JPS60171570A - System for setting decision area for classification - Google Patents

Abstract

PURPOSE:To distinguish specific objects from other kinds and to enable a system for extracting with high accuracy by deleting an overlapped part of plural clusters in a targeted decision area by using training data and test data. CONSTITUTION:Firstly, an extracted picture 22 is displayed as shown in 21 on a graphic terminal 10, from which internal organs as a bone (A) 23, lung (B) 24, stomach (C) 25, etc., are indicated. Secondly, a decision area correction processing 26 displays the feature amount of the specified organ as 34, of which an area of a material A becomes 27; B area 28; C area 29. The system reads out a decision area 40 already prepared from the training data and also stored in a decision area file 30 and overlaps it as shown in 35. The system uses the decision from which extracted material is deleted as a corrected decision area 41, and the system is stored in the file 30. Thirdly, a comparison extracting process 31 uses the area 41 in the file 30, decides whether or not the respective feature amounts of the picture 22 are within the decision area. If they are within the area, the system extracts it as a targeted material.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Application of the Invention] The present invention relates to an image analysis technique, and particularly to a method of extracting a target object with high precision using multivariate information.

[Background of the invention]

Conventionally, the maximum likelihood method is generally used as a method for classifying and extracting targets from multivariate information, such as multispectral (multiple wavelength) images.

The maximum likelihood method is a method often used in the fields of medical imaging and remote sensing, and is a method of classifying unknown data that belong to a class with the highest luminosity.

Luminous intensity is a probability density relationship when data is normally distributed, and is expressed as follows: Formula (1) %Formula% Here, S: Variance-covariance matrix The concept and disadvantages of the maximum likelihood method are explained using Figures 1 and 2. do. Assume a multispectral image as the multivariate information.

FIG. 1 shows spectral information of objects A, B, and C.

Specific wavelengths λ1 of objects A, B, and C. FIG. 2(a) shows the physical quantities at λ2 expressed in the feature space (υ1 112 coordinate system). When physical quantities are mapped to the feature space at all wavelengths, clusters are formed that have regions in the feature space according to the characteristics of the object, as shown in FIG. 2(b).

In other words, the maximum likelihood method is a method of classifying the class to which unknown data should belong to the class with the highest likelihood among the 0 classes, given the probability distribution of 9 classes consisting of n-dimensional truth data. It is.

This maximum likelihood method has the disadvantage that the classification accuracy is low when multiple clusters overlap, as in A and B in Figure 2(b), that is, when there are other substances with similar characteristics to the target substance to be extracted. there were.

[Purpose of the invention]

An object of the present invention is to provide a method for extracting a specific target object by distinguishing it from other types of objects having similar characteristics with high accuracy using multivariate information.

[Summary of the invention]

In order to achieve the above object, in the present invention, clusters overlap in the feature space of a specific target, as shown in FIG. The feature is that the classification accuracy is improved by deleting overlapping parts from the classification judgment area.

Conventional classification methods such as the maximum likelihood method use learning data to reduce the decision range, but the present invention further
This method of reducing the classification decision area using the classification target data (test data) is an effective method that is possible when there is an area in the test data whose attributes are clear from prior information.

[Embodiments of the invention]

An embodiment of the present invention will be described below.

Examples include NM, which is currently attracting attention in the field of medical imaging.
specific internal organs from R (nuclear magnetic resonance) scanner images,
For example, it is an image analysis system that automatically extracts a diseased liver.

As multivariate information of the NMR scanner image, two are used: a vertical relaxation time T1 and a horizontal relaxation time T2.

Normally, a nuclear magnetic dipole with spin points in an arbitrary direction, but when placed in a magnetic field, the dipole is oriented in the direction of the magnetic field lines (magnetization vector), and when the excitation magnetic field is removed, the magnetization vector returns to its initial steady state. Return to There are two ways to recover to this equilibrium value, which are characterized by relaxation times (Ts and T2).

The relaxation time varies depending on the type of internal organ or the degree of tissue degeneration.

The purpose of this system is to automatically extract the diseased liver. The features of a diseased liver are different from a normal liver or other organs.

However, in this case, the distribution characteristics of each object in the feature space are not clear (the variance is large), and there are many overlaps with other clusters, and there are many ambiguous regions.

The extraction process flow will be explained with reference to FIG.

FIG. 3 is a flowchart of the extraction process of diseased tissue.

