CN107064019B - The device and method for acquiring and dividing for dye-free pathological section high spectrum image - Google Patents

Abstract

A device and method for collecting and segmenting hyperspectral images of non-stained pathological slices. The middle part of the support has a slice sample platform, and the computer automatically collects and processes the hyperspectral images of unstained pathological slices to obtain the segmentation results of lesion regions; the present invention is based on tissue For the spectral difference caused by lesions, use a personal computer to synchronously control related modules, collect spectral sequence images of unstained pathological tissue sections, preprocess and superimpose to generate corresponding three-dimensional hyperspectral data, and develop based on this data combined with the current popular neural network classification ideas The spectral classification algorithm is used to identify and segment lesion areas, which speeds up the identification rate and efficiency of pathological tissue sections, avoids possible manual errors in the staining process, and reduces the time required for section production. The subjectivity brought about by discrimination can provide better assistance for pathologists to detect pathological slides.

Description

Device and method for hyperspectral image acquisition and segmentation of non-stained pathological sections

technical field

The invention belongs to the field of detection devices, and in particular relates to a device and method for rapid collection and segmentation of non-stained pathological sections based on hyperspectral data.

Background technique

Pathological examination has been widely used in clinical work and scientific research. In the clinical aspect, the postmortem pathological examination and surgical pathological examination are mainly carried out. The purpose of surgical pathological examination is to clarify the diagnosis and verify the preoperative diagnosis and improve the level of clinical diagnosis; secondly, after the diagnosis is confirmed, the next treatment plan can be determined and the prognosis can be estimated, so as to improve the level of clinical treatment. Through clinical and pathological analysis, a large amount of extremely valuable scientific research data can also be obtained.

A pathological slide is a type of pathological specimen. During production, some diseased tissues or organs are treated with various chemicals and embedding methods to make them fixed and hardened, sliced into thin slices on a microtome, adhered to glass slides, dyed with various colors, and prepared for viewing under a microscope. Check to observe pathological changes, make pathological diagnosis, and provide assistance for clinical diagnosis and treatment. Routine pathological slice making technology is the basis of pathological diagnosis, and slice quality is an important guarantee to guide doctors to make accurate diagnosis. Hematoxylin and Eosin staining, referred to as HE staining, is the most widely used staining method in biology and medical cell and histology. In the actual production process, due to the fact that some pathological technicians have not received strict and systematic work training, lack of work experience, and improper operation, the phenomenon of section falling off, uneven staining, blurring, wrinkles, and inconspicuous nuclei-cytoplasmic contrast often occur. affect the diagnosis. Therefore, if the process of making slices can be simplified, it is very meaningful to use pathological slices for auxiliary diagnosis.

The pathological process of the tissue is often accompanied by changes in the tissue structure at the cellular and subcellular levels, which leads to the migration of the tissue spectrum. By combining the traditional two-dimensional imaging technology with the spectral technology, the two-dimensional space of the observation target can be provided at the same time. Information and one-dimensional spectral information, through the analysis of spectral data, the morphological structure and chemical composition of the tissue can be obtained. Applying hyperspectral imaging technology to the field of auxiliary detection of pathological slides can effectively reduce the experience requirements for pathologists, reduce the subjectivity of pathological detection, and improve the efficiency of pathological detection.

At present, Siddiqi Anwer M and others in the world use a microspectral imager to obtain hyperspectral data of stained cervical cancer slices, and machine recognition of the slices by training the least squares support vector machine algorithm, the sensitivity and specificity can reach more than 90%. . Domestic Qingdao Institute of Optoelectronic Engineering Technology also used a similar method to identify liver cancer slices, and also achieved a good identification effect. However, both of these two methods require staining of cancer sections, which still require high experimental and experience requirements for pathologists, and are time-consuming.

Contents of the invention

In order to overcome the deficiencies in the prior art above, the object of the present invention is to provide a device and method for hyperspectral image acquisition and segmentation of non-stained pathological slices, which can realize automatic acquisition of hyperspectral images of pathological slices, and based on sample slices The three-dimensional hyperspectral data is used to train the artificial neural network algorithm to identify unstained liver cancer slice images. While reducing the complexity of traditional stained slice detection, it also reduces the subjective factors in the process of discrimination, and can provide pathologists with a diagnosis. good reference.

