WO2018132997A1 - 一种基于影像组学的生存期预测方法及装置 - Google Patents

一种基于影像组学的生存期预测方法及装置 Download PDF

Info

Publication number
WO2018132997A1
WO2018132997A1 PCT/CN2017/071665 CN2017071665W WO2018132997A1 WO 2018132997 A1 WO2018132997 A1 WO 2018132997A1 CN 2017071665 W CN2017071665 W CN 2017071665W WO 2018132997 A1 WO2018132997 A1 WO 2018132997A1
Authority
WO
WIPO (PCT)
Prior art keywords
patient
image
region
tumor
tumor region
Prior art date
Application number
PCT/CN2017/071665
Other languages
English (en)
French (fr)
Inventor
李其花
孙秋畅
李志成
宋柏霖
王梦巧
Original Assignee
中国科学院深圳先进技术研究院
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 中国科学院深圳先进技术研究院 filed Critical 中国科学院深圳先进技术研究院
Priority to PCT/CN2017/071665 priority Critical patent/WO2018132997A1/zh
Publication of WO2018132997A1 publication Critical patent/WO2018132997A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
    • G16Z99/00—Subject matter not provided for in other main groups of this subclass

Definitions

  • the present application belongs to the field of biomedical engineering, and in particular relates to a method and device for predicting survival based on image omics.
  • Image omics uses automated high-throughput data feature extraction algorithms to transform image data into high-resolution, excavable image feature data from massive data such as images, pathology, and genes. Describe organizational characteristics through these data mining. Studies have reported that imaging ensemble data can determine the response of tissue characteristics to treatment and predict the prognosis of patients. The use of imaging omics data to accurately distinguish patients or tumor phenotypes can be an effective complement to clinical parameters.
  • Imaging omics is a method of classifying patients based on their tumor phenotype in medical images. It evaluates tumor phenotypes by extracting a large number of advanced imaging features from tumor images, and then uses a reliable, reproducible methodology to synthesize these features and clinical outcomes as a potential prognostic indicator, providing a non-invasive The precise diagnosis and treatment method. Imaging omics generates a unique tumor dataset that is a quantification of a tumor phenotype that provides a higher predictive power than current clinically applied imaging metrics.
  • the imaging method is used to predict the survival time of the patient, and the complete tumor region is segmented, and the image features are extracted according to the complete tumor region, and the correlation between the imaging features of the intact tumor region and the survival period is established.
  • the imaging group-based analysis only segmented the complete tumor region, and the sub-region without considering the tumor region has different pathological features, and only extracts the imaging features of the tumor region, ignoring the sub-region of the tumor region.
  • the image features with different pathologies result in the feature of the extracted image not fully representing the region of interest, which greatly limits the number and quality of image features extracted based on the region of interest.
  • an imaging group-based survival prediction method including:
  • Image feature extraction is performed on the tumor area of each patient and the sub-area of the tumor area;
  • the redundant features in each patient image feature are removed to obtain image features after screening of each patient;
  • Another technical solution of the present application is to provide an imaging group-based survival prediction apparatus, including:
  • An acquisition module for acquiring image data of multiple patients and survival time of each patient
  • a segmentation module configured to segment a tumor region and a tumor region sub-region of each patient from image data of each patient
  • a feature extraction module configured to perform image feature extraction on a tumor region of each patient and a sub-region of the tumor region
  • a screening module for removing redundant features in each patient image feature to obtain image features after screening of each patient
  • the correlation module is configured to obtain a correlation between the image features and the survival period according to the survival period of each patient and the image characteristics of each patient after screening.
  • the image group-based survival prediction method and device provided by the present application can refine the tumor region to obtain a sub-region of the tumor region, and extract features from the tumor region and the sub-region of the tumor region, thereby obtaining a large number of image features for seeking images.
  • the relationship between characteristics and (patient) survival provides more powerful support.
  • FIG. 1 is a flowchart of a method for predicting patient lifetime based on image omics according to an embodiment of the present application
  • FIG. 2 is a structural diagram of an image omics-based patient lifetime prediction apparatus according to an embodiment of the present application
  • FIG. 3 is a structural diagram of an apparatus for predicting a patient lifetime based on an imaging group according to an embodiment of the present application
  • FIG. 4 is a schematic diagram of sub-area division of a brain tumor area according to an embodiment of the present application.
  • FIG. 5 is a structural diagram of an electronic device according to an embodiment of the present application.
  • FIG. 1 is a flowchart of a method for predicting patient lifetime based on image omics in an embodiment of the present application.
  • the spatial-temporal heterogeneity of the tumor tissue is considered, and the extracted region of interest is not limited to the tumor region, and the sub-region of the tumor region is extracted, and the image feature extraction is performed on the sub-region of the tumor region.
  • the obtained image features are more comprehensive, and the correlation between the established image features and the survival period is more accurate, which can improve the accuracy of the patient's survival prediction. Specifically, including:
  • Step 101 Obtain image data of multiple patients and the survival time of each patient.
  • the image data is a pre-treatment image of the patient, including but not limited to images including PET, CT, and MRI.
  • the lifetime is the period of time between the generation of image data and the death of the patient.
  • the image data of multiple patients and the survival time of each patient can be obtained from the hospital.
  • the image data of the patient obtained in this step is derived, for example, from 57 patient image data sets in TCIA (The Cancer Imaging Archive).
  • the number of image data of the patient is not limited in this application. Generally, the more the number, the more accurate the relationship determination is.
  • Step 102 Segment the tumor region and the tumor region of each patient from the image data of each patient, the tumor region is the entire tumor region, and the sub-region of the tumor region is a sub-region divided within the entire tumor region.
  • sub-regions of the brain tumor area include, but are not limited to, necrotic areas, enhancement areas, and edema areas, and the division results are shown in FIG.
  • the tumor region and the subregion of the tumor region constitute a region of interest, and the tumor region includes a suspected tumor region in addition to the real tumor region.
  • Pathology has confirmed that there are abnormal morphological capillaries in the edema area of the tumor, interstitial edema and scattered tumor cells infiltrating and growing in neovascular or dilated blood vessels; necrotic areas are due to tumor growth too fast, insufficient supply of nutrients, leading to tumors Internal necrosis, which can indirectly reflect the growth rate of the tumor.
  • Using the sub-region of the tumor region as the region of interest can be a detailed representative characteristic of the reactive tumor, providing a larger extraction region for the next high-throughput feature extraction.
  • Step 103 Perform image feature extraction on the tumor region of each patient and the sub-region of the tumor region.
  • the sub-regions of the tumor region have their unique pathological features. Extracting the features of these regions can more fully reflect the nature of the tumor. At the same time, further obtaining a large number of imaging features to greatly improve the accuracy of the patient's survival in the later stage. .
  • the image features extracted from the tumor region and the sub-region of the tumor region may be the same or different, which is not specifically limited in the present application.
  • Step 104 Remove redundant features in each patient image feature to obtain image features after screening of each patient. Redundant features include features that can be derived from other features and features that are unrelated to prediction.
  • Step 105 According to the survival period of each patient and the image characteristics of each patient after screening, the relationship between the image features and the survival period is obtained.
  • the tumor region can be refined to obtain a sub-region of the tumor region, and the image features are extracted from the tumor region and the sub-region of the tumor region, thereby obtaining a large number of image features, and providing a more relationship for seeking image characteristics and (patient) survival period. More support.
  • the method before step 102, the method further comprises: preprocessing the image data of each patient.
  • preprocessing the image data of each patient includes performing image registration, image smoothing, and data normalization processing on the image data of each patient.
  • the preprocessing process will be described in detail below:
  • Image registration Matching and superimposing two or more images acquired at different times, different imaging devices or different conditions (such as illuminance, imaging position and angle) to unify the coordinate system of each patient image data.
  • the specific registration methods are relative registration and absolute registration.
  • Relative registration refers to selecting one image in a multi-image as a reference image, and registering other related images with it, and the coordinate system can be arbitrarily selected during registration.
  • Absolute registration refers to first defining a control grid, and all images are registered relative to the control network, that is, each image is completed separately. The geometric correction of the image is used to achieve the unification of the coordinate system.
  • the image registration method may be selected according to requirements, which is not specifically limited in this application.
  • Image smoothing Unsmooth burrs, sharp edges, etc. may occur during image data acquisition and morphological processing. This is usually not the original feature of the object, but is artificially created and needs to be removed. Image smoothing is used to remove glitch in each patient's image data. There are many methods for smoothing image data, such as averaging filters, median filters, etc. These methods can be implemented by the prior art and will not be described in detail herein.
  • Data standardization Standardize image data of multiple patients to be acquired into unified standard image data.
  • various methods for standardizing image data such as minimum-maximum standardization, Z-score standardization, decimal scale standardization, linear transformation, and the like. These methods can be implemented by the prior art, and the following is a linear transformation as an example.
  • sample A be the image data of a patient
  • maxA is the maximum value of the sample A pixel
  • A is any original pixel value of sample A
  • A' is the normalized pixel value opposite A
  • M is the largest pixel of all patient image data. value.
  • the image features extracted in the above step 103 include, but are not limited to, histogram feature data, shape feature data, and texture feature data.
  • the histogram feature is used to describe the gray values of all pixel points of the tumor region and the sub-regions of each tumor region, including mean, median, maximum, minimum, range, energy, entropy, skewness, kurtosis, Standard deviation, variance, mean absolute difference, root mean square, etc.
  • Shape features are used to describe the three-dimensional features of the tumor area and the sub-areas of each tumor area, including volume, longest diameter, surface area, hardness, density, and other quotient, spherical imbalance, curvature, eccentricity, surface area to volume ratio, etc. .
  • the texture features are used to describe the texture features of the tumor region and the sub-regions of each tumor region to quantify the heterogeneity within the tumor, including gray level co-occurrence matrix features, gray-scale run matrix class features, gray-scale region matrix class features, neighbors Domain gray tone adjustment matrix class features, wavelet transform, Laplace transform, Gaussian transform and so on.
  • Table 1 for the features specifically included in the histogram feature data, shape feature data, and texture feature data. It should be noted that all the features in Table 1 are not extracted, and can be selected according to the type of tumor. For example, after the brain tumor is partitioned, 14 histogram features are extracted for the entire region and the internal sub-region of the brain tumor, 28 shapes. Features, 52 texture features.
  • the above step 104 can alleviate the dimensionality disaster problem in the later learning, and can also reduce the difficulty of the later learning.
  • the existing method can be used to remove the redundant features in the image features of each patient, and the specific methods include:
  • Breadth-first search enumerates all feature combinations, and breadth-first traverses the feature sub-space, starting from a vertex, and traversing the wider feature subspace around it radially.
  • Branch and bound search Add branch boundaries based on exhaustive search. For example, if it is concluded that some branches are unlikely to search for solutions that are better than the best solution currently found, then these branches can be clipped.
  • Directional search First select the N highest-scoring features as feature subsets, add them to a priority queue that limits the maximum length, take the highest-scoring subset from the queue each time, and then exhaustively add 1 to the subset. All feature sets generated after the feature are added to the queue.
  • the score of the feature is the size of the Consistency Index (CI: Harrell's Concordance Index).
  • the CI value is related to the correlation between the selected feature and the survival period. The higher the correlation, the higher the score of the feature.
  • Optimal priority search Similar to directed search, the only difference is that the length of the priority queue is not limited.
  • Sequence forward selection The feature subset X starts from the empty set, and each time a feature x is selected, the feature subset X is added, so that the feature function J(X) is optimal. To put it simply, each time you choose a feature that makes the evaluation function's value optimal, it is actually a simple greedy algorithm.
  • Sequence backward selection Starting from the feature set O, each time a feature x is removed from the feature set O, the evaluation function value is optimized after the feature x is removed.
  • the algorithm has two forms: (1) The algorithm starts from the empty set, adds L features in each round, and then removes R features from it, so that the evaluation function value is optimal. (L>R); (2) The algorithm starts from the complete set, removes R features first in each round, and then adds L features to make the evaluation function value optimal. (L ⁇ R).
  • Sequence floating selection is developed by increasing L to R selection algorithm. The difference between this algorithm and the L-to-R selection algorithm is that the L and R of the sequence floating selection are not fixed, but are “floating”. ", that is, it will change.
  • the sequence float selection has the following two variants depending on the search direction. (1) Sequence floating forward selection: The beginning of the empty set, each round selects a subset x among the unselected features, so that the evaluation function is optimal after adding the subset x, and then selects the subset z among the selected features. So that the evaluation function is optimal after culling the subset z. (2) Sequence floating backward selection: similar to SFFS, the difference is that SFBS starts from the complete set, and features are removed first in each round, and then features are added.
  • Decision tree Run C4.5 or other decision tree generation algorithm on the training sample set. After the decision tree is fully grown, run the pruning algorithm on the tree. Then the characteristics of each branch of the final decision tree are the selected feature subsets. Decision tree methods typically use information gain as an evaluation function.
  • Clustering algorithm Based on K-means clustering feature selection method, the basic idea is to use K-means clustering algorithm to determine the best classification number for each feature subset, thus deleting one of the most relevant features. .
  • the invention adopts a consistency clustering method: calculating the frequency at which two sub-samples are gathered together in the case of multiple runs, and making a visual evaluation according to the result of the coincidence rate: comparing stability and determining the optimal cluster number ( K).
  • the basic assumption is that if there is an optimal number of clusters K, the subsample stability corresponding to K will be optimal.
  • Consistent clustering uses (0-1) to describe stability.
  • We use the hierarchical clustering algorithm to measure the difference based on Pearson correlation. After 2000 re-sampling iterations, in order to determine the optimal cluster number, we first calculate the cumulative distribution function of different cluster numbers.
  • the optimal number of clusters is the value corresponding to the function convergence in the cumulative distribution function.
  • the step 102 of separately segmenting the tumor region and the tumor region of each patient from the image data of each patient includes:
  • the pre-segmented image data may be divided into tumor regions and sub-regions of the tumor region by an experienced expert according to the anatomy of the tumor region and the subregion of the tumor region.
  • the segmented brain images include edema, enhanced, necrotic, and non-enhanced areas.
  • the extracted image feature set includes, but is not limited to, a histogram feature, a shape feature, and a texture feature. Histogram features, shape features and texture features specifically include Table 1. The features specifically included in the image feature set are determined by regions of the pre-divided image data. Extracting features for each of the segmented sub-regions, for example, the extracted features include 23 low-order features: 14 gray value features (the gray value of each pixel and 6 pixel points adjacent thereto and 7 The pixel is centered with the 3 ⁇ 3 module to extract the mean gray value); 6 first-order texture features (mean, variance, skewness, peak, energy, entropy); 3 position features (X, Y, Z III) Directions).
  • the learning method described in this embodiment includes, but is not limited to, a support vector machine, a random forest, a convolutional neural network, etc., and the present application does not specifically limit the learning method.
  • the above step 102 may also implement segmentation of the tumor region and the subregion of the tumor region by manual segmentation.
  • an experienced expert divides the patient's image into a tumor region and a subregion of the tumor region according to the anatomy of the tumor region and the subregion of the tumor region. It is also possible to perform segmentation processing by combining sampling and manual processing.
  • the method further comprises: combining, for each patient, a sub-region of the partial tumor region of the patient to obtain a merged sub-region of the tumor region of the patient.
  • the glioblastoma includes a necrotic area, a reinforced area, a non-enhanced area, and an edema area, and the necrotic area, the enhanced area, and the non-enhanced area have similar pathological significance (fatal). Therefore, the necrotic area, the enhanced area, and the non-enhanced area can be combined.
  • step 103 image feature extraction is performed on the tumor region and the tumor region of each patient, and further, image feature extraction is performed on the tumor region of each patient, the sub-region of the tumor region, and the combined sub-region of the tumor region.
  • the present embodiment is more suitable for practical needs, and the feature of extracting the merged sub-region of the tumor region can reflect the nature of the tumor more comprehensively and multi-angle. At the same time, the large number of imaging features obtained in this step can greatly improve the accuracy of predicting the survival of patients in the later stage.
  • the relationship between the image features obtained by the above step 105 and the lifetime may be expressed as: the survival period is a result of linear combination of the plurality of image features after screening or all the image features after screening. Assuming that each image feature represents a factor, the lifetime is a multi-factor equation. The construction process of the equation is as follows:
  • Step 201 Randomly confirm an equation of a lifetime, which may be a single factor or a multi-factor, linear or non-linear, which is not limited in this application.
  • the equation can be expressed as:
  • Y indicates the survival period;
  • X is the image feature selected by the tumor region;
  • m 0 is the number of image features selected by the tumor region;
  • W is the coefficient of X, which is unknown;
