The paper offers a method of the big data mining to solve problems of identification of cause-and-effect relationships in changing diagnostic information on medical images with different kinds of diseases. As integrated indices of the fundus vessels and coronary heart blood vessels we have used a global set of geometric features which is supposed to be a rather complete characteristic of diagnostic images and allows to make a successful diagnosis of vascular malformations. To evaluate informativity of vascular diagnostic features based on a classification efficiency criterion and in order to form new features required to improve a diagnostics quality a method of discriminative analysis of sample data has been considered. A filtration method of invalid data is proposed using a clustering algorithm to improve a performance quality of the developed algorithm of the discriminative analysis of feature vectors.
The very large data (‘big data’) mining is a key problem of modern information
technologies. In accordance with the Forecast of the Scientific and Technology
Development of the Russian Federation for the period until 2030 approved by the
Chairman of the Government of the Russian Federation Dmitry A. Medvedev
promising research areas include the “Technologies of data processing and analysis”
including methods and techniques of collection, processing, analysis and storage of
very large volumes of information. The purpose of the paper is to develop methods
and algorithms of the bid data mining to solve problems of identification of
causeand-effect relationships in changing diagnostic information on medical images with
different kinds of diseases. It is offered to use a single approach to the analysis of
different image classes based on evaluation of combined geometric parameters of
selected areas of interest which are considered to be a basic feature set for further
diagnostic analysis [
The diagnostic image mining information technology includes a method of formation of the efficient feature space to classify a predefined image set.
A highlighting technique of diagnostically significant information on blood vessel
images is based on a new generalized mathematical model of blood vessels for two
classes of diagnostic images, i.e. the fundus blood-vessel system and coronary blood
vessels, characterized by a set of geometric parameters [
A geometric approach to formation of diagnostic features, which, unlike traditional abstract spectral-and-correlation features are well accustomed and understandable by medical professionals, demonstrate a good visual effect and take into account an object’s specific character, allows to finally increase a diagnostics efficiency.
In order to select the most successful features we use their correlation with results of expert evaluations, as well as a dispersion analysis of learning samples or the diagnostic error analysis using particular characteristics. The effectiveness of different features is evaluated for automated diagnosis problems and proper recommendations are articulated on how to use different groups of features in medical practice.
The diagnostic image mining information technology includes the following advanced techniques and algorithms: ─ the technique and the algorithm of increasing a degree of the feature informativity based on the discriminative analysis and formation of an optimal learning sample to learn a disease diagnosis expert system; ─ an estimation method of a class separatability which is not influenced by distribution of objects through classes and is independent from a used classifier; ─ the algorithm of decrease of feature space dimensions and formation of new informative features that maximize a separatability criterion based on methods of the discriminative analysis and allow to increase a diagnosis accuracy of a pathology degree; ─ a technique of optimum learning sample formation to learn a diagnostic system based on exemption of anomalous observations that will also enable to increase a disease diagnosis accuracy.
Problem-oriented and distributed software solutions of the analysis of medical and
diagnostic images are developed to detect pathological changes including software
tools for quantitative estimates of the pathology degree based on expert findings and
the proposed methods of classification. They are intended to ensure the user with an
opportunity to manage a decision-making and analysis process [
A system of classification and diagnostic researches [
Proper researches have been performed together with medical professionals from
the Medical and Stomatological University (Moscow), Ophthalmology Academic
Department, based on a digital image analysis of the fundus. Diagnostics techniques
for ophthalmic diseases have been developed based on evaluation of global vascular
characteristics (features). The paper considers geometric characteristics proposed in
[
The scatter matrix inside classes demonstrates a variety of objects with respect to mean vectors of classes: W kg1(Xk xk )(Xk xk ) , where k – class data will correspond to the mean vector xk [x1k x21k xpk ] , and g – is a total number of classes.
