A method and system for subgroup analysis of medical image features

Abstract

The present invention relates to a method and system for subgroup analysis of medical image features. First, observation data corresponding to each of a plurality of medical image samples is obtained, and the observation data includes first-class image features, second-class image features, and a response variable. Then, using the observation data corresponding to all samples as the input of an objective function, with the goal of minimizing the objective function, a first regression coefficient set for the first-level subgroup and a second regression coefficient set for the second-level sub-subgroup are calculated. Finally, subgroup analysis results for all samples are obtained based on the first regression coefficient set and the second regression coefficient set, and thus a nested hierarchical structure including sub-subgroups can be generated, classifying the samples at multiple levels, providing more similarity information, and assisting in the image diagnosis of histopathology.

Description

Technical Field

[0001] The present invention relates to the technical field of subgroup analysis, and particularly to a method and system for subgroup analysis of medical image features. Background Art

[0002] In the image analysis of histopathology, some samples have a certain subgroup structure. Different subgroups are significantly different in their effects. Therefore, different samples have differences in characteristics, states, and distributions, forming heterogeneity and no longer meeting the condition of the same distribution, which brings difficulties to image diagnosis. The subgroup analysis method can overcome this difficulty. The subgroup analysis method can be mainly divided into two categories: supervised subgroup analysis and unsupervised subgroup analysis. When there is no response variable y, it is unsupervised subgroup analysis, which mainly performs subgroup analysis based on clustering and can be subdivided into three types: distance-based clustering, density-based clustering, and model-based clustering. According to different types of the response variable y, the supervised subgroup analysis can be divided into two categories. When y is only used to label the subgroup category to which the sample belongs, it is the joint graph model estimation. When y is the regression response variable, it is the subgroup regression analysis.

[0003] However, the existing subgroup regression analysis methods all optimize the regression coefficients and impose penalties on the regression coefficients, and the obtained result is to divide the samples with the same regression coefficients into the same subgroup. The obtained result is a single structure, that is, all samples are divided into several subgroups, and a hierarchical subgroup structure with sub-subgroups cannot be generated. Summary of the Invention

[0004] The object of the present invention is to provide a method and system for subgroup analysis of medical image features, which can generate a nested hierarchical structure including sub-subgroups, classify samples at multiple levels, provide more similarity information, and assist in the image diagnosis of histopathology.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A method for subgroup analysis of medical image features, the subgroup analysis method comprising:

[0007] Obtaining the observation data corresponding to each of a plurality of medical image samples; the observation data includes the first type of image features, the second type of image features, and a response variable;

[0008] Using the observation data corresponding to all the samples as the input of an objective function, and aiming at minimizing the objective function, calculating to obtain a first regression coefficient set of the first-level subgroups and a second regression coefficient set of the second-level sub-subgroups; the first regression coefficient set includes the first regression coefficient corresponding to each of the samples; the second regression coefficient set includes the second regression coefficient corresponding to each of the samples;

[0009] Obtain the subgroup analysis results of all the samples based on the first set of regression coefficients and the second set of regression coefficients.

[0010] A medical image feature subgroup analysis system, the subgroup analysis system includes:

[0011] An observation data acquisition module, configured to acquire the observation data corresponding to each of the multiple medical image samples; the observation data includes first-class image features, second-class image features, and a response variable;

[0012] A regression coefficient calculation module, configured to use the observation data corresponding to all the samples as the input of an objective function, and take minimizing the objective function as the goal, to calculate a first set of regression coefficients for the first-level subgroups and a second set of regression coefficients for the second-level sub-subgroups; the first set of regression coefficients includes the first regression coefficient corresponding to each of the samples; the second set of regression coefficients includes the second regression coefficient corresponding to each of the samples;

[0013] A subgroup analysis module, configured to obtain the subgroup analysis results of all the samples based on the first set of regression coefficients and the second set of regression coefficients.