In order to achieve the above object, technical scheme of the present invention is:

The device for hyperspectral image acquisition and segmentation of non-stained pathological slices includes an alloy steel substrate 104, a xenon lamp light source 105 is placed on the alloy steel substrate 104, a fixed bracket 101 is vertically arranged on the alloy steel substrate 104, and the fixed bracket 101 is vertically arranged on the alloy steel substrate 104. A hyperspectral image acquisition module 102 and a sample platform 103 are arranged downward, the hyperspectral image acquisition module 102, the sample platform 103 are coaxial with the xenon lamp light source 105, and the hyperspectral image acquisition module 102 is externally connected to a computer 106;

Described hyperspectral image acquisition module 102 comprises CCD camera 110, connects computer 106 through CCD camera 110 side USB interface, and CCD camera 110 is connected with liquid crystal tunable filter 114 by relay lens 112, relay lens adjustment ring 113 and Keeping the coaxial center, the liquid crystal tunable filter 114 is connected to the computer 106 through the side USB interface 115 , and the lower part of the liquid crystal tunable filter 114 is connected to the objective lens 118 through the C-mount focusing ring 116 and the aperture 117 .

A hyperspectral image acquisition and segmentation method for non-stained pathological sections, comprising the following steps:

Step 1, device construction: place the unstained slice on the slice sample stage 104 and adjust the illumination range of the xenon lamp 103 to make the slice evenly illuminated, and set the parameters of the CCD camera 110 and the liquid crystal tunable filter 114 through the computer 106 to make the image grayscale Clear, adjust the relay lens adjustment ring 113, C-mount focus ring 116, and aperture 117 at the same time to make the slice coincide with the imaging focal plane;

Step 2, parameter setting and collecting training sample data: set the image acquisition parameters of the hyperspectral image acquisition module 102 on the computer 106, including acquisition mode, starting and ending wavelengths, wavelength resolution and exposure time, and start collecting and obtaining unstained section samples For the two-dimensional image data of each spectrum, the same acquisition parameters are kept after the acquisition is completed, and the spectral images of the dark background without illumination and the completely blank field of view under the same illumination parameters are respectively collected as background information reference;

Step 3: Data preprocessing, the black/white reference information obtained in step 2 is preprocessed for each spectral image, and then superimposed according to the sequence of spectral segments to obtain a three-dimensional hyperspectral matrix. The z-axis direction of this matrix corresponds to the spectral information of each point in the space. For each point Spectral curves are processed by wide interval reciprocal method to highlight features;

Step 4: Discriminate the training of the neural network, select the training sample slice lesion/non-lesion area wide-interval spectral curve obtained in step 3 according to the results of the stained slices as the training data, and train the spectral data discrimination network;

Step 5: Acquisition and area identification of hyperspectral data of unstained slices, collect hyperspectral data of unstained slices according to the method of step 2, perform the same preprocessing according to the method of step 3, and input the discriminant neural network trained in step 4 , to obtain the lesion/non-lesion discrimination result of each point, and the recognition result of the unstained section can be generated by combining the discrimination results of each point.

Described step three specific schemes are as follows:

The preprocessing of the spectral image adopts the method of pixel point correction. Under the same conditions as the hyperspectral image acquisition of slices, a set of hyperspectral data without illumination and a set of completely dark hyperspectral images and a set of hyperspectral images with a completely blank field of view are used. The difference between the hyperspectral image minus the dark image is compared to the difference between the blank image and the dark image, and the preprocessing formula is shown in formula 1.

