CN108765411A - A kind of tumor classification method based on image group - Google Patents

Abstract

The present invention provides a kind of tumor classification method based on image group, including：Focal area is extracted, for extracting lesion data；Heterogeneous areas is extracted, and initial foreign region all in the lesion data is extracted using clustering method, is then judged last cluster result according to class cluster connectivity on the image and number of pixels, is found out multiple heterogeneous areas；Parameter vector extracts, and extracts the parameters such as form, the state in each heterogeneous area；Parting group is extracted, and carrying out class cluster to each heterogeneous area using clustering method divides to obtain multiple classifications.New lesion parting judgement, for judging that each heterogeneous differentiation type of new lesion belongs to which kind.New lesion genotyping result, for exporting new lesion genotyping result.The present invention tumor classification method, have many advantages, such as it is non-invasive, non-intervention, repeatable, and its analyze object be based on whole lesion tissues, the comprehensive height of information.

Description

A tumor classification method based on radiomics

technical field

The invention belongs to the technical field of medical image processing, and in particular relates to a radiomics-based tumor typing method.

Background technique

With the prolongation of the average life expectancy of human beings, the threat of cancer to human beings has become increasingly prominent, and it has become the first cause of death of urban and rural residents in our country. Taking breast cancer as an example, breast cancer refers to malignant tumors occurring in breast ductal epithelium. It is one of the most common malignant tumors in women, with an incidence rate of 23/100,000, accounting for 7-10% of all kinds of malignant tumors in the whole body. Between the ages of 45 and 55, and showing an increasing trend year by year, especially Shanghai, Beijing, Tianjin and the coastal areas are the high incidence areas of breast cancer in my country. The establishment of breast cancer screening and early diagnosis system can effectively reduce the mortality rate, and it also benefits from the continuous development of molecular biology technology and the improvement of the standardization level of comprehensive diagnosis and treatment in recent years.

Tumor classification is the basis for diagnosing tumors, judging prognosis, choosing treatment methods, and conducting various researches. Therefore, establishing an accurate and effective tumor classification system is the key. So far, tumor classification is mainly morphological classification based on histopathology (ie: pathological typing). Malignant tumors are highly heterogeneous at the molecular level, and the molecular genetic changes of tumors with the same morphology may not be consistent, resulting in great differences in the prognosis and response to treatment of tumors. Therefore, researchers have also proposed methods such as genotyping, clinical typing, and immunohistochemical typing in order to more accurately judge the biological behavior of tumors, estimate prognosis, and select or study more targeted personalized treatment methods. . Several common tumor classification methods:

Pathological classification: According to pathological classification, breast cancer is divided into non-invasive carcinoma, early invasive carcinoma, and invasive carcinoma. International WHO classification into non-invasive cancer and invasive cancer.

Genotyping: Breast cancer was divided into Luminal A type, Luminal B type, HER-2 overexpression type, Basal-like basal-like type and Normal-like normal cell-like type by gene microarray analysis technology.

Clinical classification: Common clinical classifications include hormone receptor positive, HER2/neu receptor positive, triple negative breast cancer (ER, PR, HER2/neu are all negative).

Immunohistochemical typing: According to ER, PR, HER-2 and Ki-67, it was classified into luminal epithelial type, HER2 overexpression type and basal-like type.

Breast cancer is a highly heterogeneous tumor with various histological forms. Heterogeneous tumors refer to tumors that have undergone multiple divisions and proliferations during the growth process, and their daughter cells exhibit molecular biological or genetic changes, which can lead to changes in tumor growth rate, invasion ability, drug sensitivity, prognosis, etc. There are differences in all aspects. Simply put, there can be many different genotypes or subtypes of cells in the same tumor. Therefore, the same tumor can show different therapeutic effects and prognosis in different individuals, and even tumor cells in the same individual have different characteristics and differences. The heterogeneity of breast cancer determines the unique clinicopathological characteristics of each subtype and also determines the prognosis of each subtype. The occurrence and development of breast cancer is heterogeneous not only in pathology, but also in genetics and gene phenotype. Therefore, pathological type combined with immunohistochemical typing has become a routine method for breast cancer diagnosis.

Although the above classification methods have a high guiding value for the prognosis of breast cancer and the selection of treatment options, and the gene expression profile and gene microarray-based breast cancer genotyping can initially reflect the biological behavior of the tumor, but there are still There must be a problem. The existence of breast cancer heterogeneity, there are great differences in histological morphology, immunophenotype, biological behavior and treatment response. Clinicopathological diagnosis generally uses puncture or surgical biopsy, which not only causes trauma to the body, but also shows certain limitations. The pathological morphology of the same region may be different, and the degree of differentiation may vary greatly in different regions; similarly, the heterogeneity of breast cancer It may also make the molecular typing and immunohistochemical detection results obtained by gene chip technology not completely consistent. Therefore, it is of great significance to find a general classification method to guide clinical practice, provide a theoretical basis for solving tumor heterogeneity, and provide better individualized treatment options for patients.

