METHOD FOR CLASSIFYING SPECTRA OF OBJECTS HAVING COMPLEX INFORMATION CONTENT

Abstract

The invention relates to a method for classifying spectra of objects having complex information content after recording of the spectra involving the use of a method for preprocessing data and of a method, associated with the data preprocessing, for classification with the calculation of a classifier. After the recording of the spectra and the preprocessing of the spectra, a multiple classification method is thereby performed with at least two different methods for the data preprocessing of the spectra and the method, assigned to the respective data preprocessing, for classification. After the recording and the data preprocessing of the spectra, the following steps are thereby carried out: a calculation of multiple classifiers of the series per type of data preprocessing; a determination of the classifiers of the series with iterative adjustment and validation; a calculation of probabilities of the class association, with all classifiers of the series or classifiers being equally incorporated into the determination of a classification result.

Claims

1. A method for classifying spectra of objects (2) having complex information content with at least two different pieces of object information (31, 32; 51, 52, 53, 54), involving the use of a method for recording and preprocessing spectral data and a method, associated with the data preprocessing, for classification with the calculation of a classifier, characterized in that, a multiple classification method with at least two different data preprocessing methods (5, 6, 7, 8) for the spectra and the classification method (9, 10, 11, 12) assigned to the respective data preprocessing (5, 6, 7, 8) is performed after the recording and the data preprocessing (5, 6, 7, 8) of the spectra.

2. The method according to claim 1 in which the following steps are carried out after the recording and data preprocessing: a calculation of multiple classifiers of the series (13, 14, 15, 16) per type of data preprocessing (5, 6, 7, 8); a determination of the classifiers of the series (13, 14, 15, 16), iteratively calculated and validated; a calculation of probabilities for the class association; an equal incorporation of all classifiers of the series (13, 14, 15, 16) or classifiers (131, 132, 133, 134, 135; additional classifiers for the series 14, 15, 16) into the determination of a classification result (18, 30, 43, 44).

3. The method according to claim 1, in which during the setting of the number of calculated classifiers (N.sub.G) in the series (13, 14, 15, 16) for each classification group (9, 10, 11, 12), the scale of the spectral data points (v.sub.S) and the doubled half-width of the spectral regions (w.sub.S) and also the number of selected spectral regions (R.sub.S) are taken into account:
N.sub.G=v.sub.S/2w.sub.S.Math.R.sub.S  (1) wherein with the equation (I) it is ensured that each data point can be selected with equal probability.

4. The method according to claim 3 in which the data points belonging to the scale of the spectral data points (v.sub.S) are weighted.

5. The method according to claim 1 in which at least one of the spectral preprocessing methods (5, 6, 7, 8) is structured such that respectively defined characteristics become prominent and other defined characteristics are suppressed.

6. The method according to claim 5 in which at least one spectral preprocessing method (5) with equivalently defined and equally weighted characteristics is added to at least one of the spectral preprocessing methods (6, 7, 8) with differently defined characteristics for evaluation.

7. The method according to claim 1 in which the preprocessed spectra (4) are designed as a training set (19) and multiple classifiers of the series (13, 14, 15, 16) or classifiers (131, 132, 133, 134, 135, etc.) are defined and validated.

8. The method according to claim 1 in which at least one method of unsupervised classification and/or supervised classification is used to select spectral regions R.sub.S or individual wavelength ranges and for subsequent analysis.

9. The method according to claim 1 in which a neural network method and/or a linear wavelet transform method is used as a method for classification in the classification groups (9, 10, 11, 12).

10. The method according to claim 1 in which preprocessed spectra (5, 6, 7, 8) are classified using optical molecular spectroscopy, preferably absorption, emission, scattering, or UV/vis, NIR, IR absorption, fluorescence, Raman.

11. The method according to claim 1 in which raw spectra (25), baseline corrections (26), normalizations (27), derivatives, covariance, and/or Raman spectra (28) are used for the data preprocessing (5, 6, 7, 8).

12. The method according to claim 1 in which a calculation of a median (30) or performance of a cluster analysis is carried out for the evaluation of the classifiers of the series (13, 14, 15, 16) for a classification result (18).

13. The method according to claim 1 in which the following steps are carried out: acquisition and recording of the spectra (4) by means of at least one optical device having at least one spectrometer and/or additional detectors; generating digitized signals and storing the recorded spectra (4) in storage units of the classification units/groups (9, 10, 11, 12) of an evaluation unit; spectral preprocessing (5, 6, 7, 8), in that the recorded and stored spectra (4) are individually preprocessed in the individual storage units and the associated digitized evaluated signals are made available for further processing; separating the preprocessed spectra as a training set 19 and as a test set 29; configuring and using the preprocessed spectra as a training set 19 and a test set 29 separate from the training set 19; calculating the classifiers of the series (13, 14, 15, 16; 131, 132, 133, 134, 135) of the integrated individual classification methods, with an incorporation of iterative methods and a validation in the classification groups (9, 10, 11, 12); classifying the preprocessed spectra of the training set with all classifiers of the series; placing the spectra of the training set in a class of object information with an expression of a probability for the class association; calculating a classification result by calculating the median or by performing a cluster analysis to show the probability result of the training set object information associated with a class; classifying the preprocessed spectra of the test set (24) with all classifiers of the series (13, 14, 15, 16); placing the spectra in a class of object information (31, 32) with an expression of a probability for the class association; calculating a classification result (18, 30; 43, 44) by calculating the median (30) or by performing a cluster analysis (43, 44) to show the probability result of the object information (31, 32) associated with a class.

14. The method according to claim 1 in which optionally bird eggs, preferably chicken eggs, are used as objects (2) and the binary information about the female egg sex (31) and about the male egg sex (32) is used as object information.