  • X1, Xi and Xn are the sub-regions of the tumor region.
  • the image features; m 1 , m i and m n are the number of image features selected by the sub-regions of the tumor region;
  • W1, Wi and Wn are the coefficients of X1, Xi and Xn, respectively, and are unknown;
  • z represents a constant, For unknowns.
  • W, W1, Wi, and Wn are constant coefficients.
  • a multi-parameter COX logistic regression model can be used to construct a linear multi-factor equation, and the gradient descent algorithm is used to seek the optimal parameter combination of the above equations, that is, the unknown value is calculated.
  • the equation for constructing a good lifetime may also be represented by a time-survival probability curve, and the generation process of the time-survival probability curve includes:
  • step 101 For a plurality of patients in step 101, the equation values are calculated by the constructed equations, respectively. For example, if 59 cases are included in step 101, that is, including 59 patients, the image features after screening of the 59 cases need to be substituted into the constructed equation, and a total of 59 equation values are obtained.
  • the patient's survival status is recorded. If the patient has died at the final follow-up, the patient's survival status is assigned a value of 1. If the patient is alive or lost after the last follow-up, the patient's survival status is assigned a value of zero.
  • step 2) assuming that the intermediate value corresponds to the third patient, the results obtained in step 2) are as shown in Table 2.
  • an embodiment of the present invention is also provided with an image omics based patient survival prediction apparatus, as described in the following embodiments. Since the principle of solving the problem of the device is similar to the method for predicting the survival of the patient based on phenotype, the implementation of the device can be referred to the implementation of the method, and the repeated description is not repeated.
  • FIG. 2 is a structural diagram of an image omics-based patient lifetime prediction apparatus according to an embodiment of the present application.
  • the device can be implemented in a smart terminal, such as a mobile phone, a tablet computer, a computer, or the like by a logic circuit, or can implement functions of various components by software in a functional module manner, and run on the smart terminal.
  • the device includes:
  • the obtaining module 201 is configured to acquire image data of multiple patients and a survival period of each patient.
  • the segmentation module 202 is configured to segment the tumor region of each patient and the subregion of the tumor region from the image data of each patient.
  • the segmentation module 203 is specifically configured to extract image features from the pre-segmented image data; using the extracted image features as a training set, the pre-processed image data of each patient is segmented by a learning method to obtain a tumor. Subregions of the region and the tumor region.
  • the feature extraction module 203 is configured to perform image feature extraction on the tumor region of each patient and the sub-region of the tumor region.
  • the image features include histogram feature data, shape feature data, and texture feature data.
  • the screening module 204 is configured to remove redundant features in each patient image feature to obtain image features after screening of each patient.
  • the association module 205 is configured to obtain a correlation between the image features and the survival period according to the survival period of each patient and the image features after screening of each patient.
  • the imaging group-based survival prediction device can refine the tumor region to obtain a sub-region of the tumor region, and perform image feature extraction on the tumor region and the sub-region of the tumor region, thereby obtaining a large number of image features for seeking image features. Provide more support for the relationship with (patient) survival.
  • the image omniscience based patient lifetime prediction apparatus further includes a preprocessing module 206 for preprocessing the image data of each patient.
  • the pre-processing module 206 is specifically configured to perform image registration, image smoothing, and data normalization processing on image data of each patient.
  • the imaging ensemble-based patient lifetime prediction apparatus further includes a merging module 207, for each patient, combining the sub-regions of the partial tumor regions of the patient to obtain the The combined subregion of the patient's tumor area.
  • the feature extraction module 203 is further configured to perform image feature extraction on the tumor region of each patient, the sub-region of the tumor region, and the merged sub-region of the tumor region.
  • the expression of the relationship between the image features and the lifetime obtained by the association module 205 is:
  • Y indicates the survival period
  • X indicates the image features selected from the tumor region
  • m 0 indicates the number of image features selected from the tumor region
  • W indicates the coefficient of X
  • X1, Xi, and Xn indicate the image features selected from the subregions of the tumor region.
  • m 1 , m i and m n represent the number of image features selected from the sub-regions of the tumor region
  • W1, Wi and Wn represent the coefficients of X1, Xi and Xn, respectively
  • z represents a constant.
  • the present application further provides an electronic device, which may be a desktop computer, a tablet computer, a mobile terminal, etc., and the embodiment is not limited thereto.
  • the electronic device may refer to the implementation of the method of Embodiment 1 and the apparatus described in Embodiment 2, and the content thereof is incorporated herein, and the details are not described again.
  • FIG. 5 is a schematic structural diagram of a system of an electronic device 500 according to an embodiment of the present application.
  • the electronic device 500 can include a central processing unit 100 and a memory 140 including computer readable instructions; the memory 140 is coupled to the central processing unit 100.
  • the computer readable instructions when executed, cause the processor to: acquire image data of a plurality of patients and a lifetime of each patient; and segment tumor regions and tumor regions of each patient from image data of each patient Sub-region; image feature extraction is performed on the tumor region of each patient and the sub-region of the tumor region; redundant features in each patient image feature are removed to obtain image features after screening of each patient; according to each patient's survival period and each patient The selected image features obtained the correlation between image features and lifetime.
  • the image feature data includes histogram feature data, shape feature data, and texture feature data.
  • the method further comprises: preprocessing the image data of each patient, wherein the preprocessing includes image registration, image smoothing, and data standardization processing. .
  • the method further comprises: combining, for each patient, a sub-region of a part of the tumor region of the patient to obtain a tumor region of the patient The merging sub-region; image feature extraction of the tumor region and the tumor region sub-region of each patient further extracts image features for the tumor region of each patient, the sub-region of the tumor region, and the combined sub-region of the tumor region.
  • the imaging ensemble-based patient lifetime prediction device can be configured separately from the central processing unit 100, for example, the imaging ensemble-based patient lifetime prediction device can be configured as a chip connected to the central processing unit 100.
  • the function of the imaging life prediction device based on angiography is realized by the control of the central processing unit.
  • the electronic device 500 may further include: a communication module 110, an input unit 120, an audio processing unit 130, a display 160, and a power source 170. It should be noted that the electronic device 500 does not have to include all the components shown in FIG. 5; in addition, the electronic device 500 may further include components not shown in FIG. 5, and reference may be made to the prior art.
  • central processor 500 also sometimes referred to as a controller or operational control, may include a microprocessor or other processor device and/or logic device that receives input and controls various components of electronic device 500. The operation of the part.
  • the memory 140 may be, for example, one or more of a buffer, a flash memory, a hard drive, a removable medium, a volatile memory, a non-volatile memory, or other suitable device.
  • the above-mentioned information related to the failure can be stored, and a program for executing the related information can be stored.
  • the central processing unit 100 can execute the program stored by the memory 140 to implement information storage or processing and the like.
  • the functions of other components are similar to those of the existing ones and will not be described here.
  • the various components of electronic device 500 may be implemented by special purpose hardware, firmware, software, or a combination thereof without departing from the scope of the invention.
  • the embodiment of the present application further provides a computer readable program, wherein when the program is executed in an electronic device, the program causes the computer to perform image group based analysis in the electronic device as described in Embodiment 1 above. Patient survival prediction method.
  • the embodiment of the present application further provides a storage medium storing a computer readable program, wherein the computer readable program causes the computer to execute the angiography based patient lifetime prediction method described in Embodiment 1 above in the electronic device.
  • embodiments of the present application can be provided as a method, system, or computer program product.
  • the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment in combination of software and hardware.
  • the application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) including computer usable program code.
  • the computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture comprising the instruction device.
  • the apparatus implements the functions specified in one or more blocks of a flow or a flow and/or block diagram of the flowchart.
  • These computer program instructions can also be loaded onto a computer or other programmable data processing device such that a series of operational steps are performed on a computer or other programmable device to produce computer-implemented processing for execution on a computer or other programmable device.
  • the instructions provide steps for implementing the functions specified in one or more of the flow or in a block or blocks of a flow diagram.