Contents of the invention

Aiming at the problems existing in the prior art, the present invention provides a radiomics-based tumor typing method, which is called an imaging typing method. The tumor typing method of the present invention has the advantages of non-invasiveness, non-intervention, repeatability, etc., and The analysis object is based on all lesion tissues, and the information is comprehensive. this invention:

(1) A radiomics-based tumor classification method was proposed, a heterogeneous region extraction method was proposed, and an image classification group was constructed;

(2) A classification method for judging new lesions based on the image classification group is proposed, which can be compared with other classification methods such as the existing pathological classification, clinical classification, and molecular classification. The present invention provides a comparison of effects compared with molecular typing.

The technical solution is as follows:

A radiomics-based approach to tumor typing that includes:

Lesion area extraction, used to extract accurate lesion data, the process includes finding seed points and lesion area growth;

Heterogeneous region extraction, using the clustering method to extract all the initial heterogeneous regions in the precise lesion data, and then judging the final clustering results according to the connectivity and number of pixels of the clusters on the image, to find out multiple heterogeneous regions area;

Parameter vector extraction, for the extracted heterogeneous area data set, extract the shape, state and other parameters of each heterogeneous area to form a parameter vector matrix;

Typing group extraction, for each row of the parameter vector matrix, that is, a heterogeneous area, the clustering method is used to divide each heterogeneous area into clusters, and multiple categories are obtained after clustering, which are recorded as typing groups;

The new lesion type determination is used to determine which type each heterogeneous type of the new lesion belongs to. After the new lesion area extraction and the new lesion heterogeneous area extraction, according to the classification group extraction results, determine the type of each new lesion. Which type of heterogeneous area belongs to;

The typing result of the new lesion is used to output the typing result of the new lesion, and the sum of each heterogeneous distinguishing vector of the new lesion is the typing result of the new lesion.

Finding the seed point in the lesion area extraction includes:

(1) Image binarization: distinguish the foreground and background in the original image;

(2) Remove noise points: open operation to remove white noise in the image, and close operation to remove small holes in the image;

(3) distance calculation: the distance from the non-zero pixel in the image to the nearest 0 pixel after step (1) and step (2);

(4) Acquisition of centroid: Find the coordinates of the point corresponding to the maximum distance among all distances, that is, the seed point.

The growth of the lesion area includes:

The seed point to find; take the seed point as the center, consider the 8 neighboring pixels of the seed point, if the seed point satisfies the growth criterion, merge the neighboring pixels with the seed point in the same area; when all the points belong to, the growth ends.

The growth guidelines include:

Criterion 1: Determine whether the absolute value of the pixel value difference between the neighboring pixel and the current pixel is less than the difference threshold Ts;

Criterion 2: Judging whether the pixel value of the neighboring pixels meets the pixel threshold Tp required for growth;

The heterogeneous region extraction step includes,

(1) Based on the input lesion data, find pixel coordinates and gray values with non-zero gray values in the image, and organize them into a list of triplets;

(2) Use the clustering algorithm to cluster the list of triples, test the number of clusters from 2 to 10 respectively, use the elbow rule to determine the optimal number of clusters, and output the preliminary clustering results;

(3) According to the above clustering results, judge whether each cluster is connected. If not, split the cluster into several largest connected sub-clusters; whether the connection is judged is based on whether the pixels in the cluster are adjacent on the plane. Or there is an adjacent path;

(4) For each of the above-mentioned connected clusters, judge the total number of pixels in the cluster. If the total number is less than a certain empirical value, discard the cluster; otherwise, keep it as the final set of heterogeneous regions.

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

The classification determination of the new lesion includes taking a similarity function between the heterogeneous area of the new lesion and the center points of multiple categories obtained from the extraction result of the typing group. The more likely the heterogeneous area belongs to the cluster, the center point here is the center of gravity calculated based on the vectors of all heterogeneous areas in the cluster.

The typing vector of each heterogeneous area of the new lesion is , the typing result of each heterogeneous area of the new lesion is expressed as a Boolean vector, the corresponding position belonging to this class is 1, and the other is 0, which is called the typing vector.

