Clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma based on dictionary learning

Abstract

The present invention provides a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma based on dictionary learning. The construction method includes: collecting medical imaging data and clinical statistical indicator data of patients after surgery for hepatocellular carcinoma; performing data annotation on the medical imaging data to obtain lesion regions of interest; performing feature selection on the lesion regions of interest to obtain optimal features; performing dictionary learning expression on the optimal features to obtain a sparse expression of the optimal features; and constructing a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma using a multivariate logistic regression algorithm based on the sparse expression of the optimal features and the clinical statistical indicator data, as well as the activated partial thromboplastin time. The present invention can solve the technical problem that there is no imaging omics-assisted diagnosis model specifically established for microvascular invasion of hepatocellular carcinoma in the prior art.

Description

Technical Field

[0001] The present invention relates to the technical field of nuclear magnetic resonance medical image processing, and in particular to a clinical auxiliary diagnosis model for hepatocellular carcinoma microvascular invasion based on dictionary learning. Background Art

[0002] Current clinical data indicates that hepatocellular carcinoma (HCC) is the sixth most common cancer worldwide and the third leading cause of cancer-related death. Therefore, accurate diagnosis and treatment of HCC are crucial.

[0003] Currently, the main treatments for hepatocellular carcinoma include surgical resection and liver transplantation. However, clinical data show that the five-year recurrence rate for patients treated with these two treatments remains as high as 70% and 35% respectively. Multiple studies have shown that microvascular invasion (MVI) is a major risk factor for early recurrence of hepatocellular carcinoma after surgery. This is typically confirmed through histopathology of surgical or biopsy specimens. Accurately predicting MVI preoperatively would aid in the development of treatment strategies and potentially prolong patient survival.

[0004] Radiomics is an emerging machine learning-based intelligent analysis method for medical images. It utilizes a series of data mining algorithms to analyze high-throughput imaging features. By building artificial intelligence models with optimal features, it can play a vital role in disease diagnosis and prognosis. However, currently, there is no radiomics model specifically developed for microvascular invasion in hepatocellular carcinoma. The effectiveness and accuracy of other existing models for preoperative prediction of microvascular invasion remain controversial.

[0005] Therefore, in existing clinical medical work, there is an urgent need for a clinical auxiliary diagnostic model for microvascular invasion of hepatocellular carcinoma. Summary of the Invention

[0006] In response to the shortcomings of the existing technology, the present invention proposes a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma based on dictionary learning to solve the technical problem that there is no imaging omics auxiliary diagnosis model specifically established for microvascular invasion of hepatocellular carcinoma in the existing technology.

[0007] The technical solution adopted in the present invention is as follows:

[0008] A clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma based on dictionary learning is constructed by:

[0009] Obtain medical imaging data and clinical statistical indicator data of patients undergoing surgery for hepatocellular carcinoma;

[0010] Annotate medical imaging data to obtain the region of interest of the lesion;

[0011] Perform feature selection on the lesion region of interest to obtain the optimal features;

[0012] Perform dictionary learning on the optimal features to obtain the optimal feature sparse expression;

[0013] Based on the optimal feature sparse expression and clinical statistical indicator data, a multivariate logistic regression algorithm was used to construct a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma.

[0014] Furthermore, the medical imaging data includes multi-sequence magnetic resonance imaging, and the multi-sequence magnetic resonance imaging includes T1-weighted imaging, T2-weighted imaging, arterial phase, venous phase, hepatobiliary phase and apparent diffusion coefficient.

[0015] Furthermore, clinical statistical indicators include: gender, age, whether hepatitis B is positive and whether there is cirrhosis.

[0016] Furthermore, after the medical imaging data are labeled, data preprocessing is performed to obtain the lesion region of interest; data preprocessing includes: matching and interpolating all medical imaging data according to physical space distance, and then normalizing the interpolated medical imaging data.