[0014] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0015] The present invention provides a medical image feature subgroup analysis method and system. First, acquire the observation data corresponding to each of the multiple medical image samples, where the observation data includes first-class image features, second-class image features, and a response variable. Then, use the observation data corresponding to all the samples as the input of an objective function, and take minimizing the objective function as the goal, to calculate a first set of regression coefficients for the first-level subgroups and a second set of regression coefficients for the second-level sub-subgroups. Finally, obtain the subgroup analysis results of all the samples based on the first set of regression coefficients and the second set of regression coefficients, and thus can generate a nested hierarchical structure including sub-subgroups, classify the samples at multiple levels, provide more similarity information, and assist in the image diagnosis of histopathology. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0017] Figure 1 It is a flowchart of the subgroup analysis method provided in Embodiment 1 of the present invention;

[0018] Figure 2 It is the flowchart of the feature extraction step provided by Embodiment 1 of the present invention;

[0019] Figure 3 It is the system block diagram of the subgroup analysis system provided by Embodiment 2 of the present invention. Detailed implementation manners

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0021] The object of the present invention is to provide a medical image feature subgroup analysis method and system with a hierarchical structure, which can generate a nested hierarchical structure including sub-subgroups, classify samples at multiple levels, provide more similarity information, and assist in the image diagnosis of histopathology.

[0022] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation manners.

[0023] Embodiment 1:

[0024] The existing subgroup regression analysis method mainly uses a mixture regression model. Given the independent variable x, the response variable y follows a certain mixture distribution f(y; x), as follows:

[0025]

[0026] where K 0 is the order of the mixture distribution; π k is the proportion of the mixture distribution, satisfying non-negativity and normalization; f(·) is the density function of a class of distributions; θ k (x) = h(x T β k ), h(·) is the link function, β k is the regression coefficient; φ k is the dispersion coefficient.

[0027] In the case of obtaining the observed samples (x 1 , y 1 )......(x n , y n ), optimization is to maximize the following objective function to obtain the result:

[0028]

[0029] Subsequently, an improved method based on fused penalty was proposed, which can automatically identify the number of subgroups and perform division:

[0030]

[0031] where β is the regression coefficient; μ i is the intercept; p(·, λ) is a concave penalty function, where λ is the regularization penalty parameter used to adjust the penalty strength. Each sample corresponds to an intercept term parameter, and the penalty term penalizes the difference in the intercept terms between samples. If it is compressed to zero, it is considered that these two samples belong to the same subgroup. This method proves that the heterogeneous analysis method using fused penalty can estimate the number of subgroups and the corresponding regression coefficients and achieve subgroup division.

[0032] Subsequently, there was another study. Based on the existing fused penalty method, the penalty object was extended from the intercept term to all regression coefficients:

[0033]

[0034] First, assume that each sample corresponds to a set of regression coefficients. Using this penalty method, samples with equal regression coefficients can be divided into the same subgroup.

[0035] In existing medical image detection, taking lung adenocarcinoma as an example, image features can be divided into two categories. The first category is relatively intuitive features with clear clinical meanings, which are visual image features that clinicians consider particularly important for disease diagnosis based on experience. Such features are of great significance in the prognosis of lung adenocarcinoma, but the limited number of features may result in the loss of some useful information. The second category is the features generated after feature extraction and smoothing using digital imaging software. The characteristics are that they do not rely on doctor experience and have the characteristics of high-dimensionality, and can obtain more potential information. However, such features have no obvious medical significance and will also contain noise. And the above existing methods, whether inputting the first category of features, the second category of features, or both, will only generate the result of a single-layer subgroup structure, that is, all input features are regarded as having the same status, and the subgroup structure divided by each feature is the same, and a hierarchical subgroup structure with sub-subgroups cannot be generated. And if the existing method is used twice to input the first category of features and the second category of features respectively, it will also lead to the non-existence of a corresponding inclusion relationship in the obtained results and may also bring the problem of unclear subgroup boundaries.

