Medical image dataset oversampling method and device based on positive instance synthesis

Abstract

The present application relates to a kind of medical image data set oversampling method and equipment based on positive instance synthesis, the method includes the following steps: for preprocessed medical image data set, it is divided into positive subset and negative subset, for each feature in negative subset, set the maximum value upper limit of feature and feature minimum value lower limit as feature boundary;By generating the value between the maximum value upper limit and feature minimum value lower limit for each feature, synthesize positive instance in feature boundary, update newly synthesized positive instance to make it distribute in positive instance space;Convert updated positive instance to original space, complete oversampling.Compared with prior art, the present application can be properly handled for high imbalance, extreme imbalance and absolute imbalance data, using the synthetic oversampling with positive and negative to generate diverse and adaptive synthetic instance, also has good performance for multiclass imbalance problem.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information processing, and particularly relates to a medical image data set oversampling method and device based on positive instance synthesis. BACKGROUND

[0002] With the continuous development of computer science and information science, data greatly promotes the progress of science and social development. Data plays an important role in various fields, and the scale of data is growing at an explosive speed. How to mine and use massive data has attracted widespread attention from academia and industry. Data classification is a widely used data analysis form, which aims to discover the association rules hidden behind data and categories, and divides the instances according to their distribution in the feature space. However, in the real world data, all data sets have more or less class skew distribution, and the class with fewer instances is called minority class. The minority class is often the aspect of interest to researchers. For example, in the field of business finance, risk management, precision marketing, fraud transaction identification, and public opinion analysis are needed. In the medical field, disease prediction, health detection, and cancer cell detection are needed. In the biotechnology field, gene detection and DNA detection are needed. In the industrial field, automatic driving and pedestrian re-identification are needed. The data distribution of the data sets required by the above-mentioned applications will be uneven in actual situations. Therefore, there is an urgent need for a technology to deal with the classification problem of these unbalanced data sets.

[0003] The fundamental purpose of the study of data imbalance is to mine effective positive instance data from an imbalanced dataset. In the field of computer vision and machine learning, resampling methods are commonly used to preprocess imbalanced datasets. Therefore, research based on sampling methods also has very important theoretical significance and application value. Generally, imbalanced datasets are a major problem in health applications. Medical data classification faces the problem of imbalanced data sample quantity, where at least one class forms only a small number of data, but this is a disadvantage of most machine learning algorithms. The class labels of medical datasets are mostly imbalanced. When the dataset is imbalanced, existing classification algorithms usually perform poorly in the case of minority classes. In order to solve the class imbalance problem, the patent proposes a medical image dataset oversampling method based on positive instance synthesis. In order to verify the effectiveness of the oversampling method mentioned in the patent, some real clinical medical image datasets are selected. For example: HAM10000 dataset. This dataset is a multi-source dermoscopy image dataset of 7 common skin lesions: including nevus cells (nv), melanoma (mel), benign keratosis (bkl), basal cell carcinoma (bcc), actinic keratosis (akiec), V blood vessel disease (vasc) and skin fibroma (df). The dataset contains 1015 images with an imbalance ratio as high as 47:1. Chest Xray14 dataset. This dataset contains 14 lung diseases. This chest X-ray dataset has disease labels from 30805 unique patients. It is also a real medical image imbalanced dataset. And Retinal Fundus Multi-disease Image Dataset retinal fundus multi-disease image dataset. The oversampling method proposed in the patent is completed on these real medical image imbalanced datasets.

[0004] Although existing sampling methods show good results in many experiments. But many methods are by copying positive instances near the boundary between classes, which can cause the problem of adding the same instances still exists. And some methods have the main drawback of losing valuable information and increasing the opportunity of overfitting. How to do a good classification from an imbalanced dataset is still a challenging problem.

[0005] In summary: imbalanced data is ubiquitous in the real world, and large-scale datasets usually exhibit long-tail label distribution. Especially for critical applications related to safety or health, such as autonomous driving and medical diagnosis, data is inherently severely imbalanced. In order to further address this challenge, it is crucial to understand the different characteristics of class-imbalance learning.

[0006] The existing sampling technology cannot well deal with the extremely unbalanced data, and there is currently a lack of oversampling method to solve the problem that the traditional sampling method easily loses the information contained in the negative class instances when solving the class imbalance problem. SUMMARY

[0007] The present application aims to overcome the defects of the prior art and provide a medical image data set oversampling method and device based on positive class instance synthesis to solve or partially solve the problem that the traditional sampling method easily loses the information contained in the negative class instances when solving the class imbalance problem.