Abstract

本申请提供了一种基于影像组学的生存期预测方法及装置，其中，方法包括获取多个患者的影像数据及各患者的生存期；从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域；对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取；将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征；根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。本申请通过划分出肿瘤区域及肿瘤区域的子区域，对肿瘤区域及肿瘤区域的子区域进行特征提取，从而获得大量的影像特征，为寻求影像特征和患者生存期的关系提供更多有力支持。

Description

一种基于影像组学的生存期预测方法及装置 技术领域
本申请属于生物医学工程领域，特别涉及一种基于影像组学的生存期预测方法及装置。
背景技术
近年来，随着模式识别工具的增加和肿瘤个体化治疗的发展，影像组学应运而生。影像组学从影像、病理、基因等海量数据中利用自动化高通量的数据特征提取算法将影像数据转化为具有高分辨率的可挖掘的影像特征数据。通过这些数据挖掘描述组织特性。有研究报道，影像组学数据可以判断组织特性对治疗的反应，并预测患者的预后。应用影像组学数据对患者或肿瘤表型进行精确区分，可以成为对临床参数的一种有效补充。
影像组学是一种可以根据患者医学影像中的肿瘤表型对患者进行分类的方法。它通过对肿瘤图像提取大量先进的影像特征来评估肿瘤表型，然后使用可靠，重现性好的方法论，将这些特征与临床结果进行综合分析作为潜在的预后指标，从而提供一种非侵入性的精准诊疗方法。影像组学生成一个独特的肿瘤数据集，它是一个肿瘤表型的量化，可以提供比目前临床应用的成像指标更高的预测能力。
现有技术采用影像组学方法预测患者生存期的方法中分割出了完整的肿瘤区域，根据该完整的肿瘤区域提取了影像特征，建立了完整肿瘤区域影像特征和生存期之间的关联性。
发明内容
现有技术中，基于影像组学的分析只分割出了完整的肿瘤区域，没有考虑肿瘤区域的子区域具有不同的病理特征，另只提取了肿瘤区域的影像特征，忽略了肿瘤区域的子区域具有不同的病理的影像特征，从而导致所提取的影像特征不能全面的代表感兴趣区域的特点，极大的限制了基于感兴趣区域所提取的影像特征的数量与质量的问题。
为了解决上述技术问题，本申请的一技术方案为提供一种基于影像组学的生存期预测方法，包括：
获取多个患者的影像数据及各患者的生存期；
从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域；
对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取；
将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征；
根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
本申请另一技术方案为提供一种基于影像组学的生存期预测装置，包括：
获取模块，用于获取多个患者的影像数据及各患者的生存期；
分割模块，用于从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域；
特征提取模块，用于对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取；
筛选模块，用于将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征；
关联模块，用于根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
本申请提供的基于影像组学的生存期预测方法及装置能够细化肿瘤区域得到肿瘤区域的子区域，对肿瘤区域及肿瘤区域的子区域进行特征提取，从而获得大量的影像特征，为寻求影像特征和(患者)生存期的关系提供更多有力支持。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本申请实施例的基于影像组学的患者生存期预测方法的流程图；
图2为本申请实施例的基于影像组学的患者生存期预测装置的结构图；
图3为本申请实施例的基于影像组学的患者生存期预测装置的结构图；
图4为本申请实施例的脑瘤区域的子区域划分的示意图；
图5为本申请实施例的电子设备的结构图。
具体实施方式
为了使本申请的技术特点及效果更加明显，下面结合附图对本申请的技术方案做进一步说明，本申请也可有其他不同的具体实例来加以说明或实施，任何本领域技术人员在权利要求范围内做的等同变换均属于本申请的保护范畴。
在本说明书的描述中，参考术语“一个实施例”、“一个具体实施例”、“一些实施例”、“例如”、“示例”、“具体示例”或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本申请的至少一个实施例或示例中。在本说明书中，对上述术语的示意性表述不一定指的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任何的一个或多个实施例或示例中以合适的方式结合。各实施例中涉及的步骤顺序用于示意性说明本申请的实施，其中的步骤顺序不作限定，可根据需要作适当调整。
实施例1
如图1所示，图1为本申请实施例的基于影像组学的患者生存期预测方法的流程图。本实施例中，考虑了肿瘤组织空间-时间异质性，提取出的感兴趣区域不局限于肿瘤区域，还提取出了肿瘤区域的子区域，并对肿瘤区域的子区域进行了影像特征提取，使得获得的影像特征更全面，建立的影像特征与生存期之间的关联关系精度更高，能够提高患者生存期预测的准确度。具体的，包括：
步骤101：获取多个患者的影像数据及各患者的生存期。
多个患者的影像数据为患者的影像集合，例如表示为V＝{vi,i＝1,…,N}，N为患者的个数，每个体数据vi表示一个影像样本。
本步骤中所述的多个患者患有同一种肿瘤疾病，如脑瘤。影像数据为患者进行治疗前的影像图像，包括但不限于包括PET，CT及MRI等影像。生存期为影像数据产生至患者死亡之间的时间段。实施时，可从医院获取多个患者的影像数据及各患者的生存期。
本步骤中获得的患者的影像数据例如来源于TCIA(The Cancer Imaging Archive，癌症影像存档)中的57例患者影像数据集。本申请对患者的影像数据个数不做限定，一般情况下，个数越多，关联关系确定越准确。
步骤102：从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域，肿瘤区域为整个肿瘤区域，肿瘤区域的子区域为整个肿瘤区域内部划分的子区域。以脑瘤为例，脑瘤区域的子区域包括但不限于坏死区、增强区及水肿区，划分结果如图4所示。
肿瘤区域及肿瘤区域的子区域构成感兴趣区域，肿瘤区域除了包括真实肿瘤区域外，还包括疑似肿瘤区域。