15. The method according to claim 1 in which optionally tissue samples, for example brain tumors, are used as objects (2), and four characteristics (3) with: characteristic (51): healthy tissue, characteristic (52): tumor tissue with tumor grade I and II according to the histological classification model of the World Health Organization (WHO), characteristic (53): tumor tissue with tumor grade III and IV according to the WHO, and characteristic (54): necrotic tissue are used and selected and defined as object information

16. The method according to claim 1 in which, after passing through at least two classification methods (9, 10, 11, 12), each provided with a preceding data preprocessing (5, 6, 7, 8) using different spectral data, with at least one determined classifier (131, 141, 151, 161), at least two determined classifiers (131, 141, 151, 161) are collectively obtained and used for the evaluation (17) and the subsequent determination of a probability result (18) with regard to the predefined different pieces of object information (31, 32; 51, 52, 53, 54), wherein the probability result (18) is outputted so that a conclusion at least about the object information (31, 32; 51, 52, 53, 54) determined to have a highest value is rendered possible.

17. An apparatus for classifying spectra (4) of objects (2) having complex information content, in which apparatus the aforementioned method according to claim 1 is implemented, comprising at least the following units: at least one detecting optical device having at least one spectrometer and/or additional detectors for the acquisition and recording of the spectra (4); a unit for generating digitized signals in the form of data points by which the spectra are manifested (4); storage units for storing the recorded spectra (4) in the classification units/groups (9, 10, 11, 12) of an evaluation unit comprising the classification units; units for spectral preprocessing (5, 6, 7, 8) in which the recorded spectra (4) are individually preprocessed in the individual storage units and the associated digitized evaluated signals are made available for further processing; training sets (19) for configuring and using the preprocessed spectra (25, 26, 27, 28); at least one classification unit for the groups (9, 10, 11, 12) for calculating the classifiers of the series (13, 14, 15, 16), with an incorporation of iterative methods and a validation in the classification units; test sets (29) for classifying the preprocessed spectra (25, 26, 27, 28) with all classifiers of the series (13, 14, 15, 16); a unit for placing the preprocessed spectra (25, 26, 27, 28) in at least one binary class object information (31, 32; 51, 52, 53, 54) with an expression of a probability for the class association; an evaluation unit for calculating the classification result (18, 30; 43, 44), for example, in the form of the median (30) or performing a cluster analysis (43, 44) for deter mining the probability result of at least one of the predefined pieces of object information (31, 32; 51, 52, 53, 54).

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The invention is explained by means of exemplary embodiments with the aid of drawings.

[0076] Thereby:

[0077] FIG. 1 shows a schematic block diagram of a method according to the invention for classifying spectra of objects having complex information content, in particular optical molecular spectra, for assigning and determining binary object information, wherein the method is embodied as a multiple classification method;

[0078] FIG. 2a shows a schematic illustration of an individual classification method with spectral preprocessing: raw spectra, wherein all characteristics are incorporated equally (equally-sized circles);

[0079] FIG. 2b shows a schematic illustration of an individual classification method with spectral preprocessing: linear baseline correction, with a favored, large characteristic circle for the fluorescence intensity and multiple less relevant, small characteristic circles;

[0080] FIG. 2c shows a schematic illustration of an individual classification method with spectral preprocessing: normalization, with a favored, large characteristic circle for the fluorescence profile and multiple less relevant, small characteristic circles;

[0081] FIG. 2d shows a schematic illustration of an individual classification method with spectral preprocessing: Raman spectra, with a favored, large characteristic circle for the molecular composition and multiple less relevant, small characteristic circles;

[0082] FIG. 3a shows a schematic illustration of the spectra assigned to the raw spectrum classification method according to FIG. 2a, wherein the dotted-line spectrum is assigned to the female chicken-egg spectrum;

[0083] FIG. 3b shows a schematic illustration of the spectra assigned to the linear baseline correction classification method according to FIG. 2b, wherein the dotted-line spectrum is assigned to the female chicken-egg spectrum;

[0084] FIG. 3c shows a schematic illustration of the spectra assigned to the normalization classification method according to FIG. 2c, wherein the dotted-line spectrum is assigned to the female chicken-egg spectrum;

[0085] FIG. 3d shows a schematic illustration of the spectra assigned to the Raman spectrum classification method according to FIG. 2d, wherein the dotted-line spectrum is assigned to the female chicken-egg spectrum;

[0086] FIG. 4a shows a schematic illustration of the sequence of a classification with determination of a classifier according to the prior art;

[0087] FIG. 4b shows a flowchart for the multiple classification method according to the invention having a training set and test set in an algorithmic connection, with classification-result designing from a large number of classifiers;

[0088] FIG. 5 shows a schematic illustration of a probability/classifier quantity bar graph for twenty classifiers according to FIG. 1 for a chicken egg used as an example, wherein the bars located above the dashed line cut-off are assigned to the female sex;

[0089] FIG. 6 shows a top view of the bar graph for an egg according to FIG. 5 and a possible view on a display;

[0090] FIG. 7 shows a probability (female)/classifier quantity graph with the specification of the calculated median for an egg with the 20 classifiers according to FIG. 6, wherein a cut-off with a thick dashed line is at 0.5 for the probability and a median value with a thin dashed line is at approximately 0.72 for the probability for “female,” so that the sex of the egg can be identified as female;

[0091] FIG. 8 shows a schematic illustration of a probability/classifier quantity bar graph for an egg for optionally 120 classifiers, wherein the end regions of the bars located above the thick dashed line—cut-off—are assigned to the female sex of an egg and the end regions of the bars located below the thick dashed cut-off are assigned to the male sex of an egg;

[0092] FIG. 9 shows a graph of the calculated median, illustrated as a dashed line, for an egg detected as female with the 120 classifiers according to FIG. 8 in a probability/classifier quantity graph.