Among them, R is the transmittance spectral value converted by preprocessing, I _im is the gray value of the original hyperspectral image at each wavelength, I _bl is the gray value of each wavelength image in complete darkness, and I _wh is the illuminated blank field of view The gray value of each wavelength image below, and then a method of wide-interval derivative is adopted, that is, the derivative is obtained by increasing the interval of the independent variable, as shown in formula 2:

In the formula, λ represents the wavelength, and Δλ represents the wavelength interval;

The step four, the specific process is as follows:

(1) Establish a discriminative neural network model

The input information represents the spectral information of a pixel, followed by the convolutional layer and the maximum pooling layer to the right to calculate a series of feature images, and the output layer is obtained after the feature images are classified. The whole network consists of input layer, convolutional layer C1, max pooling layer S1, convolutional layer C2, max pooling layer S2, fully connected layer F and output layer. The sample size of the input layer is (n1,1), where is the number of n1 bands. The first hidden convolutional layer C1 uses 24 kernel functions of size k1×1 to filter n1×1 input data. The hidden convolutional layer C1 contains 24×n2×1 nodes, where n2=n1-k1+1. There are 24×(k1+1) trainable parameters between the input layer and the convolutional layer C1, the maximum pooling layer S1 is the second hidden layer, the size of the kernel function is (k2,1), and the maximum pooling layer S1 contains There are 24×n3×1 nodes, where n3=n2/k2, and this layer has no parameters. The convolutional layer C2 contains 24×n4×1 nodes, the kernel function is (k3,1) where n4=n3-k3+1, there is 24×(k3+1) between the maximum pooling layer S1 and the convolutional layer C2 a trainable parameter. The maximum pooling layer S2 contains 24×n5×1 nodes, the size of the kernel function is (k4,1), where n5=n4/k4, and this layer has no parameters. The fully connected layer F contains n6 nodes, and there are (24×n6+1) training parameters between this layer and the maximum pooling layer S2. The final output layer board contains n7 nodes, and there are 24×n6×1 nodes between this layer and the complete connection layer F, and 24×n6×1×n7 training parameters. The convolutional neural network classifier with the above parameters is used to distinguish hyperspectral pixels, where n1 is the number of spectral channels, n7 is the number of output data categories, n2, n3, n4, n5 are the dimensions of the feature image, and n6 is The dimensionality of the fully connected layer.

(2) Forward propagation

The deep convolutional neural network used is a 5-layer structure, plus the input and output layer can also be regarded as 7 layers, expressed as (L+1) layer, where L=6, contains n1 input units in the input layer, in The output layer contains n7 output units, and the hidden layers are C1, S1, C2, S2 and F layers. Suppose _xi is the input of layer i and the output of layer (i-1), we can calculate _xi+1 as:

x _i+1 = f _i (u _i ) (Formula 3)

in

is the weight matrix of the i-th layer acting on the input data, b _i is the additional Bayesian vector of the i-th layer, f _i ( ) is the activation function of the i-th layer, and the hyperbolic function tanh(u) is selected as the convolutional layer The activation functions of C1, C2 and the complete connection layer F, take the maximum function max(u) as the activation function of the maximum pooling layers S1 and S2. The classifier is used to perform multi-classification of the data, the output category is n7, and the regression model of n7 class labels is defined as follows:

The output vector y=x _L+1 of the output layer represents the probability of all categories in the current iteration.

(3) Backpropagation

In the backward propagation stage, the trained parameters are adjusted and updated using the descending gradient method. Each parameter is determined by minimizing the cost function and calculating the partial derivative of the cost function. The loss function used is defined as follows:

Among them, m is the training sample size and Y is the output size. is the j-th value of the actual output y ⁽ⁱ⁾ of the i-th training sample, and the vector size is n7. The expected output Y ⁽ⁱ⁾ of the i-th sample, the probability value of the label is 1, if it is another category, the probability value is 0. 1{j=Y ⁽ⁱ⁾ } means that if j is equal to the i-th training The expected class label of the sample, its value is 1; otherwise, its value is 0. We added a negative sign in front of J(θ) to make the calculation more convenient.

The loss function obtains the partial derivative of u _i ;

where ° denotes element-wise multiplication. f′(u _i ) can be expressed simply as

Therefore, at each iteration, an update will be performed:

To judge the training parameters, α is the learning factor (α=0.01), and

Since θ _i contains W _i and b _i , and

in

After many iterations of training, the regression of the cost function is getting smaller and smaller, which also means that the actual output is getting closer to the expected output. When the difference between the actual output and the expected output is small enough, the iteration stops. Finally, the training A good deep convolutional neural network model can be used for hyperspectral image classification.