病理学已证实：肿瘤水肿区内有异常形态的毛细血管，间质水肿和散在的肿瘤细胞在新生血管或扩张血管的浸润生长；坏死区则是因肿瘤生长过快，养料供应不够，导致肿瘤内部坏死，它可以间接的反应肿瘤的增长速度。将肿瘤区域的子区域作为感兴趣区域可以详细的具有代表性的反应肿瘤所具有的性质特征，为下一步的高通量特征提取提供了更大的提取区域。
步骤103：对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取。
肿瘤区域的子区域有其独特的病理特征，提取这些区域的特征可更全面的反应出肿瘤的性质，同时，进一步获得大量的影像学特征数据以极大的提高后期预测患者生存期的准确性。
肿瘤区域、肿瘤区域的子区域提取的影像特征可以相同，也可以不同，本申请对此不作具体限定。
步骤104：将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征。冗余特征包括可以通过其他特征推演出来的特征及与预测无关的特征。
步骤105：根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
本实施例能够细化肿瘤区域得到肿瘤区域的子区域，对肿瘤区域及肿瘤区域的子区域进行影像特征提取，从而获得大量的影像特征，为寻求影像特征和(患者)生存期的关系提供更多有力支持。
一些实施方式中，上述步骤102之前还包括：对各患者的影像数据进行预处理。
因患者的影像数据获取参数不同，通过本步骤的处理能够标准化患者的影像数据。具体的，对各患者的影像数据进行预处理包括对各患者的影像数据进行图像配准、图像平滑及数据标准化处理。下面将对预处理过程进行详细的说明：
图像配准：将不同时间，不同成像设备或不同条件(如照度，摄像位置和角度)获取的两幅或多幅影像进行匹配，叠加，从而统一各患者影像数据的坐标系。
具体的配准方法有相对配准和绝对配准。相对配准是指选择多影像中的一张影像作为参考影像，将其它的相关影像与之配准，配准时可任意选择坐标系统。绝对配准是指先定义一个控制网格，所有的影像相对于该控制网络来进行配准，也就是分别完成各影 像的几何校正来实现坐标系的统一。具体实施时，可根据需求选择图像配准方法，本申请对此不作具体限定。
图像平滑：在影像数据获取、形态学处理过程中，可能会产生不平滑的毛刺、锋利的边缘等情况。这通常不是物体原始特征，而是人为造成的，需要去除，图像平滑用于去除各患者影像数据中的毛刺。影像数据平滑的方法有很多种，例如均值滤波器，中值滤波器等，这些方法均可通过现有技术实现，此处不再详细叙述。
数据标准化：标准化即将获取的多个患者的影像数据变为统一标准的影像数据。实施时，影像数据标准化的方法有多种，例如最小-最大标准化，Z-score标准化，小数定标标准化，线性变换等等。这些方法均可通过现有技术实现，下面以线性变换为例进行说明。
设样本A为一患者的影像数据，maxA为样本A像素的最大值，A为样本A的任意一原像素值，A′为与A相对的标准化的像素值，M为所有患者影像数据像素最大值。
其公式为：A′＝(A/maxA)×(M+100)。
一些实施方式中，上述步骤103中提取的影像特征包括但不限于直方图特征数据，形状特征数据及纹理特征数据。
直方图特征用于描述肿瘤区域及各个肿瘤区域的子区域的所有像素点的灰度值，包括均值，中值，最大值，最小值，极差，能量，熵，偏斜度，峰度，标准差，方差，平均绝对差，均方根等等。
形状特征用于描述肿瘤区域及各个肿瘤区域的子区域的三维特征，包括体积，最长径，表面积，硬度，密度，等周商，球形不均衡度，曲率，偏心率，表面积体积比等等。
纹理特征用于描述肿瘤区域及各个肿瘤区域的子区域的纹理特征来量化肿瘤内部的异质性，包括灰度共生矩阵类特征，灰度游程矩阵类特征，灰度尺寸区域矩阵类特征，邻域灰度调差矩阵类特征，小波变换，拉普拉斯变换，高斯变换等等。
直方图特征数据，形状特征数据及纹理特征数据具体包括的特征参见表一。需要说明的是，表一中的所有特征并非都得提取，可根据肿瘤类型进行选取，例如，脑瘤分区后，对脑瘤整个区域及内部子区域分别提取14个直方图特征，28个形状特征，52个纹理特征。
表一
Figure PCTCN2017071665-appb-000001
Figure PCTCN2017071665-appb-000002
Figure PCTCN2017071665-appb-000003
Figure PCTCN2017071665-appb-000004
上述步骤104可以减轻后期学习中的维数灾难问题，也可以降低后期学习的难度，实施时，可采用现有的方法去除各患者影像特征中的冗余特征，具体的方法包括：
1)广度优先搜索：枚举了所有的特征组合，广度优先遍历特征子空间，即从一个顶点开始，辐射状地优先遍历其周围较广的特征子空间；
2)分支限界搜索：在穷举搜索的基础上加入分支限界。例如：若断定某些分支不可能搜索出比当前找到的最优解更优的解，则可以剪掉这些分支。
3)定向搜索：首先选择N个得分最高的特征作为特征子集，将其加入一个限制最大长度的优先队列，每次从队列中取出得分最高的子集，然后穷举向该子集加入1个特征后产生的所有特征集，将这些特征集加入队列。特征的得分即特征与生存期的一致性指标(CI:Harrell’s Concordance Index)值的大小，CI值的高低由所选特征和生存期的相关性有关，相关性越高，特征的分数越高。
4)最优优先搜索：与定向搜索类似，唯一的不同点是不限制优先队列的长度。
5)序列前向选择：特征子集X从空集开始，每次选择一个特征x加入特征子集X，使得特征函数J(X)最优。简单说就是，每次都选择一个使得评价函数的取值达到最优的特征加入，其实就是一种简单的贪心算法。
6)序列后向选择：从特征全集O开始，每次从特征集O中剔除一个特征x，使得剔除特征x后评价函数值达到最优。
7)增L去R选择算法：该算法有两种形式:(1)算法从空集开始，每轮先加入L个特征，然后从中去除R个特征，使得评价函数值最优。(L>R)；(2)算法从全集开始，每轮先去除R个特征，然后加入L个特征，使得评价函数值最优。(L<R)。
8)序列浮动选择：序列浮动选择由增L去R选择算法发展而来，该算法与增L去R选择算法的不同之处在于：序列浮动选择的L与R不是固定的，而是“浮动”的，也就是会变化的。序列浮动选择根据搜索方向的不同，有以下两种变种。(1)序列浮动前向选择：空集开始，每轮在未选择的特征中选择一个子集x，使加入子集x后评价函数达到最优，然后在已选择的特征中选择子集z，使剔除子集z后评价函数达到最优。 (2)序列浮动后向选择：与SFFS类似，不同之处在于SFBS是从全集开始，每轮先剔除特征，然后加入特征。
9)决策树：在训练样本集上运行C4.5或其他决策树生成算法，待决策树充分生长后，再在树上运行剪枝算法。则最终决策树各分支处的特征就是选出来的特征子集了。决策树方法一般使用信息增益作为评价函数。
10)遗传算法：首先随机产生一批特征子集，并用评价函数给这些特征子集评分，然后通过交叉、突变等操作繁殖出下一代的特征子集，并且评分越高的特征子集被选中参加繁殖的概率越高。这样经过N代的繁殖和优胜劣汰后，种群中就可能产生了评价函数值最高的特征子集。
11)聚类算法：基于K均值聚类的特征选择方法，其基本思想就是对每一个特征子集利用K均值聚类算法确定其最佳分类数，从而删除掉相关性较大的特征之一。本发明采用的是一致性聚类方法：计算在多次运行的情况下两个子样本聚集在一起的频率，并根据一致率的结果做出视觉评价：比较稳定性和确定最优聚类数(K)。基本假设是，如果存在最优聚类数K，K所对应的子样本稳定性会达到最佳。一致性聚类是用(0-1)来描述稳定。我们用层次聚类算法与基于皮尔森相关的不同性进行测量，经过2000次重采样迭代，为了确定最佳聚类数，我们首先计算不同聚类数的累积分布函数。最佳聚类数即为累积分布函数中函数收敛时所对应的数值。
一些实施方式中，上述步骤102从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域的过程包括：
从预先分割好的影像数据中提取影像特征集；将该提取出的影像特征集训练分类器，通过该分类器来分割(预处理后的)各患者的影像数据以得到肿瘤区域及肿瘤区域的子区域。
实施时，预先分割好的影像数据可由有经验的专家按照肿瘤区域及肿瘤区域的子区域的解剖结构将患者的影像数据划分成肿瘤区域和肿瘤区域的子区域。以脑瘤为例，分割好的脑部影像包括水肿区域、增强区域、坏死区域及非增强区域。
提取的影像特征集包括但不限于直方图特征，形状特征，纹理特征。直方图特征，形状特征及纹理特征具体包括的内容如表一。影像特征集具体包括的特征由预先分割好的影像数据各区域确定。对各分割出的子区域分别提取特征，例如提取的特征包括23个低阶特征：14个灰度值特征(每个像素点和与它相邻的6个像素点的灰度值及以这7个 像素点为中心用3×3模块提取的均值灰度值)；6个一阶纹理特征(均值，方差，偏斜度，峰值，能量，熵)；3个位置特征(X,Y,Z三个方向)。
本实施方式中所述的学习方法包括但不限于支持向量机，随机森林，卷积神经网络等等，本申请对学习方法不做具体限定。
具体实施时，上述步骤102还可采用手动分割方式实现肿瘤区域及肿瘤区域的子区域的分割。例如，让有经验的专家按照肿瘤区域及肿瘤区域的子区域的解剖结构将患者的影像划分为肿瘤区域及肿瘤区域的子区域。亦或采样人工和计算机处理结合的方式进行分割处理。
一实施方式中，上述步骤102之后还包括：针对每一患者，对该患者的部分肿瘤区域的子区域进行合并处理以得到该患者的肿瘤区域的合并子区域。
实施时，肿瘤区域的子区域合并方式有很多种，例如随机合并、具有相似病理意义的肿瘤区域的子区域合并，本申请对具体的合并方法不作限定。以脑部胶质母细胞瘤为例，该胶质母细胞瘤包括坏死区、增强区、非增强区及水肿区，因坏死区、增强区和非增强区具有相似病理意义(致命性)，所以可以将坏死区、增强区和非增强区进行合并。
上述步骤103对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取进一步为：对各患者的肿瘤区域、肿瘤区域的子区域及肿瘤区域的合并子区域进行影像特征提取。
本实施方式更适用于实际需求，提取肿瘤区域的合并子区域的特征可以更全面的、多角度的反应出肿瘤的性质。同时这一步获得的大量影像学特征可以极大的提高后期预测患者生存期的准确性。
一些实施方式中，上述步骤105得到的影像特征与生存期之间的关联关系可以表示为：生存期为筛选后的多个影像特征或全部筛选后的影像特征进行线性组合的结果。假设每个影像特征代表一个因子，则生存期就是一个多因子方程，该方程的构建流程如下：
步骤201：随机确认一个生存期的方程，该方程可以是单因子或多因子，线性或非线性，本申请对此不作限定。例如，方程可表示为：
Y＝W*X+W1*X1+...+Wi+Xi+...+Wn*Xn+z，
其中，
Figure PCTCN2017071665-appb-000005
Figure PCTCN2017071665-appb-000006
Y表示生存期；X为肿瘤区域筛选出的影像特征；m0为肿瘤区域筛选出的影像特征个数；W为X的系数，为未知量；X1、Xi及Xn为肿瘤区域的子区域筛选出的影像特征；m1、mi及mn为肿瘤区域的子区域筛选出的影像特征个数；W1、Wi及Wn分别为X1、Xi及Xn的系数，为未知量；z表示常数，为未知量。W、W1、Wi及Wn为常数系数。
具体构建方程时，可选用多参数COX逻辑回归模型构建一个线性的多因子方程，并用梯度下降的算法寻求上述方程的最优参数组合，即计算得到未知量的值。
步骤202：分别将各患者筛选后的影像特征值及生存期代入步骤201中的方程，建模得到影像特征的系数W＝{Wi,i＝1,…,m}及常数z，将得到的影像特征的系数及常数z代入至步骤201中得到构建好的生存期的方程。
进一步的，构建好的生存期的方程还可由时间-生存概率曲线表示，该时间-生存概率曲线的生成过程包括：
1)对于步骤101中多个患者，分别通过构建好的方程计算方程值。例如，步骤101中包括59个病例样本，即包括59个患者，则需要将这59个病例筛选后的影像特征代入构建好的方程中，共得到59个方程值。
2)提取计算出的方程值中的中间值，将该中间值为界，计算出的方程值若大于或等于中间值则赋为1，计算出的方程值若小于中间值则赋为0。
同时，记录患者的生存状态，若最后随访时患者已死亡，则该患者生存状态赋值为1，若最后随访时患者活着或者失访，则患者的生存状态赋值为0。
继续步骤1)中的例子，假设中间值对应第3个患者，则步骤2)得到的结果如表二，
表二：
Figure PCTCN2017071665-appb-000007
Figure PCTCN2017071665-appb-000008
3)将各患者方程值的赋值及其对应患者的生存期、患者生存状态作为输入，通过SPSS软件绘制得到时间-生存概率的曲线。
实施例2
基于同一发明构思，本申请实施例中还提供了一种基于影像组学的患者生存期预测装置，如下面的实施例所述。由于该装置解决问题的原理与基于影像组学的患者生存期预测方法相似，因此该装置的实施可以参见方法的实施，重复之处不再赘述。
如图2所示，图2为本申请实施例的基于影像组学的患者生存期预测装置的结构图。该装置可以通过逻辑电路实现运行于智能终端，例如手机、平板电脑、计算机等设备中，或者以功能模块的方式由软件实现各部件的功能，运行于所述智能终端上。具体的，该装置包括：
获取模块201，用于获取多个患者的影像数据及各患者的生存期。
分割模块202，用于从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域。详细的说，分割模块203具体用于从预先分割好的影像数据中提取影像特征；将该提取出的影像特征作为训练集，通过学习方法来分割预处理后的各患者的影像数据以得到肿瘤区域及肿瘤区域的子区域。
特征提取模块203，用于对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取。详细的说，影像特征包括直方图特征数据，形状特征数据及纹理特征数据。