[0093] FIG. 10a shows a schematic illustration of an individual classification method with spectral preprocessing: raw spectra, wherein all characteristics are incorporated equally, with eight selected spectral regions R.sub.S1 to R.sub.S8 for a scale of the spectral data points over the entire spectral range between the wavenumbers 570 cm.sup.−1 and 2750 cm.sup.−1 for determining the number of classifiers per data preprocessing.

[0094] FIG. 10b shows an enlarged section (I-I) of the data point graph in the predefined spectral region R.sub.S8 according to FIG. 10a;

[0095] FIG. 10c shows an enlarged section (I-I) of the data point graph in the spectral region R.sub.S8 with a specification of the weighting of data points in the range of the region R.sub.S8 according to FIG. 10a and FIG. 10b;

[0096] FIG. 11a shows a graph of the calculated median, illustrated as a dashed line, for an egg detected as male with the 120 classifiers similar to FIG. 8 in a probability/classifier quantity graph,

[0097] FIG. 11b shows a first histogram illustration of the interdependency of the number of elements in the cluster and the focus of the cluster; and

[0098] FIG. 11c shows a second histogram illustration of the interdependency of the number of elements in the cluster and the focus of the cluster.

DETAILED DESCRIPTION

[0099] FIG. 1 and FIGS. 2a, 2b, 2c, 2d are considered together below.

[0100] FIG. 1 shows, in a schematic block diagram, a method 1 according to the invention for classifying spectra 4 of an object 2 having complex information content with at least two/binary and different pieces of object information/characteristics, in particular optical molecular spectra 4, for assigning object information/characteristics 3 for a probabilistic definition, for example, of binary object information 31, 32.

[0101] Bird eggs, for example chicken eggs, can be used as the objects 2 being examined, and the characteristic 31 for the female egg sex and the characteristic 32 for the male egg sex, for example, can be searched for and defined as binary object information 3.

[0102] The method 1 according to the invention for carrying out the classification is described below.

[0103] For this purpose, a block-wise sequence of the method 1 according to the invention is shown in FIG. 1.

[0104] In the method 1 for classifying spectra 4 of objects 2 having complex information content with at least two different pieces of object information, the calculation of a classifier occurs after the recording, involving the use of a method for preprocessing data and a method, associated with the data preprocessing, for classification.

[0105] According to the invention, following the recording and data preprocessing of spectra 4, a multiple classification method with at least two different methods of data preprocessing 5, 6, 7, 8 of the spectra 4 and the method, assigned to the respective data preprocessing 5, 6, 7, 8, for classification in the groups 9, 10, 11, 12 is carried out to determine multiple, for example five, classifiers per group 9, 10, 11, 12, that is, a large number of classifiers overall, for example twenty (five classifiers/group×four groups) classifiers 131, 132, 133, 134, 135, etc., for the series 14, 15, 16.

[0106] The following steps are thereby carried out following recording and data preprocessing of the spectra, wherein the steps refer to FIG. 1: [0107] a calculation of five classifiers of the series 13, 14, 15, 16 per type of data preprocessing 5, 6, 7, 8 so that ultimately twenty classifiers 131, 132, 133, 134, 135, etc. are determined; [0108] a determination of the five classifiers of the series 13, 14, 15, 16, iteratively adjusted and validated; [0109] a calculation of probabilities for the class association; [0110] an equal incorporation of all five classifiers of the series 13, 14, 15, 16 and/or classifiers 131, 132, 133, 134, 135, etc. into the determination of a classification result 18, for example, in the form of a median 30.

[0111] During the setting and/or determination of the number of classifiers N.sub.G to be calculated in the series 13, 14, 15, 16 in relation to the groups 9, 10, 11, 12, a scale of the spectral data points v.sub.S and a doubled half-width w.sub.S of spectral regions R.sub.S and also a number of selected spectral regions R.sub.S are factored into the following equation (I):

[00002]$\begin{array}{cc}{N}_G=\frac{v_S}{2w_S.Math.R_S}& \left(I\right)\end{array}$

wherein with the equation (I) it is ensured that each data point v.sub.S can be selected with equal probability.

[0112] For a total of twenty classifiers of the four series 13, 14, 15, 16 with N.sub.G (13), N.sub.G (14), N.sub.G (15), and N.sub.G (16) according to FIG. 1 and according to FIG. 5, FIG. 6, and FIG. 7, the following parameters, for example, are predefined for the entire spectral range from 500 cm.sup.−1 to 2750 cm.sup.−1.

[0113] Scope of the spectral data points v.sub.S in a predefined total spectral range of 500 cm.sup.−1 to 2750 cm.sup.−1 with v.sub.S=800;

Number of selected spectral regions R.sub.S with R.sub.S=8;
Width W of the spectral regions R.sub.S with W=2.Math.w.sub.S=5, that is, there can be twenty data points
v.sub.S in one region R.sub.S. The half-width w.sub.S is therefore w.sub.S=2.5.

[0114] According to FIGS. 2b, 2c, 2d, the spectral preprocessing methods 6, 7, 8 can be structured such that respectively defined characteristics are favored and become prominent, and other characteristics are suppressed.

[0115] In the spectral preprocessing 5 of the raw spectra 25 according to FIG. 2a, all incorporated characteristics can be treated and preprocessed in an equal manner.

