CN116994754A - Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images - Google Patents

Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images Download PDF

Info

Publication number
CN116994754A
CN116994754A CN202310919538.1A CN202310919538A CN116994754A CN 116994754 A CN116994754 A CN 116994754A CN 202310919538 A CN202310919538 A CN 202310919538A CN 116994754 A CN116994754 A CN 116994754A
Authority
CN
China
Prior art keywords
pancreatic
diabetes
pancreas
features
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310919538.1A
Other languages
Chinese (zh)
Inventor
祝徐
祁志卫
苏云杉
陈祥
张莉
岳昆
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yunnan University YNU
Original Assignee
Yunnan University YNU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Yunnan University YNU filed Critical Yunnan University YNU
Priority to CN202310919538.1A priority Critical patent/CN116994754A/en
Publication of CN116994754A publication Critical patent/CN116994754A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/30—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for calculating health indices; for individual health risk assessment
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/54—Extraction of image or video features relating to texture
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/80—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level
    • G06V10/806—Fusion, i.e. combining data from various sources at the sensor level, preprocessing level, feature extraction level or classification level of extracted features
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Medical Informatics (AREA)
  • Health & Medical Sciences (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • Public Health (AREA)
  • Data Mining & Analysis (AREA)
  • Biomedical Technology (AREA)
  • Epidemiology (AREA)
  • Pathology (AREA)
  • Primary Health Care (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a visual assessment method for pancreatic lesions of diabetics based on CT images, which is characterized in that pancreatic CT images of a plurality of diabetics and non-diabetics are collected to form a pancreatic CT image data set, the pancreatic CT image data set is divided into a training set and a verification set, key features of diabetes are extracted based on the pancreatic CT image data set, a biomarker model is constructed, key feature vectors of diabetes of all persons in the training set are extracted, decision boundaries of the biomarker model are trained for a plurality of times as training samples, verification is carried out by adopting the verification set, and a group of feature fusion coefficients and bias values with highest accuracy are selected; and for a person needing to carry out visual assessment of pancreatic lesions, extracting the value of each pixel point diabetes key feature of a pancreatic parenchymal region and calculating to obtain a biomarker, and then generating a pancreatic heterogeneity clustering label graph and calculating a heterogeneity score. The invention can clearly present the pancreatic lesion condition and realize quantitative evaluation on the pancreatic lesion degree.

