CN108765411B

CN108765411B - Tumor typing system based on image omics

Info

Publication number: CN108765411B
Application number: CN201810566664.2A
Authority: CN
Inventors: 栗伟; 于鲲; 冯朝路; 郭志伟; 王东杰; 赵大哲
Original assignee: Northeastern University China
Current assignee: Northeastern University China
Priority date: 2018-06-05
Filing date: 2018-06-05
Publication date: 2021-01-12
Anticipated expiration: 2038-06-05
Also published as: CN108765411A

Abstract

The invention provides a tumor typing system based on imaging omics, which comprises: extracting a focus area for extracting focus data; heterogeneous region extraction, namely extracting all initial heterogeneous regions in the focus data by adopting a clustering method, and then judging a final clustering result according to the connectivity of the cluster on the image and the number of pixels to find out a plurality of heterogeneous regions; extracting parameter vectors, namely extracting the parameters such as the form, the state and the like of each heterogeneous region; and (4) carrying out classification group extraction, namely carrying out cluster classification on each heterogeneous region by adopting a clustering method to obtain a plurality of classes. And (4) new focus classification judgment, which is used for judging the type of each heterogeneous region classification of the new focus. And the new focus typing result is used for outputting the new focus typing result. The tumor typing system has the advantages of being non-invasive, repeatable and the like, and the analysis object of the tumor typing system is based on all lesion tissues, so that the information comprehensiveness is high.

Description

Tumor typing system based on image omics

Technical Field

The invention belongs to the technical field of medical image processing, and particularly relates to a tumor typing system based on image omics.

Background

With the prolonging of the average life span of human beings, the threat of cancer to human beings is increasingly prominent, and the cancer becomes the first cause of death of urban and rural residents in China. Taking breast cancer as an example, the breast cancer refers to malignant tumor generated on mammary duct epithelium, is one of the most common malignant tumors of women, has the morbidity of 23/10 ten thousands, accounts for 7-10% of various malignant tumors of the whole body, has the peak of attack between 45-55 years old, and tends to rise year by year, especially in Shanghai, Jing, jin and coastal areas, which are high-incidence areas of breast cancer in China. The breast cancer screening and early diagnosis system can effectively reduce the mortality rate, and is more benefited by the ever-developing molecular biology technology in recent years and the improvement of the standardized level of comprehensive diagnosis and treatment.

Tumor typing is the basis for diagnosing, judging prognosis, selecting treatment methods and conducting various researches on tumors, and therefore, the establishment of an accurate and effective tumor classification system is key. To date, classification of tumors has been primarily based on histopathological morphological classification (i.e., pathotyping). Malignant tumors are highly heterogeneous at the molecular level, and the molecular genetic changes of tumors with the same morphology are not necessarily consistent, resulting in a large difference in the prognosis and response to treatment of the tumors. Therefore, researchers have also proposed methods such as genotyping, clinical typing, immunohistochemical typing, etc. to more accurately determine the biological behavior of tumors, estimate prognosis, and select or study more targeted personalized therapies. Several methods of tumor typing are common:

pathological typing: breast cancer is classified into non-invasive cancer, early-stage invasive cancer and invasive cancer according to pathological classification. International WHO organization is classified as non-invasive cancer and invasive cancer.

Genotyping: the breast cancer is divided into Luminal A type, Luminal B type, HER-2 overexpression type, Basal-like Basal matrix type and Normal-like cell type by adopting a gene microarray analysis technology.

And (3) clinical typing: the clinically common classifications include hormone receptor positive, HER2/neu receptor positive, triple negative breast cancer (ER, PR, HER2/neu are all negative).

Immunohistochemical typing: the classification into luminal epithelial, HER2 overexpressing and basal patterns was based on ER, PR, HER-2 and Ki-67.

Breast cancer is a highly heterogeneous group of tumors with diverse histological morphologies. Heterogeneous tumor refers to the fact that during the growth process of tumor, after multiple divisions and proliferations, the daughter cells of the tumor show changes in molecular biology or genes, so that the growth rate, invasion capacity, sensitivity to drugs, prognosis and other aspects of the tumor are different. In short, many cells of different genotypes or subtypes may be present in the same tumor. Therefore, the same tumor can show different treatment effects and prognosis in different individuals, and even the tumor cells in the same individual have different characteristics and differences. The heterogeneity of breast cancer determines the clinical pathological characteristics characteristic of each subtype and also determines the prognosis of each subtype. The occurrence and development of breast cancer are heterogeneous not only in the pathological aspect, but also in the aspects of heredity, genetic phenotype and the like, so that the combination of the pathological type and immunohistochemical typing becomes the conventional means for breast cancer diagnosis at present.