Elements of the scatter matrix between classes B is counted by the formula:
bij kg1 nk xik xi xjk xj , i, j 1, p , xi 1 n k1 nk xik – is a mean feature
g
value of - feature in all classes, nk – is a number of objects in k - class,
xik 1 nk nmk1 xikm – is the mean feature value in class k , and xikm – is a value of i
feature for m -object in k - class. The matrixes W and B contain all basic information
about interrelationships inside and between classes. In order to obtain the class
separability criterion some number is to be associated to these matrixes. This number
should be increased with the increase of scattering between classes or with the
decrease of scattering inside classes. For this purpose the following criterion is more
frequently used [
Identification of correlation matrix Identification of scatter matrix Computation of separability criteria Feature selection to form new features Classification using selected features Computation of structural coefficients
Analysis of discriminative functions
Identification of correlation matrix and scatter matrixes for new features Computation of criteria for new features Classification using new features
A number of studies of vascular malformations have been undertaken for diabetic retinopathy (DR) of 151 patients suffered with diabetes (D) based on the digital image analysis of the fundus. After processing of images the sample amounted to 8175 measurements including 1490 arteriols of the 1st-order, 2345 arteriols of the 2nd-order, 1960 venules of the 1st-order, and 2380 venules of the 2nd-order. Medical professionals separately consider venules and arteriolas of the 1st and 2nd order (GROUP 1 – GROUP 4) since different tendencies of vessel changes can be observed in these GROUPS at different pathology stages. Figure 2 shows some examples of diagnostic images of the fundus for different stages of diabetes (1D, 4D) and a measurement sketch of the vascular system that gives an order of the vessel examination accepted by ophthalmologists. 4 D
At the examination of features it was concluded that there are two high-correlated characteristic groups. The first group includes those features which describe path parameters (e.g. the path straightness and tortuosity), and the second group includes the features characterizing a vessel radius function (e.g. the radius tortuosity and a beads-factor). Figure 3 gives values of the separability criteria of single features for each of the four GROUPS (venules and arteriols of the 1st and 2nd orders).
Figure 3 shows that maximum feature values belong to the vessel features which are not particularly significant for medical professionals in diagnosing. Besides, in proper pathological conditions each group may have vessels which do not comply with the given pathology (for instance, the norm feature). It may be thus concluded that the sample may contain some noisy data.
In order to eliminate noise in the sample it is proposed to filter down an original sample using the clustering approach of k-means. Each GROUP was divided into 5 clusters in accordance with original clusters inside each of the GROUPS. Feature vectors which were not included into a correct cluster were removed from the sample. Values of individual separability criteria for features inside the GROUPS after filtration are given in Figure 4.
Analyzing Figure 4 we may conclude that in the sample obtained after filtration the
features with the greatest separability criterion are the characteristics which are
considered by medical professionals as specifically diagnostically significant in visual
diagnosing of pathology that corresponds to an information letter [
The classification error was evaluated for the GROUPS obtained thereby. The error
was estimated by means of the U-approach [
Analyzing the obtained findings it may be concluded that filtration allows not only to increase feature separability criteria and decrease the classification error, but also to identify diagnostically significant features. Researches made on four GROUPS of blood vessels have shown that it is important for each GROUP to have its own set of diagnostic features that is proved by clinical research studies of medical practitioners. For example, the mean diameter of the blood vessel acts differently for venules and arteriols under pathological changes. The research results have showed that the application of the proposed feature formation technique allowed not only to eliminate invalid data, but also led to a reduction in errors of classification. This resulted to the increase of the separability criterion by 39% for GROUP 1, by 42% - for GROUP 2, by 24% - for GROUP 3 and by 39% - for GROUP 4. Besides, we have obtained some additional information based on the used features such as their informativity and have identified relationships between some features.
A new technology has been developed for the data mining to identify cause-andeffect relationships of changing diagnostic information on medical images with different kinds of diseases including the filtration of invalid data and the discriminative analysis based on maximizing the separability criterion.
The algorithm based on the selection of features with the biggest separability criterion and on a complete enumeration followed by new features maximizing this criterion has been developed. Separability criteria are formed using the scattering matrixes between and inside classes. As a result of the discriminative analysis the best features have been defined for each group of vessels based on the separability criterion. It is shown that each of the four GROUPS of vessels has efficiently used its own set of global geometric features that has been proved by clinical researches. The classification error was calculated for each group of vessels before and after the algorithm performance. It is shown that the proposed technology of the feature space analysis by groups, including the algorithm of filtration of sampled data and the algorithm of formation of the efficient feature space, allowed to increase the classification efficiency of vessels by “normal-feature” classes and different degrees of “pathology”. The classification error was hereby reduced to 2.3%-3.5% for different pathology GROUPS.
This work was partially supported by the Ministry of education and science of the Russian Federation in the framework of the implementation of the Program of increasing the competitiveness of SSAU among the world’s leading scientific and educational centers for 2013-2020 years; by the Russian Foundation for Basic Research grants (# 14-01-00369-a, # 14-07-97040-p_ povolzh'e_a, # 15-29-03823, # 15-29-07077); by the ONIT RAS program # 6 “Bioinformatics, modern information technologies and mathematical methods in medicine” 2015.