[0008] The object of the present application can be achieved by the following technical solutions:

[0009] In one aspect of the present application, a medical image data set oversampling method based on positive class instance synthesis is provided, comprising the following steps:

[0010] The pre-processed medical image data set is divided into a positive class subset and a negative class subset, and for each feature in the negative class subset, the maximum upper limit of the feature and the minimum lower limit of the feature are set as the feature boundary;

[0011] By randomly generating a value between the maximum upper limit and the minimum lower limit of each feature, a positive class instance is synthesized within the feature boundary, and the newly synthesized positive class instance is updated to be distributed in the positive class instance space;

[0012] The updated positive class instance is converted to the original space to complete the oversampling.

[0013] As a preferred technical solution, the acquisition of the pre-processed data set comprises the following steps:

[0014] A real medical image unbalanced data set is acquired, and dimension reduction processing is performed on the unbalanced data set to obtain a dimension-reduced data set;

[0015] The dimension-reduced data set is normalized to obtain the pre-processed data set.

[0016] As a preferred technical solution, the dimension reduction processing of the medical image unbalanced data set is realized by the following formula:

[0017] X new =X T ξX

[0018] In the formula, ξ represents linear transformation, X represents a high-dimensional feature vector, and X new represents a feature vector mapped to a low dimension.

[0019] As a preferred technical scheme, the normalization processing of the unbalanced data set comprises the following steps:

[0020] The vector of the maximum value of the feature and the vector of the minimum value of the feature in the medical image data set after dimension reduction are obtained, and the normalization processing is realized based on the vector of the maximum value of the feature and the vector of the minimum value of the feature.

[0021] As a preferred technical scheme, the normalization processing of the medical image unbalanced data set is realized by using the following formula:

[0022] D norm =(D-Min D )÷(Max D -Min D )

[0023] In the formula, D norm represents the data set after normalization, D represents the data set after dimension reduction, Max D represents the vector of the maximum value of the feature, and Min D represents the vector of the minimum value of the feature.

[0024] As a preferred technical scheme, the following formula is used to realize the updating of the newly synthesized positive class instance so that it is distributed in the positive class instance space:

[0025]

[0026] In the formula, S u represents the generated instance after updating, S represents the generated instance before updating, dir is the direction vector of the generated instance to the negative class neighborhood, and m represents the distance between the negative class and the generated instance.

[0027] As a preferred technical scheme, the following formula is used to realize the conversion of the updated positive class instance to the original space:

[0028]

[0029] In the formula, D new represents the new data set converted back to the original space, represents the vector of the maximum value of the feature after updating, represents the vector of the minimum value of the feature after updating, and D norm represents the data set after normalization.

[0030] As a preferred technical scheme, the negative class subset is a non-convex set.

[0031] In another aspect of the present application, an electronic device is provided, comprising one or more processors and a memory having stored therein one or more programs, the one or more programs including instructions for performing the above-described medical image dataset oversampling method based on positive instance synthesis.

[0032] In another aspect of the present application, a computer-readable storage medium is provided, comprising one or more programs for execution by one or more processors of an electronic device, the one or more programs including instructions for performing the above-described medical image dataset oversampling method based on positive instance synthesis.

[0033] Compared with the prior art, the present application has the following advantages:

[0034] (1) Compared with the traditional sampling method for processing unbalanced dataset, the present application synthesizes new positive instances within the feature boundary by setting the maximum upper limit and minimum lower limit of the feature in the negative class as the feature boundary, comprehensively considers the influence of positive data and negative data in the dataset, and solves or partially solves the problem that the traditional sampling method easily loses the information contained in the negative instances when solving the class imbalance problem.

[0035] (2) For extremely unbalanced datasets, the present application can be appropriately operated to achieve good positive and negative class balance. For the case of non-convex negative class data space, the present application can also generate synthesized instances suitable for positive class data space, has strong robustness, and for the data space of positive class and negative class, the present application points out where to generate synthesized instances, which well solves the dataset imbalance problem and promotes the research on unbalanced dataset classification problem. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 Flowchart of the medical image dataset oversampling method based on positive instance synthesis in Example 1. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present application will be described below in conjunction with the accompanying drawings of the embodiments of the present application, which are clear and complete. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should fall within the scope of protection of the present application.

[0038] Example 1

[0039] As Figure 1The embodiment provides a medical image dataset oversampling method based on positive instance synthesis, and comprises the following steps.

[0040] In step S1, some real unbalanced datasets are obtained, and the datasets are selected to cover different size and dimension ranges.

[0041] In step S2, the obtained datasets are subjected to dimension reduction processing.

[0042] The feature vectors in the training unbalanced dataset G are mainly subjected to dimension reduction processing, and a linear transformation matrix ξ is used to map the sample vectors in the high-dimensional training set to low-dimensional modes. Linear transformation X new = X T ξX.