Compared with the prior art, the present invention uses a personal computer to synchronously control related modules based on the spectral differences caused by tissue lesions, collects spectral sequence images of unstained pathological tissue sections, preprocesses and superimposes them to generate corresponding three-dimensional hyperspectral data, and based on This data is combined with the current popular neural network classification idea to develop a spectral classification algorithm to identify and segment lesion areas, which speeds up the identification speed and efficiency of pathological tissue slices. After adopting this device, pathologists do not need the traditional HE staining process when making pathological slices, avoiding possible manual errors in the staining process, reducing the time required for slice production, and using machine algorithms for automatic discrimination, reducing the manual discrimination. The resulting subjectivity can provide better assistance for pathologists to detect pathological slides.

Description of drawings

Figure 1 is a schematic diagram of the overall structure of the device of the present invention.

Fig. 2 is a schematic structural diagram of the hyperspectral image acquisition module of the present invention.

Fig. 3 is the physical figure of the device of the present invention.

Fig. 4 is a screenshot of the supporting collection software of the present invention.

Fig. 5 is a flow chart of the slice spectral data identification algorithm of the present invention.

Fig. 6 is a schematic diagram of the structure of the spectral data discrimination neural network used in the present invention.

Fig. 7a is the multi-point original spectrum of extracted tissue slices according to the present invention; Fig. 7b is the spectrum curve after processing by the wide-interval derivative method.

Fig. 8 is a comparison between the prediction image of the unstained slice model of the present invention and the segmentation results of the dyed slice at the same site.

Fig. 9 is a comparison between the prediction image of the unstained slice model of the present invention and the recognition result of the support vector machine algorithm.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

Referring to Figures 1 to 3, the device for hyperspectral image acquisition and segmentation of non-stained pathological slices includes an alloy steel substrate 104, on which a xenon lamp light source 105 is placed, and a fixed bracket is vertically arranged on the alloy steel substrate 104 101, the fixed bracket 101 is provided with a hyperspectral image acquisition module 102 and a sample platform 103 from top to bottom, the hyperspectral image acquisition module 102, the sample platform 103 and the xenon light source 105 are coaxial, and the hyperspectral image acquisition module 102 is externally connected to a computer 106;

Described hyperspectral image acquisition module 102 comprises CCD camera 110, connects computer 106 through CCD camera 110 side USB interface, and CCD camera 110 is connected with liquid crystal tunable filter 114 by relay lens 112, relay lens adjustment ring 113 and Keeping the coaxial center, the liquid crystal tunable filter 114 is connected to the computer 106 through the side USB interface 115 to realize synchronous control with the CCD camera 110, and the lower part of the liquid crystal tunable filter 114 passes through the C-port focus ring 116, the aperture 117 and the objective lens 118 Connected, the relay lens 112 and the objective lens 118 are equipped with a focus ring for adjusting the focal plane and an aperture ring for controlling the amount of incoming light, and the slices can be adjusted by adjusting the relay lens adjustment ring 113, the C-mount focus ring 116, and the aperture 117. The sliced sample is placed on the sample platform 104 to achieve accurate focus and proper exposure. A xenon light path lens group is arranged on the side of the sample stage. The xenon light path lens group is controlled by a xenon light source drive circuit. The liquid crystal tunable filter and the scientific research camera are connected to the computer.

A hyperspectral image acquisition and segmentation method for non-stained pathological sections, comprising the following steps:

Step 1, device construction: place the unstained slice on the slice sample stage 104 and adjust the illumination range of the xenon lamp 103 to make the slice evenly illuminated, and set the parameters of the CCD camera 110 and the liquid crystal tunable filter 114 through the computer 106 to make the image grayscale Clear, adjust the relay lens adjustment ring 113, C-mount focus ring 116, and aperture 117 at the same time to make the slice coincide with the imaging focal plane;

Step 2, with reference to Fig. 4, parameter setting gathers training sample data: the image acquisition parameter of hyperspectral image acquisition module 102 is set on computer 106, comprises acquisition mode, start, stop wavelength, wavelength resolution and exposure time, start to gather and obtain For the two-dimensional image data of each spectrum of the unstained section sample, keep the same acquisition parameters after the acquisition is completed, and collect the spectral images of the dark background without light and the completely blank field of view under the same light parameters as background information reference;