Description

Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images
Technical Field
The invention belongs to the technical field of image histology, and particularly relates to a visual evaluation method for pancreatic lesions of diabetics based on CT images.
Background
Diabetes is a metabolic disease characterized by hyperglycemia, mainly caused by insulin secretion defect, pancreas is the only organ of human endocrine insulin, and when abnormal sugar metabolism occurs in human body, the pancreas can be fibrosed, fat and other diseases, and in severe cases, pancreatic cancer can be caused. If the treatment is not carried out in time, the pancreatic lesion degree can be continuously aggravated and various diseases are induced, so that the human health is endangered. Histological examination (e.g., tissue aspiration biopsy) is currently the gold standard for diagnosing pancreatic lesions, is an invasive procedure, and presents problems such as interpretation errors, repeatable difficulties, and the like. At present, a computer tomography (Computed Tomography, CT) technology is generally adopted clinically, pancreatic lesion degree is judged by manually reading pancreatic CT images and assisting clinical characterization of patients, but the method has poor diagnosis accuracy for early lesions, a professional doctor is required to review a large number of CT images for manual diagnosis to judge the pancreatic lesion degree, meanwhile, due to irregular pancreatic shapes, large differences of shapes, sizes, positions and the like among different image slices, and the problems of fuzzy boundary and the like caused by similar pancreatic and surrounding tissue density, the problem of careless leakage is unavoidable only by manual diagnosis. Thus, there is a need to create a computer-assisted means to assist physicians in visualizing and quantitatively assessing the extent of lesions in a patient's pancreas lesions.
The known disease diagnosis method based on the cluster analysis technology only clusters the clinical characterization of the patient, does not analyze the pathological changes of the physiological tissues by using CT images, and is difficult to accurately evaluate the disease progress of the patient. For example, yu Rui (Yu Rui. Principal component factor and cluster analysis of patients with chronic obstructive pulmonary disease with double bronchodilators [ D ]. Chongqing medical university, 2022) the principal component factor analysis is performed by selecting 15 characteristic indexes of chronic obstructive pulmonary disease, converting into 9 common factors, and cluster analysis is performed according to the common factors, classifying the patients into 3 kinds, namely, patients with mild, moderate and severe pulmonary function impairment respectively, and then analyzing prognosis conditions for different patients respectively. Wang Jiangxia (Wang Jiangxia, yang Lixia, wei Ruixian, mi Deng sea. Discussion of the characteristics of the symptoms of diabetic nephropathy based on data mining [ J ]. Chinese medicine research, 2022,35 (6): 75-79.) the symptoms, pulse conditions, tongue conditions of diabetic nephropathy patients are subjected to cluster analysis, 3 core combinations are obtained, and the core combinations are analyzed, so that a basis is provided for the diagnosis and treatment of diabetic nephropathy by a clinician.
Image histology is a multidisciplinary technique that uses non-invasive imaging modalities to analyze the complex features of a disease without performing biopsies or other invasive procedures to extract pathological features. The pancreatic imaging characteristics of the diabetics can be extracted from the images in a high throughput manner, and a model is built to assist clinical diagnosis, so that the workload of doctors can be effectively reduced, the diagnosis accuracy is improved, and the risk of trauma estimated by the patients in a prognosis manner and required medical resources are reduced.
The known pancreatic disease diagnosis and evaluation method based on image histology does not comprehensively consider visualization of pancreatic lesion areas and quantitative evaluation of pancreatic lesions, and is difficult to know development of pancreatic lesions in time, so that diagnosis and treatment effects of the method are poor and interpretability is poor. For example, yuan Mengyi (Yuan Mengyi. Pancreatic cystic tumor classification based on image histology feature fusion [ D ]. Zhejiang university of industry, 2020) proposes a pancreatic cystic tumor classification method based on image histology feature fusion, which comprehensively quantifies pathological features of pancreatic cystic tumors by extracting different types of features such as gray scale, texture, geometry, clinical characterization and the like of CT images of pancreatic cystic tumors, and classifies the pancreatic cystic tumors according to a joint dissimilarity feature matrix, but the method can only classify benign and malignant tumors of pancreatic cystic tumors and fails to quantitatively evaluate pancreatic heterogeneity of patients. The method can obtain standard quantitative diabetes risk assessment results, but cannot realize pancreas region visualization, and is difficult to further analyze pancreatic lesion conditions and disease progression of diabetics.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provides a visual assessment method for pancreatic lesions of diabetics based on CT images, which is characterized in that key characteristics of diabetes contained in the pancreatic CT images are extracted, a pancreatic heterogeneity cluster tag map is generated based on a biomarker model, and a heterogeneity score is calculated, so that pancreatic lesions are clearly presented, and quantitative assessment is realized on the degree of the pancreatic lesions.
In order to achieve the aim of the invention, the visual assessment method for pancreatic lesions of diabetics based on CT images comprises the following steps:
s1: collecting pancreas CT images of a plurality of diabetics and non-diabetics, and constructing a pancreas CT image data set C= { C 1 ,c 2 ,…,c N }，c n Representing an nth Zhang Yixian CT image, n=1, 2, …, N representing the number of images in the pancreatic CT image dataset; for each pancreas CT image c n Labeling, if the corresponding person is a diabetic patient, labeling f n =1, otherwise tag f n =0; at each pancreatic CT image c n The outline of the pancreas is outlined in a dot mode to obtain a pancreas parenchymal region, and then a corresponding labeling image I is generated n Labeling image I n The pixel value of the pixel points which belong to the pancreas parenchymal area is 1, and the pixel value of the pixel points which do not belong to the pancreas parenchymal area is 0; dividing pancreas CT image dataset C into training set C train And validation set C test ；
S2: extracting key characteristics of diabetes based on pancreas CT image dataset C, specifically comprising the following steps:
s2.1: setting M texture features according to actual needs, and for each pancreas CT image C in the pancreas CT image data set C n Respectively according to the corresponding marked image I n Extracting to obtain pancreas parenchymal region, and extracting each texture feature value in pancreas parenchymal region to form texture feature vector x n Then, texture feature vectors of all pancreas CT images are used as row vectors to form a texture feature matrix X with the size of M multiplied by N;
s2.2: normalizing the texture feature matrix X, mapping all feature values to a range of 0-1 to obtain a normalized texture feature matrix X * ；
S2.3: for texture feature matrix X * Performing dimension reduction operation to obtain a dimension-reduced texture feature matrix X';
s2.4: the method for screening the key characteristics of diabetes based on the characteristic weights comprises the following steps:
s2.4.1: weights W of all features j′ Initializing to 0, j '=1, 2, …, U representing the number of texture features in the texture feature matrix X', and initializing the number of iterations t=1;
s2.4.2: randomly selecting a diabetic patient a from the pancreas CT image data set, and calculating the distance between the diabetic patient a and other personnel b on the texture feature j':
wherein ,xa [j′]、x b [j′]Texture feature vector x representing diabetic patient a and person b, respectively a 、x b The value, max, of the texture feature j j′ 、min j′ Respectively representing the maximum value and the minimum value of the texture features j' of all people;
s2.4.3: for diabetes patient a, the distance from diabetes patient a is selected from other diabetes patientsSmall diabetics b 1 Screening out the person with smallest distance from diabetes patient a from the persons with non-diabetes patient as the nearest neighbor of the same kind 2 As heterogeneous nearest neighbors, the feature weights W are then updated using the following formula j′ ：
S2.4.4: judging whether the iteration times t is less than t max ，t max Representing a preset maximum number of iterations, if yes, proceeding to step S2.4.5, otherwise proceeding to step S2.4.6;
s2.4.5: let t=t+1, return to step S2.4.2;
s2.4.6: taking the feature weight obtained in the last iteration as the weight of the corresponding texture feature, selecting the first K features with the largest weight from the U features in the texture feature matrix X' as key features of diabetes, and setting the value of K according to actual conditions;
s3: a biomarker model is constructed, and the decision boundary is expressed as follows:
wherein ,a diabetes key feature vector representing an input sample, < +.>Values representing key features of kth diabetes, k=1, 2, …, K, f k The characteristic fusion coefficient corresponding to the kth diabetes key characteristic is represented, and B represents a bias value;
when (when)Indicating that the person to whom the input sample corresponds is diabetic, otherwise not;
extracting training set C from texture feature matrix X train The key feature vector of diabetes of each person is used as a training sample to carry out multi-round training on the decision boundary of the biomarker model, and feature fusion coefficients and bias values obtained by each round of training are recorded; then extracting verification set C from texture feature matrix X test The key feature vectors of diabetes of each person are respectively predicted whether each person is a diabetic patient by adopting each group of feature fusion coefficients and offset values, the accuracy is counted, and a group of feature fusion coefficients and offset values with highest accuracy are used as the feature fusion coefficients f for final use k And a bias value B;
s4: for the personnel who need to carry out visual assessment of pancreatic lesions, pancreatic CT images of the personnel are acquiredMarking a pancreatic parenchymal region, and extracting K values of key characteristics of diabetes from each pixel point in the pancreatic parenchymal region>The individual pixel biomarker Y is then calculated according to the following formula:
dividing pixel points in a pancreatic parenchyma region into D categories according to a biomarker by adopting a k-means clustering algorithm, setting the value of D according to actual needs, and setting the pixel points in each category as one color so as to obtain a pancreatic heterogeneity clustering label graph;
s5: dividing a pancreatic heterogeneity clustering tag map into three sub-regions in a horizontal direction by adopting an equal proportion cutting mode, wherein the three sub-regions respectively represent a head part, a body part and a tail part, and calculating to obtain pancreatic heterogeneity scores RTFscore (m) of each sub-region, wherein m=1, 2 and 3, and the specific method is as follows:
according to the classification of the pixel points obtained in the step S4, respectively for each sub-region in the current sub-regionObtaining connected domains by pixel points of the types, deleting the connected domains with the number of the pixel points smaller than a threshold value from the obtained connected domains, screening the connected domain with the maximum number of the pixel points from all the rest connected domains of the types, and recording the number of the pixel points as s m,max The heterogeneity score RTFscore (m) for this sub-region is then calculated using the following formula:
wherein S represents a pancreatic CT imageTotal number of pixels in the parenchymal region of the pancreas.
The invention discloses a visual assessment method for pancreatic lesions of diabetics based on CT images, which comprises the steps of collecting pancreatic CT images of a plurality of diabetics and non-diabetics to form a pancreatic CT image data set, dividing the pancreatic CT image data set into a training set and a verification set, extracting key features of diabetes based on the pancreatic CT image data set, constructing a biomarker model, extracting key feature vectors of diabetes of each person in the training set, taking the key feature vectors as training samples to train decision boundaries of the biomarker model for multiple rounds, then adopting the verification set to verify, and selecting a group of feature fusion coefficients and bias values with highest accuracy; and for a person needing to carry out visual assessment of pancreatic lesions, extracting the value of each pixel point diabetes key feature of a pancreatic parenchymal region and calculating to obtain a biomarker, and then generating a pancreatic heterogeneity clustering label graph and calculating a heterogeneity score.