Its beneficial effect is:

The present invention proposes a new tumor classification method, which considers comprehensive information, targets all lesion tissues, and has the advantages of non-invasive, non-intervention, and good repeatability. For some highly heterogeneous tumors, such as breast cancer, its accurate classification is conducive to guiding clinical practice, providing a theoretical basis for solving tumor heterogeneity, and providing better individualized treatment options for patients.

Description of drawings

Figure 1 is the flow chart of image typing model construction, (a) determination of new lesion typing, (b) image typing group construction

Figure 2 is the construction process of the image typing group;

Fig. 3 is the flow chart of obtaining the seed point of the lesion area;

Figure 4 is a flowchart of the growth of the lesion area;

Fig. 5 is a flowchart of heterogeneous region extraction;

Figure 6 is a schematic diagram of the elbow rule;

Figure 7 is a diagram of the heterogeneous region extraction results, (a) the preliminary results are obtained based on the data A _lesion clustering, (b) the cluster connectivity judgment is divided into multiple sub-regions;

Fig. 8 is an example figure of classification group division;

Figure 9 is an example of lesion typing data.

Detailed ways

The construction process of the image classification model of the present invention is shown in Fig. 1 . This process is divided into two processes: the process of image typing group construction and the process of new lesion typing, in which the operations of extracting lesion areas and extracting heterogeneous areas are the same in the above two processes.

1. The process of image typing group construction

The construction process of the image typing group is shown in Figure 1(b). Based on a large number of breast cancer ROI (Region of Interest) image data, first extract the lesion area data A, and then extract heterogeneous areas for data A, extract texture, dynamics, shape, statistics and other parameters for each heterogeneous area to construct multi-dimensional The similarity clustering method is used to divide these heterogeneous regions into K sets, each set of clusters is recorded as an image typing group, and the center of K sets is recorded as an image typing center, as shown in Figure 2. The four processes are described below.

1) Lesion area extraction

The original input ROI image data is obtained by identifying the lesion area in the image with the assistance of a doctor, and then manually marking the ROI area, that is, marking the approximate area of the lesion with a rectangle. Based on such ROI image data (expressed as a data matrix), first calculate the center point, that is, the data matrix A is as follows:

Among them, m and n represent the height and width of the ROI respectively, and p _ij represents the gray value of the pixel. For each region, a seed point must be designated as the starting point of growth, and the seed point obtains the centroid process in the ROI, as shown in Figure 3.

(1) Image binarization: distinguish the foreground from the background. That is, all pixels greater than the threshold T ₀ in matrix A are 255, and all pixels smaller than T ₀ are 0. T0 takes the average value of the data in A.

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

(3) Distance calculation: the distance from the non-zero pixel in image A to the nearest 0 pixel after step 1 and step 2

(4) Acquisition of the centroid: find the coordinates of the point corresponding to the maximum distance among all the distances, that is, the centroid of the ROI, which is the seed point we are looking for.

Based on the above seed points, region growing is carried out to obtain precise lesion regions. The process is shown in Figure 4.

The steps of lesion area growth are as follows:

(1) The searched seed point is set to (x ₀ , y ₀ ), and the mask matrix (mask) is established. The size of the matrix and the ROI are the same. Except for the point (x ₀ , y ₀ ), all others are set to 0, and the mask[ x ₀ ,y ₀ ]=1;

(2) With (x ₀ , y ₀ ) as the center, consider the 8 neighboring pixels (x, y) of (x ₀ , y ₀ ), if (x ₀ , y ₀ ) satisfies the growth criterion (2.1 and 2.2 criteria) , merge (x,y) with (x ₀ ,y ₀ ) (in the same area), and mask[x,y]=1 and push the point into the stack;

-2.1 Criterion: determine whether the absolute value of the pixel value difference between the neighboring pixel and the current pixel is less than the difference threshold Ts, and take Ts=20 in the implementation process;

– 2.1 Criterion: Judging whether the pixel value of neighboring pixels meets the pixel threshold Tp required for growth, the implementation process Tp=pixel average value-20;

(3) Take a pixel from the stack, treat it as (x ₀ , y ₀ ) and return to step (2);

(4) When the stack is empty, it means that all the points in the ROI have ownership, and the growth ends

Based on the acquired mask matrix mask, make mask and data matrix A The lesion data matrix A _lesion is obtained by operation, namely:

in The operator means: for all points that are zero in the mask, the corresponding position in A _lesion is also set to zero, and the position data in A _lesion corresponding to non-zero points is the corresponding position data in A.

2) Heterogeneous region extraction

Based on the extracted lesion data A _lesion , use a clustering method (such as K-means) to extract all the initial heterogeneous regions in the lesion, and then determine the final clustering result according to the connectivity of the clusters on the image and the number of pixels , the extraction process is shown in Figure 5.