(1) Input of known information Vertical/horizontal relaxation time Ti for diseased liver, etc.

Input the known information 5 of 1゛2.

(2) Creating training data: Mapping the input known information to the feature space (step 4)
).

In other words, using the known information shown in FIG. 4, the target determination area 13 of the diseased liver and other information, such as normal liver ill 4°, and the feature amounts of other organs 15 are mapped onto the feature space formed from Tl and T2. . Although the feature space shown in FIG. 4 is expressed in two dimensions, it generally constitutes a multidimensional feature space (three or more dimensions). (Here, we assumed two dimensions to facilitate the explanation of the decision area reduction method. We did not use the third feature, proton density information ρ, and only the relaxation times Tl and T2 were adopted as features.) First of all. The target judgment area 13 mapped onto the feature space is extracted.

In this case, as shown in FIG. 4, the feature area (normal liver 14, other organs 15) outside the target judgment area 13 is removed from the judgment area because it degrades the accuracy of the training data. This is shown in the figure.The corrected target judgment area 18 is obtained by deleting the feature values of the normal liver fi 19 and other organs 20 from the judgment area. Instruct.

(3) Storing judgment area data: As described above, the target judgment area data instructed by the human from the terminal 6 is stored in the judgment area file 7.

(++) Target object classification/extraction processing: Here, two-stage processing is performed (step 6).

FIG. 6 shows a detailed processing flow of target object classification extraction processing.

First, extract the image 22 from the graphic terminal]0 to 21
Display as follows. Among these, bones (A) 23, lungs (B)
24. Indicates organs such as the stomach (C) 25.

Second, in the judgment region modification process 26, the feature amounts of the designated organs are displayed as shown in 34. Among these, the area of substance A is 27. The area of B is 28. The area of C is 29. 34
The judgment area 40, which has already been created from the 1 to 12 matching data and stored in the judgment area file 30, is read out and superimposed on the display as shown in 35. Then, the area obtained by deleting the areas A, B, and C other than the extraction target from the judgment area is defined as the corrected judgment area 4]. This is stored in the judgment area file 30. The above processing adds a modification to the judgment region based on the standard feature distribution of each substance created from the training data to match the image-specific feature distribution created from the extraction target image 22.

Thirdly, in the comparison extraction process 3j, the target judgment area 41 in the judgment area file 30 is used to determine whether the feature amount of each point of the extraction target image 22 is within the judgment area, and if it is within the area, it is extracted as the target substance. do.

(S) Output of extraction results The extracted results clearly indicate the extracted portion 33 (lesioned liver) and are output to the graphic terminal IO or the magnetic tape 11.

According to this system, a lesioned part of a specific organ can be automatically extracted from unknown images from a large number of NMR images.

As described above, this method: (1) 1- When creating training data (judgment area), ambiguous areas are deleted from the judgment area.

(2) In addition to (1), in order to compensate for variations in the characteristics of images to be extracted, the distribution of feature amounts outside the target determination area is deleted from the feature space.

The classification extraction accuracy is improved by removing ambiguous areas through this process.

In the above process, there is no problem even if only one area is deleted.

The above extraction process is also effective using a satellite image in which multivariate data is composed of multispectral (multiple wavelength) data.

〔Effect of the invention〕

According to the present invention, classification and extraction accuracy can be improved by using a decision region reduction method that uses not only training data but also test data to delete ambiguous regions within the target decision region, that is, parts where multiple clusters overlap. It can be converted to F.

[Brief explanation of drawings]

Figure 1 is a diagram showing the spectral distribution characteristics of the object, Figure 2 is a diagram showing the mapping result to the feature space for the maximum likelihood method, Figure 3 is a diagram showing the extraction processing flow of a lesion target, Figure 4 is NM
FIG. 5 is a diagram showing the mapping of the R information data to the feature space, FIG. 5 is a diagram showing a judgment area from which ambiguous regions have been deleted, and FIG. 6 is a diagram showing a detailed flow of target object classification/extraction processing. 33...Extracted part of the diseased liver. 1st mouth 2nd figure u1 3rd mouth 40 y5 Showa z figure

Claims (1)

[Claims] In the process of extracting a target object from multivariate data,
When determining the judgment area of a Mekuri object from the distribution characteristics of multivariate data mapped to the feature space, areas of distribution characteristics due to other objects that overlap with the judgment area are deleted from the multivariate data to be processed. A classification judgment area setting method characterized by: