Method for constructing reproducible phenotype descriptor of medical image region of interest

Abstract

The present application relates to image feature analysis technology, in particular to a kind of medical image region of interest reproducible phenotype descriptor construction method.It is the construction method of imageology phenotype capable of being reproducible on test data set.It includes: from medical image training data set, test data set is segmented out region of interest, according to segmented region of interest, imageology feature is extracted by imageology feature extraction method;According to the imageology feature of medical image data set, image feature is clustered by feature clustering method, generates the image phenotype of strong correlation in class feature, weak correlation between classes feature;According to image phenotype, reproducible image phenotype is identified on medical image training data set, test data set by regression method, the feature importance sequence of reproducible image phenotype is calculated, the image phenotype of feature importance sequence consistent on above data set is determined, feature vector is calculated according to image phenotype, and the construction of phenotype descriptor is completed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image feature analysis, and particularly relates to a method for constructing a reproducible phenotype descriptor of a region of interest in medical images. BACKGROUND

[0002] In recent years, feature analysis of a region of interest in medical images has gradually become an important research field. Through extraction and analysis of imaging features of a region of interest in a medical image dataset, imaging features provide a new research idea for revealing microscopic information that cannot be observed by the naked eye.

[0003] However, whether imaging features can be reproduced on a test dataset is an important challenge in this research field. Due to the bias of different observers in segmenting a region of interest in a medical image, and the difference in imaging feature extraction methods and parameters, it is difficult for imaging feature analysis methods to be reproduced on a wide range of test datasets, thereby hindering the development of research in this field. Therefore, the present imaging feature analysis research field is to propose a reproducible imaging phenotype descriptor. SUMMARY

[0004] The present application provides a method for constructing a reproducible phenotype descriptor of a region of interest in medical images, which is an imaging phenotype construction method that can be reproduced on a test dataset.

[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions, comprising:

[0006] Segmenting a region of interest from a medical image training dataset and a test dataset, and extracting imaging features according to the segmented region of interest through an imaging feature extraction method;

[0007] Clustering imaging features through a feature clustering method according to the imaging features of the medical image dataset, to generate an imaging phenotype with strong intra-class feature correlation and weak inter-class feature correlation;

[0008] Identifying an imaging phenotype that can be reproduced on the medical image training dataset and the test dataset through a regression method according to the imaging phenotype, calculating a feature importance sequence of the reproducible imaging phenotype, determining an imaging phenotype with a consistent feature importance sequence on the above datasets, calculating a feature vector according to the imaging phenotype, and completing the construction of the phenotype descriptor.

[0009] Further, the segmentation of a region of interest from a medical image training dataset and a test dataset, and the extraction of imaging features according to the segmented region of interest through an imaging feature extraction method comprises:

[0010] The method for segmenting the region of interest on the medical image training dataset and the test dataset is a manual segmentation or automatic segmentation method.

[0011] The image feature extraction on the segmented region of interest is an imageomics feature extraction method.

[0012] According to the segmented region of interest, first-order statistical features, second-order texture features, and high-order image features are extracted from each region of interest on the medical image.

[0013] Further, the image features are clustered by a feature clustering method according to the image features of the medical image dataset to generate an image phenotype with strong correlation within the class and weak correlation between the classes, which includes:

[0014] All the image features extracted from the region of interest of the medical image dataset form a sample-feature matrix.

[0015] The correlation between the image features is calculated from the sample-feature matrix by a feature clustering method.

[0016] According to the correlation calculation results between the features, the feature clustering is completed according to the correlation intensity. The features with strong correlation are clustered into one class, and each cluster represents an image phenotype, thereby generating an image phenotype with strong correlation within the class and weak correlation between the classes.

[0017] Further, according to the image phenotype, the reproducible image phenotype on the medical image training dataset and the test dataset is identified by a regression method, which includes:

[0018] The significant image phenotype is identified on the medical image training dataset by a regression method, and the significant image phenotype is identified on the medical image test dataset by a regression method.