The heterogeneous region extraction steps are as follows:

(1) Based on the input lesion data A _lesion , find pixel coordinates and gray values with non-zero gray values in the image, and organize them into a triplet list L;

(2) Use a clustering algorithm (such as K-means) to cluster the data L, test the number of different clusters from 2 to 10, and use the elbow rule to determine the optimal number of clusters K _best (as shown in 6); and output the preliminary results of clustering, as shown in Figure 7(a), the preliminary results are obtained based on the data A _lesion clustering, where the optimal value of K is 2; among them, the schematic diagram of the elbow rule in Figure 6 , where the abscissa is the K value, and the ordinate is the clustering objective function (or cost function). Select the abscissa of the point with the largest front and rear slope changes in the image as the optimal K value, as shown in the figure, K=3 is the desired value.

(3) According to the above clustering results, it is judged whether each cluster is connected. If it is not connected, the cluster is split into several largest connected sub-clusters; whether the connection is judged is based on whether the pixels in the cluster are adjacent on the plane. Or there is an adjacent path;

(4) For each of the above connected clusters, determine the total number of pixels in the cluster, and if the total number is less than 10 (experience value, can be adjusted according to the actual disease type), discard the cluster, otherwise keep it as the final set of heterogeneous regions A heterogeneous region in Φ As shown in Figure 7(b), the cluster connectivity judgment is divided into multiple sub-regions, and the clusters with less than 10 pixels are deleted. The number of clusters here is 4, that is, the A _lesion matrix is converted into 4 data matrices

3) Parameter vector extraction

For the extracted heterogeneous area data set Φ, extract each heterogeneous area parameters such as shape and shape, among which The parameters extracted in this application include texture parameters, dynamic parameters, statistical parameters, morphological parameters and clinical parameters. The calculation methods of the above parameters are shown in Table 1.

Table 1. Calculation method of heterogeneous region parameters

After the calculation of all the above parameters is completed, the sum of the dimensions is 62 dimensions. The number of corresponding parameters extracted according to different tumor types is also different, and the obtained dimension is also different, which is recorded as D (D=62). Assuming that the number of heterogeneous regions is H, a parameter vector ω is extracted from each heterogeneous region, and the size of the extracted parameter vector matrix Ω is D*H, namely:

4) Typing group extraction

Each row of the heterogeneous area data Ω of the parameter vector above represents a heterogeneous area. This application uses a clustering method to divide each heterogeneous area into clusters. Here, the clustering method uses K-means (where K is an empirical value). K categories are obtained after clustering, and the center vector of each category is: Y=(C ₁ , C ₂ ,..., C _K ), as shown in Figure 8:

Each center vector here is the center of gravity calculated according to all heterogeneous region vectors in the cluster.

2. The process of typing new lesions

Judging which typing group a new heterogeneous area (parameter vector v) belongs to can be calculated by the following formula. The category with the largest similarity is taken as the category group of the heterogeneous area.

Where: 1≤k≤K

Sim is a measure of the similarity function between v and each center point. Usually, the Euclidean distance is used here. The smaller the distance, the greater the possibility that the heterogeneous area belongs to the cluster.

Based on the above calculation method, the typing type of each heterogeneous area represents a Boolean vector S:

S=(0 ₁ ,...,1 _k ′,...,0 _K )

That is, the corresponding position belonging to the k' class is 1, and the others are 0, which is called the classification vector. A lesion ROI region has multiple 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, then the lesion typing vector is the sum of all heterogeneous distinguishing vectors, namely:

Three, method validation

The verification data set size of the present invention is 636 breast cancer ROI data, which are divided according to molecular types: 162 cases of LA type, 162 cases of LB type, 153 cases of HER-2 type, and 159 cases of Basic-like type. These types have been verified by biopsy pathology Divide these data into two sets, one part is the training set of 477 cases, and the rest is used as the test set of 159 cases. The typing data extracted from each lesion is shown in Figure 9, and k=10 is taken here.

The GBDT, XGBboost and MLP data classification models were selected, the molecular typing was used as the classification label, the image typing vector was used as the model input, and the prediction accuracy was verified by cross-testing, as shown in Table 2.

Table 2 Prediction accuracy

serial number 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%

According to the above test results, the division of image typing and molecular typing reached the highest comparison accuracy of 83.13%, and the 16.87% error came from two aspects: one is that there is an error in the classification model itself; the other is that molecular typing is an accurate division (One of the four categories), depending on the type of local biopsy tissue, and the imaging classification is based on the classification of the overall lesion, which is imprecise and determined based on the actual characteristics of different heterogeneous areas. Therefore, the second type of error does not It cannot explain that there is a problem with the analysis results of image typing.

