JP4985480B2 - Method for classifying cancer cells, apparatus for classifying cancer cells, and program for classifying cancer cells - Google Patents

Description

  The present invention relates to a method for classifying cancer cells, a device for classifying cancer cells, and a program for classifying cancer cells, and more particularly to self-organization of protein data acquired by a laser scanning cytometer (LSC). The present invention relates to a method for classifying cancer cells using a conversion map (SOM), an apparatus for classifying cancer cells, and a program for classifying cancer cells.

  In the fields of medicine, pharmacy, and biology, cancer tissue and cells are removed from patients and analyzed at the tissue and cell level for cancer testing, treatment policy determination, and treatment and drug development. Has been done. Cancers can be classified to some extent based on their morphology (determination of tissue type), but even if cancers of the same tissue type are removed from the same site, the speed of progression and response to therapeutic agents Often it is not uniform and a more detailed classification is required in the medical field than the histological classification determined by morphological features.

  Regarding the examination and classification of malignant cells such as cancer, there are those disclosed in the following literature.

  Patent Document 1 describes a system and method for detecting malignant cells from an image obtained with a microscope and a CCD camera by using a classification device. Accordingly, cancer cells can be detected from visual information such as cell shape, but cancer cells having the same shape cannot be classified.

  Patent Document 2 describes that a b coordinate value of a Lab color space is obtained from imaged cell image data, and the cell is classified using a feature parameter related to the color of the cell using the coordinate value. However, in this method, the tissue is classified according to the color and shape of the cells, but the cells having the same shape cannot be classified.

  Patent Document 3 describes using a self-organizing map as a method for determining a health state and displaying the health state on the self-organizing map, but does not classify cells.

  Patent Document 4 describes that an image acquired by MRI can be easily searched and accessed using a self-organizing map for an image acquired by MRI. It is not intended for images, and features of cancer cells cannot be extracted and classified.

There is a method called cytometry as a method for quantifying intracellular substances. This is a method of quantifying the size of cells and intracellular substances by irradiating laser to cells that become samples stained with molecular probes such as fluorescent dyes alone, fluorescent dye-labeled antibodies and DNA, and measuring the amount of fluorescence emitted. Is what you do. In recent years, a laser scanning cytometer (LSC) has been developed as one of the cytometry techniques, and it has become possible to measure the degree of aggregation of intracellular substances. This LSC is used to acquire data on a wide variety of cancer cells, conduct analysis focusing on the amount of protein and its degree of aggregation, and attempt to classify cancer cells that are difficult to classify with a microscope. No effective classification method has been found so far.
Special table 2001-512824 gazette JP 2004-340738 A JP 2003-263502 A JP 2006-235971 A

  In conventional classification of cancer cells using images taken with a CCD camera, only visual information about the cancer cells can be obtained and cancer cells can be detected, but feature extraction is limited. It was. In addition, cancer cells having almost the same shape could not be further classified. For this reason, it has been required to obtain more appropriate information based on the characteristics of cancer cells so that more detailed classification of cancer cells having the same shape can be performed based on the characteristics.

  The present invention has been made to solve the above-mentioned problems, and a method for classifying cancer cells according to the present invention is performed by processing data about cancer cells extracted by a laser scanning cytometer. A method for classifying cells, wherein data of protein mass and protein aggregation degree are cut out as necessary data from data acquired by a laser scanning cytometer for cancer cells, and a scatter diagram is created. The input data is input to the input layer of a network having an input layer and a map layer, each of which is a collection of neurons, and is input between the input layer and the map layer. A self-organizing map is formed by learning about connection weights and performing topological mapping, and the formed self-organizing map is formed. And carrying out the classification of cancer cells to determine the characteristics of the cancer cells by the distribution form of the protein in, it is made of.