[0116] The preprocessed spectra 4 are configured as a training set 24 according to FIG. 4b, and multiple classifiers of the series 13, 14, 15, 16, or for example the classifiers 131, 132, 133, 134, 135 specifically for series 13 etc., are iteratively determined and validated.

[0117] After passing through at least two classification methods 9, 10, 11, 12, each provided with a preceding data preprocessing 5, 6, 7, 8 using different spectra, with at least one determined classifier 131, 141, 151, 161, according to FIG. 1 at least two of the determined classifiers 131, 141, 151, 161 can be collectively obtained and used for the evaluation and the subsequent determination of a probability result 18 with regard to the predefined different pieces of object information 31, 32, wherein the probability result 18 is outputted so that a conclusion at least about the object information 31 or 32 determined to have a highest value is rendered possible.

[0118] At least one method of supervised classification and/or unsupervised classification can be used to select spectral regions R.sub.S or individual wavelength ranges/wavenumber ranges and for subsequent analysis.

[0119] The subsequent analysis can be a linear discriminant analysis or a non-linear discriminant analysis.

[0120] However, a neural network method and/or a linear wavelet transform method can also be used as a method for classification in the groups 9, 10, 11, 12.

[0121] The spectra 4 from optical molecular spectroscopy, such as absorption, emission, scattering, or UV/vis, NIR, IR absorption, fluorescence, Raman, can be classified using the method according to the invention.

[0122] For the data preprocessing methods 5, 6, 7, 8 shown in FIG. 1 raw spectra 25, baseline corrections 26, normalizations 27, derivatives, covariance, and/or a principal component analysis/Raman spectra 28 can be defined and used. A data preprocessing is constituted by the formation of digital signals, that is, of data points, which when strung together result in the respective calculated spectral curve of 25, 26, 27, 28, and which can thus be assigned to the different individual classification methods used.

[0123] For the evaluation of the classifiers of the series 13, 14, 15, 16 for a classification result 18, a calculation of a median 30 (FIG. 7, FIG. 9, FIG. 11a) for an object 2 or performance (FIG. 11a, FIG. 11b) of a cluster analysis can be provided.

[0124] The known k-means cluster analysis can be used as an example of an evaluation. In FIG. 11a, at least two clusters are thereby predefined for “male” and “female.” The cluster to which the most elements, that is, probabilities, are assigned, defines the sex (FIG. 11b—male). A group of compiled elements with similar properties (here: probabilities classification results) is referred to as a cluster.

[0125] FIG. 11 a shows a curve of the interdependency between the probability for a “male” characteristic and a classifier quantity of 120 classifiers. In the plot of the sorted probabilities, it is apparent that the median 30 lies just above the cut-off 42 of 0.5 on the probability coordinate. The egg 2 is therefore just barely classified as male. The known k-means cluster analysis leads to a clearer result in this case.

[0126] For this purpose, FIG. 11b shows a first histogram illustration 43 of the interdependency between the number of elements in the cluster and the focus of the cluster, and FIG. 11c shows a second histogram illustration 44 of the interdependency between the number of elements in the cluster and the focus of the cluster, wherein the obtained classification results are regarded as elements.

[0127] To this end, two clusters are formed in FIG. 11 b, the foci of which are 0.84 and 0.17. Because more elements, that is, classification results, are assigned to the cluster with the focus of 0.84, the egg 2 can be evaluated as being clearly male.

[0128] If five clusters are selected in the histogram illustration 44 according to FIG. 11c, this result is also confirmed. Here, 65 calculated probability values are classified as male, wherein the cluster (number 4=#4) with the strongest value of 0.96 for “male” also contains the most elements according to FIG. 11a. The sex of the egg 2 is therefore clearly “male.”

[0129] This also applies equally and similarly for the cluster analysis if the sex of the egg 2 is determined to be female.

[0130] The method 1 according to the invention can be achieved by means of the following steps, with the use of hardware components of an accompanying apparatus: [0131] acquisition and recording of the spectra 4 by means of at least one optical device having at least one spectrometer and/or additional detectors; [0132] generating digitized signals, the data points, and storing the detected spectra 4 in storage units of the classification units of an evaluation unit; [0133] spectral preprocessing 5, 6, 7, 8, in that the stored spectra 4 composed of the data points v.sub.S are individually evaluated in the individual storage units and the associated digitized evaluated signals are made available for further processing; [0134] configuring or forming the pretreated spectra 25, 26, 27, 28 as a training set 19 and a test set 29 separate therefrom; [0135] calculating the classifiers of the series 13, 14, 15, 16 in the form of individual classifiers 131, 132, 133, 134, 135 of a series 13, etc. of the integrated individual classification methods 9, 10, 11, 12, with an incorporation of iterative methods and a validation in the classification units/groups; [0136] classifying the evaluated spectra 25, 26, 27, 28 of the test set 24 with all classifiers of the series 13, 14, 15, 16; [0137] placing the spectra 25, 26, 27, 28 in a class of object information with an expression of a probability for the class association; [0138] calculating the median 30 or performing a previously indicated cluster analysis to show the probability result in the form of a classification result 18 of a piece of object information associated with a class.

[0139] The construction of classifiers with regard to the series 13, 14, 15, 16 by means of the training set 19 is verified using a test set 29 having, for example, maximally 30% of the spectra (dashed line to the classifiers 13, 14, 15, 16) according to FIG. 4b, so that a classified test set 24 (dashed line from the classifiers of the series 13, 14, 15, 16 to the classified test set 24) is obtained as a result.

[0140] It should also be noted that, by their very nature, the recorded in ovo spectra 4 are generally highly variable. This is caused on the one hand by the inherent variability of biological systems and on the other hand by the sensitivity of Raman spectroscopic measurements.

[0141] External interference of a systematic and random nature results in a high variability of the spectral characteristics and is thus superimposed on the sex-relevant information.

[0142] Furthermore, in the method of Raman spectroscopy, fluorescent light is also present which likewise contains molecular information, but which is also superimposed on the normally much weaker Raman spectroscopic molecular information about the composition of the examined object.

[0143] In FIG. 2a, a schematic illustration of the raw spectra 25 is shown as one of all integrated individual classification methods in FIG. 1 for the four individual classification methods.

[0144] According to FIG. 2a, FIG. 2b, FIG. 2c, and FIG. 2d, at least four classes of signals or defined characteristics can generally be formed: [0145] molecular composition 20, [0146] fluorescence intensity 21, [0147] fluorescence profile 22, and [0148] variation of physical parameters 23,
wherein these defined characteristics 20, 21, 22, 23 are embodied as equally sized, outlined circles and/or circles with differently dashed outlines for the purpose of visual illustration in FIG. 3a, FIG. 3b, FIG. 3c, and FIG. 3d.

[0149] Underlying the classifiers is one mathematical expression each for separating the signals according to the object information 3 (31 female, 32 male).

[0150] Three classes/characteristics 20, 21, 22 of the four classes/characteristics 20, 21, 22, 23 contain sex-relevant information. However, it is not possible to eliminate the variation 23 of the physical parameters from the spectra 4 in such a way that no or only a minor loss of information occurs in the three other classes 20, 21, 22. Thus, because of the equivalence of all of the defined characteristics, the raw spectra 25 have the highest content of all information, but also the highest content of interference. By adding at least one of the indicated data preprocessing methods, for example 26, from the data preprocessing methods 26, 27, 28 with differently evaluated characteristics, the interference is reduced. By using additional data preprocessing methods 27, 28, the original interference is minimized or even eliminated.

[0151] FIG. 2b shows that the in ovo spectra 4 recorded as digital signals are subjected to a linear baseline correction 26, wherein the fluorescence intensity signal 21 (large circle) stands out. At the same time, signals 23 for physical parameters in the spectra are suppressed (small circle). Due to the typically significant intensity differences between fluorescence signals and Raman signals, the information about the molecular composition 20 (small circle) recedes into the background. However, the fluorescence intensity 21 (large circle) itself is potential marker for the sex detection, since male embryos frequently, but not always, exhibit a biochemical blood composition that has a higher fluorescence intensity 21 than female embryos or the blood of female embryos.

[0152] FIG. 2c shows that, by means of the method of spectral normalization 27, for example by means of vector normalization or area normalization, variations in the fluorescence intensity 21 (small circle) can be compensated for and the random influences of physical parameters 23 (small circle) can be minimized. This preferably allows the fluorescence profile 22 (large circle) to be emphasized. At the same time, only a few pieces of information for the molecular composition 20 (small circle), based on the Raman signals, are minimized. Since the fluorescence profile 22, that is, the spectral characteristics of the fluorescence, is determined by the molecular composition 20, sex-relevant information can be emphasized.

[0153] FIG. 2d shows that a most complete possible correction of what is referred to as the background of Raman spectra 28 results in an exclusive emphasis on the Raman bands, that is, on the information about the molecular structure and composition 20 (large circle) of the examined object 2.

[0154] In FIGS. 3a, 3b, 3c, and 3d, one schematic illustration each of the spectra assigned to the individual classification methods (based on the relative wavenumber) is shown in reference to FIGS. 2a, 2b, 2c, 2d.

[0155] At least the spectral preprocessing method 5 with equivalently defined characteristics is added to at least one of the spectra preprocessing methods 6, 7, 8 with differently defined characteristics for the purpose of evaluation.

[0156] FIG. 4b shows a flowchart for the multiple classification method according to the invention having a training set and test set in local separation, but in an algorithmic connection, with a classification-result design from a large number of classifiers. The respective spectral preprocessing is thereby structured thereby such that respectively defined characteristics become more clearly prominent and other characteristics are more strongly suppressed. The spectra 4 preprocessed in such a manner are, according to FIG. 4b, then configured as a variable training set 19, wherein multiple series 13, 14, 15, 16 of classifiers, for example specifically in one series 131, 132, 133, 134, 135, etc., are calculated. Typically, all classifiers of the series 13, 14, 15, 16 are iteratively calculated and validated. Variability means that, without additional preprocessing methods, any desired spectra can be selected for each classifier being calculated. In this manner, multiple classifiers 131, 132, 133, 134, 135 can for example be defined for the first series 13, etc. This also applies for the other series 14, 15, 16. According to FIG. 4b, the selected spectra of the test set 29, for example 30%, are subsequently classified to form a classified test set 24 using all classifiers. The placement of the spectra 4; 25, 26, 27, 28 in a defined class of characteristics (male, female) thereby preferably takes place as a score, or in an expression of a probability for the class association. To obtain a single statement from the classifiers of the series 13, 14, 15, 16 or 131, 132 133, 134, 135, etc., the relations among the classifiers 13, 14, 15, 16; 131, 132, 133, 134, 135 are determined. A simple way of doing so is the calculation of the median 30 or, as stated previously for example, the performance of a cluster analysis.

[0157] In the flow chart illustrated in FIG. 4b, a comparison of each classified spectrum from each form of preprocessing with the characteristic takes place in the node provided with the reference numeral 45={circle around (1)}. The result is thereby only outputted as “true” or “false.”

Example

[0158] The training set 19 comprises 100 spectra. Of these, 60 are selected for the calculation of the classifiers. If four methods of data preprocessing 5, 6, 7, 8 are used, there are 60×4=240 classified spectra. From the comparison with the list of characteristics there thus result 240 statements of either “true” or “false.” This result is, for example, achieved in the set 129th iteration step.