[0017] Furthermore, multiple imaging omics features were extracted from multiple-sequence magnetic resonance images, and feature selection was performed on the multiple imaging omics features using variance threshold method, univariate selection method and LASSO regression in turn to screen out the optimal features.

[0018] Furthermore, the optimal features include high grayscale area emphasis, kurtosis, small area high grayscale emphasis, skewness, regional entropy, large correlation low grayscale emphasis, and grayscale inhomogeneity.

[0019] Furthermore, the optimal feature sparse expression is obtained, including:

[0020] Perform feature dictionary learning on all optimal features to obtain a unit dictionary that can express all features;

[0021] The optimal features are re-expressed using the unit dictionary to obtain the optimal feature sparse representation.

[0022] Furthermore, the activation function of the multi-factor logistic regression algorithm is the Sigmoid activation function

[0023] Furthermore, when constructing a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma using a multivariate logistic regression algorithm, activated partial thromboplastin time was added as a factor for logistic regression analysis.

[0024] Furthermore, when optimizing the auxiliary diagnosis model, data from external centers are contacted for external verification.

[0025] It can be seen from the above technical solution that the beneficial technical effects of the present invention are as follows:

[0026] 1. During the training process, the clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma combines multi-sequence magnetic resonance imaging, clinical statistical indicator data, and activated partial thromboplastin time to construct a multimodal heterogeneous model. The model comprehensively learns from different sources and heterogeneous data, increasing the reliability and accuracy of the model's prediction results.

[0027] 2. Through the machine learning algorithm of dictionary learning, the dense input features of the optimal features are converted into sparse features. The sparse expression of radiomics features will help improve the overall generalization ability of the model and further enhance the reliability of the model prediction results.

[0028] 3. Further improving the reliability of the auxiliary diagnosis model through forward-looking and external verification can make the clinical promotion and application prospects of the auxiliary diagnosis model better. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0030] Figure 1 The figure is a schematic diagram of the auxiliary diagnosis model construction process according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following embodiments of the technical solution of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore only examples and are not intended to limit the scope of protection of the present invention.

[0032] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0033] Example

[0034] This example provides a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma based on dictionary learning. The model is constructed in the following manner:

[0035] 1. Obtain medical imaging data and clinical statistical indicator data of patients after surgery for hepatocellular carcinoma

[0036] Medical images are multi-sequence magnetic resonance imaging (MRI) images, including six sequences: T1-weighted imaging, T2-weighted imaging, arterial phase, venous phase, hepatobiliary phase, and apparent diffusion coefficient. Clinical statistical indicators include: gender, age, hepatitis B virus positivity, and the presence of cirrhosis. Data acquisition methods are not limited and can be implemented in any feasible manner within existing technologies, such as directly accessing existing data from a medical institution's database via a computer.

[0037] All patients who underwent surgery for hepatocellular carcinoma were divided into a microvascular invasion group and a normal group. In a specific embodiment, the microvascular invasion group includes patients who underwent surgery and whose hepatocellular carcinoma recurred as confirmed by histopathology of biopsy specimens.

[0038] 2. Label the medical imaging data to obtain the region of interest of the lesion

[0039] In some embodiments, a region of interest (ROI) is delineated from medical imaging data of patients undergoing surgery for hepatocellular carcinoma for subsequent high-throughput feature extraction. The delineation method is not limited and can be performed manually by an expert.

[0040] In some embodiments, after data annotation, the medical image data can also be preprocessed. This preprocessing includes matching and interpolating all medical image data based on physical spatial distances, and then normalizing the interpolated medical image data. In a specific embodiment, the DICOM spacing is interpolated to (1mm*1mm*1mm). Preprocessing the medical image data can improve the overall generalization ability of the model.

[0041] In some embodiments, a complete standard database of microvascular invasion of patients after hepatocellular carcinoma surgery is established through data annotation, which can maximize information mining of microvascular invasion images and identify the features most relevant to hepatocellular carcinoma recurrence.