The invention has the following beneficial effects:
(1) The invention safely, noninvasively and accurately acquires a series of key features most relevant to diabetes in pancreatic images through diabetes key feature extraction, biomarker model construction and the like, and solves the defects of low accuracy, poor interpretability and the like in the prior art and provides a new solution for diagnosis of other diseases by constructing image biomarkers and analyzing pathological changes of pancreatic tissues of patients.
(2) The invention provides a pancreas lesion area visualization method based on key feature clustering, which is used for generating a pancreas heterogeneity clustering label graph by clustering key features in pancreas CT images of diabetics, clearly and intuitively displaying pancreas lesion areas of the diabetics, and solving the problem that pancreas lesions are difficult to observe by naked eyes in the traditional pancreas image field.
(3) The invention provides an evaluation method of pancreatic heterogeneity, provides a new strategy for quantitatively evaluating the pathological change degree of pancreatic regions, quantitatively analyzes the pathological change degree of pancreatic regions of patients through pancreatic heterogeneity scoring functions, accurately measures the pathological change degree of the pancreas in a noninvasive manner, and reduces the risk of trauma in prognosis evaluation of patients.
Drawings
FIG. 1 is a flow chart of an embodiment of a method for visual assessment of pancreatic lesions of a diabetic patient based on CT images in accordance with the present invention;
FIG. 2 is a flow chart of extracting key features of diabetes mellitus according to the present invention;
FIG. 3 is a flow chart of a method for selecting key features of diabetes based on feature weights in the present invention;
FIG. 4 is an original pancreatic CT image of a patient;
FIG. 5 is a schematic view of a region of the pancreas identified by the CT image of the pancreas of FIG. 4;
FIG. 6 is a graph of pancreatic heterogeneity cluster tags generated in this example;
fig. 7 is a heterogeneity score plot of the pancreatic heterogeneity cluster tag plot in this example.
Detailed Description
The following description of the embodiments of the invention is presented in conjunction with the accompanying drawings to provide a better understanding of the invention to those skilled in the art. It is to be expressly noted that in the description below, detailed descriptions of known functions and designs are omitted here as perhaps obscuring the present invention.
Examples
FIG. 1 is a flowchart of an embodiment of a method for visual assessment of pancreatic lesions of a diabetic patient based on CT images according to the present invention. As shown in fig. 1, the method for visually evaluating pancreatic lesions of diabetics based on CT images comprises the following specific steps:
s101: obtaining a pancreatic CT image sample set:
collecting pancreas CT images of a plurality of diabetics and non-diabetics, and constructing a pancreas CT image data set C= { C 1 ,c 2 ,…,c N }，c n N Zhang Yixian CT images are represented, n=1, 2, …, N represents the number of images in the pancreatic CT image dataset. For each pancreas CT image c n Labeling, if the corresponding person is a diabetic patient, labeling f n =1, otherwise tag f n =0. At each pancreatic CT image c n The outline of the pancreas is outlined in a dot mode to obtain a pancreas parenchymal region, and then a corresponding labeling image I is generated n Labeling image I n The pixel value of the pixel belonging to the pancreatic parenchymal region is 1, and the pixel value of the pixel not belonging to the pancreatic parenchymal region is 0. Dividing pancreas CT image dataset C into training set C train And validation set C test 。
S102: extracting key characteristics of diabetes:
and extracting key characteristics of diabetes based on the pancreas CT image data set C. Fig. 2 is a flow chart of extracting key features of diabetes mellitus in the present invention. As shown in fig. 2, the specific steps of extracting the key features of diabetes in the invention include:
s201: extracting texture features:
setting M texture features according to actual needs, and for each pancreas CT image C in the pancreas CT image data set C n Respectively according to the corresponding marked image I n Extracting to obtain pancreas parenchymal region, and extracting each texture feature value in pancreas parenchymal region to form texture feature vector x n And then, taking the texture feature vectors of all the pancreas CT images as row vectors to form a texture feature matrix X with the size of M multiplied by N. In this embodiment, the texture feature matrix is generated by free open source software LIFEx for image feature processing, and the texture features include intensity features, position intensity features, intensity histogram features, and gray level symbiotic momentMatrix characteristics, gray run length matrix characteristics, neighborhood gray difference matrix characteristics and gray size area matrix characteristics.
S202: texture feature normalization:
different features in the texture feature matrix X have different dimensions and dimension units, so that all feature values are mapped to a range of 0-1 through a normalization method to reduce the influence of distribution differences among different features on a subsequent biomarker model, and the normalized texture feature matrix X is obtained * . The normalized calculation formula in this embodiment is:
wherein μ and σ represent the standard deviation and average value of the texture feature data, respectively.
S203: dimension reduction of texture features:
to eliminate the normalized texture feature matrix X * Too high dimensions cause the problem of overfitting of the model, and the texture feature matrix X is needed * And performing dimension reduction operation to obtain a dimension-reduced texture feature matrix X'. The specific method for implementing dimension reduction of the texture feature matrix in the embodiment comprises the following steps:
traversing texture feature matrix X * Feature vector of each texture feature in (a)i=1, 2, …, M, calculating its feature vector +.>Correlation coefficient ρ between ij ，j＝1,2,…,M&j+.i. The correlation coefficient ρ in the present embodiment ij The calculation formula of (2) is as follows:
wherein ,representing feature vector +.>And feature vector->Covariance between>Respectively represent feature vector +>Feature vector->Standard deviation of (2).
The larger absolute value of the correlation coefficient indicates the stronger correlation between feature vectors, so if |ρ ij When the I is larger than the preset threshold, the correlation between the two eigenvectors is too strong, and the contribution of the two eigenvectors to diabetes prediction is not obviously different from the contribution of only one eigenvector selected, so that the eigenvectors are removedOtherwise, do nothing.
S204: key feature selection for diabetes:
in the invention, the relevance of different texture features and diabetes is calculated, different weights are given to different features, the larger the feature weight is, the more relevant the feature is to the diabetes, otherwise, the weaker the relevance of the feature to the diabetes is. FIG. 3 is a flow chart of a method for selecting key features of diabetes based on feature weights in the present invention. As shown in fig. 3, the specific steps of the invention for selecting the key characteristics of diabetes based on the characteristic weights comprise:
s301: initializing parameters:
weights W of all features j′ Initialized to 0, j' =12, …, U represents the number of texture features in the texture feature matrix X'. The number of iterations t=1 is initialized.
S302: calculating texture feature distance:
randomly selecting a diabetic patient a from the pancreas CT image data set, and calculating the distance between the diabetic patient a and other personnel b on the texture feature j':
wherein ,xa [j′]、x b [j′]Texture feature vector x representing diabetic patient a and person b, respectively a 、x b The value, max, of the texture feature j j′ 、min j′ Representing the maximum and minimum values of the texture feature j' among all persons, respectively.
S303: updating the feature weights:
for diabetes patient a, screening diabetes patient b with minimum distance from diabetes patient a from other diabetes patients 1 Screening out the person with smallest distance from diabetes patient a from the persons with non-diabetes patient as the nearest neighbor of the same kind 2 As heterogeneous nearest neighbors, the feature weights W are then updated using the following formula j′ ：
S304: judging whether the iteration times t is less than t max ，t max Indicating a preset maximum number of iterations, if yes, step S305 is entered, otherwise step S306 is entered.
S305: let t=t+1, return to step S302.
S306: screening for key features of diabetes:
and taking the feature weight obtained in the last iteration as the weight of the corresponding texture feature, selecting the first K features with the largest weight from the U 'features in the texture feature matrix X' as key features of diabetes, and setting the value of K according to actual conditions.
S103: building and training a biomarker model:
in step S102, K mutually independent key characteristics of diabetes are screened out, and based on the key characteristics, a biomarker model is constructed, and the biomarker model predicts whether a sample has diabetes probability by fitting a decision boundary. Decision boundaries of the biomarker model in the present invention are expressed as follows:
wherein ,a diabetes key feature vector representing an input sample, < +.>Values representing key features of kth diabetes, k=1, 2, …, K, f k And B represents the bias value.
When (when)It indicates that the person to whom the input sample corresponds is a diabetic patient, otherwise not.
Extracting training set C from texture feature matrix X train The key feature vector of diabetes of each person is used as a training sample to train the decision boundary of the biomarker model for multiple rounds, and feature fusion coefficients and bias values obtained by each round of training are recorded. Then extracting verification set C from texture feature matrix X test The key feature vectors of diabetes of each person are respectively predicted whether each person is a diabetic patient by adopting each group of feature fusion coefficients and offset values, the accuracy is counted, and a group of feature fusion coefficients and offset values with highest accuracy are used as the feature fusion coefficients f for final use k And a bias value B.
The loss function is very important for biomarker model training, and in general, pancreatic CT image data of a diabetic patient is far more than pancreatic CT image data of a non-diabetic patient, so that the prediction accuracy of the model on the non-diabetic patient is low. Therefore, in this embodiment, a loss weight is added to the diabetes prediction loss function, so as to improve the situation that the prediction accuracy is not high due to the imbalance of the data of the diabetic patient and the non-diabetic patient in the pancreatic image data set. The loss function calculation formula adopted in the present embodiment is as follows:
wherein Q represents the number of people in the current training batch, y q Indicating whether person q in the current training batch is a real tag for diabetics,and (3) representing the prediction result of the biomarker model on the personnel in the current training batch, wherein alpha represents a preset loss weight coefficient, and 0 < alpha < 1.
After the loss function is calculated, a random gradient descent method is adopted to iteratively update the characteristic fusion coefficient f in the biomarker model k And a bias value B.
S104: generating a pancreatic heterogeneity cluster tag map:
for the personnel who need to carry out visual assessment of pancreatic lesions, pancreatic CT images of the personnel are acquiredMarking a pancreatic parenchymal region, and extracting K values of key characteristics of diabetes from each pixel point in the pancreatic parenchymal region>The individual pixel biomarker Y is then calculated according to the following formula:
the method comprises the steps of dividing pixel points in a pancreatic parenchymal region into D categories according to a biomarker by adopting a k-means clustering algorithm, setting the value of D according to actual needs, and setting the pixel points in each category as a color so as to obtain a pancreatic heterogeneity clustering label graph. In the invention, similar pixels are classified into the same region by classification so as to display the pancreatic lesion region. The k-means clustering algorithm is a common classification algorithm, and the specific process is not described here again.