[0043] Wherein, ξ represents linear transformation, X represents a high-dimensional feature vector, and X new represents a feature vector after being mapped to low dimension.

[0044] In step S3, the dataset obtained in step S2 is subjected to normalization processing. Before generating a synthetic instance, the dataset needs to be converted. The conversion process is referred to as dataset normalization in the application, and the normalization formula used is as follows:

[0045] D norm = (D-Min D ) ÷ (Max D -Min D )

[0046] Through the formula, all the datasets are converted into normalized datasets.

[0047] In the formula, D norm represents a normalized dataset, D represents a dimension-reduced dataset, Max D represents a maximum value vector of features, and Min D represents a minimum value vector of features.

[0048] In step S4, the feature boundary of the normalized dataset is set.

[0049] The traditional sampling method directly copies the positive instances to balance the unbalanced dataset, and ignores all information of the negative instances. The method provided by the application comprehensively uses the information provided by the negative instances. By finding a decision boundary, in step S4, after the normalized dataset in step S3, D norm is divided into two subsets of positive and negative classes, A norm and B norm become normalized subsets of the positive and negative classes respectively. For each feature, the maximum value and the minimum value are set as the upper limit u f and the lower limit lf .

[0050] where: A norm represents the positive class normalized subset, B norm represents the negative class normalized subset, u f represents the upper limit of the feature maximum value, l f represents the lower limit of the feature minimum value.

[0051] Step S5, generate synthetic positive class instances within the feature boundary.

[0052] By generating synthetic instances in the region of the positive class, the data space of the negative class is prevented from expanding. And by calculating the Euclidean distance, the decision boundary is distributed as much as possible in the positive class instance space. For each feature, a value between l f and u f is randomly generated. The synthetic instance is generated as a vector with all elements located within the feature boundary. Therefore, the generated instances can be as much as possible scattered in the data space of the minority class positive class.

[0053] Step S6, update the synthetic instances.

[0054] Although the instances are generated within the feature boundary, they can be placed outside the data space of the negative class. For example, if the negative class is not convex, the space defined by the feature boundary can overlap with the negative class data space. Further updating of these generated instances is needed. The purpose is to distribute all instances in the positive class instance space. We need to update these generated instances by the update formula. Its update formula is as follows:

[0055] S u = S + dir x m

[0056] where: S u represents the updated generated instance, S represents the generated instance before updating, dir is the direction vector of the generated instance to the negative class neighborhood, and m represents the magnitude, which represents the distance between the negative class and the generated instance.

[0057] Step S7, convert the final generated instance back to the original space.

[0058] Convert the updated synthetic instance back to the original space by the inverse decoding formula. Its direction decoding formula is as follows:

[0059]

[0060] where, D new represents the new data set converted back to the original space, represents the updated feature maximum value vector, represents the updated feature minimum value vector, Dnorm represent the normalized dataset.

[0061] To test the classification effect of the sampling method, the embodiment uses three real medical image imbalanced datasets selected from HAM10000, Chest-Xray14 and Retinal Fundus Multi-disease Image Dataset. The HAM10000 dataset (multi-source skin symptom dataset) is composed of 10015 dermoscopy images, which can be used as a training set for machine learning. The cases collected by the dataset include a representative collection of all important diagnostic categories in the field of pigmented lesions. The Chest-Xray14 dataset (chest X-ray dataset) is composed of 112120 X-ray images with disease labels from 30805 unique patients. In order to create these labels, natural language processing is used to mine disease categories from related radiology report texts. The accuracy of the labels is expected to exceed 90%, and is suitable for weakly supervised learning. Chest X-ray is one of the most common and cost-effective medical image examinations. However, the clinical diagnosis of chest X-ray can be challenging, sometimes more difficult than diagnosis by chest CT imaging. The lack of a large publicly available dataset with annotations means that it is still very difficult to implement clinically relevant computer-aided detection and diagnosis (CAD) using chest X-rays in real-world medical settings, and a major obstacle to creating a large X-ray image dataset is the lack of resources to label so many images. The Retinal Fundus Multi-disease Image Dataset dataset includes 3200 fundus images taken using three different fundus cameras, of which 46 cases are annotated by consensus of two senior retina specialists. The number of people worldwide with impaired vision is estimated to be 2.2 billion, of which at least 1 billion people suffer from vision impairment that can be prevented or not yet addressed. The world faces huge challenges in eye care, including inequalities in coverage and quality of preventive, treatment and rehabilitation services. Early detection and diagnosis of eye lesions can prevent vision impairment. The existing sampling techniques cannot well handle the case of extremely imbalanced data, and in solving the problem of class imbalance, the information contained in the negative class instances is easily lost.