Although the classification method has higher guiding value for prognosis judgment of breast cancer and selection of treatment schemes, and gene expression profile and gene chip-based breast cancer genotyping can preliminarily reflect biological behaviors of tumors, certain problems still exist. The presence of breast cancer heterogeneity, which varies greatly in tissue morphology, immunophenotype, biological behavior, and therapeutic response. Clinical pathological diagnosis generally adopts puncture or surgical biopsy, not only causes trauma to the body, but also shows certain limitation, pathological forms in the same area may be different, and differentiation degrees in different areas may be greatly different; similarly, breast cancer heterogeneity may also make the molecular typing and immunohistochemical detection results obtained by gene chip technology not completely consistent. Therefore, it is expected that a universal typing method can be found for clinical guidance, a theoretical basis is provided for solving the heterogeneity of tumors, and a better individualized treatment scheme is provided for patients.

Disclosure of Invention

The tumor typing system has the advantages of being non-invasive, repeatable and the like, and the analysis object of the tumor typing system is based on all lesion tissues, so that the information comprehensiveness is high. The invention comprises the following steps:

(1) a tumor typing system based on image omics is provided, a heterogeneous region extraction method is provided, and an image typing group is constructed;

(2) the image typing group-based typing system for judging new focuses is provided, and can be compared with other typing methods such as existing pathological typing, clinical typing, molecular typing and the like for judgment. The invention provides a comparison of the effect of comparison with molecular typing.

The technical scheme is as follows:

a system for imaging omics based tumor typing comprising:

extracting a focus area for extracting accurate focus data, wherein the process comprises the steps of searching for a seed point and growing the focus area;

heterogeneous region extraction, namely extracting all initial heterogeneous regions in the accurate focus data by adopting a clustering method, and then judging a final clustering result according to the connectivity of the cluster on the image and the number of pixels to find out a plurality of heterogeneous regions;

extracting parameter vectors, namely extracting the parameters such as the form, the state and the like of each heterogeneous region aiming at the extracted heterogeneous region data set to form a parameter vector matrix;

performing typing group extraction, namely performing cluster division on each heterogeneous region by adopting a clustering method aiming at each row of a parameter vector matrix, namely a heterogeneous region, obtaining a plurality of classes after clustering, and recording the classes as typing groups;

and (4) new focus classification judgment, which is used for judging the type of each heterogeneous region classification of the new focus. After the new focus is extracted in a new focus area and a new focus heterogeneous area, judging which type each heterogeneous area of the new focus belongs to according to the typing group extraction result;

and the new focus typing result is used for outputting the new focus typing result, and the sum of each heterogeneous region typing vector of the new focus is the new focus typing result.

The searching for the seed point in the extraction of the focus area comprises the following steps:

(1) image binarization: distinguishing the foreground from the background in the original image;

(2) removing noise points: opening operation is carried out to remove white noise in the image, and closing operation is carried out to remove small holes in the image;

(3) and (3) distance calculation: the distance from the nonzero pixel point in the image to the nearest 0 pixel after the processing of the step (1) and the step (2);

(4) obtaining a mass center: and finding the coordinates of the point corresponding to the maximum distance in all the distances, namely the seed point.

The focal region growing comprises:

seed points to be searched; taking the seed point as a center, considering 8 neighborhood pixels of the seed point, and if the seed point meets the growth criterion, combining the neighborhood pixels and the seed point in the same region; when all the points have attribution, the growth is finished.

The growth criteria include:

criterion 1: judging whether the absolute value of the pixel value difference between the neighborhood pixel and the current pixel is smaller than a difference threshold Ts;

criterion 2: judging whether the pixel value of the neighborhood pixel meets a pixel threshold Tp required by growth;

the heterogeneous region extracting step includes the steps of,

(1) based on the input focus data, finding out pixel coordinates and gray values with gray values different from 0 in the image, and organizing into a three-tuple list;

(2) clustering the triple list by using a clustering algorithm, respectively testing the number of clusters different from 2 to 10, determining the optimal number of clusters by using an elbow rule, and outputting a clustering preliminary result;

(3) judging whether each cluster is connected or not according to the clustering result, and if not, splitting the cluster into a plurality of maximum connected sub-clusters; judging whether the connection is carried out or not according to whether the pixel points in the cluster are adjacent on the plane or whether an adjacent passage exists or not;

(4) and judging the total number of pixels in the cluster according to each connected cluster, if the total number is smaller than a certain check value, discarding the cluster, and if not, reserving the cluster as a final heterogeneous region set.