[0036] To address the above problems, this embodiment is used to provide a medical image feature subgroup analysis method, which belongs to a supervised subgroup regression analysis method when y is the regression response variable. As Figure 1 shown, the subgroup analysis method includes:

[0037] S1: Obtain the observation data corresponding to each of the multiple medical image samples; the observation data includes the first type of image features, the second type of image features, and the response variable;

[0038] Specifically, S1 may include:

[0039] (1) For each medical image sample, perform tissue detection to obtain the medical image sample and the response variable of the sample;

[0040] Taking the image obtained after tissue detection as the medical image sample, taking lung adenocarcinoma as an example, the response variable is FEV1 (the percentage of the volume of air that can be forcibly exhaled from the lungs within 1 second before bronchodilation in the patient compared to the normal reference value).

[0041] (2) Process the sample to obtain a labeled image and an unlabeled image respectively; the labeled image is the image obtained after manually adding the cell type within the area on the sample; the unlabeled image is the original sample;

[0042] After performing tissue detection to obtain the medical image sample, the labeled image refers to the visual attributes that are considered important for disease grading and diagnosis and are easy to see based on the experience of clinicians, and adding the image of the cell type within the area on the medical image sample, but this method requires the participation of pathologists, has limited features, and consumes a large amount of time. The unlabeled image refers to the original medical image sample without annotation.

[0043] (3) Extract features from the labeled image to obtain the first type of image features; extract features from the unlabeled image to obtain the second type of image features.

[0044] More specifically, as Figure 2 shown, extracting features from the labeled image to obtain the first type of image features includes: using the deep learning cell classification tool ConvPath to identify cell types in the labeled image, specifically, tumor cells, stromal cells, and lymphocytes can be identified to obtain a cell type map; using OpenCV to extract features such as the perimeter and area of each cell in the cell type map to obtain the first type of image features. The ConvPath software integrates CNN (Convolutional Neural Network), which will use the category with the highest probability as the predicted cell type for output to identify cell types. OpenCV is an open-source library for image recognition and can perform various functions such as edge recognition, area extraction, and color recognition.

[0045] Feature extraction is performed on the unlabeled images to obtain the second type of image features, including: segmenting the unlabeled images to obtain multiple sub-images; using CellProfiler to perform feature extraction and smoothing on a randomly selected number of sub-images to obtain the second type of image features. Specifically, 20 to 40 sub-images can be randomly selected, and CellProfiler is used to perform feature extraction and smoothing on these sub-images to obtain the second type of image features. CellProfiler is an open-source software containing multiple modules that can implement different functions of image processing.

[0046] S2: Using the observation data corresponding to all the samples as the input of the objective function, aiming to minimize the objective function, calculate the first regression coefficient set of the first-layer subgroup and the second regression coefficient set of the second-layer sub-subgroup; the first regression coefficient set includes the first regression coefficient corresponding to each sample; the second regression coefficient set includes the second regression coefficient corresponding to each sample;

[0047] Before S2, the subgroup analysis method of this embodiment includes:

[0048] (1) Establish an observation model corresponding to the observation data; the observation model is the association relationship between the first type of image features, the second type of image features, and the response variable;

[0049] The observation model includes:

[0050]

[0051] where, y i is the response variable; x i is the first type of image features, x i =(x i1 ,…,x iq ) T ; β i is the first regression coefficient of q dimensions; z i is the second type of image features, z i =(z i1 ,...,z ip ) T ; γ i is the second regression coefficient of p dimensions; ∈ i is the residual term, which satisfies E(∈ i ) = 0, Var(∈ i ) = σ 2 ; n is the total number of samples.

[0052] (2) Establish an objective function based on the observation model.