[0062] The method for obtaining the positive and negative comprehensive sampling by the method can solve the data imbalance problem. The method can also make reasonable operations for extremely imbalanced data sets. Meanwhile, the method solves the neglect of the negative class method in the traditional sampling method, and comprehensively considers the influence of the positive class data and the negative class data in the imbalanced data set. Imbalanced data sets can be found everywhere in daily life, such as biomedical, financial analysis, and industrial fields. The method can well solve the problem of imbalanced data sets and well promote the research on the classification problem of imbalanced data sets. Compared with the existing sampling technology, the embodiment can make appropriate operations for extremely imbalanced data sets; the embodiment can generate synthetic instances suitable for the positive class data space for non-convex negative class data space and the like; and the embodiment points out where to generate synthetic instances for the data space of the positive class and the negative class.

[0063] The data set obtained by the method can be used as a training set for machine learning to train a neural network model, and is suitable for weakly supervised learning. The trained neural network model can effectively diagnose pigmentary lesions. By using the method, a large X-ray image data set can be created with fewer sample instances. By inputting a human eye image into the neural network trained by the training set obtained by the method, prevention and diagnosis of eye lesions can be achieved.

[0064] Embodiment 2

[0065] The embodiment provides an electronic device, including one or more processors and a memory, the memory has one or more programs stored therein, and the one or more programs include instructions for executing the medical image data set oversampling method based on positive instance synthesis as described in embodiment 1.

[0066] Embodiment 3

[0067] The embodiment provides a computer-readable storage medium, including one or more programs for one or more processors of an electronic device to execute, and the one or more programs include instructions for executing the medical image data set oversampling method based on positive instance synthesis as described in embodiment 1.

[0068] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for oversampling medical image datasets based on positive class instances, characterized in that, Includes the following steps: For the preprocessed medical image dataset, it is divided into a positive class subset and a negative class subset. For each feature in the negative class subset, the upper limit of the maximum value and the lower limit of the minimum value of the feature are set as feature boundaries. By randomly generating values ​​between the upper limit of the maximum value and the lower limit of the minimum value of each feature, positive class instances are synthesized within the feature boundary, and the newly synthesized positive class instances are updated to be distributed in the positive class instance space. The updated positive class instance is then transferred to the original space to complete the oversampling. Obtaining the preprocessed dataset involves the following steps: Obtain a real imbalanced dataset of medical images, perform dimensionality reduction on the imbalanced dataset, and obtain the dimensionality-reduced dataset. For the dimensionality-reduced dataset, the preprocessed dataset is obtained through normalization. The following formula is used to update the newly synthesized positive class instances so that they are distributed throughout the positive class instance space: In the formula, This indicates the instance generated after the update. S This refers to the instance generated before the update. dir It is the direction vector from the generated instance to the neighborhood of the negative class. m This represents the distance between the negative class and the generated instance.

2. The method for oversampling medical image datasets based on positive class instances as described in claim 1, characterized in that, The following formula is used to perform dimensionality reduction on the imbalanced medical image dataset: In the formula, Let X represent a linear transformation, and let X represent a high-dimensional eigenvector. This represents the feature vector after mapping to a lower dimension.

3. The method for oversampling medical image datasets based on positive class instances as described in claim 1, characterized in that, Normalization of imbalanced datasets includes the following steps: Obtain the vectors of the maximum and minimum features in the dimensionality-reduced dataset, and perform normalization based on the vectors of the maximum and minimum features.

4. The method for oversampling medical image datasets based on positive class instances as described in claim 1, characterized in that, The imbalanced dataset is normalized using the following formula: In the formula, This represents the normalized dataset. D This represents the dataset after dimensionality reduction. The vector representing the maximum eigenvalue. This represents the vector of eigenvalue minimums.

5. The method for oversampling medical image datasets based on positive class instances as described in claim 1, characterized in that, The updated positive class instance is converted to the original space using the following formula: In the formula, This indicates returning to the original space with a new dataset. This represents the vector of the maximum value of the updated feature. This represents the updated feature minimum vector. This represents the normalized dataset.

6. The method for oversampling medical image datasets based on positive class instances as described in claim 1, characterized in that, The negative subset is a non-convex set.

7. An electronic device, characterized in that, include: One or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the medical image dataset oversampling method based on positive class instance synthesis as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that, Includes one or more programs executable by one or more processors of an electronic device, said one or more programs including instructions for performing the oversampling method for medical image datasets synthesized based on positive class instances as described in any one of claims 1-6.

Images