In summary, the image typing method of the present invention has the advantages of non-invasive, non-interventional, and good repeatability, and the analysis object is based on the whole lesion tissue, not limited to the local area (such as puncture, gene slice, etc.), and the information is comprehensive. It is a new and advanced tumor classification method.

Claims (8)

1. a kind of tumor classification method based on image group, which is characterized in that including：

Focal area is extracted, and for extracting accurately lesion data, flow includes finding seed point and focal area growth；

Heterogeneous areas is extracted, and initial foreign region all in the accurate lesion data is extracted using clustering method, then root Judge last cluster result according to class cluster connectivity on the image and number of pixels, finds out multiple heterogeneous areas；

Parameter vector extracts, and for the heterogeneous areas data set of extraction, extracts the parameters such as form, the state in each heterogeneous area, Composition parameter vector matrix；

Parting group is extracted, heterogeneous to each using clustering method for every a line of parameter vector matrix, i.e. a heterogeneous area Area carries out class cluster division, obtains multiple classifications after cluster, is denoted as parting group；

New lesion parting judgement, for judging that each heterogeneous differentiation type of new lesion belongs to which kind, new lesion is by new disease Stove extracted region is extracted according to parting group as a result, judging that each heterogeneous area belongs to new lesion with after the extraction of new lesion heterogeneous areas Any class parting；

New lesion genotyping result, for exporting new lesion genotyping result, new lesion each heterogeneous differentiation type vector and be new disease The genotyping result of stove.

2. a kind of tumor classification method based on image group according to claim 1, which is characterized in that the focal area Seed point is found in extraction includes：

(1) image binaryzation：Foreground and background in original image is distinguished；

(2) noise spot is removed：Opening operation removes the white noise in image, and closed operation removes the small cavity in image；

(3) distance calculates：Distance of the non-zero pixels point to 0 nearest pixel in image after step (1) and step (2) processing；

(4) barycenter obtains：The coordinate of the point corresponding to maximum distance, i.e. seed point are found in all distances.

3. a kind of tumor classification method based on image group according to claim 2, which is characterized in that the focal area Growth includes：

The seed point of searching；Centered on seed point, 8 neighborhood territory pixels of seed point are considered, if seed point meets growth criterion, Neighborhood territory pixel is merged with the seed point in the same area；When point has ownership, growth to terminate.

4. a kind of tumor classification method based on image group according to claim 3, which is characterized in that the growth criterion Including：

Criterion 1：Judge whether the pixel value absolute value of the difference of neighborhood territory pixel and current pixel is less than difference threshold Ts；

Criterion 2：Judge whether the pixel value size of neighborhood territory pixel meets the pixel threshold Tp needed for growth.

5. a kind of tumor classification method based on image group according to claim 1, which is characterized in that the heterogeneous areas Extraction step includes,

(1) input lesion data are based on, gray value in image non-zero pixel coordinate and gray value is found, is organized into a ternary Groups List；

(2) ternary Groups List is clustered using clustering algorithm, is tested respectively from 2~10 different class cluster numbers, utilizes elbow Portion's rule determines optimal class cluster number, and exports cluster PRELIMINARY RESULTS；

(3) judge whether be connected in each class cluster for above-mentioned cluster result, be split into such cluster if be not connected to Several maximum connection subclass clusters；Judge whether connection whether is abutted in the plane according to pixel in class cluster or in the presence of neighbour Connect road；

(4) it is directed to the class cluster of each above-mentioned connection, judges sum of all pixels in class cluster, if sum is less than a certain empirical value, just Such cluster is abandoned, is otherwise retained, is gathered as final heterogeneous area.

6. a kind of tumor classification method based on image group according to claim 1, which is characterized in that the parameter vector Extraction includes parametric texture, kinetic parameter, statistical parameter, morphological parameters and clinical parameter.

7. a kind of tumor classification method based on image group according to claim 1, which is characterized in that the new lesion Parting judgement includes the heterogeneous areas of new lesion, and the central point that multiple classifications that result obtains are extracted to parting group takes similar letter Number, similar function takes Euclidean distance here, and the possibility that such cluster is belonged to apart from the heterogeneous area of the smaller expression new lesion is bigger； Here central point is according to the center of gravity that all heterogeneous area's vectors calculate in such cluster.

8. a kind of tumor classification method based on image group according to claim 1, which is characterized in that the new lesion is every A heterogeneous differentiation type vector is, the genotyping result in each heterogeneous area of new lesion is expressed as a boolean vector, belongs to such pair It is 1 to answer position, other are 0, referred to as parting vector.