[0159] In the node designated by the reference character 46={circle around (2)}, an evaluation of the classified spectra takes place with regard to a set criterion or multiple set criteria. An accuracy bound or a maximum number of iterative steps, for example, serve as criteria. The criteria can linked by an AND or OR logical operation.

Example

[0160] Of the 240 possible statements, 205 are “true” and 35 are “false.” There thus results a correctness of 85% for the training set.

[0161] Before the classification begins, the following are set as criteria:

1. correctness >80% and
2. a maximum number of iterations: 1000.
That is, after the completed

[0162] 129th iteration step <1000 [0163] and with
an obtained correctness of 85%>predefined correctness.

[0164] In the case of a logical AND operation, it is possible to arrive at “bad” (wherein the classifiers are stored as a best intermediate result, however) and in the case of a logical OR operation at “good.”

[0165] If the number of the predefined classifiers being defined has been reached at the junction indicated by the reference character 47={circle around (4)}, all classifiers (each of which has namely led to the best result at node {circle around (2)}=45) are passed to the validation of the entire training set 19.

Example

[0166] It is predefined that 30 classifiers per data preprocessing 5, 6, 7, 8 are to be calculated and result in a multiple classification. Thus, 30×4=120 classifiers are passed to validation.

[0167] At the node/comparison junction indicated by the reference character 48={circle around (3)}, a collective evaluation of the classification of all spectra in the training set 19 takes place according to the leave-one-out or cross-validation method.

[0168] In the event of a “passed test,” the classifiers are passed to the classification of the “unknown” spectra of the test set 29.

[0169] If the test is not passed, a classification according to the predefined criteria is not possible.

[0170] At the node/comparison junction indicated by the reference character 49={circle around (3)}, a final evaluation of the classification of the spectra in the test set 24 is performed with the aid of the known characteristics of the spectra.

Example

[0171] The test set 29 comprises 50 spectra. These spectra were respectively classified with 120 classifiers, that is, 120 probabilities for the class association are assigned to each spectrum. From this, the association with a class follows according the median or cluster analysis. This is the result of the multiple classification for each individual spectrum. If, for example, 41 of the 50 spectra are correctly classified, this results in a correctness of 82% for the entire test set 24.

[0172] From the comparison with the list of characteristics, the method 1 of multiple classification created in such a manner is conclusively evaluated. The method is thus created and can then be used for spectra without knowledge of the characteristics.

[0173] FIG. 5 shows a schematic illustration of a probability/classifier quantity bar graph 38 for 20 classifiers according to FIG. 1 for displaying an egg 2 identified as female. The unshaded end regions/front faces 33 of bars 34 can thereby be associated with the female sex above a certain line the cut-off 42 and the bars 35 with the shaded end regions/front faces 36 can be associated with the male sex below the cut-off 42. In FIG. 5, the cut-off value is at 0.5 and the median 30 therein has a value of 0.72. Therefore, the egg 2 is clearly classified as “female.”

[0174] FIG. 6 shows the top view related to the perspective bar graph, which view is shown on a color display as a classification result image 37 with the majority of unshaded end regions/front faces 33 for the “female” object information 31. On the color display, the unshaded front faces 33 can be embodied in red and the shaded front faces 36 can be embodied in blue, so that a color visual display of the evaluation of the sex can also be made.

[0175] The shaded squares can thereby be depicted in a blue color and the unshaded squares in a red color. The few blew squares indicate the male object information 32. The more numerous red squares indicate the female object information 31. Since the red squares are more numerous, the sex of the incubated chicken egg 2 can be identified as a female characteristic 31.

[0176] FIG. 7 shows an illustration of the calculated median 30 at 10 classifiers in relation to the quantity of the twenty classifiers for an egg 2 according to FIG. 1, with a series of five classifiers 131, 132, 133, 134, 135 each per group for four groups 9, 10, 11, 12. In the bar graph and median illustration, 17 classifiers thereby result in the display of a female egg 2. The overall classification result 18 can be stated using the calculated median 30.

[0177] FIG. 8 shows a schematic illustration of an additional exemplary probability/classifier quantity bar graph 39 for 120 classifiers, illustrated as bars, for the display of an egg 2 identified as female. The cut-off 42 once again shows the boundary between the “male” characteristic and the “female” characteristic. Here, too, the bars 34 (31) ending above the cut-off 42 are shown unshaded on their front faces, and the bars 35 (32) ending below the cut-off 42 are shown shaded.

[0178] For an egg 2 with a male sex characteristic 32, a different bar graph can be embodied, wherein in this case the front faces located above the cut-off 42 of the embodied bars 35, being in the majority compared to the unshaded front faces of the bars 34, are shaded (not shown).

[0179] FIG. 9 shows an illustration of the calculated median 30 in relation to the quantity of 120 total classifiers according to the bar graph in FIG. 8 in a probability/classifier quantity graph for an egg 2 with a female sex characteristic 31, with classifiers sorted by ascending points. Here, the median 30 is at half of the 120 determined classifiers and has a probability value of 0.95.

[0180] The classification units/groups 9, 10, 11, 12 contained in an evaluation unit for defining the object information in the form of binary sex characteristics 31, 32 female or male of fertilized and unincubated and incubated eggs 2 function as follows:

[0181] The functional principle will now be explained.

[0182] After the spectral preprocessing 5, 6, 7, 8, multiple classifiers of the series 13, 14, 15, 16 are calculated from each class 25, 26, 27, 28. The definition of the classifier series 13, 14, 15, 16 takes place according to an algorithm which, in a kind of tandem method, first selects spectral regions R.sub.S from the coordinate of the relative wavenumbers and then classifies the intensity values of the selected regions R.sub.S by means of discriminant analysis.