  An apparatus for classifying cancer cells according to the present invention is an apparatus for classifying cancer cells by processing data about cancer cells extracted by a laser scanning cytometer, and the laser scanning cytometer for cancer cells. Scatter plot creating means for cutting out the data of protein mass and protein aggregation degree as necessary data from the data acquired in step 1 to create a scatter plot, and input data for lowering the data indicating the scatter plot and creating input data A self-organizing map by creating a topological mapping by learning about the connection weight between the input layer and the map layer, having a network comprising an input layer and a map layer, each of which is a collection of neurons; And a self-organizing map forming means for forming

  In addition, the program for classifying cancer cells according to the present invention is a program for classifying cancer cells on a computer by processing data on cancer cells extracted by a laser scanning cytometer. However, from the data obtained with a laser scanning cytometer for cancer cells, the data of protein mass and protein aggregation degree are cut out as necessary data to create a scatter diagram, and the data indicating the scatter diagram is reduced in dimension. Create input data, input the input data to an input layer of a network having an input layer and a map layer, each of which is a collection of neurons, and learn about connection weights between the input layer and the map layer to obtain topological A self-organizing map is formed by mapping.

  In the present invention, cancer cell data acquired with a laser scanning cytometer can be extracted and classified by performing mapping using learning using a self-organizing map, and the same shape can be obtained. Even cancer cells can be classified in more detail based on the characteristics of the cells.

  In the classification of cancer cells according to the present invention, cells are classified using a self-organizing map (SOM) as a classification technique based on data analysis. The SOM is one of unsupervised learning models in a neural network, and automatically finds out the similarity of input data by learning, and forms a topological map in which similar inputs are arranged near the map layer. When forming a topological map, image data representing the relationship between the amount of protein and the degree of aggregation is created from cancer cell data extracted by a laser scanning cytometer (LSC), and the distance information is mapped by learning to classify cells. I do.

The classification of cancer cells according to the present invention classifies cells by performing analysis processing on data acquired from a sample of cells by detection with a laser scanning cytometer. First, a laser scanning cytometer will be described.
[Laser scanning cytometer]
A laser scanning cytometer (hereinafter referred to as LSC) is a device that detects scattered light and fluorescence emitted from a cell by irradiating a cell sample with a laser. As the sample, cells cultured or pasted on a slide glass are used. These are fluorescently labeled (stained) by a method according to the intracellular substance (DNA, protein, non-peptide substance, drug, etc.) to be measured. The cells are scanned with a laser, the brightness of the fluorescence emitted from the cells is detected with a photosensor, and the detected signal is converted to measure the size of the cell, the amount of intracellular material, localization (location information), etc. it can.

In LSC, the amount of fluorescence is measured in units of minute pixels, and when measuring the amount, area, etc. of the fluorescent material, position information is also acquired at the same time. The amount of fluorescence of all the pixels in the region recognized as one cell is added to obtain the amount of fluorescence (Integral) in one cell, and the pixel showing the highest amount of fluorescence among them is defined as the maximum pixel (Max pixel). In other words, the maximum pixel is an index representing the aggregation degree of the fluorescent substance, that is, the aggregation degree of the intracellular substance to be measured. The cancer cell cycle can be easily identified by the value of the maximum pixel value.
[Cell samples]
Antibodies have the property of reacting specifically with antigens. If this property is utilized, the intracellular substance can be fluorescent immunostained by antigen-antibody reaction. In the LSC, measurement is performed on cells attached on a cell array as shown in FIG. 1 or cultured on the cell array. The cell array has a large number (for example, 50) of concave spots having a diameter of about 2 mm on a slide glass, and the hatched portion in the figure is raised by printing with non-fluorescent ink. The portion where the spots are arranged is about 16 mm long and 30 mm wide. There are usually about 100 to 500 cancer cells in one spot of the cell array. In the conventional method, one type of cell is measured with one slide glass. However, data of many cancer cells can be acquired at a time by using a cell array.
[LSC data]
In the present invention, LSC2 (manufactured by Olympus Corporation) is used as the LSC. Data obtained by LSC measurement is data in text format, and an example is shown in FIG. The data in the form of FIG. 2 will be described briefly. The first part represents LSC device settings, in which the type of laser, a certain range of cells on the cell array, and the like are set. Subsequent to the device setting data, data such as No, position information, diameter, area, circumference, and DNA index (DI) related to each cell are described.

  When performing the analysis for cell classification, if the initial value is used for the device setting, the data for each cell may be considered without using the device setting data. Further, regarding the data about each cell, there is unnecessary data (referred to as garbage) in which cells overlap each other. However, the unnecessary data is deleted and necessary data is acquired.

  In order to obtain the necessary data, data is extracted for each spot in the cell array by extracting the necessary data such as protein mass from the text format data in a selected format using a data extraction program (for example, AWC: Any Where Cyte). Is extracted and the data is stored in csv format. An example of the data in the csv format is as shown in FIG.