Step 3: Data preprocessing, the black/white reference information obtained in step 2 is preprocessed for each spectral image, and then superimposed according to the sequence of spectral segments to obtain a three-dimensional hyperspectral matrix. The z-axis direction of this matrix corresponds to the spectral information of each point in the space. For each point Spectral curves are processed by wide interval reciprocal method to highlight features;

Described step three specific implementation schemes are as follows:

The preprocessing of the spectral image adopts the method of pixel point correction. Under the same conditions as the hyperspectral image acquisition of slices, a set of hyperspectral data without illumination and a set of completely dark hyperspectral images and a set of hyperspectral images with a completely blank field of view are used. The difference between the hyperspectral image minus the dark image is compared to the difference between the blank image and the dark image, and the preprocessing formula is shown in formula 1.

Among them, R is the transmittance spectral value converted by preprocessing, I _im is the gray value of the original hyperspectral image at each wavelength, I _bl is the gray value of each wavelength image in complete darkness, and I _wh is the illuminated blank field of view The gray value of each wavelength image below, and then a method of wide-interval derivative is adopted, that is, the derivative is obtained by increasing the interval of the independent variable, as shown in formula 2:

In the formula, λ represents the wavelength, and Δλ represents the wavelength interval;

Step 4: Training of the discriminative neural network. According to the results of the stained slices, the wide-interval spectral curves of the lesion/non-lesion regions of the training sample slices obtained in step 3 are selected as training data, and the spectral data discriminant network is trained. The network parameter training algorithm is shown in Figure 5

As shown, the specific process is as follows:

(1) Establish a discriminative neural network model as shown in Figure 6,

The input information represents the spectral information of a pixel, followed by the convolutional layer and the maximum pooling layer to the right to calculate a series of feature images, and the output layer is obtained after the feature images are classified. The whole network consists of input layer, convolutional layer C1, max pooling layer S1, convolutional layer C2, max pooling layer S2, fully connected layer F and output layer. The sample size of the input layer is (n1,1), where is the number of n1 bands. The first hidden convolutional layer C1 uses 24 kernel functions of size k1×1 to filter n1×1 input data. The hidden convolutional layer C1 contains 24×n2×1 nodes, where n2=n1-k1+1. There are 24×(k1+1) trainable parameters between the input layer and the convolutional layer C1, the maximum pooling layer S1 is the second hidden layer, the size of the kernel function is (k2,1), and the maximum pooling layer S1 contains There are 24×n3×1 nodes, where n3=n2/k2, and this layer has no parameters. The convolutional layer C2 contains 24×n4×1 nodes, the kernel function is (k3,1) where n4=n3-k3+1, there is 24×(k3+1) between the maximum pooling layer S1 and the convolutional layer C2 a trainable parameter. The maximum pooling layer S2 contains 24×n5×1 nodes, the size of the kernel function is (k4,1), where n5=n4/k4, and this layer has no parameters. The fully connected layer F contains n6 nodes, and there are (24×n6+1) training parameters between this layer and the maximum pooling layer S2. The final output layer board contains n7 nodes, and there are 24×n6×1 nodes between this layer and the complete connection layer F, and 24×n6×1×n7 training parameters. The convolutional neural network classifier with the above parameters is used to distinguish hyperspectral pixels, where n1 is the number of spectral channels, n7 is the number of output data categories, n2, n3, n4, and n5 are the dimensions of the feature image, and n6 is The dimensionality of the fully connected layer.

(2) Forward propagation

The deep convolutional neural network used is a 5-layer structure, plus the input and output layer can also be regarded as 7 layers, expressed as (L+1) layer, where L=6, contains n1 input units in the input layer, in The output layer contains n7 output units, and the hidden layers are C1, S1, C2, S2 and F layers. Suppose _xi is the input of layer i and the output of layer (i-1), we can calculate _xi+1 as:

x _i+1 = f _i (u _i ) (Formula 3)

in

is the weight matrix of the i-th layer acting on the input data, b _i is the additional Bayesian vector of the i-th layer, f _i ( ) is the activation function of the i-th layer, and the hyperbolic function tanh(u) is selected as the convolutional layer The activation functions of C1, C2 and the complete connection layer F, take the maximum function max(u) as the activation function of the maximum pooling layers S1 and S2. The classifier is used to perform multi-classification of the data, the output category is n7, and the regression model of n7 class labels is defined as follows:

The output vector y=x _L+1 of the output layer represents the probability of all categories in the current iteration.