[0183] In a comparison with the training data for the class association, another selection of spectra classes and the classification of the intensity values occur in a repeated step. This cycle is repeated iteratively until an accuracy that can no longer be improved is reached, wherein the stopping criterion can be predefined.

[0184] The risk of overtraining, and therefore reaching high instabilities, grows as the number of spectral classes 25, 26, 27, 28 used for the classification increases. It is therefore desirable to use only a few (3 to maximally 20) spectral classes to create the classifier series 13, 14, 15, 16. However, because the sex-relevant information is distributed, albeit varyingly, across the entire spectral range, essential spectral information would actually remain unused if only one classifier were to be created. For this reason, it is expedient that multiple (10 to 20) classifiers in the series 13, 14, 15, 16 are calculated per group of data preprocessing 5, 6, 7, 8.

[0185] This has the advantage that, on the one hand, the accuracy of the classification is improved, solely based on the fact that the greatest possible amount of spectral information is incorporated, and that on the other hand the robustness, that is, the stability, is increased since multiple classifiers of the series 13, 14, 15, 16 support the assignment and individual erroneous assignments are compensated for.

[0186] The hardware units assigned to the classifications operate identically for all four groups 9, 10, 11, 12. Thus, instead of the four units controlled in parallel, it is also possible to use only one which creates the series 13, 14, 15, 16 of the classifiers serially in a predefined order. During the setting of the number of calculated classifiers N.sub.G in the series 13, 14, 15, 16 for each group 9, 10, 11, 12, the scale of the spectral data points v.sub.S and the doubled half-width of the spectral regions w.sub.S and also the number of selected spectral regions R.sub.S must be taken into account:

[00003]$\begin{array}{cc}{N}_G=\frac{v_S}{2w_S.Math.R_S}& \left(I\right)\end{array}$

With the equation (I), it is ensured that each data point v.sub.S can be selected with equal probability.

[0187] In FIG. 10a and FIG. 10b, twenty classifiers per data preprocessing 25 are shown using the example of the raw spectra (intensity/wavenumber curves). For this purpose, FIG. 10 shows an enlarged section of the associated male intensity/wavenumber curves. The number of the spectral data points v.sub.S covers the entire spectral range between 500 cm.sup.−1 and 2750 cm.sup.−1. The number of the selected spectral regions R.sub.S in FIG. 10a is R.sub.S=8, with R.sub.S1, R.sub.S2, R.sub.S3, R.sub.S4, R.sub.S5, R.sub.S6, R.sub.S7, and R.sub.S8.

[0188] According to equation (I), twenty classifiers N.sub.G can be calculated therefrom for the raw spectrum 25. With four data preprocessing methods 25, 26, 27, 28, this means a total of 80 classifiers generated (20 classifiers/group×4 groups).

[0189] According to the enlarged section in FIG. 10b, the data points v.sub.S can also be additionally weighted. For this purpose, a weighting chart 40 is shown in FIG. 10c, from which it is evident that the highest weighting value is assigned to the middle data point 41.

[0190] This can be performed both with the male spectrum and also with the female spectrum.

[0191] The evaluation 17 and the classification of the results assigned to the classifiers of the series 13, 14, 15, 16 are carried out in an evaluation unit and conducted until a classification result 18 (30) is produced.

[0192] Ultimately, a classification result 18 is outputted in the form of the median 30, which in the sex determination of chicken eggs represents the binary sex information 31, 32 (male or female) with the highest probability.

[0193] In general, the method according to the invention can be completed with the following detailed steps: [0194] acquisition and recording of the spectra by means of at least one optical device having at least one spectrometer and/or additional detectors; [0195] generating digitized signals in the form of data points, and storing the detected spectra in storage units of classification units of an evaluation unit; [0196] spectral preprocessing, in that the stored spectra are individually evaluated in the individual storage units and the associated digitized evaluated signals are made available for further processing; [0197] separating the preprocessed spectra as a training set and as a test set; [0198] configuring the preprocessed spectra as a training set and a test set separate from the training set;

[0199] wherein according to the invention at least a [0200] calculation of the classifiers of the series of the integrated individual classification methods, with an incorporation of iterative methods and a validation in the classification groups; [0201] classification of the evaluated spectra of the training set with all classifiers of the series; [0202] placement of the spectra of the training set in a class of object information with an expression of a probability for the class association; [0203] calculation of the median or performance of a cluster analysis to show the probability result of the training set object information associated with a class; [0204] classification of the evaluated spectra of the test set with all classifiers of the series; [0205] placement of the spectra of the test set in a class of object information with an expression of a probability for the class association; and [0206] calculation of the median or performance of a cluster analysis to show the probability result/classification result of the test set object information associated with a class.

[0207] are carried out.

[0208] It should also be noted that, by their very nature, the recorded spectra 4 are generally highly variable. This is based on the one hand on the inherent variability of biological systems and on the other hand on the sensitivity of Raman spectroscopic measurements. External interference of a systematic and random nature results in a high variability of the spectral characteristics and is thus superimposed on the characteristic-relevant information. Furthermore, in the method of Raman spectroscopy, fluorescent light is also present which likewise contains molecular information, but which is also superimposed on the normally much weaker Raman spectroscopic molecular information about the composition of the examined object 2.