[0019] The significant image phenotypes on the medical image training dataset and the test dataset are compared to identify the image phenotype that can be reproduced on the dataset.

[0020] Further, the feature importance sequence of the reproducible image phenotype is calculated to determine the image phenotype with consistent feature importance sequence on the above dataset, and the feature vector is calculated according to the image phenotype to complete the construction of the phenotype descriptor, which includes:

[0021] According to the medical image training dataset and the test dataset, the importance of the features that constitute each image phenotype is sorted to obtain the feature importance sequence of each image phenotype.

[0022] According to the feature importance sequence of each image phenotype obtained from the above medical image training dataset and test dataset, an image phenotype (PatternDescripor) with consistent feature importance sequence is identified;

[0023] According to the image phenotype (PatternDescripor), a feature vector is calculated, and the construction of the phenotype descriptor is completed, including:

[0024] The feature vector generated in the feature clustering step of the image phenotype (PatternDescripor) is calculated, and the construction of the phenotype descriptor is calculated according to the following formula:

[0025] PatternDescripor = f1x1 + f2x2 + f3x3 +... + f n x n

[0026] Wherein, PatternDescripor represents the phenotype descriptor corresponding to the image phenotype, x1, x2, x3,..., x n represent the feature vector of the image phenotype corresponding to the phenotype descriptor, f1, f2, f3,..., f n is each image feature in the image phenotype.

[0027] Compared with the prior art, the present application has the following advantages.

[0028] The present application: through the image feature extraction of the region of interest segmented on the medical image training dataset and test dataset, the correlation between the image features is clustered by the feature clustering method to generate the image phenotype, the image phenotype which can be reproduced on the training dataset and test dataset is identified, the image phenotype with consistent feature importance sequence of the reproducible image phenotype on the above dataset is determined, and the phenotype descriptor is constructed according to the feature vector. BRIEF DESCRIPTION OF DRAWINGS

[0029] The present application will be further described below in combination with the drawings and specific embodiments. The protection scope of the present application is not limited to the following descriptions.

[0030] Figure 1 The reproducible phenotype descriptor construction flowchart of the lung CT image lesion tissue provided by the embodiment of the present application.

[0031] Figure 2 The schematic diagram of the manually segmented region of interest of the lung CT image lesion tissue provided by the embodiment of the present application.

[0032] Figure 3The lung CT image lesion tissue manual segmentation region of interest provided by the embodiment of the present application is illustrated by an imageomics method for image feature extraction.

[0033] Figure 4 The lung CT image data set feature clustering (taking a principal component analysis method as an example) provided by the embodiment of the present application is illustrated.

[0034] Figure 5 The lung CT image lesion tissue reproducible image phenotype (taking P<0.05 as a standard) provided by the embodiment of the present application is illustrated.

[0035] Figure 6 The lung CT image data set image phenotype feature importance sequence provided by the embodiment of the present application is illustrated. DETAILED DESCRIPTION

[0036] The reproducible phenotype descriptor of the lung CT image lesion tissue provided by the embodiment of the present application is constructed according to the imageomics method for extracting image features from the region of interest on the training data set and the test data set of the lung CT image, the feature clustering method and the logistic regression method are used to construct the image phenotype on the training data set and the test data set of the lung CT image and generate the feature importance sequence of the reproducible image phenotype, the image phenotype with consistent feature importance sequence on the above data set is identified to construct the reproducible phenotype descriptor of the lung CT image lesion tissue.

[0037] The construction of the reproducible phenotype descriptor of the lung CT image lesion tissue is described in detail below in combination with the drawings.

[0038] Figure 1 The reproducible phenotype descriptor of the lung CT image lesion tissue provided by the embodiment of the present application is constructed by the flowchart.

[0039] REFERENCE Figure 1 In step S101, the lesion tissue of the lung CT image is segmented by the manual segmentation method of the region of interest from the lung CT image training data set.

[0040] Here, the lesion tissue boundary of the lung CT image manually segmented by the artificial method is marked as a white region.

