fields are randomly selected, the morphological parameters such as the diameter, nuclear area and nuclear-cytoplasmic ratio of 50 hepatocytes in each field are measured, and the average value is calculated, the inflammatory cells are counted by using the cell counting algorithm, the number of inflammatory cells per unit area is calculated, and then the pathological section characteristics corresponding to the pathological section are obtained.

[0066] It should be further explained that in the specific implementation process, the process of constructing the multi-modal joint grading diagnosis model based on the liver image features, serum biochemical marker features and pathological section features includes:

[0067] The liver image features, serum biochemical marker features and pathological section features obtained through the previous steps are merged, the feature vectors corresponding to the liver image features, serum biochemical marker features and pathological section features are connected end to end in a ratio of 1:1:1, a comprehensive feature vector is constructed, and the comprehensive feature vector is used as the input data for subsequent model construction. The comprehensive feature vector can comprehensively reflect the multi-aspect information of the influence of schistosomiasis on liver fibrosis;

[0068] The support vector machine algorithm is selected to construct the multi-modal joint grading diagnosis model, all sample data are divided into a training set and a test set in a ratio of 8:2, the training set is used for parameter learning and optimization of the model, and the test set is used to evaluate the generalization ability and prediction performance of the model;

[0069] It should be noted that the proportions of various samples in the training set and the test set are consistent with the proportions in the original data set, so as to ensure the balance and representativeness of the data division;

[0070] The penalty parameter and the kernel function parameter of the multi-modal joint grading diagnosis model are set;

[0071] wherein the penalty parameter and the kernel function parameter are denoted as C and D respectively, C = 2, and D = 0.1;

[0072] The penalty parameter C controls the tolerance of the multi-modal joint hierarchical diagnosis model to classification errors in the training process, the greater the value of C, the lower the penalty of the multi-modal joint hierarchical diagnosis model to classification errors, which may lead to overfitting; the smaller the value of C, the higher the tolerance of the multi-modal joint hierarchical diagnosis model to classification errors, which may lead to underfitting; the kernel function parameter D determines the width of the radial basis kernel function, which affects the complexity and generalization ability of the model;

[0073] The grid search algorithm is combined with 5-fold cross-validation to optimize the penalty parameter C, the grid search algorithm performs exhaustive search within the pre-set parameter range, and the value range of C is set as follows:

[0074] C = [0.1, 1, 2, 5, 10];

[0075] The value range of the kernel function parameter D is set as [0.01, 0.05, 0.1, 0.5, 1];

[0076] Each time 4 data are selected from the training set as the training subset, and 1 data is selected as the validation subset, and the accuracy of the multi-modal joint hierarchical diagnosis model on the validation subset under different parameter combinations is calculated;

[0077] The parameter combination with the highest accuracy is selected as the optimal parameter, and the optimization is iterated until the accuracy of the multi-modal joint hierarchical diagnosis model on the training set reaches more than 95%, and the accuracy on the test set is stable at more than 90%, and the final multi-modal joint hierarchical diagnosis model is constructed.

[0078] It should be further pointed out that, in the specific implementation process, the process of detecting liver fibrosis of schistosomiasis related patients by the multi-modal joint hierarchical diagnosis model includes:

[0079] For the schistosomiasis related patients who need to be detected for liver fibrosis, the same equipment and parameters as when the multi-modal joint hierarchical diagnosis model is constructed are used for multi-modal data acquisition;

[0080] That is, using an ultrasonic diagnostic instrument equipped with a 5MHz probe, a gain of 50dB, and a TGC curve of 20dB, 30dB, and 40dB in the near field, the middle field, and the far field, respectively, to collect liver multi-section dynamic images;

[0081] That is, using a 128-slice spiral CT device, with tube voltage 120 kV, tube current 200 mA, layer thickness 1 mm, pitch 1.25, the parameters are scanned, and the patient is fasting for more than 6 hours before scanning and oral contrast agent; Using a 3.0T superconducting magnetic resonance imaging system, according to the T1WI sequence: TR=500ms, TE15ms, T2WI sequence: TR=4000ms, TE=100ms, the parameters of MRI scanning are performed;

[0082] Fasting collection of 8ml of patient's venous blood, centrifugation at 3000r / min for 15 minutes to separate serum, and detection of serum biochemical markers such as hyaluronic acid, laminin, type III procollagen and type IV collagen by enzyme-linked immunosorbent assay;