The parameter vector extraction includes texture parameters, kinetic parameters, statistical parameters, morphological parameters, and clinical parameters.

The classification judgment of the new focus comprises the steps that the heterogeneous region of the new focus and the central points of a plurality of categories obtained by the classification group extraction result take similar functions, the similar functions take Euclidean distances, and the smaller the distance, the higher the possibility that the heterogeneous region of the new focus belongs to the cluster is. Where the center point is the center of gravity calculated from all the heterogeneous region vectors in the cluster.

The classification vector of each heterogeneous region of the new focus is represented as a Boolean vector, the corresponding position belonging to the class is 1, and the other positions are 0, and the classification vector is called as a classification vector.

The beneficial effects are as follows:

the invention provides a novel tumor typing method and an image typing system, which take the comprehensive information into consideration, aim at all lesion tissues, and have the advantages of no wound, no intervention, good repeatability and the like. The accurate classification of some highly heterogeneous tumors, such as breast cancer, is beneficial to guiding clinical treatment, provides a theoretical basis for solving the heterogeneity of the tumors, and provides a better individualized treatment scheme for patients.

Drawings

FIG. 1 is a flow chart of image classification model construction;

FIG. 2 is a process of image classification group construction;

FIG. 3 is a flow chart of lesion area seed point acquisition;

FIG. 4 is a flowchart of focal zone growth;

FIG. 5 is a flow chart of heterogeneous region extraction;

FIG. 6 is a schematic illustration of an elbow rule;

FIG. 7 is a diagram showing the result of heterogeneous region extraction;

FIG. 8 is an exemplary diagram of a partition group division;

FIG. 9 is an example of lesion typing data

Detailed Description

The image classification model construction process of the present invention is shown in FIG. 1. The process is divided into two processes: the image classification group construction process and the new focus classification process, wherein the operations of extracting focus areas and extracting heterogeneous areas are consistent in technology in the two processes.

First, image classification group construction process

The image typing group construction process is shown in FIG. 1 (b). Based on a large amount of breast cancer Region of Interest (ROI) image data, focus Region data a is extracted first, then heterogeneous regions are extracted for the data a, parameters such as texture, power, form and statistics are extracted for each heterogeneous Region to construct a multidimensional vector, a similarity clustering method is adopted to divide the heterogeneous regions into K sets, each cluster set is recorded as an image classification group, and K set centers are recorded as image classification centers, as shown in fig. 2. These four processes are described below, respectively.

1) Focal region extraction

The originally input ROI image data is obtained by identifying a lesion region in an image with the aid of a doctor and then marking the ROI region manually, i.e., marking an approximate region of the lesion with a rectangle. Based on such ROI image data (represented as a data matrix), the center point is first calculated, i.e., data matrix a is as follows:

where m and n represent the height and width of the ROI, respectively, p_ijThe gray value of the pixel point is represented. For each region, a seed point is first designated as a starting point for growth, and the seed point acquires the centroid process in the ROI, as shown in fig. 3.

(1) Image binarization: the foreground and background are distinguished. I.e. all of matrix a are greater than threshold T₀All of the pixels of (a) are 255,

less than T₀Is all 0. T0 averages the data in a.

(2) Removing noise points: and the open operation removes white noise in the image, and the closed operation removes small holes in the image.

(3) And (3) distance calculation: the distance from the nonzero pixel point in the image A to the nearest 0 pixel after the processing of the step 1 and the step 2

(4) Obtaining a mass center: the coordinates of the point corresponding to the maximum distance, i.e. the centroid of the ROI, i.e. the seed point we find, are found among all distances.

Based on the seed points, region growing is performed to obtain an accurate lesion region, and the process is shown in fig. 4.

The focal region growing steps are as follows:

(1) the seed point found is set to (x)₀，y₀) Establishing a mask matrix (mask) having a size consistent with the ROI, and dividing by (x)₀，y₀) Except for the dots, all the others are set to 0, mask [ x ]₀，y₀]＝1；

(2) With (x)₀，y₀) As a center, consider (x)₀，y₀) If (x) is (x, y) of 8 neighboring pixels₀，y₀) Satisfying growth criteria (2.1 and 2.2 criteria), and (x, y) is compared with (x)₀，y₀) Merge (in the same region) while mask [ x, y ]]1 and pressing the point into the stack;

-2.1 criteria: judging whether the absolute value of the pixel value difference between the neighborhood pixel and the current pixel is smaller than a difference threshold Ts, and taking Ts as 20 in the implementation process;

-2.1 criteria: judging whether the pixel value of the neighborhood pixel meets a pixel threshold Tp required by growth or not, wherein the implementation process Tp is pixel average-20;

(3) taking a pixel from the stack and treating it as (x)₀，y₀) Returning to the step (2);

(4) when the stack is empty, all points in the ROI have attribution, and the growth is finished

Based on the obtained mask matrix, the mask and the data matrix A are made

Calculating to obtain a focus data matrix A_lesionNamely:

wherein

The operator represents: all masPoint of zero in k, A_lesionThe middle corresponding position is also set to be zero, and the non-zero point corresponds to A_lesionThe middle position data is corresponding position data in A.

2) Heterogeneous region extraction

Based on extracted focus data A_lesionAll initial heterogeneous regions in the lesion are extracted by a clustering method (such as K-means), and then a final clustering result is determined according to the connectivity of the clusters on the image and the number of pixels, wherein the extraction process is shown in FIG. 5.

The heterogeneous region extraction step is as follows:

(1) based on input lesion data A_lesionFinding out pixel coordinates and gray values with gray values different from 0 in the image, and organizing a triple list L;

(2) clustering the data L by using a clustering algorithm (such as K-means), respectively testing the number of different clusters from 2 to 10, and determining the optimal number K of clusters by using an elbow rule_best(principle shown in fig. 5); and outputs the clustering preliminary result, as shown in fig. 7 (a); fig. 5 is a schematic diagram of an elbow rule, where the abscissa is a K value and the ordinate is a clustering objective function (or cost function). The abscissa of the point with the largest change in the front and rear slopes in the image is selected as the optimal K value, and the K value is obtained as 3 in the figure.

(3) Judging whether each cluster is communicated or not according to the clustering result, and if not, dividing the cluster into a plurality of maximum communicated sub-clusters; judging whether the connection is carried out or not according to whether the pixel points in the cluster are adjacent on the plane or whether an adjacent passage exists or not;

(4) and (3) aiming at each connected cluster, judging the total number of pixels in the cluster, if the total number is less than 10 (an empirical value can be adjusted according to actual disease species), discarding the cluster, and if not, reserving the cluster as a heterogeneous region in the final heterogeneous region set phi

As shown in fig. 7 (b).

3) Parameter vector extraction

Extracting each hetero for the extracted hetero region data set phiZone(s)

The form, shape, etc. of (A), wherein

The parameters extracted by the method comprise texture parameters, kinetic parameters, statistical parameters, morphological parameters and clinical parameters, and the parameter calculation method is shown in table 1.

TABLE 1 heterogeneous region parameter calculation method

After all the parameters are calculated, the sum of the dimensions of the parameters totals 62 dimensions, the number of corresponding parameters extracted according to different tumor types is also different, the obtained dimensions are also different, and the obtained dimension is marked as D (D is 62). Assuming that the number of the heterogeneous regions is H, extracting a parameter vector omega from each heterogeneous region, wherein the extracted parameter vectors form a matrix omega with the size of D x H, namely:

4) fractional group extraction

Each line of the parameter vector heterogeneous region data omega represents a heterogeneous region, and the clustering method is adopted to perform cluster division on each heterogeneous region, so thatThe method of clustering uses the K-means (where K is an empirical value). Obtaining K categories after clustering, wherein the central vector of each category is as follows: y ═ C₁，C₂，...，C_K) As shown in fig. 8:

each center vector here is the center of gravity calculated from all the heterogeneous region vectors in the cluster.

Second, the new lesion typing process

The decision to which grouping a new heterogeneous region (parameter vector v) belongs can be calculated as follows. And taking the class with the maximum similarity as the class group of the heterogeneous region.

Wherein: k is more than or equal to 1 and less than or equal to K

Sim is a similarity function measuring v and each central point, and generally, the euclidean distance is taken as the distance, and the smaller the distance, the more likely the heterogeneous region belongs to the cluster.

Based on the above calculation method, the type of each heterogeneous region represents a boolean vector S:

S＝(0₁，...，1_k′，...，O_K)

that is, the corresponding positions belonging to the k' class are 1, and the others are 0, and are called as the typing vectors. A lesion ROI area has a plurality of heterogeneous regions, and each heterogeneous region can be calculated to obtain a typing vector. Assuming that there are t heterogeneous regions in a lesion ROI region, the lesion typing vector is the sum of the heterogeneous region typing vectors, i.e.:

third, method verification

The invention verifies 636 breast cancer ROI data in the size of a data set, and the data set is divided according to molecular typing: 162 cases of LA type, 162 cases of LB type, 153 cases of HER-2 type, 159 cases of Basic-like type, which have been confirmed by biopsy pathology; these data were divided into two sets, one set was training set 477 cases, and the remaining set was testing set 159 cases, and the extracted typing data for each lesion is shown in fig. 9, where k is 10.

GBDT, XGBboost and MLP data classification models were selected, molecular typing was used as classification labels, image typing vectors were used as model inputs, and prediction accuracy was verified by cross-testing, as shown in table 2.

TABLE 2 prediction accuracy

Serial number	Model (model)	Training set accuracy	Test set accuracy
				1	GBDT	72.2％	83.13％
2	XGBboost	71.35	81.23％
				3	MLP	74.33％	82.5％

From the above test results, the image typing and molecular typing achieved a maximum accuracy of 83.13% control, and 16.87% of the error causes are from two aspects: firstly, the classification model has errors; the second is that molecular typing is precise classification (one of four types) which depends on the type of local biopsy tissue, while image typing is the typing classification of the whole focus, is not precise classification, and is judged by integrating the actual characteristics of different heterogeneous regions, so the second type of error reasons cannot explain that the analysis result of the image typing has problems.

In conclusion, the image typing system has the advantages of being noninvasive, non-invasive and good in repeatability, an analysis object is based on all lesion tissues and is not limited to local parts (such as puncture, gene section and the like), the information comprehensiveness is high, and the image typing system is a new and advanced tumor typing method.

Claims

1. A system for imaging omics based tumor typing, comprising:

heterogeneous region extraction, namely extracting all initial heterogeneous regions in the accurate focus data by adopting a clustering method, and then judging a final clustering result according to the connectivity of the clusters on the image and the number of pixels to find out a plurality of heterogeneous regions;

extracting parameter vectors, namely extracting the form and state parameters of each heterogeneous region aiming at the extracted heterogeneous region data set to form a parameter vector matrix;

the new focus classification judgment is used for judging the type of each heterogeneous region classification of the new focus; after the new focus is extracted in a new focus area and a new focus heterogeneous area, judging which type each heterogeneous area of the new focus belongs to according to the typing group extraction result; and the new focus typing result is used for outputting the new focus typing result, and the sum of each heterogeneous region typing vector of the new focus is the new focus typing result.

2. The system of claim 1, wherein the finding of seed points in lesion region extraction comprises:

3. The system of claim 2, wherein the focal region growing comprises:

4. The system of claim 3, wherein the growth criteria comprises:

criterion 2: and judging whether the pixel value of the neighborhood pixels meets the pixel threshold Tp required by growth.

5. The system of claim 1, wherein the heterogeneous region extracting step comprises,

(3) judging whether each cluster is connected or not according to the clustering preliminary result, and if not, splitting the cluster into a plurality of maximum connected sub-clusters; judging whether the connection is carried out or not according to whether the pixel points in the cluster are adjacent on the plane or whether an adjacent passage exists or not;

6. The system of claim 1, wherein the parameter vector extraction comprises texture parameters, kinetic parameters, statistical parameters, morphological parameters, and clinical parameters.

7. The system of claim 1, wherein the classification of the new lesion comprises a heterogeneous region of the new lesion and a central point of a plurality of classes obtained from the classification group extraction result, wherein the similar function is a euclidean distance, and a smaller distance indicates that the heterogeneous region of the new lesion is more likely to belong to the class of clusters; where the center point is the center of gravity calculated from all the heterogeneous region vectors in the cluster.

8. The system of claim 1, wherein the new lesion is classified into each heterogeneous region according to a classification vector, the classification result of each heterogeneous region of the new lesion is represented as a boolean vector, the corresponding position belonging to the class is 1, and the others are 0, and the classification vector is referred to as a classification vector.