(3) Backpropagation

In the backward propagation stage, the trained parameters are adjusted and updated using the descending gradient method. Each parameter is determined by minimizing the cost function and calculating the partial derivative of the cost function. The loss function used is defined as follows:

Among them, m is the training sample size and Y is the output size. is the j-th value of the actual output y ⁽ⁱ⁾ of the i-th training sample, and the vector size is n7. The expected output Y ⁽ⁱ⁾ of the i-th sample, the probability value of the label is 1, if it is another category, the probability value is 0. 1{j=Y ⁽ⁱ⁾ } means that if j is equal to the i-th training The expected class label of the sample, its value is 1; otherwise, its value is 0. We added a negative sign in front of J(θ) to make the calculation more convenient.

The loss function obtains the partial derivative of u _i ;

where ° denotes element-wise multiplication. f′(u _i ) can be expressed simply as

Therefore, at each iteration, an update will be performed:

To judge the training parameters, α is the learning factor (α=0.01), and

Since θ _i contains W _i and b _i , and

in

After many iterations of training, the regression of the cost function is getting smaller and smaller, which also means that the actual output is getting closer to the expected output. When the difference between the actual output and the expected output is small enough, the iteration stops. Finally, the training A good deep convolutional neural network model can be used for hyperspectral image classification.

Step 5: Acquisition and area identification of hyperspectral data of unstained slices, collect hyperspectral data of unstained slices according to the method of step 2, perform the same preprocessing according to the method of step 3, and input the discriminant neural network trained in step 4 , to obtain the lesion/non-lesion discrimination result of each point, and the recognition result of the unstained section can be generated by combining the discrimination results of each point.

Validation of the device and method using isotopic stained/unstained sections

In order to verify the feasibility of this device and method, taking liver cancer slices as an example, we used the stained/unstained slices of the same site provided by Mengchao Hepatobiliary Hospital of Fujian Medical University for system verification. The slice production process is as follows: the diseased tissue is obtained from the liver cancer patient, soaked in formalin solution and fixed, the tissue is completely dehydrated and embedded in paraffin, cut out a thin slice with a microtome, and two slices of the same thickness are cut from the same tissue Thin sections, one was stained with conventional hematoxylin-eosin staining method to make H&E stained section (for theoretical control), and the other was placed directly on the glass slide and dewaxed with xylene to make unstained tissue section. Hyperspectral image acquisition conditions are: xenon lamp light intensity 30mW/cm ² , wavelength range 400-718nm, wavelength interval 3nm.

For this batch of samples, we selected the derivative spectrum with a wavelength interval of 177nm for further distinction between components. The original spectrum is shown in Figure 7a and the curve after processing by the wide-interval derivative method is shown in Figure 7b. The processed standard data As the input training of the neural network classification algorithm, the neural network parameters are trained, and then the standard slices are segmented and verified. The prediction results are shown in Figure 8. At the same time, the segmentation results are compared with the segmentation results of the traditional convolution support vector machine algorithm, and the comparison results are shown in Figure 8. As shown in 9, the accuracy calculation formula shown in formula 13 is adopted to calculate the cancer region segmentation accuracy of the two algorithms, and the results are shown in Table 1

Table 1 Comparison of classification accuracy between deep convolutional neural network model and support vector machine

It can be seen from the experimental results that this device can effectively collect hyperspectral images of unstained liver cancer slices, and the supporting software algorithm has a better result for cancer region segmentation. The process reduces the interference of artificial factors in the traditional method, and can provide better diagnostic assistance for pathologists.