[0053] To simultaneously identify and estimate the double-layer subgroup structure, a penalized objective function is proposed. The objective function Q(β, γ) includes:

[0054]

[0055] where β j is the first regression coefficient corresponding to the j-th sample; β m is the first regression coefficient corresponding to the m-th sample; γj is the second regression coefficient corresponding to the j-th sample; γ m is the second regression coefficient corresponding to the m-th sample; λ 1 is the first penalty parameter; λ 2 is the second penalty parameter; p(·, λ) is the penalty function, which can adopt the MCP penalty. The λ in p(·, λ) can take λ 1 , or can also take λ 2 .

[0056] Taking the observed data corresponding to all samples as the input of the objective function, and aiming to minimize the objective function, the set of first regression coefficients of the first-layer subgroup and the set of second regression coefficients of the second-layer sub-subgroup can be calculated.

[0057] It is known that are the estimated values of the regression coefficient sets β and γ:

[0058]

[0059] where arg min is the value of the variable when Q(β, γ) reaches the minimum; Furthermore, the first regression coefficient and the second regression coefficient corresponding to each sample are obtained.

[0060] Preferably, S2 may include:

[0061] (1) For the calculation of the MCP penalty function, the ADMM algorithm is used to optimize the objective function to obtain the optimized objective function;

[0062] The optimized objective function is as follows:

[0063]

[0064] s.t. v jm = β j - β m , ω jm = γ j - γ m ;

[0065] where,

[0066] (2) Using the observed data corresponding to all samples as the input of the optimized objective function, aiming to minimize the optimized objective function, iterative estimation is performed through the augmented Lagrangian method to calculate the first regression coefficient set of the first-layer subgroup and the second regression coefficient set of the second-layer sub-subgroup.

[0067] In the model , two regression coefficients are defined, β i defines the first-layer subgroup structure, and γ i defines the more refined second-layer sub-subgroup structure, which is nested under the first layer. During the calculation process, by giving different status to β and γ in the penalty term of the optimization objective function Q(β, γ), the fusion penalty and the sparse group penalty are combined to make the output result have the structure of this double-layer nested relationship, so as to achieve the hierarchical subgroup estimation structure. Specifically, through the penalty function form The first penalty is a common penalty for both, and the second only penalizes one, so as to form a hierarchical structure in the output, which can be used to further analyze the characteristics of patients in medical image analysis.

[0068] S3: Obtain the subgroup analysis results of all the samples based on the first regression coefficient set and the second regression coefficient set.

[0069] S3 may include:

[0070] (1) Determine whether the first regression coefficient corresponding to the j-th sample is equal to the first regression coefficient corresponding to the m-th sample to obtain the first judgment result; j = 1, 2,..., n; m = 1, 2,..., n; and j ≠ m; n is the total number of samples;

[0071] (2) If the first judgment result is yes, then the j-th sample and the m-th sample belong to the same subgroup; and determine whether the second regression coefficient corresponding to the j-th sample is equal to the second regression coefficient corresponding to the m-th sample to obtain the second judgment result;

[0072] (3) If the second judgment result is yes, then the j-th sample and the m-th sample belong to the same sub-subgroup of the same subgroup; if the second judgment result is no, then the j-th sample and the m-th sample belong to different sub-subgroups of the same subgroup;

[0073] (4) If the first judgment result is no, then the j-th sample and the m-th sample belong to different subgroups.

[0074] Substitute the two types of image features and the response variable obtained into the objective function Q(β, γ) described above, and use the ADMM algorithm to optimize the objective function, then the estimated value of the regression coefficient of each sample can be obtained. According to the similarities and differences of the regression coefficients of each sample, the samples can be divided into different subgroups and sub-subgroups.

[0075] This embodiment first proposes an analysis method with a double-layer subgroup structure. By establishing a new model A penalty is proposed for this model The objective function Q(β, γ) is used to simultaneously identify and give the coefficients β and γ of each subgroup and sub-subgroup in the nested double-layer subgroup structure. This method is improved on the existing subgroup analysis methods. It can generate a nested hierarchical structure containing sub-subgroups in a single estimation, classify samples at multiple levels, and provide more similarity information. It is a new method that can be used for cancer heterogeneity analysis.