[0042] 3. Select features of the lesion region of interest to obtain the optimal features

[0043] In a specific embodiment, the feature selection of the lesion region of interest includes: extracting 1409 types of imaging features, a total of 8454, from six sequences, namely T1-weighted imaging, T2-weighted imaging, arterial phase, venous phase, hepatobiliary phase and apparent diffusion coefficient, respectively, and performing feature selection on multiple imaging features using variance threshold method, univariate selection method and LASSO regression (Least Absolute Shrinkage and Selection Operator). First, 2659 features are selected from 8454 features using the variance threshold method, then 36 features are selected from the 2659 features using the univariate selection method, and finally, 7 optimal features are selected from the 36 features using LASSO regression, including: High Gray Level Zone Emphasis, Kurtosis, Small Area High Gray Level Emphasis, Skewness, Zone Entropy, Large Dependence Low Gray Level Emphasis. The optimal features screened out in this step include four categories: first-order statistics, shape, texture, and filtering.

[0044] The need for optimal feature selection stems from the fact that the optimal imaging features for predicting microvascular invasion, as screened from thousands of high-throughput omics features, are overly dense, exposing the risk of overfitting the auxiliary diagnosis model. In this step, the variance thresholding method, univariate selection, and LASSO regression are sequentially used to improve the computational efficiency and accuracy of feature selection.

[0045] 4. Perform dictionary learning on the optimal features to obtain the optimal feature sparse expression

[0046] When training an auxiliary diagnosis model, a key consideration in selecting features is sparseness, as many columns in the matrix are irrelevant to the learning task. Removing unused columns can significantly reduce the difficulty of the learning task. Another form of sparsity exists: the matrix may contain many zero elements, but these zero elements do not appear in entire rows or columns. Having such a sparse representation improves the efficiency of the learning task. Furthermore, sparse matrices offer efficient storage methods, minimizing storage and computational overhead. Therefore, for optimal features, if the given data is non-sparse, it can be converted to a sparse representation, simplifying the learning task.

[0047] In some embodiments, dictionary learning (Dictionary Learning) is performed on the optimal features, and the original dense input features are converted into sparse features through the machine learning algorithm of dictionary learning.

[0048] The optimal feature sparse expression mainly includes two steps: first, the feature dictionary learning is performed on all the optimal features to find the unit dictionary that can express all the features; in a specific implementation method, given a data set {x1, ..., x m}, the dictionary learning algorithm is as follows:

[0049]

[0050] In the above formula, is the dictionary matrix, k represents the vocabulary size of the dictionary, It's data The first term in the formula is given by a i Reconstruct x i , the second term is to let a i As sparse as possible.

[0051] Next, we perform sparse feature expression, which involves re-expressing the original optimal features using the learned unit dictionary. This sparse expression of radiomics features will improve the overall generalization ability of the model and further enhance the reliability of the auxiliary diagnosis model.

[0052] 5. Based on the optimal feature sparse expression and clinical statistical index data, a multivariate logistic regression algorithm was used to construct a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma

[0053] When constructing an auxiliary diagnosis model based on the optimal feature sparse expression and clinical statistical indicator data, a multi-factor logistic regression algorithm is used. In a specific embodiment, the activation function of the multi-factor logistic regression algorithm is a Sigmoid activation function, which can increase the nonlinear classification ability of the model. The auxiliary diagnosis model constructed has a logistic regression classifier expression as follows:

[0054]

[0055] In the above formula, Y represents the classifier output, E represents the natural number e, b0 represents the bias term bias, b1 represents the parameters that the model needs to train, and x includes the optimal feature sparse expression and clinical statistical indicator data.