Claims (2)

1. the method for carrying out Image Acquisition and segmentation using the acquisition of dye-free pathological section high spectrum image and segmenting device, special Sign is that the acquisition of dye-free pathological section high spectrum image and segmenting device include alloy steel substrate (104), alloy steel substrate (104) it is equipped with xenon source (105), is also vertically installed on alloy steel substrate (104) fixed bracket (101) on, fixed bracket (101) it is provided with high spectrum image acquisition module (102) and example platform (103), high spectrum image acquisition module from top to bottom (102), example platform (103) and xenon source (105) are concentric, high spectrum image acquisition module (102) external computer (106)；

The high spectrum image acquisition module (102) includes CCD camera (110), passes through CCD camera (110) side USB interface External computer (106), CCD camera (110) pass through relaying camera lens (112), relaying camera adjusting ring (113) and liquid crystal tunable Optical filter (114) is connected and keeps concentric, and liquid crystal tunable filter (114) passes through side USB interface (115) and computer (106) it is connected, liquid crystal tunable filter (114) lower part passes through C mouthfuls of focusing rings (116), aperture (117) and object lens (118) phase Even；

The following steps are included:

Step 1: device is built: the slice that will be unstained is placed on example platform (103) and adjusts xenon source (105) illumination Range to be sliced uniform illumination, and passes through computer (106) setting CCD camera (110) and liquid crystal tunable filter (114) Parameter make image grayscale clear, while adjust relaying camera adjusting ring (113), C mouthfuls of focusing rings (116), aperture (117) make It must be sliced and be overlapped with imaging focal plane；

Step 2: parameter setting and the acquisition of training sample data: in setting high spectrum image acquisition module on computer (106) (102) image acquisition parameter, including acquisition mode, starting, termination wavelength, wavelength resolution and time for exposure, start to acquire Acquisition, which is unstained, is sliced each spectrum two-dimensional image data of sample, and the acquisition parameter maintained like after the completion of acquisition acquires respectively The completely black background of no light and with each spectrum picture in the complete blank visual field under illumination parameter as background information refer to；

Step 3: data prediction, by spectrum after being pre-processed using black/white reference information obtained by step 2 to each spectrum picture Section superimposition obtains three-dimensional bloom spectrum matrix, this matrix z-axis direction corresponds to every, space spectral information, makees to each point curve of spectrum The processing of wide interval derivative method is with prominent features；

Step 4: differentiating the training of neural network, and three gained training sample of selecting step is sliced lesion/non-lesion region wide interval The curve of spectrum differentiates network as training data, training spectroscopic data；Detailed process is as follows:

(1) it establishes and differentiates neural network model

Input information represents the spectral information of a pixel, toward the right followed by convolutional layer and maximum pond layer for calculating one Series of features picture obtains output layer after feature image classification；Whole network includes input layer, convolutional layer C1, maximum pond layer S1, Convolutional layer C2, maximum pond layer S2, complete articulamentum F and output layer；The sample-size of input layer is (n1,1), and wherein n1 is spectrum Number of active lanes；Convolutional layer C1 is first hidden layer, uses the input of 24 kernel function filtering n1 × 1 having a size of k1 × 1 Data；Convolutional layer C1 includes 24 × n2 × 1 node, wherein n2=n1-k1+1；Have 24 between input layer and convolutional layer C1 × (k1+1) it is a can training parameter, maximum pond layer S1 is second hidden layer, and the size of kernel function is (k2,1), maximum pond layer S1 packet Containing 24 × n3 × 1 node, wherein n3=n2/k2, this layer do not have parameter；Convolutional layer C2 includes 24 × n4 × 1 section Point, kernel function are (k3,1) wherein n4=n3-k3+1, have 24 between maximum pond layer S1 and convolutional layer C2 × (k3+1) is a trains Parameter；Maximum pond layer S2 includes 24 × n5 × 1 node, and kernel function size (k4,1), wherein n5=n4/k4, this layer do not have There is parameter；Complete articulamentum F includes n6 node, has (24 × n6+1) a training to join between this layer and maximum pond layer S2 Number；Last output layer contains n7 node, has 24 × n6 × 1 node, 24 × n6 × 1 between this layer and complete articulamentum F × n7 training parameter；The convolutional neural networks classifier of above-mentioned parameter is established for distinguishing EO-1 hyperion pixel, wherein n1 is light Number of active lanes is composed, n7 is exported data category number, and n2, n3, n4, n5 are the dimension of feature image respectively, and n6 is completely to connect Connect the dimension of layer；

(2) propagated forward

The depth convolutional neural networks used are 5 layers of structures, in addition input and output layer can also regard 7 layers as, are expressed as (L+1) layer, Wherein L=6 includes n1 input unit in input layer, includes n7 output unit in output layer, hidden layer C1, S1, C2, S2 and F layers；Assuming that x_iIt is the output of i-th layer input (i-1) layer, we can calculate x_i+1Are as follows:

x_i+1=f_i(u_i) (formula 3)

Wherein

It is i-th layer of weight matrix coefficient for acting on input data, b_iIt is i-th layer of additional Bayes's vector, f_i() is I-th layer of activation primitive selects hyperbolic function tanh (u) as the activation primitive of convolutional layer C1, C2 and complete articulamentum F, It is maximized activation primitive of the function max (u) as maximum pond layer S1 and S2；Data are carried out by more classification using classifier, it is defeated Classification is n7 out, and n7 category regression model is defined as follows:

The output vector y=x of output layer_L+1Indicate the probability in current iteration all categories；

(3) back-propagating

In the back-propagating stage, the parameter that training obtains is updated using Decent Gradient Methods adjustment, by using minimum cost letter Cost function partial derivative is counted and calculated to determine each parameter, the loss function used is defined as follows:

Wherein, m is training sample amount, and Y is output quantity；It is i-th training sample reality output y⁽ⁱ⁾Jth time value, to Taking measurements is n7；The desired output Y of i-th of sample⁽ⁱ⁾, the probability value of label is 1, and if it is other classifications, then probability value is 0； 1 { j=Y⁽ⁱ⁾It is meant that if j is equal to the expectation category of i-th training sample, its value if is 1；Otherwise, its value is 0；Increase negative sign before J (θ), is more facilitated so that calculating；

Loss function is to u_iLocal derviation is asked to obtain；

Wherein ° expression element multiplication；f′(u_i) can be expressed simply as

Therefore, in iteration each time, update will be all executed, θ is all parameter sets in neural network, θ_iInclude W_iAnd b_i:

For training of judgement parameter, α is Studying factors, α=0.01, and

Due to θ_iInclude W_iAnd b_i, and

Wherein:

By multiple repetitive exercise, the recurrence of cost function is smaller and smaller, this also means that it is actual output with it is desired defeated It becomes closer to out, when actual output is sufficiently small with desired output difference, iteration stopping, finally, trained depth Convolutional neural networks model may be used for the image classification of EO-1 hyperion；

Step 5: being unstained and be sliced acquisition and the region recognition of high-spectral data, be unstained cutting according to the mode of step 2 The acquisition of piece high-spectral data does identical pretreatment, the differentiation mind that the training of input step four finishes according to the mode of step 3 Through network, show that every lesion/non-lesion differentiates as a result, differentiate that result produces the identification knot of dye-free slice at comprehensive every Fruit.

2. according to claim 1 adopted using the acquisition of dye-free pathological section high spectrum image and segmenting device progress image Collection and the method for segmentation, which is characterized in that step three concrete scheme is as follows:

The pretreatment of spectrum picture uses pixel correction method, is sliced high spectrum image under the same conditions in same acquisition, makes With the high-spectral data of one group of no light complete darkness and the high spectrum image in one group of complete blank visual field, with the height of corresponding wavelength Spectrum picture subtracts difference of the difference than upper blank image and dark image of dark image, and pretreatment formula is as shown in formula 1；

Wherein, R is by the Optical transmission spectrum value of pretreatment conversion, I_imIt is the gray value under each wavelength of original high spectrum image, I_blIt is each wavelength image gray value under the conditions of complete darkness, I_whIt is the gray value for having each wavelength image under the illumination blank visual field, with A kind of method for using wide interval derivative afterwards, i.e. increase independent variable interval are differentiated, as shown in formula 2:

λ indicates that wavelength, Δ λ indicate wavelength interval in formula.