[0209] On the basis of these preliminary remarks and FIG. 1, a further exemplary embodiment with more than two characteristics of object information will be explained. FIG. 1 shows in the schematic block diagram a method 1 according to the invention for classifying spectra 4 of an object 2 having complex information content, in particular optical molecular spectra 4, for assigning object information/characteristics 3 for a probabilistic definition, for example, of binary object information 31, 32 or of four pieces of object information 51, 52, 53, 54.

[0210] Tissue samples, for example, brain tumors, can also be used and applied as the objects 2 to be examined, and in place of the binary characteristic information 31, 32, four different characteristics 3 can for example also be selected and defined with 51, 52, 53, 54, for example [0211] characteristic 51: healthy tissue, [0212] characteristic 52: tumor tissue with tumor grade I and II according to the histological classification model of the World Health Organization (WHO), [0213] characteristic 53: tumor tissue with tumor grade III and IV according to the WHO, and [0214] characteristic 54: necrotic tissue.

[0215] The acquisition and recording of the backscatter radiation from the tissue sample occurs by means of at least one optical device as described, for example, in the publication DE 10 2014 010 150 A1. The recorded backscattering spectra 4 are digitized and stored in an evaluation unit. The data preprocessing occurs, for example, through three different methods 5, 6, 7; the data sets thereby obtained can, for example, respectively contain raw spectra, normalized spectra, and spectra with a non-linear baseline correction, wherein the stored spectra are individually evaluated in the individual storage units and the associated digitized evaluated signals are made available for further processing. The preprocessed spectra are configured as a training set, wherein according to the invention, a calculation is performed of the classifiers of the series of the integrated individual classification methods, with an incorporation of iterative methods and a validation in the classification units. Furthermore, the classification of the evaluated spectra of the test set takes place with all classifiers of the series, and the placement of the tissue spectra in a class of object information takes place according to the characteristics 51 through 54 with an expression of a probability for the class association. The classification is evaluated by the calculation of the mean or by means of a cluster analysis, and the probability result/classification result of the test set object information associated with a class is shown. This means that, for each recorded spectrum of a tissue sample, a score is calculated by means of multiple classification, which score lies in one of the 4 probability ranges according to set cut-offs, which ranges correspond to the histological findings of the following characteristics: 51—healthy/52—WHO I, II/53—WHO III IV/54—necrosis.

[0216] An apparatus for classifying spectra 4 of objects 2 having complex information content, preferably with objects in the form of chicken eggs 2 for a definition of binary egg information 31, 32 female or male in which apparatus the aforementioned method is implemented and which is to a large extent embodied in accordance with the block diagram (box drawing) in FIG. 1, can comprise at least the following units: [0217] at least one detecting optical device having at least one spectrometer and/or additional detectors for the acquisition and recording of the spectra 4; [0218] a unit for generating digitized signals in the form of data points by which the spectra are manifested 4; [0219] storage units for storing the recorded spectra 4 in the classification units/groups 9, 10, 11, 12 of an evaluation unit comprising the classification units; [0220] units for spectral preprocessing 5, 6, 7, 8 in which the stored spectra 4 are individually evaluated in the individual storage units and the associated digitized evaluated signals are made available for further processing; [0221] training sets 19 for configuring and using the preprocessed spectra 4; 25, 26, 27, 28; [0222] at least one classification unit for the groups 9, 10, 11, 12 for calculating the classifiers of the series 13, 14, 15, 16 for the integrated individual conventional classification methods 25, 26, 27, 28, with an incorporation of iterative methods and a validation in the classification units; [0223] test sets 29 for classifying the evaluated spectra 4 with all classifiers of the series 13, 14, 15, 16; [0224] a unit for placing the spectra 4 in a, for example, binary class—male 32 or female 31—of object (egg) information with an expression of a probability for the class association; [0225] an evaluation unit for calculating the classification result 18, for example, in the form of the median 30 or after performing a cluster analysis to show the probability result of the object (egg) information 31, 32 associated, for example, with one of the binary classes female or male.

[0226] A similar apparatus can be constructed for the multiple classification method with the four characteristics 51, 52, 53, 54 or with additional predefined characteristics.

LIST OF REFERENCE NUMERALS

[0227] 1 Method in a box drawing [0228] 2 Object/egg [0229] 3 Object information/characteristics [0230] 4 Recorded spectra [0231] 5, 6, 7, 8 Preprocessing [0232] 9, 10, 11, 12 Classification [0233] 13, 14, 15, 15, 16 Series of classifiers [0234] 131, 132, 133, 134, 135, 136 Classifier [0235] 17 Evaluation [0236] 18 Classification result [0237] 19 Training set [0238] 20 Molecular composition [0239] 21 Fluorescence intensity [0240] 22 Fluorescence profile [0241] 23 Variation of physical parameters [0242] 24 Classified test set [0243] 25, 26, 27, 28 Preprocessed spectra [0244] 29 Test set with preferably 30% of the spectra selected [0245] 30 Median [0246] 31 Female object information/characteristic [0247] 32 Male object information/characteristic [0248] 33 Unshaded front face, assigned to the female sex [0249] 34 Bar for the female sex [0250] 35 Bar for the male sex [0251] 36 Shaded front face, assigned to the male sex [0252] 37 Classification result image [0253] 38, 39 Bar graph [0254] 40 Weighting chart [0255] 41 Middle data point in the region of a spectrum curve [0256] 42 Set cut-off [0257] 43 First histogram of the cluster analysis [0258] 44 Second histogram of the cluster analysis [0259] 45 ={circle around (1)} comparison node [0260] 46 ={circle around (2)} comparison node [0261] 47 ={circle around (5)} comparison node [0262] 48 ={circle around (3)} comparison node [0263] 49 ={circle around (4)} comparison node [0264] 50 Classifier according to the prior art [0265] 51, 52, 53, 54 Characteristic