[0083] Under the guidance of ultrasound, about 1.5-2cm of liver tissue samples are obtained from the edge of the right lobe of the liver using an 18G puncture needle, and after fixation, dehydration, transparency, wax immersion and embedding, pathological sections with a thickness of 4μm are prepared, hematoxylin-eosin (HE) staining and Masson staining are performed;

[0084] The ultrasound image is subjected to two-dimensional wavelet transform filtering, adaptive histogram equalization and cubic spline interpolation processing, the image resolution is resampled to 512x512 pixels, the CT and MRI images are subjected to gray scale normalization processing, the CT value or MRI signal intensity is mapped to the [0,1] interval, and a rigid registration algorithm based on mutual information is used for spatial standardization registration;

[0085] The serum biochemical marker data is standardized according to the detection method characteristics, with the mean and standard deviation of the marker in healthy people as the reference, after digitization of the pathological sections, non-local mean filtering algorithm is used for noise reduction, color calibration card is used for color calibration, and brightness is adjusted to 80 and contrast is adjusted to 1.2;

[0086] According to the feature extraction method when constructing the multi-modal joint grading diagnosis model, the features of the new data are extracted, that is, the improved ResNet-50 convolutional neural network is used to extract the ultrasound image features, the region growing and morphological analysis algorithm is used to extract the CT image features, the 3D-U-Net deep learning model is used to extract the MRI image features, the principal component analysis combined with factor analysis dimension reduction algorithm is used to extract the serum biochemical marker features, and the professional pathological image analysis software is used to extract the pathological section features;

[0087] The multi-modal feature vectors corresponding to the liver image features, serum biochemical marker features and pathological section features of the new patient are input into the trained multi-modal joint grading diagnosis model, and then the grading results of liver fibrosis detection are output, including F0-F4 levels.

[0088] It needs to be further explained that in the specific implementation process, the process of verifying the detection result of the liver fibrosis detection, dynamically optimizing the multi-modal joint grading diagnosis model based on the detection result, and evaluating the model indication of the multi-modal joint grading diagnosis model after the dynamic optimization includes:

[0089] The liver fibrosis grading results output by the multi-modal joint grading diagnosis model are summarized, covering the F0-F4 grading results corresponding to each patient. Other relevant clinical data of the patient are also collected synchronously, including symptoms such as fatigue, abdominal distension, signs such as liver and spleen size, and other medical examination results such as liver function indicators and etiology examination indicators.

[0090] The grading results of the multi-modal joint grading diagnosis model are compared in detail with the patient's clinical symptoms, signs, and other medical examination results. If the multi-modal joint grading diagnosis model diagnoses a patient with F3 fibrosis, obvious liver function abnormalities and hepatosplenomegaly are present, which are consistent with the model grading results, then it is determined to be verified. If there is a large deviation, such as the multi-modal joint grading diagnosis model diagnosing early fibrosis F1, but the patient already has severe liver function damage, then it is marked as a result to be verified.

[0091] The key performance indicators of the model are continuously monitored using the sliding window method, including model accuracy, sensitivity, and specificity. A sliding window containing the last 5 optimization results is set, and the average value and standard deviation of each key performance indicator in the sliding window are calculated.

[0092] If the average value of the model accuracy, sensitivity, and specificity in the sliding window improves by less than 1% and the standard deviation is greater than 5%, it is determined that the model optimization effect is not good, and the multi-modal joint grading diagnosis model is trapped in a local optimum.

[0093] When the model accuracy in the training set is higher than 95%, but the model accuracy in the test set is lower than 80%, and the difference in loss function values between the training set and the test set is greater than 0.5, it is determined that the multi-modal joint grading diagnosis model has overfitting phenomenon.

[0094] If the accuracy of the training set and the test set is less than 70% and the loss function value is greater than 0.75, it indicates that the model has underfitting problem.

[0095] If the model training speed is too slow, increase the learning rate;

[0096] If the model is in oscillation, reduce the learning rate;

[0097] If the model generalization ability does not meet the standard when using 5-fold cross-validation, adjust the cross-validation fold number to 7 or 3, and retrain and verify the model;

[0098] After each dynamic optimization is completed, the key performance indicators of the model are evaluated;

[0099] If the model accuracy is improved by at least 2% in each of the last 10 optimizations, the accuracy difference between the training set and the test set is always within 10%, the sensitivity is greater than 85%, and the specificity is greater than 90%, it is determined that the model performance is in a stable improvement state, otherwise, no determination is made; when the model performance is in a stable improvement state, stop dynamic optimization and determine the final model parameters.

[0100] It should be further pointed out that, in the specific implementation process, the process of determining whether to adjust the optimization parameters of the dynamic optimization of the multi-modal joint hierarchical diagnosis model based on the model indicators includes:

[0101] According to the historical performance of the multi-modal joint hierarchical diagnosis model in the training and verification process and the actual needs of clinical application, corresponding threshold values are set for each model indicator of the multi-modal joint hierarchical diagnosis model;

[0102] For example, set the improvement threshold of the model accuracy to at least 2% in each of the last multiple optimizations, the accuracy difference threshold between the training set and the test set to not more than 10%, the sensitivity threshold to greater than 85%, and the specificity threshold to greater than 90%, and set an acceptable range for the loss function value difference, such as not more than 0.5;

[0103] Compare and analyze the monitored model indicators with the corresponding set threshold values, as follows:

[0104] When the model accuracy of the multi-modal joint hierarchical diagnosis model is improved by at least 2% in each of the last 10 optimizations, the accuracy difference between the training set and the test set is always within 10%, the sensitivity is greater than 85%, and the specificity is greater than 90%, it is determined that the optimization parameters of the current multi-modal joint hierarchical diagnosis model are not adjusted during dynamic optimization;

[0105] When the multi-modal joint hierarchical diagnosis model is in a local optimum, overfitting occurs or there is an underfitting problem, the optimization parameters of the current multi-modal joint hierarchical diagnosis model during dynamic optimization are adjusted.

[0106] The above embodiments are only used to illustrate the technical method of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical method of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical method of the present application.

Claims

1. A method for constructing a liver fibrosis detection and grading model for schistosomiasis, characterized in that, The method comprises the following steps: Step S1: Collecting multi-modal liver sample data related to schistosomiasis, and standardizing the multi-modal liver sample data to generate corresponding standard multi-modal data; Step S2: Processing the different standard multi-modal data by corresponding processing methods to obtain liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis; Step S3: Constructing a multi-modal joint grading diagnosis model based on the liver image features, serum biochemical marker features and pathological section features, and detecting liver fibrosis of patients with schistosomiasis through the multi-modal joint grading diagnosis model; Step S4: Verifying the detection results of the liver fibrosis detection, dynamically optimizing the multi-modal joint grading diagnosis model based on the detection results, and evaluating the model indications of the multi-modal joint grading diagnosis model after the dynamic optimization is completed; Step S5: Based on the model indications, determining whether to adjust the optimization parameters of the dynamic optimization of the multi-modal joint grading diagnosis model; The process of collecting multi-modal liver sample data related to schistosomiasis and standardizing the multi-modal liver sample data to generate corresponding standard multi-modal data comprises: Selecting a number of diagnosed patients and a number of healthy control groups in a schistosomiasis epidemic area, and setting the selection ratio of diagnosed patients to healthy control groups; Configuring an ultrasonic diagnostic instrument, setting the gain values of the TGC curve in the near field, the middle field and the far field respectively, collecting liver multi-slice dynamic images, and performing two-dimensional wavelet transform filtering, adaptive histogram equalization and cubic spline interpolation processing on the liver multi-slice dynamic images; Scanning the liver multi-slice dynamic images of the diagnosed patients with schistosomiasis by a spiral CT device, and performing image gray normalization and rigid registration processing after scanning, using a superconducting magnetic resonance imaging system to perform MRI scanning according to the set parameters, and performing gray normalization and rigid registration after scanning; Then the standardized ultrasonic images of the diagnosed patients are obtained, which include MRI images and CT images; Collecting several milliliters of venous blood of the diagnosed patients, centrifuging the serum, detecting specific biochemical markers and performing standardization conversion to obtain corresponding serum biochemical marker data, and puncturing liver pathological section samples under ultrasonic guidance, and performing non-local mean filtering, color calibration and brightness contrast adjustment after staining and digitization.

2. The method for constructing a liver fibrosis detection and grading model for schistosomiasis according to claim 1, characterized in that, The process of processing the different standard multi-modal data by corresponding processing methods to obtain liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis comprises: An improved convolutional neural network is constructed, the standardized ultrasonic images are divided into a training set and a test set in a ratio of 7:3, and the ultrasonic image features in the standardized ultrasonic images are extracted by using an Adam optimizer and a data enhancement technique; The CT liver region in the standardized ultrasound image is extracted by Otsu threshold segmentation and region growing algorithm, the CT value, morphological and blood vessel related features are calculated, the MRI image is divided into training set and test set in the ratio of 6:4 by using 3D-U-Net model, the image is trained and normalized by using Adagrad optimizer to extract features; The multi-modal feature vectors corresponding to the serum biochemical marker data are constructed, and the fibrous tissue, liver cells and inflammatory cells are identified in the Masson and HE staining sections by using pathological image analysis software, and the fiber ratio, cell morphology parameters and inflammatory cell number are calculated; The liver image features, serum biochemical marker features and pathological section features representing the influence of schistosomiasis on liver fibrosis are summarized.

3. The method for constructing a liver fibrosis detection and grading model for schistosomiasis according to claim 2, characterized in that, The process of constructing a multi-modal joint grading diagnosis model based on liver image features, serum biochemical marker features and pathological section features includes: The liver image features, serum biochemical marker features and pathological section feature vectors are spliced into comprehensive feature vectors in the ratio of 1:1:1, and an initial multi-modal joint grading diagnosis model is constructed by using support vector machine algorithm and radial basis kernel function; Obtain sample data, and divide the sample data into training set and test set in the ratio of 8:2; Set the penalty parameter and kernel function parameter of the initial multi-modal joint grading diagnosis model; Use grid search algorithm combined with 5-fold cross validation to optimize the penalty parameter and kernel function parameter in the value range until the training set accuracy is above 95% and the test set accuracy is stable above 90%, and apply the initial multi-modal joint grading diagnosis model to detect liver fibrosis of patients with schistosomiasis and output the grading results.

4. The method for constructing a liver fibrosis detection and grading model for schistosomiasis according to claim 3, characterized in that, The process of detecting liver fibrosis of patients with schistosomiasis by multi-modal joint grading diagnosis model includes: Collect the relevant data of different modalities of patients with schistosomiasis who need to be detected for liver fibrosis, and standardize all the relevant data to obtain different multi-modal standard data of the patients, and the liver image features, serum biochemical marker features and pathological section features obtained after feature processing of different multi-modal standard data; Extract the multi-modal feature vectors corresponding to the liver image features, serum biochemical marker features and pathological section features respectively, and input all the multi-modal feature vectors into the multi-modal joint grading diagnosis model which has been trained, and output the grading results of liver fibrosis detection of the related patients, including F0-F4 levels.

5. The method for constructing a liver fibrosis detection and grading model for schistosomiasis according to claim 4, characterized in that, Verify the detection results of liver fibrosis detection, dynamically optimize the multi-modal joint grading diagnosis model based on the detection results, and evaluate the model indicators of the multi-modal joint grading diagnosis model after dynamic optimization. The liver fibrosis grading results output by the multi-modal joint grading diagnosis model are summarized, and the liver fibrosis grading results are compared and analyzed with the patient's clinical symptoms, signs and other examination results, and then the detection results are verified. The multi-modal joint grading diagnosis model is dynamically optimized by using the sliding window method, and then the model accuracy and sensitivity of the multi-modal joint grading diagnosis model are monitored. The mean and standard deviation of the last 5 results are calculated, and the model indication of the multi-modal joint grading diagnosis model is evaluated.

6. The method for constructing a liver fibrosis detection and grading model for schistosomiasis according to claim 5, characterized in that, The process of determining whether to adjust the optimization parameters for dynamically optimizing the multi-modal joint grading diagnosis model includes: The model accuracy and sensitivity of the multi-modal joint grading diagnosis model are monitored and recorded. The indication threshold is set, and the current model indication is compared with the indication threshold. If the accuracy is continuously improved by at least 2% for 10 times, the accuracy difference between the training set and the test set is within 10%, the sensitivity is more than 85%, and the specificity is more than 90%, it is determined to stop adjusting the optimization parameters. If the threshold is not met, such as slow index improvement, overfitting or underfitting, the optimization parameters continue to be adjusted.

7. The method for constructing a liver fibrosis detection and grading model for schistosomiasis according to claim 6, characterized in that, For the multi-modal joint grading diagnosis model trained by the gradient descent algorithm, the learning rate is increased when the training is slow, the cross-validation fold is adjusted when the oscillation is reduced and the generalization ability of cross-validation is decreased, and the multi-modal joint grading diagnosis model is retrained and verified.

Images