[0076] Embodiment 2:

[0077] This embodiment is used to provide a medical image feature subgroup analysis system, as Figure 3 shown. The subgroup analysis system includes:

[0078] An observation data acquisition module M1, configured to acquire the observation data corresponding to each of the multiple medical image samples; the observation data includes first-class image features, second-class image features, and a response variable;

[0079] A regression coefficient calculation module M2, configured to use the observation data corresponding to all the samples as the input of the objective function, and take the minimum value of the objective function as the goal to calculate the first regression coefficient set of the first-layer subgroup and the second regression coefficient set of the second-layer sub-subgroup; the first regression coefficient set includes the first regression coefficient corresponding to each of the samples; the second regression coefficient set includes the second regression coefficient corresponding to each of the samples;

[0080] A subgroup analysis module M3, configured to obtain the subgroup analysis results of all the samples based on the first regression coefficient set and the second regression coefficient set.

[0081] In each embodiment of this specification, the key points described are the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple. For the relevant parts, refer to the description in the method part.

[0082] Specific examples are used in this article to elaborate on the principles and implementation manners of the present invention. The descriptions of the above embodiments are only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manners and application scopes. In summary, the content of this specification should not be construed as a limitation to the present invention.

Claims

1. A method for subgroup analysis of medical image features, characterized in that, the subgroup analysis method includes: obtaining the observation data corresponding to each of a plurality of medical image samples; the observation data includes a first type of image feature, a second type of image feature, and a response variable; taking the observation data corresponding to all the samples as the input of an objective function, and aiming to minimize the objective function, calculating to obtain a first set of regression coefficients of the first-layer subgroup and a second set of regression coefficients of the second-layer sub-subgroup; the first set of regression coefficients includes the first regression coefficient corresponding to each sample; the second set of regression coefficients includes the second regression coefficient corresponding to each sample; obtaining the subgroup analysis results of all the samples based on the first set of regression coefficients and the second set of regression coefficients; the specific process of obtaining the observation data corresponding to each of a plurality of medical image samples includes: for each medical image sample, detecting the tissue to obtain the medical image sample and the response variable of the sample; processing the sample to obtain a labeled image and an unlabeled image respectively; the labeled image is the image obtained after artificially adding the cell types in the region on the sample; the unlabeled image is the original sample; extracting the first type of image feature from the labeled image; extracting the second type of image feature from the unlabeled image; the specific process of taking the observation data corresponding to all the samples as the input of an objective function and aiming to minimize the objective function to calculate and obtain a first set of regression coefficients of the first-layer subgroup and a second set of regression coefficients of the second-layer sub-subgroup includes: using the ADMM algorithm to optimize the objective function to obtain an optimized objective function; taking the observation data corresponding to all the samples as the input of the optimized objective function, and aiming to minimize the optimized objective function, performing iterative estimation through the augmented Lagrangian method to calculate and obtain a first set of regression coefficients of the first-layer subgroup and a second set of regression coefficients of the second-layer sub-subgroup.

2. The subgroup analysis method according to claim 1, characterized in that, the specific process of extracting the first type of image feature from the labeled image includes: using the cell classification tool ConvPath of deep learning to identify cell types in the labeled image to obtain a cell type map; using OpenCV to extract the perimeter and area of each type of cell in the cell type map to obtain the first type of image feature.

3. The subgroup analysis method according to claim 1, characterized in that, the specific process of extracting the second type of image feature from the unlabeled image includes: segmenting the unlabeled image to obtain a plurality of segmented sub-images; using CellProfiler to extract features and smooth randomly selected several of the segmented sub-images to obtain the second type of image feature.

4. The subgroup analysis method according to claim 1, characterized in that, Before calculating the first regression coefficient set of the first-layer subgroups and the second regression coefficient set of the second-layer sub-subgroups with the observed data corresponding to all the samples as the input of the objective function and aiming to minimize the objective function, the subgroup analysis method includes: Establish an observation model corresponding to the observed data; the observation model is the association relationship between the first type of image features, the second type of image features, and the response variable; Establish an objective function based on the observation model.

5. The subgroup analysis method according to claim 4, wherein, the observation model includes: Among them, y i is the response variable; x i is the first type of image feature; β i is the first regression coefficient; z i is the second type of image feature; γ i is the second regression coefficient; ∈ i is the residual term; n is the total number of samples.

6. The subgroup analysis method according to claim 5, wherein, the objective function includes: Among them, β j is the first regression coefficient corresponding to the j-th sample; β m is the first regression coefficient corresponding to the m-th sample; γ j is the second regression coefficient corresponding to the j-th sample; γ m is the second regression coefficient corresponding to the m-th sample; λ 1 is the first penalty parameter; λ 2 is the second penalty parameter; p(·, λ) is the penalty function.

7. The subgroup analysis method according to claim 1, wherein, the obtaining the subgroup analysis results of all the samples based on the first regression coefficient set and the second regression coefficient set specifically includes: Judging whether the first regression coefficient corresponding to the j-th sample is equal to the first regression coefficient corresponding to the m-th sample to obtain a first judgment result; j = 1, 2,..., n; m = 1, 2,..., n; and j ≠ m; n is the total number of the samples; If the first judgment result is yes, the j-th sample and the m-th sample belong to the same subgroup; and judging whether the second regression coefficient corresponding to the j-th sample is equal to the second regression coefficient corresponding to the m-th sample to obtain a second judgment result; If the second judgment result is yes, the j-th sample and the m-th sample belong to the same sub-subgroup of the same subgroup; if the second judgment result is no, the j-th sample and the m-th sample belong to different sub-subgroups of the same subgroup; If the first judgment result is no, the j-th sample and the m-th sample belong to different subgroups.

8. A medical image feature subgroup analysis system, wherein, the subgroup analysis system includes: An observed data acquisition module, configured to acquire the observed data corresponding to each of multiple medical image samples; the observed data includes the first type of image features, the second type of image features, and the response variable; A regression coefficient calculation module, configured to calculate the first regression coefficient set of the first-layer subgroups and the second regression coefficient set of the second-layer sub-subgroups with the observed data corresponding to all the samples as the input of the objective function and aiming to minimize the objective function; the first regression coefficient set includes the first regression coefficient corresponding to each of the samples; the second regression coefficient set includes the second regression coefficient corresponding to each of the samples; A subgroup analysis module, configured to obtain the subgroup analysis results of all the samples based on the first regression coefficient set and the second regression coefficient set; The obtaining the observed data corresponding to each of multiple medical image samples specifically includes: For each medical image sample, detecting the tissue to obtain the medical image sample and the response variable of the sample; Processing the sample to obtain a labeled image and an unlabeled image respectively; the labeled image is the image obtained after artificially adding the cell type in the region on the sample; the unlabeled image is the original sample; Extract features from the labeled images to obtain the first type of image features; extract features from the unlabeled images to obtain the second type of image features; Specifically, taking the observation data corresponding to all the samples as the input of the objective function, and aiming to minimize the objective function, calculating to obtain the first regression coefficient set of the first-layer subgroup and the second regression coefficient set of the second-layer sub-subgroup includes: Use the ADMM algorithm to optimize the objective function to obtain the optimized objective function; Taking the observation data corresponding to all the samples as the input of the optimized objective function, and aiming to minimize the optimized objective function, perform iterative estimation through the augmented Lagrangian method to calculate the first regression coefficient set of the first-layer subgroup and the second regression coefficient set of the second-layer sub-subgroup.

Images