S105: pancreatic subregion heterogeneity assessment:
dividing a pancreatic heterogeneity clustering tag map into three sub-regions in a horizontal direction by adopting an equal proportion cutting mode, wherein the three sub-regions respectively represent a head part, a body part and a tail part, and calculating to obtain pancreatic heterogeneity scores RTFscore (m) of each sub-region, wherein m=1, 2 and 3, and the specific method is as follows:
according to the classification of the pixel points obtained in the step S104, the connected domains are obtained for the pixel points of each class in the current subarea, the connected domains with the number of the pixel points smaller than the threshold value (1 in the embodiment) are deleted from the obtained connected domains, the connected domain with the largest number of the pixel points is screened from the connected domains of all the remaining classes, and the number of the pixel points is recorded as S m,max The heterogeneity score RTFscore (m) for this sub-region is then calculated using the following formula:
wherein S represents a pancreatic CT imageTotal number of pixels in the parenchymal region of the pancreas.
In order to better illustrate the technical effects of the invention, the invention is experimentally verified by adopting a specific example. In this example, a visual evaluation of pancreatic lesions was performed using CT images of the upper abdomen of a patient with type 2 diabetes in a hospital. 196 patients with clinically confirmed type 2 diabetes were selected in this example and identifiedThe middle upper abdomen CT images of 62 non-diabetic patients selected randomly in period are marked with pancreas parenchymal areas to form a pancreas CT image data set C= { C 1 ,c 2 ,…,c 258 }. Fig. 4 is an original pancreatic CT image of a patient. Fig. 5 is a schematic view of a region of the pancreas identified by the pancreatic CT image shown in fig. 4.
Diabetes key features are next extracted based on the pancreatic CT image dataset C. Firstly, extracting a substantial pancreas region image from a pancreas CT image of a pancreas CT image data set C, then introducing the substantial pancreas region image into LIFEx software in batches to generate a VOI, automatically calculating a texture feature matrix X of the pancreas parenchyma, and extracting 98 texture features in total. Table 1 is a texture feature data table in this embodiment.
Feature class Feature quantity
Intensity characteristics 18
Location intensity feature 16
Intensity histogram feature 17
Gray level co-occurrence matrix features 15
Gray scale run length matrix features 13
Neighborhood gray scale difference matrix features 5
Gray scale area matrix features 14
Totals to 98
TABLE 1
As shown in table 1, the texture features in this embodiment can be divided into 7 groups of Intensity features (Intensity Based, IB), location Intensity features (Local Intensity Based, LI), intensity histogram features (Intensity Histogram, IH), gray Level Co-occurrence Matrix, GLCM), gray Level run length matrix features (Gray Level Run Length Matrix, GLRLM), neighborhood Gray Level difference matrix features (Neighbouring Gray Tone Difference Matrix, NGTDM), gray Level size area matrix features (Gray Level Size Zone Matrix, GLSZM). Then normalizing the texture features to obtain a normalized texture feature matrix X * . Table 2 is a table of the normalized texture feature matrix portion data in this embodiment.
TABLE 2
The normalized texture feature matrix X is then processed * And (5) performing dimension reduction. In this embodiment, a threshold delta of the correlation coefficient is set to be 0.990, and a normalized texture feature matrix X is calculated * And (3) the correlation coefficient rho between the feature vectors in the step (c), and selecting a feature vector pair with the absolute value rho of the correlation coefficient being larger than delta to obtain a feature matrix X' after dimension reduction, which comprises 34 feature components after dimension reduction. Table 3 is a partial data table of the texture feature matrix after the dimension reduction in the present embodiment.
TABLE 3 Table 3
In this embodiment, the K value is set to 7, the weights of all texture feature components are calculated based on the reduced-dimension texture feature matrix X', and the first 7 features with the largest weights are selected as key features of diabetes. Table 4 is a data table of the key feature matrix portion of diabetes in this example.
TABLE 4 Table 4
Training the biomarker model to obtain a feature fusion coefficient f k And a bias value B. Table 5 is a table of biomarker model parameters obtained by training in this example.
TABLE 5
Then, for the personnel needing to carry out visual assessment of pancreatic lesions, collecting pancreatic CT images of the personnel and marking pancreatic parenchymal areas, respectively extracting the values of K key diabetes mellitus characteristics from each pixel point in the pancreatic parenchymal areas, and adopting a characteristic fusion coefficient f obtained by training k And calculating the biomarker of each pixel point by the offset value B. Table 6 is a table of biomarker portion data for each pixel point in the pancreatic parenchymal region in this example.
TABLE 6
In this embodiment, the clustering number d=3 is set, and the k-means clustering algorithm is adopted to divide the pixel points in the pancreatic parenchyma region into 3 categories according to the biomarker. Fig. 6 is a pancreatic heterogeneity cluster tag map generated in this example. The pixels of each category in fig. 6 are respectively colored differently to distinguish them.
The pancreatic heterogeneity cluster tag map is then partitioned into Head (Head), body (Body), tail (Tail) and the heterogeneity score for each portion is calculated separately. Fig. 7 is a heterogeneity score plot of the pancreatic heterogeneity cluster tag plot in this example. Table 7 is a table of heterogeneity scores of pancreatic heterogeneity cluster tag plots for all samples in this example.
Head RTFscore Body RTFscore Tail RTFscore
Image data 1 0.68 0.87 0.96
Image data 2 0.89 0.67 0.68
Image data 3 0.72 0.40 0.82
Image data 4 0.68 0.80 0.59
…… …… …… ……
Image data 258 0.31 0.88 0.66
TABLE 7
While the foregoing describes illustrative embodiments of the present invention to facilitate an understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, but is to be construed as protected by the accompanying claims insofar as various changes are within the spirit and scope of the present invention as defined and defined by the appended claims.