[0056] The inventors of this application have found through research that the recurrence of hepatocellular carcinoma is also related to the results of certain physiological index tests of postoperative patients. Hepatocellular carcinoma patients are often accompanied by abnormal coagulation indicators. The inventors of this application collected laboratory biochemical tests related to coagulation, including thrombin time (TT), prothrombin time (PT), and activated partial thromboplastin time (APTT) to verify whether abnormal coagulation function is related to the occurrence of hepatocellular carcinoma. Studies have shown that the activated partial thromboplastin time (APTT) is positively correlated with the occurrence of hepatocellular carcinoma. Analysis of its mechanism shows that this is because: tumor cells directly produce a variety of procoagulant activities (PCA) and proinflammatory cytokines. PCA can activate the coagulation system, leading to the consumption of coagulation factors, causing the body to be in a pre-bleeding state; and proinflammatory cytokines (including tumor necrosis factor-a and interleukin-1B) can reduce the activation of the protein C system, promote endogenous coagulation, and lead to prolonged APTT.

[0057] In combination with the above mechanism analysis, in some embodiments, when using a multifactor logistic regression algorithm to construct a clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma, the activated partial thromboplastin time is used as a factor for logistic regression analysis to obtain another clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma with higher prediction accuracy.

[0058] In some embodiments, when fine-tuning the trained auxiliary diagnosis model, external validation can be performed using data from external centers. This forward-looking, external validation can further enhance the reliability of the auxiliary diagnosis model and enhance its clinical application prospects.

[0059] During the training process, the clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma implemented in this paper combines multi-sequence magnetic resonance imaging, clinical statistical indicator data and activated partial thromboplastin time to construct a multimodal heterogeneous model. The model comprehensively learns data from different sources and heterogeneous data, which increases the reliability and accuracy of the model's prediction results.

[0060] The machine learning algorithm of dictionary learning converts the dense input features of the optimal features into sparse features. The sparse expression of imaging omics features will be conducive to improving the overall generalization ability of the model and further enhancing the reliability of the model prediction results.

[0061] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. A clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma based on dictionary learning, characterized by: The method for constructing the auxiliary diagnosis model includes: Obtaining medical imaging data and clinical statistical index data of patients undergoing surgery for hepatocellular carcinoma, as well as the patient's activated partial thromboplastin time; the medical imaging data comprises multi-sequence magnetic resonance imaging, including T1-weighted imaging, T2-weighted imaging, arterial phase, venous phase, hepatobiliary phase, and apparent diffusion coefficient; the clinical statistical indexes comprise: gender, age, whether hepatitis B is positive, and whether cirrhosis is present; Annotate medical imaging data to obtain the region of interest of the lesion; Feature selection is performed on the lesion region of interest to obtain the optimal features of microvascular invasion of hepatocellular carcinoma; the optimal features include high grayscale area emphasis, kurtosis, small area high grayscale emphasis, skewness, regional entropy, large correlation low grayscale emphasis, and grayscale inhomogeneity; Perform dictionary learning on the optimal features to obtain the optimal feature sparse expression; A clinical auxiliary diagnosis model for microvascular invasion of hepatocellular carcinoma was constructed using a multivariate logistic regression algorithm based on the optimal feature sparse expression, clinical statistical index data, and the patients' activated partial thromboplastin time.

2. The auxiliary diagnosis model according to claim 1, characterized in that: After the medical image data are labeled, data preprocessing is performed to obtain the lesion region of interest; the data preprocessing includes: matching and interpolating all medical image data according to physical space distance, and then normalizing the interpolated medical image data.

3. The auxiliary diagnosis model according to claim 1, characterized in that: Multiple imaging omics features were extracted from multiple sequence magnetic resonance imaging images, and feature selection was performed on the multiple imaging omics features using variance threshold method, univariate selection method and LASSO regression in turn to screen out the optimal features.

4. The auxiliary diagnosis model according to claim 1, characterized in that Get the optimal feature sparse representation, including: Perform feature dictionary learning on all optimal features to obtain a unit dictionary that can express all features; The optimal features are re-expressed using the unit dictionary to obtain the optimal feature sparse representation.

5. The auxiliary diagnosis model according to claim 1, characterized in that: The activation function of the multi-factor logistic regression algorithm is the Sigmoid activation function.

6. The auxiliary diagnosis model according to claim 1, characterized in that: When tuning the auxiliary diagnosis model, contact the data of external centers for external verification.

Images