  In the data shown in FIG. 3, the amount of protein (Integral) and the degree of protein aggregation (Max Pixel) are used for cell classification. The amount of protein is the amount of protein present in the cancer cell, and the degree of aggregation is an amount representing how the protein is present in the cancer cell. The degree of protein aggregation can be newly measured by using LSC.

Next, a scatter diagram is created from the cut out data as shown in FIG. The scatter diagram is such that data is extracted for each spot of the cell array, and the distribution characteristic is expressed by using the amount of protein and the degree of aggregation as parameters for a specific protein, and examples thereof are shown in FIGS. ). Each point in the figure represents data on cancer cells. In these scatter plots, it can be seen that there is a difference in the shape, slope, density, etc. of the cluster of points due to the difference in protein, but when different proteins have similar shapes, the same type of cancer cell, the same type of protein However, there are various shapes such as different shapes, and data cannot be classified simply by looking at the scatter diagram. Therefore, data is classified using SOM.
[Self-Organizing Map (SOM)]
(1) Structure of SOM network FIG. 5 shows the structure of the SOM network. Each network consists of two layers, an input layer and a map layer, each of which is a collection of neurons. There is no neuron connection in the layer, and the neuron is fully connected between the input layer and the map layer. In the map layer, neurons are usually arranged in two dimensions in order to visually see the output.

Input vector x (t) = [x ₁ (t),..., X _i (t),..., X _n (t)] at time t
, The neuron of the map layer has a connection weight x _j (t) = [w _j1 (t),..., W _ji (t) _,.
(W _ji is a load between the i-th neuron in the input layer and the j-th neuron in the map layer.) The input from the input layer is received through, and learning described later is repeated according to the learning algorithm. As a result, nearby neurons in the map layer react to similar inputs. That is, the network performs topological mapping.
(2) Learning algorithm in SOM Topological mapping is performed by learning about the connection load between neurons between the input layer and the map layer. The learning algorithm is shown below.
a. Network initialization The initial value of the bond weight between the input layer and the map layer is set by a random number.
b. Input of input vector Input vector x = (x ₁ ,..., X _i ,..., X _n ) is input to the input layer.
c. Calculating the distance between the input vector and the connection weight vector Calculate the distance between the input vector and the connection weight vector of each neuron in the map layer. The distance d _j between the coupling weight vector of the j-th neuron of the input vector and map layer

で与えられる。
ｄ．勝者ニューロンの決定
距離ｄ_ｊが最小となるニューロン、すなわち入力ベクトルに最も近い結合荷重ベクトルをもつマップ層でニューロンを選択する。このニューロンを勝者ニューロンと呼び、ｊ^＊とする。
ｅ．結合荷重と各パラメータの更新
勝者ニューロンとその近傍領域内の全てのニューロンの結合荷重を
Δｗ_ｊｉ＝αｈ（ｊ，ｊ^＊）（ｘ−ｗ_ｊｉ）・・・・・・（２）
に基づいて更新する。この近傍関数ｈ（ｊ，ｊ^＊）は

Given in.
d. Determining the winner neuron Select the neuron with the smallest distance d _j , that is, the map layer with the connection weight vector closest to the input vector. This neuron is called the winner neuron and is designated j ^* .
e. Connection weight and update of each parameter The connection weights of the winner neuron and all the neurons in the neighborhood are expressed as follows: Δw _ji = αh (j, j ^* ) (x−w _ji ) (2)
Update based on. This neighborhood function h (j, j ^* ) is

で定義される。αは学習率係数であり、学習の経過とともに減少させるようにする。同様に、σも学習の経過とともに減少させる。
ｆ．反復
ｂに戻り、ｂからｅの過程を反復する。がん細胞、蛋白質についてＬＳＣで取得されたデータを分散図データとし処理して得られた入力データについてＳＯＭによる学習を行い、その結果として、がん細胞の特性を表すマップ図が得られる。
〔システムの構成〕
ＬＳＣで取得されたデータを処理して入力データとしＳＯＭによる学習を行う過程をフローで示すと図６のようになる。ＬＳＣで取得されデータ処理を行って得られた散布図のデータについて、データ量を減らすために特徴を保持したまま低次元化する平滑化処理、グレースケール変換を行い、標本化を行う。この処理により例えば１０×１０＝１００次元まで次元を落とし、入力ベクトルを作成する。入力ベクトルの各要素は標本化後の散布図の画像の１ブロックに対応し、０〜２５５の値をもつ。