筛选模块204，用于将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征。
关联模块205，用于根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
本申请提供的基于影像组学的生存期预测装置能够细化肿瘤区域得到肿瘤区域的子区域，对肿瘤区域及肿瘤区域的子区域进行影像特征提取，从而获得大量的影像特征，为寻求影像特征和(患者)生存期的关系提供更多有力支持。
一些实施方式中，如图3所示，基于影像组学的患者生存期预测装置还包括预处理模块206，用于对各患者的影像数据进行预处理。预处理模块206具体用于对各患者的影像数据进行图像配准、图像平滑及数据标准化处理。
一些实施方式中，复请参阅图3，基于影像组学的患者生存期预测装置还包括合并模块207，用于针对每一患者，对该患者的部分肿瘤区域的子区域进行合并处理以得到该患者的肿瘤区域的合并子区域。
特征提取模块203进一步用于对各患者的肿瘤区域、肿瘤区域的子区域及肿瘤区域的合并子区域进行影像特征提取。
一些实施方式中，关联模块205得到的影像特征与生存期之间的关联关系的表达公式为：
Y＝W*X+W1*X1+...+Wi+Xi+...+Wn*Xn+z，
其中，
Figure PCTCN2017071665-appb-000009
Figure PCTCN2017071665-appb-000010
Y表示生存期，X表示肿瘤区域筛选出的影像特征，m0表示肿瘤区域筛选出的影像特征个数，W表示X的系数，X1、Xi及Xn表示肿瘤区域的子区域筛选出的影像特征，m1、mi及mn表示肿瘤区域的子区域筛选出的影像特征个数，W1、Wi及Wn分别表示X1、Xi及Xn的系数，z表示常数。
实施例3
本申请还提供一种电子设备，该电子设备可以是台式计算机、平板电脑及移动终端等，本实施例不限于此。在本实施例中，该电子设备可以参照实施例1的方法的实施及实施例2所述的装置，其内容被合并于此，重复之处不再赘述。
如图5所示，图5为本申请实施例的电子设备500的系统构成示意图。该电子设备500可以包括中央处理器100和包括计算机可读指令的存储器140；存储器140耦合到中央处理器100。
计算机可读指令在被执行时使所述处理器执行以下操作：获取多个患者的影像数据及各患者的生存期；从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域；对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取；将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征；根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
其中，所述影像特征数据包括直方图特征数据，形状特征数据及纹理特征数据。
从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域之前还包括：对各患者的影像数据进行预处理，其中，预处理包括图像配准、图像平滑及数据标准化处理。
从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域之后还包括：针对每一患者，对该患者的部分肿瘤区域的子区域进行合并处理以得到该患者的肿瘤区域的合并子区域；对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取进一步为对各患者的肿瘤区域、肿瘤区域的子区域及肿瘤区域的合并子区域进行影像特征提取。
在另一个实施方式中，基于影像组学的患者生存期预测装置可以与中央处理器100分开配置，例如可以将基于影像组学的患者生存期预测装置配置为与中央处理器100连接的芯片，通过中央处理器的控制来实现基于影像组学的患者生存期预测装置的功能。
如图5所示，该电子设备500还可以包括：通信模块110、输入单元120、音频处理单元130、显示器160、电源170。值得注意的是，电子设备500也并不是必须要包括图5中所示的所有部件；此外，电子设备500还可以包括图5中没有示出的部件，可以参考现有技术。
如图5所示，中央处理器500有时也称为控制器或操作控件，可以包括微处理器或其他处理器装置和/或逻辑装置，该中央处理器100接收输入并控制电子设备500的各个部件的操作。
其中，存储器140，例如可以是缓存器、闪存、硬驱、可移动介质、易失性存储器、非易失性存储器或其它合适装置中的一种或更多种。可储存上述与失败有关的信息，此外还可存储执行有关信息的程序。并且中央处理器100可执行该存储器140存储的该程序，以实现信息存储或处理等。其他部件的功能与现有类似，此处不再赘述。电子设备500的各部件可以通过专用硬件、固件、软件或其结合来实现，而不偏离本发明的范围。
本申请实施例还提供一种计算机可读程序，其中当在电子设备中执行所述程序时，所述程序使得计算机在所述电子设备中执行如上面实施例1所述的基于影像组学的患者生存期预测方法。
本申请实施例还提供一种存储有计算机可读程序的存储介质，其中所述计算机可读程序使得计算机在电子设备中执行上面实施例1所述的基于影像组学的患者生存期预测方法。
本领域内的技术人员应明白，本申请的实施例可提供为方法、系统、或计算机程序产品。因此，本申请可采用完全硬件实施例、完全软件实施例、或结合软件和硬件方面的实施例的形式。而且，本申请可采用在一个或多个其中包含有计算机可用程序代码的计算机可用存储介质(包括但不限于磁盘存储器、CD-ROM、光学存储器等)上实施的计算机程序产品的形式。
本申请是参照根据本申请实施例的方法、设备(系统)、和计算机程序产品的流程图和/或方框图来描述的。应理解可由计算机程序指令实现流程图和/或方框图中的每一流程和/或方框、以及流程图和/或方框图中的流程和/或方框的结合。可提供这些计算机程序指令到通用计算机、专用计算机、嵌入式处理机或其他可编程数据处理设备的处理器以产生一个机器，使得通过计算机或其他可编程数据处理设备的处理器执行的指令产生用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的装置。
这些计算机程序指令也可存储在能引导计算机或其他可编程数据处理设备以特定方式工作的计算机可读存储器中，使得存储在该计算机可读存储器中的指令产生包括指令装置的制造品，该指令装置实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能。
这些计算机程序指令也可装载到计算机或其他可编程数据处理设备上，使得在计算机或其他可编程设备上执行一系列操作步骤以产生计算机实现的处理，从而在计算机或其他可编程设备上执行的指令提供用于实现在流程图一个流程或多个流程和/或方框图一个方框或多个方框中指定的功能的步骤。
以上所述仅用于说明本申请的技术方案，任何本领域普通技术人员均可在不违背本申请的精神及范畴下，对上述实施例进行修饰与改变。因此，本申请的权利保护范围应视权利要求范围为准。