Claims (5)

1. A visual evaluation method of pancreatic lesions of diabetics based on CT images is characterized by comprising the following steps:
s1: collecting pancreas CT images of a plurality of diabetics and non-diabetics, and constructing a pancreas CT image data set C= { C 1 ,c 2 ,…,c N }，c n Representing an nth Zhang Yixian CT image, n=1, 2, …, N representing the number of images in the pancreatic CT image dataset; for each pancreas CT image c n Labeling, if the corresponding person is a diabetic patient, labeling f n =1, otherwise tag f n =0; at each pancreatic CT image c n The outline of the pancreas is outlined in a dot mode to obtain a pancreas parenchymal region, and then a corresponding labeling image I is generated n LabelingImage I n The pixel value of the pixel points which belong to the pancreas parenchymal area is 1, and the pixel value of the pixel points which do not belong to the pancreas parenchymal area is 0; dividing pancreas CT image dataset C into training set C train And validation set C test ；
S2: extracting key characteristics of diabetes based on pancreas CT image dataset C, specifically comprising the following steps:
s2.1: setting M texture features according to actual needs, and for each pancreas CT image C in the pancreas CT image data set C n Respectively according to the corresponding marked image I n Extracting to obtain pancreas parenchymal region, and extracting each texture feature value in pancreas parenchymal region to form texture feature vector x n Then, texture feature vectors of all pancreas CT images are used as row vectors to form a texture feature matrix X with the size of M multiplied by N;
s2.2: normalizing the texture feature matrix X, mapping all feature values to a range of 0-1 to obtain a normalized texture feature matrix X * ；
S2.3: for texture feature matrix X * Performing dimension reduction operation to obtain a dimension-reduced texture feature matrix X';
s2.4: the method for screening the key characteristics of diabetes based on the characteristic weights comprises the following steps:
s2.4.1: weights W of all features j′ Initializing to 0, j '=1, 2, …, U representing the number of texture features in the texture feature matrix X', and initializing the number of iterations t=1;
s2.4.2: randomly selecting a diabetic patient a from the pancreas CT image data set, and calculating the distance between the diabetic patient a and other personnel b on the texture feature j':
wherein ,xa [j′]、x b [j′]Texture feature vector x representing diabetic patient a and person b, respectively a 、x b The value, max, of the texture feature j j′ 、min j′ Respectively represent all peopleMaximum and minimum values of texture features j' for the member;
s2.4.3: for diabetes patient a, screening diabetes patient b with minimum distance from diabetes patient a from other diabetes patients 1 Screening out the person with smallest distance from diabetes patient a from the persons with non-diabetes patient as the nearest neighbor of the same kind 2 As heterogeneous nearest neighbors, the feature weights W are then updated using the following formula j′ ：
S2.4.4: judging whether the iteration times t is less than t max ，t max Representing a preset maximum number of iterations, if yes, proceeding to step S2.4.5, otherwise proceeding to step S2.4.6;
s2.4.5: let t=t+1, return to step S2.4.2;
s2.4.6: taking the feature weight obtained in the last iteration as the weight of the corresponding texture feature, selecting the first K features with the largest weight from the U features in the texture feature matrix X' as key features of diabetes, and setting the value of K according to actual conditions;
s3: a biomarker model is constructed, and the decision boundary is expressed as follows:
wherein ,a diabetes key feature vector representing an input sample, < +.>Values representing key features of kth diabetes, k=1, 2, …, K, f k The characteristic fusion coefficient corresponding to the kth diabetes key characteristic is represented, and B represents a bias value;
when (when)Indicating that the person to whom the input sample corresponds is diabetic, otherwise not;
extracting training set C from texture feature matrix X train The key feature vector of diabetes of each person is used as a training sample to carry out multi-round training on the decision boundary of the biomarker model, and feature fusion coefficients and bias values obtained by each round of training are recorded; then extracting verification set C from texture feature matrix X test The key feature vectors of diabetes of each person are respectively predicted whether each person is a diabetic patient by adopting each group of feature fusion coefficients and offset values, the accuracy is counted, and a group of feature fusion coefficients and offset values with highest accuracy are used as the feature fusion coefficients f for final use k And a bias value B;
s4: for the personnel who need to carry out visual assessment of pancreatic lesions, pancreatic CT images of the personnel are acquiredMarking a pancreatic parenchymal region, and extracting K values of key characteristics of diabetes from each pixel point in the pancreatic parenchymal region>The individual pixel biomarker Y is then calculated according to the following formula:
dividing pixel points in a pancreatic parenchyma region into D categories according to a biomarker by adopting a k-means clustering algorithm, setting the value of D according to actual needs, and setting the pixel points in each category as one color so as to obtain a pancreatic heterogeneity clustering label graph;
s5: dividing a pancreatic heterogeneity clustering tag map into three sub-regions in a horizontal direction by adopting an equal proportion cutting mode, wherein the three sub-regions respectively represent a head part, a body part and a tail part, and calculating to obtain pancreatic heterogeneity scores RTFscore (m) of each sub-region, wherein m=1, 2 and 3, and the specific method is as follows:
according to the classification of the pixel points obtained in the step S4, respectively solving the connected domains for the pixel points of each class in the current subarea, deleting the connected domains with the number of the pixel points smaller than the threshold value from the obtained connected domains, screening the connected domains with the maximum number of the pixel points from the connected domains of all the remaining classes, and recording the number of the pixel points as S m,max The heterogeneity score RTFscore (m) for this sub-region is then calculated using the following formula:
wherein S represents a pancreatic CT imageTotal number of pixels in the parenchymal region of the pancreas.
2. The method according to claim 1, wherein the texture features in step S2.1 include intensity features, location intensity features, intensity histogram features, gray level co-occurrence matrix features, gray level run length matrix features, neighborhood gray level difference matrix features, and gray level size area matrix features.
3. The method for visual assessment of pancreatic lesions of diabetic patients according to claim 1, wherein the texture feature matrix X is used in step S2.3 * The specific method for reducing the dimension is as follows:
traversing texture feature matrix X * Feature vector of each texture feature in (a)Calculating the feature vector +.>Correlation coefficient ρ between ij ，j＝1,2,…,M&j is not equal to i; if |ρ ij If the I is larger than the preset threshold, eliminating the feature vector +.>Otherwise, do nothing.
4. A method for visual assessment of pancreatic lesions in a diabetic patient according to claim 3, wherein said correlation coefficient ρ ij The calculation formula of (2) is as follows:
wherein ,representing feature vector +.>And feature vector->Covariance between>Respectively represent feature vector +>Feature vector->Standard deviation of (2).
5. The method for visual assessment of pancreatic lesions in diabetic patients according to claim 1, wherein the loss function calculation formula adopted for the biomarker model training in step S3 is as follows:
wherein Q represents the number of people in the current training batch, y q Indicating whether person q in the current training batch is a real tag for diabetics,and (3) representing the prediction result of the biomarker model on the personnel in the current training batch, wherein alpha represents a preset loss weight coefficient, and 0 < alpha < 1.
CN202310919538.1A 2023-07-25 2023-07-25 Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images Pending CN116994754A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310919538.1A CN116994754A (en) 2023-07-25 2023-07-25 Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310919538.1A CN116994754A (en) 2023-07-25 2023-07-25 Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images