Defined by α is a learning rate coefficient, and is decreased with the progress of learning. Similarly, σ is decreased with the progress of learning.
f. Return to iteration b and repeat the process from b to e. Learning by SOM is performed on the input data obtained by processing the data acquired by the LSC for cancer cells and proteins as scatter diagram data, and as a result, a map diagram representing the characteristics of the cancer cells is obtained.
[System configuration]
FIG. 6 is a flowchart showing a process of learning by SOM by processing data acquired by LSC and using it as input data. The scatter diagram data obtained by the data processing obtained by the LSC is sampled by performing a smoothing process and a gray scale conversion for reducing the dimensions while maintaining the characteristics in order to reduce the data amount. By this processing, for example, the dimension is reduced to 10 × 10 = 100 dimensions, and an input vector is created. Each element of the input vector corresponds to one block of the image of the scatter diagram after sampling, and has a value of 0 to 255.

  The map layer in the SOM network is, for example, a two-dimensional map of 20 × 20 = 400 neurons, and each neuron in the map layer has a 100-dimensional connection weight vector. This corresponds to the entire image of the scatter diagram after sampling.

  A device for classifying cancer cells by processing data acquired by the LSC and using it as input data to perform learning by SOM has the configuration shown in FIG. This apparatus includes an input data generation unit 10 that roughly processes text data acquired by an LSC to be input data for SOM generation, and an SOM generation unit 20 that learns about the input data. .

  The input data generation unit 10 includes a data cutout unit that cuts out necessary data from text data acquired by the LSC, a data conversion unit that lowers the cutout data and converts it into cvs data, and converted cvs data A scatter diagram creating unit for creating a scatter diagram from the scatter diagram, and an input data generating unit for reducing the dimensionality and sampling of the data of the scatter diagram as an input signal.

The SOM generation unit 20 has a configuration of a neural network that inputs and learns the input data generated by the input data generation unit 10, and includes an input layer and a map layer, each of which is a collection of neurons. The SOM is formed by performing topological mapping by learning about the coupling load between the two.
[Example of analysis]
It shows about the result of having learned by SOM about five kinds of cancer cells and four kinds of proteins. The cancer cells are KURAMOCHI, MKN1, MS-1, SW-13, and WiDr, and the proteins of each cancer cell are L-26, LCA, RET, and p27. Learning by SOM is performed 15 times for 6 scatterplot images per type of protein, which is defined as one cycle, and is repeated 30 cycles. Learning about one type of protein is called sub-learning, and the number of times is called sub-learning. Each parameter in learning and the number of learning were set as shown in Table 1.

また、学習終了後に得られたマップを図８〜１２に示す。マップ図では、１種類のがん細胞についての４種類の蛋白質データをそれぞれについての勝者ニューロンとして表示している。４種類の蛋白質をＡ（Ｌ−２６），Ｂ（ＬＣＡ），Ｃ（ＲＥＴ），Ｄ（ｐ２７）で表し、１種類の蛋白質データについての６枚のデータにそれぞれ番号１〜６を対応づけている。図中、Ｘは２種類以上の蛋白質に対する勝者ニューロンを示しており、例えばＡ１は蛋白質Ｌ−２６に関する１番目のデータの勝者ニューロンを示している。

Moreover, the map obtained after completion | finish of learning is shown to FIGS. In the map diagram, four types of protein data for one type of cancer cell are displayed as winner neurons for each type. Four types of proteins are represented by A (L-26), B (LCA), C (RET), and D (p27), and numbers 1 to 6 are associated with six pieces of data for one type of protein data, respectively. Yes. In the figure, X indicates a winner neuron for two or more types of proteins. For example, A1 indicates a winner neuron of the first data regarding the protein L-26.

  X is seen in all maps resulting from mapping. This indicates that the data of the scatter plots are similar for two or more types of proteins, and it can be seen from this result that a characteristic protein exists. For example, in FIG. 8, A1 to A4, D1, and D2 are mapped to different locations from X, but this means that there is a difference in the scatter diagram even for the same type of cancer cell data. Show.