Claims (10)

  1. 一种基于影像组学的患者生存期预测方法，其特征在于，包括：
    获取多个患者的影像数据及各患者的生存期；
    从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域；
    对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取；
    将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征；
    根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
  2. 如权利要求1所述的基于影像组学的患者生存期预测方法，其特征在于，从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域之前还包括：
    对各患者的影像数据进行预处理，其中，预处理包括图像配准、图像平滑及数据标准化处理。
  3. 如权利要求1所述的基于影像组学的患者生存期预测方法，其特征在于，所述影像特征包括直方图特征数据，形状特征数据及纹理特征数据。
  4. 如权利要求1所述的基于影像组学的患者生存期预测方法，其特征在于，从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域之后还包括：
    针对每一患者，对该患者的部分肿瘤区域的子区域进行合并处理以得到该患者的肿瘤区域的合并子区域；
    对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取进一步为对各患者的肿瘤区域、肿瘤区域的子区域及肿瘤区域的合并子区域进行影像特征提取。
  5. 如权利要求1所述的基于影像组学的患者生存期预测方法，其特征在于，影像特征与生存期之间的关联关系的表达公式为：
    Y＝W*X+W1*X1+...+Wi+Xi+...+Wn*Xn+z，
    其中，
    Figure PCTCN2017071665-appb-100001
    Figure PCTCN2017071665-appb-100002
    Y表示生存期，X表示肿瘤区域筛选出的影像特征，m0表示肿瘤区域筛选出的影像特征个数，W表示X的系数，X1、Xi及Xn表示肿瘤区域的子区域筛选出的影像特征，m1、mi及mn表示肿瘤区域的子区域筛选出的影像特征个数，W1、Wi及Wn分别表示X1、Xi及Xn的系数，z表示常数。
  6. 一种基于影像组学的患者生存期预测装置，其特征在于，包括：
    获取模块，用于获取多个患者的影像数据及各患者的生存期；
    分割模块，用于从各患者的影像数据中分别分割出各患者的肿瘤区域及肿瘤区域的子区域；
    特征提取模块，用于对各患者的肿瘤区域及肿瘤区域的子区域进行影像特征提取；
    筛选模块，用于将各患者影像特征中的冗余特征去除以得到各患者筛选后的影像特征；
    关联模块，用于根据各患者的生存期及各患者筛选后的影像特征得到影像特征与生存期之间的关联关系。
  7. 如权利要求6所述的基于影像组学的患者生存期预测装置，其特征在于，还包括预处理模块，用于对各患者的影像数据进行预处理，其中，预处理包括图像配准、图像平滑及数据标准化处理。
  8. 如权利要求6所述的基于影像组学的患者生存期预测装置，其特征在于，所述影像特征包括直方图特征数据，形状特征数据及纹理特征数据。
  9. 如权利要求6所述的基于影像组学的患者生存期预测装置，其特征在于，还包括合并模块，用于针对每一患者，对该患者的部分肿瘤区域的子区域进行合并处理以得到该患者的肿瘤区域的合并子区域；
    所述特征提取模块进一步用于对各患者的肿瘤区域、肿瘤区域的子区域及肿瘤区域的合并子区域进行影像特征提取。
  10. 如权利要求6所述的基于影像组学的患者生存期预测装置，其特征在于，所述关联模块得到的影像特征与生存期之间的关联关系的表达公式为：
    Y＝W*X+W1*X1+...+Wi+Xi+...+Wn*Xn+z，
    其中，
    Figure PCTCN2017071665-appb-100003
    Figure PCTCN2017071665-appb-100004
    Y表示生存期，X表示肿瘤区域筛选出的影像特征，m0表示肿瘤区域筛选出的影像特征个数，W表示X的系数，X1、Xi及Xn表示肿瘤区域的子区域筛选出的影像特征，m1、mi及mn表示肿瘤区域的子区域筛选出的影像特征个数，W1、Wi及Wn分别表示X1、Xi及Xn的系数，z表示常数。
PCT/CN2017/071665 2017-01-19 2017-01-19 一种基于影像组学的生存期预测方法及装置 WO2018132997A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2017/071665 WO2018132997A1 (zh) 2017-01-19 2017-01-19 一种基于影像组学的生存期预测方法及装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2017/071665 WO2018132997A1 (zh) 2017-01-19 2017-01-19 一种基于影像组学的生存期预测方法及装置

Publications (1)

Publication Number Publication Date
WO2018132997A1 true WO2018132997A1 (zh) 2018-07-26

Family

ID=62907545

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/071665 WO2018132997A1 (zh) 2017-01-19 2017-01-19 一种基于影像组学的生存期预测方法及装置

Country Status (1)

Country Link
WO (1) WO2018132997A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104252570A (zh) * 2013-06-28 2014-12-31 上海联影医疗科技有限公司 一种海量医学影像数据挖掘系统及其实现方法
CN105005714A (zh) * 2015-06-18 2015-10-28 中国科学院自动化研究所 一种基于肿瘤表型特征的非小细胞肺癌预后方法
CN105653858A (zh) * 2015-12-31 2016-06-08 中国科学院自动化研究所 一种基于影像组学的病变组织辅助预后系统和方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104252570A (zh) * 2013-06-28 2014-12-31 上海联影医疗科技有限公司 一种海量医学影像数据挖掘系统及其实现方法
CN105005714A (zh) * 2015-06-18 2015-10-28 中国科学院自动化研究所 一种基于肿瘤表型特征的非小细胞肺癌预后方法
CN105653858A (zh) * 2015-12-31 2016-06-08 中国科学院自动化研究所 一种基于影像组学的病变组织辅助预后系统和方法

Similar Documents

Publication Publication Date Title
CN106815481B (zh) 一种基于影像组学的生存期预测方法及装置
Kallenberg et al. Unsupervised deep learning applied to breast density segmentation and mammographic risk scoring
US10049450B2 (en) High-throughput adaptive sampling for whole-slide histopathology image analysis
WO2017092615A1 (zh) 一种计算机辅助诊断系统及方法
CN112768072B (zh) 基于影像组学定性算法构建癌症临床指标评估系统
US10783627B2 (en) Predicting cancer recurrence using local co-occurrence of cell morphology (LoCoM)
CN105653858A (zh) 一种基于影像组学的病变组织辅助预后系统和方法
WO2018107371A1 (zh) 图像搜索系统及方法
CN112927799B (zh) 融合多示例学习和多任务深度影像组学的生存期分析系统
Gheonea et al. Diagnosis system for hepatocellular carcinoma based on fractal dimension of morphometric elements integrated in an artificial neural network
Soleymanifard et al. Multi-stage glioma segmentation for tumour grade classification based on multiscale fuzzy C-means
US20230056839A1 (en) Cancer prognosis
Cao et al. An automatic breast cancer grading method in histopathological images based on pixel-, object-, and semantic-level features
Saueressig et al. Exploring graph-based neural networks for automatic brain tumor segmentation
Ullah et al. Evolutionary Model for Brain Cancer-Grading and Classification
Tyagi et al. LCSCNet: A multi-level approach for lung cancer stage classification using 3D dense convolutional neural networks with concurrent squeeze-and-excitation module
Gu et al. A novel method for breast mass segmentation: from superpixel to subpixel segmentation
WO2018132997A1 (zh) 一种基于影像组学的生存期预测方法及装置
CN113763343B (zh) 基于深度学习的阿兹海默症检测方法及计算机可读介质
CN115690056A (zh) 基于her2基因检测的胃癌病理图像分类方法及系统
Majumder et al. Digital image analysis for early diagnosis of cancer: Identification of pre-cancerous state
Liu et al. Cuts: A fully unsupervised framework for medical image segmentation
Levenson et al. Advancing Precision Medicine: Algebraic Topology and Differential Geometry in Radiology and Computational Pathology
Asaturyan et al. Hierarchical framework for automatic pancreas segmentation in MRI using continuous max-flow and min-cuts approach
EP4016448A1 (en) Image analysis method for improved clinical decision making

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17892581

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 20/11/2019)

122 Ep: pct application non-entry in european phase

Ref document number: 17892581

Country of ref document: EP

Kind code of ref document: A1