[0041] The specific process of step S101 is as follows: the lung CT image is read by a medical image visualization software such as ITK-SNAP, the boundary of the region of interest is manually outlined by the visualization software after the region of interest is artificially determined, and the manual segmentation of the region of interest is completed.

[0042] In step S102, according to the imaging features of the lung CT image data set, the correlation between the features is clustered by a feature clustering method, and the features with strong correlation are clustered into one class, and the image phenotype with strong correlation within the class and weak correlation between the classes is generated.

[0043] Here, the imaging feature extraction can be completed using the imageomics software. The clustering method will automatically cluster the correlation of the imaging features to generate the image phenotype.

[0044] In step S103, according to the constructed image phenotype, the image phenotype reproducible on the training data set and the test data set is identified by the logistic regression method, the feature importance sequence of the above reproducible image phenotype is calculated, the image phenotype with consistent feature importance sequence on the training data set and the test data set is identified, and the feature vector is calculated according to the image phenotype to construct the reproducible image phenotype descriptor.

[0045] Here, according to the logistic regression method, the features within each image phenotype obtained in step S102 are statistically analyzed to identify and determine the image phenotype reproducible on the training data set and the test data set.

[0046] Further, the above reproducible image phenotype is sorted by feature importance, and the feature importance sequence of each image phenotype on the lung CT image training data set and test data set is compared to identify the image phenotype with consistent feature importance sequence.

[0047] Further, the method further comprises:

[0048] According to the image phenotype with consistent feature importance sequence, the feature vector corresponding to the image phenotype obtained by the feature clustering in step S102, and according to formula (1), the image phenotype descriptor is calculated:

[0049] PatternDescripor=f1x1+f2x2+f3x3+...+f n x n (1)

[0050] Wherein, PatternDescripor represents the reproducible imaging phenotype descriptor proposed in the patent, x1, x2, x3,..., x n represents the feature vector of the image phenotype corresponding to the phenotype descriptor, f1, f2, f3,..., f n is each image feature in the image phenotype.

[0051] Here, the corresponding imaging features extracted from the region of interest are brought into formula (1) to obtain the value of the reproducible imaging phenotype descriptor PatternDescripor on the region of interest.

[0052] Figure 2 A schematic diagram of manual segmentation of a region of interest of lung CT image lesion tissue provided by an embodiment of the present application.

[0053] Referring to Figure 2 , a(1), b(1) represent an original CT image, a(2), b(2) represent a manually segmented region of interest of lung CT image lesion tissue, and the boundary of the region of interest is represented in white.

[0054] Figure 3 A schematic diagram of image feature extraction by an imageomics method provided by an embodiment of the present application.

[0055] Figure 4 A schematic diagram of feature clustering (taking a principal component analysis method as an example) of a lung CT image dataset provided by an embodiment of the present application.

[0056] Referring to Figure 4 , (a) is an image phenotype obtained by feature clustering of a lung CT image training dataset; and (b) is an image phenotype obtained by feature clustering of a lung CT image test dataset.

[0057] Figure 5 A schematic diagram of reproducible image phenotypes of lesion tissue on a lung CT image provided by an embodiment of the present application.

[0058] Referring to Figure 5 , (a) represents that, in the lung CT image training dataset, image phenotypes Dim.13, Dim.23, Dim.29, Dim.18, and Dim.14 are significant phenotypes (P<0.05) of lung CT image lesions, and (b) represents that, in the test dataset, image phenotypes Dim.22, Dim.29, Dim.21, Dim.19, and Dim.3 are significant phenotypes (P<0.05) of lung CT image lesions.

[0059] As Figure 5 can be seen, the reproducible image phenotype in the above dataset is Dim.29.

[0060] Figure 6 A schematic diagram of a feature importance sequence of an image phenotype on a lung CT image dataset provided by an embodiment of the present application.

[0061] Referring to Figure 6 , the reproducible image phenotype Dim.29 has consistency in the feature importance sequence in the training dataset (a) and in the test dataset (b);

[0062] As Figure 6It can be seen that, in the reproducible image phenotypes identified on the above data sets, the feature importance sequence has consistency is image phenotype Dim.29. According to the feature clustering (for example, the principal component analysis method), the feature vector of the image phenotype Dim.29 is obtained, and the corresponding feature value extracted from the region of interest is completed to construct the reproducible image phenotype descriptor.

[0063] It can be understood that the above specific description of the present application is only used to illustrate the present application and is not limited to the technical solutions described in the embodiments of the present application. Those skilled in the art should understand that the present application can still be modified or replaced equivalently to achieve the same technical effect; as long as the use needs are met, it is within the protection scope of the present application.

Claims

1. A method for constructing reproducible phenotypic descriptors for regions of interest in medical images, characterized in that: Regions of interest are segmented from medical image training and test datasets, and image features are extracted based on the segmented regions of interest using image feature extraction methods. Based on the imaging features of the medical image training dataset and test dataset, the image features are clustered using a feature clustering method to generate image phenotypes with strong intra-class feature correlation and weak inter-class feature correlation. Based on the image phenotypes, reproducible image phenotypes on medical image training and test datasets are identified using regression methods. The feature importance sequence of the reproducible image phenotypes is calculated, and image phenotypes with consistent feature importance sequences on the above datasets are determined. Feature vectors are calculated based on the image phenotypes to complete the construction of the phenotype descriptor. Based on the image phenotypes, the reproducible image phenotypes identified using regression methods on the medical image training dataset and test dataset include: Significant image phenotypes were identified in the medical image training dataset and in the medical image test dataset using regression methods. By comparing significant image phenotypes on the medical image training dataset and test dataset, the image phenotypes reproduced on the dataset are identified. The process of calculating the reproducible image phenotypic feature importance sequence, determining image phenotypes with consistent feature importance sequences across the above dataset, and constructing feature vectors based on the image phenotypes to complete the phenotypic descriptor construction includes: For the reproducible image phenotype, based on the medical image training dataset and test dataset, Calculate the importance ranking of the features that make up each image phenotype to obtain the feature importance sequence for each image phenotype; Based on the feature importance sequences of each image phenotype obtained from the above medical image training dataset and test dataset, identify image phenotypes with consistent feature importance sequences; Calculating feature vectors based on the image phenotype and constructing the phenotype descriptor includes: Calculate the feature vectors generated by the image phenotype in the feature clustering step, and construct the phenotype descriptor according to the following formula: in, This represents the phenotypic descriptor corresponding to the image phenotype. This represents the feature vector of the image phenotype corresponding to the phenotype descriptor. These are the various image features in the image phenotype.

2. The method for constructing a reproducible phenotypic descriptor for a region of interest in medical images according to claim 1, characterized in that: The step of segmenting regions of interest from medical image training datasets and test datasets, and extracting image features based on the segmented regions of interest using image feature extraction methods, includes: The method for segmenting the region of interest in the medical image training dataset and test dataset can be either manual segmentation or automatic segmentation. Extracting image features from the segmented region of interest is a radiomics feature extraction method; Based on the segmented regions of interest, first-order statistical features, second-order texture features, and higher-order imaging features are extracted from each region of interest in the medical image.

3. The method for constructing a reproducible phenotypic descriptor for a region of interest in medical images according to claim 1, characterized in that: Based on the imaging features of the medical image training dataset and test dataset, the image features are clustered using a feature clustering method to generate image phenotypes with strong intra-class feature correlation and weak inter-class feature correlation, including: The sample-feature matrix is ​​composed of all the imaging features extracted from the regions of interest in the medical image training dataset and the test dataset. The correlation between the imaging features is calculated from the sample-feature matrix using a feature clustering method; Based on the correlation calculation results between the features, feature clustering is performed according to the correlation strength. Features with strong correlation are clustered into one class, and each cluster represents an image phenotype, thereby generating image phenotypes with strong intra-class feature correlation and weak inter-class feature correlation.

Images