  When acquiring cancer cell data, the time required for acquiring cell cycle and protein data may differ, and from this, L-26 indicated by A1 to A4 and p27 indicated by D1 and D2 are different. It is considered that this is characteristic protein data that is different from others because it changes with time in cancer cells KURAMOCHI.

  Table 2 shows characteristic proteins that change with time in each cancer cell based on the results for the five types of cancer cells shown in FIGS.

このように、がん細胞の試料からＬＳＣにより取得されたデータについてＳＯＭでの学習による結果として得られたマップ図によりがん細胞ごとに時間的に変化することで他と区別される特徴的な蛋白質が存在することがわかり、これにより同種のがん細胞を特徴的な蛋白質により分類できることが示される。

As described above, the data acquired by the LSC from the sample of cancer cells is distinguished from others by changing with time for each cancer cell by the map obtained as a result of learning by SOM. It can be seen that proteins exist, indicating that the same type of cancer cells can be classified by characteristic proteins.

The present invention is a method for classifying cancer cells by forming a self-organizing map as a result of learning by SOM on cancer cell data acquired by LSC,
It is also an apparatus for classifying cancer cells, which includes an input data generation unit that processes text data acquired by the LSC to be input data for SOM generation, and an SOM generation unit that learns about the input data.

  Furthermore, performing the necessary processing on the cancer cell data acquired by the LSC and then learning with the self-organizing map is configured as a device equipped with an original data processing means, and is executed by a general-purpose computer. Data that shows the scatter diagram can be created by cutting out the protein mass and protein aggregation data as necessary data from the data acquired with a laser scanning cytometer for cancer cells. The input data is input to the input layer of the network having the input layer and the map layer, each of which is a collection of neurons, and learning about the coupling load between the input layer and the map layer is performed. A self-organizing map is formed by performing topological mapping.

It is the figure which looked at the cell array from the upper part. It is a figure which shows the example of the data by LSC. It is a figure which shows the example of the data which cut out and converted the data of FIG. 2 to csv format. It is a figure which shows the example of a scatter diagram. It is a figure which shows the structure of a SOM network. FIG. 6 is a flowchart illustrating a process of forming an SOM map from LSC data according to the present invention. It is a figure which shows roughly the structure of the apparatus which forms a SOM map from the data of LSC by this invention. It is a figure which shows the example of the produced SOM map. It is a figure which shows the example of the produced SOM map. It is a figure which shows the example of the produced SOM map. It is a figure which shows the example of the produced SOM map. It is a figure which shows the example of the produced SOM map.

Explanation of symbols

10 Input data generation unit 20 SOM creation unit

Claims (3)

A method of classifying cancer cells by processing data about cancer cells extracted by a laser scanning cytometer,
Cutting out the data of protein mass and protein aggregation degree as necessary data from the data acquired with a laser scanning cytometer for cancer cells, and creating a scatter plot,
Reducing the data representing the scatter diagram to create input data;
The input data is input to an input layer of a network having an input layer and a map layer, each of which is a collection of neurons, and learning is performed on the connection weight between the input layer and the map layer to perform topological mapping. Forming an organizational map;
Classifying the cancer cells by distinguishing the characteristics of the cancer cells according to the protein distribution in the self-organizing map formed,
A method for classifying cancer cells characterized by comprising: An apparatus for classifying cancer cells by processing data about cancer cells extracted by a laser scanning cytometer,
Scatter plot creation means for cutting out data on protein mass and protein aggregation degree as necessary data from data acquired with a laser scanning cytometer for cancer cells,
Input data creating means for creating input data by reducing the data representing the scatter diagram;
Self having a network composed of an input layer and a map layer, each of which is a collection of neurons, and forming a self-organizing map by performing topological mapping by learning about connection weights between the input layer and the map layer An organizing map forming means;
An apparatus for classifying cancer cells, characterized by comprising: A program for performing on a computer to classify cancer cells by processing data about cancer cells extracted by a laser scanning cytometer, which was obtained with a laser scanning cytometer Cut out the data of protein mass and protein aggregation degree as necessary data, create a scatter diagram, reduce the data showing the scatter diagram to create input data, and each input data is a collection of neurons A self-organizing map is formed by inputting into an input layer of a network having an input layer and a map layer, learning about a coupling load between the input layer and the map layer, and performing topological mapping. A program for classifying cancer cells characterized by.

Images