Publications (1)

Publication Number Publication Date
CN116994754A true CN116994754A (en) 2023-11-03

Family

ID=88520792

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310919538.1A Pending CN116994754A (en) 2023-07-25 2023-07-25 Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images

Country Status (1)

Country Link
CN (1) CN116994754A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117911722A (en) * 2024-03-19 2024-04-19 陕西中医药大学 Artificial intelligence-based tongue image feature extraction method for diabetic patients
CN117911722B (en) * 2024-03-19 2024-06-04 陕西中医药大学 Artificial intelligence-based tongue image feature extraction method for diabetic patients

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117911722A (en) * 2024-03-19 2024-04-19 陕西中医药大学 Artificial intelligence-based tongue image feature extraction method for diabetic patients
CN117911722B (en) * 2024-03-19 2024-06-04 陕西中医药大学 Artificial intelligence-based tongue image feature extraction method for diabetic patients

Similar Documents

Publication Publication Date Title
Cameron et al. MAPS: a quantitative radiomics approach for prostate cancer detection
US10255679B2 (en) Visualization and quantification of lung disease utilizing image registration
US8811724B2 (en) Classification of medical diagnostic images
JP5159242B2 (en) Diagnosis support device, diagnosis support device control method, and program thereof
EP3528261A1 (en) Prediction model for grouping hepatocellular carcinoma, prediction system thereof, and method for determining hepatocellular carcinoma group
US20180053297A1 (en) Methods and Apparatuses for Detection of Abnormalities in Low-Contrast Images
US20230154620A1 (en) Apparatus and method for assisting reading of chest medical images
US20210077009A1 (en) Predicting response to therapy for adult and pediatric crohn&#39;s disease using radiomic features of mesenteric fat regions on baseline magnetic resonance enterography
JP5456132B2 (en) Diagnosis support device, diagnosis support device control method, and program thereof
CN113017674B (en) EGFR gene mutation detection method and system based on chest CT image
Wu et al. Early identification of ischemic stroke in noncontrast computed tomography
Leichter et al. Improved mammographic interpretation of masses using computer-aided diagnosis
Lin et al. Using deep learning in ultrasound imaging of bicipital peritendinous effusion to grade inflammation severity
Guru et al. Brain tumor identification and classification using a novel extraction method based on adapted alexnet architecture
Lu et al. A Review of the Role of Ultrasound Radiomics and Its Application and Limitations in the Investigation of Thyroid Disease
Zhang et al. Application of artificial intelligence in respiratory medicine
CN116994754A (en) Diabetic pancreatic lesion visual assessment method based on CT (computed tomography) images
Karwoski et al. Processing of CT images for analysis of diffuse lung disease in the lung tissue research consortium
WO2022153100A1 (en) A method for detecting breast cancer using artificial neural network
Keikha et al. Breast Cancer Detection Using Deep Multilayer Neural Networks
Gonçalves et al. Radiomics in head and neck cancer
US10854338B2 (en) Predicting breast cancer responsiveness to hormone treatment using quantitative textural analysis
Wang et al. Diagnosis of Laryngopharyngeal Reflux Disease Based on Gray and Texture Changes of Laryngoscopic Images
Bharathi Neural Network based Earlier Stage Lung Cancer Prediction Scheme with Differential Learning Assistance
Chen et al. Study on Predicting Clinical Stage of Patients with Bronchial Asthma Based on CT Radiomics

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination