阅读说明：本技术 基于巴氏染色方式的dna倍体定量分析方法及系统 (DNA ploid quantitative analysis method and system based on Papanicolaou staining mode ) 是由 杨志明 于 2020-12-22 设计创作，主要内容包括：一种基于巴氏染色方式的细胞图像进行DNA倍体定量分析方法及系统,对采用巴氏染色方式进行染色得到的细胞玻片扫描,对得到的细胞图像中的单个细胞采用设置的积分光密度分析模型处理,得到单个细胞的积分光密度值,采用设置的细胞检测模型对所有单个细胞检测后,从检测为正常的所有单个细胞中选取部分单个细胞作为样本细胞,将统计出的样本细胞的积分光密度均值作为对照值,根据计算出该细胞玻片中的所有细胞的DNA指数,得到该细胞玻片的DNA定量分析结果。其中,所述分析模型采用注意力机制及特征金字塔机制的回归神经网络,将巴氏染色下的单个细胞对应的福尔根染色下积分光密度数值作为回归目标训练得到的。本发明保证DNA倍体定量分析的准确性。(Scanning a cell slide obtained by staining in a Papanicolaou staining mode, processing single cells in the obtained cell image by using a set integral optical density analysis model to obtain an integral optical density value of the single cells, detecting all the single cells by using the set cell detection model, selecting part of the single cells from all the single cells detected to be normal as sample cells, taking a counted integral optical density mean value of the sample cells as a control value, and obtaining a DNA quantitative analysis result of the cell slide according to the calculated DNA indexes of all the cells in the cell slide. The analysis model is obtained by training by taking an integral optical density numerical value under Fowler root staining corresponding to a single cell under Papanicolaou staining as a regression target by adopting a regression neural network of an attention mechanism and a characteristic pyramid mechanism. The invention ensures the accuracy of DNA ploid quantitative analysis.)

1. A DNA ploid quantitative analysis method based on a cell image of a Papanicolaou staining mode is characterized by comprising the following steps:

scanning a cell slide obtained by staining in a Papanicolaou staining mode to obtain a cell image;

processing the single cells in the cell image by adopting a set integral optical density analysis model to obtain an integral optical density value of the single cells, wherein the integral optical density analysis model is obtained by training by adopting an integral optical density value under Fowler root dyeing corresponding to the single cells under the Papanicolaou dyeing as a regression target by adopting a regression neural network of an attention mechanism and a characteristic pyramid mechanism;

after all single cells are detected by the set cell detection model, selecting a part of single cells from all normal single cells as sample cells, calculating the DNA indexes of all cells in the cell slide by taking the counted integrated optical density mean value of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide according to the calculated DNA indexes of all cells in the cell slide.

2. The method of claim 1, wherein the integrated optical density analysis model employs an EfficientNet network structure comprising: a convolutional layer, an attention mechanism layer, a feature gold sub-tower layer and a regression layer.

3. The method of claim 2, wherein the attention mechanism layer is used for performing attention processing on a nuclear region in the input single cell by the integrated densitometric analysis model during training;

the characteristic pyramid layer is used for fusing cell characteristics of different scales during processing.

4. The method of claim 1, wherein said regression layer employs a Huber loss function for regression, comprising:

where L is the result of the calculated Huber loss function, y represents the true value of the normalized integrated optical density under the Fowler root staining of the cells, f (x) is the predicted value under the Papanicolaou staining of the cells, x represents the input single cell image, and δ is a predefined threshold.

5. The method of claim 2, wherein the convolutional layer is configured to convolve features of individual cells in the input pap-stained cell image;

and the regression layer comprises a convolution layer and a global maximum pooling layer, the activation function adopts a Relu activation function, and the output is an integrated optical density value under the Fowler root staining corresponding to a single cell in the cell image under the Papanicolaou staining.

6. The method of claim 1, further comprising, prior to said training resulting in the integrated densitometric model, preparing a training data set for training:

preparing cell slides according to a set number, wherein the cell slides comprise negative cell slides and positive cell slides;

staining cells in the cell slide by using Papanicolaou staining and scanning the cells to obtain a cell image, then staining the cells by using fern root after fading the cells and scanning the cells to obtain a cell image of the same cell sample under the fern root staining and the Papanicolaou staining respectively;

the integrated optical density values under feulgen staining corresponding to individual cells under papanicolaou staining were obtained.

7. A system for carrying out quantitative analysis on DNA ploidy based on cell images in a Papanicolaou staining mode is characterized by comprising the following components: an acquisition unit, a processing unit and a DNA ploid quantitative analysis unit, wherein,

the acquisition unit is used for scanning the cell slide obtained by staining in a Papanicolaou staining mode to obtain a cell image;

the processing unit is used for processing the single cells in the cell image by adopting a set integral optical density analysis model to obtain an integral optical density value of the single cells, wherein the integral optical density analysis model is obtained by adopting a regression neural network of an attention machine system and a characteristic pyramid mechanism and training an integral optical density value under Fowler root dyeing corresponding to the single cells under the Pasteur dyeing as a regression target;

and the DNA ploid quantitative analysis unit is used for selecting part of single cells from all normal single cells as sample cells after detecting all the single cells by adopting the set cell detection model, calculating the DNA indexes of all the cells in the cell slide by taking the counted integrated optical density mean value of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide.

8. The system of claim 7, wherein the processing unit is further configured to apply the integrated densitometry model using an EfficientNet network architecture comprising: a convolutional layer, an attention mechanism layer, a feature gold sub-tower layer and a regression layer.

Technical Field

The invention relates to a medical cell image processing technology, in particular to a method and a system for quantitatively analyzing deoxyribonucleic acid (DNA) ploidy based on a cell image in a Papanicolaou staining mode.

Background

Cancer is a chromosomal disease, and carcinogens, rare genetic disorders, and sporadic mitotic errors can produce aneuploidy and subsequently cause tumors. Studies have shown that changes in the cellular genome and the appearance of aneuploid cells in specimens are early events in cancer development and can serve as markers for tumor detection. The quantitative analysis technology of DNA ploidy is adopted to determine and analyze the DNA content of cell nucleus and ploidy condition, which is an important method in the present malignant tumor diagnosis and is widely applied to various cytological examinations, and the process is as follows: manufacturing a cell slide based on the obtained cell specimen; after cell images are collected on the cell slide, the collected cell images are analyzed to obtain cell areas, and DNA ploidy quantitative analysis is carried out on the cell areas to obtain quantitative analysis results. The specimen may be a scraped specimen of a surface such as a cervix, an oral cavity, or the like, a specimen excreted in sputum, urine, or the like, a specimen punctured with body fluid or a tumor such as a thoracic cavity, an abdominal cavity, or the like, a specimen of an endoscopic brush of a digestive tract, a respiratory tract, or a tissue print, or the like.

The DNA non-integral times appear and is related to the canceration rate and the precancerous lesion hyperplasia degree, researches show that the change of cell genome and the appearance of the aneuploid cells of a specimen are important indexes for canceration of the cells, and the DNA content and the ploid condition of the cells are determined and analyzed by a DNA ploid quantitative analysis technology, so that the method is one of the important methods for diagnosing the malignant tumors at present and is widely applied to various cytological examinations.

At present, a Feulgen-stabilized DNA quantitative analysis method is generally used for staining cells to obtain a cell image for analysis, wherein the Feulgen staining is a staining method capable of specifically displaying DNA, the higher the DNA content is, the darker the color of the cell nucleus after staining is, but the staining technology is long in time consumption, and the problems that glandular cells cannot be identified, too much manual work is involved, the analysis time is too long and the like exist in the DNA polyploid quantitative analysis method based on the cell image. The integrated optical density is an important value for measuring the nuclear DNA content in images of cells under feulgen staining and is calculated as follows:

wherein λ is₀Represents the average value of background pixels, and represents the average value of incident light intensity when light passes through the background area, λ_iThe value of the ith pixel in the cell nucleus is shown, and n is the number of pixel points in the cell nucleus.

In order to overcome the problems, when the DNA ploid is quantitatively analyzed, a papanicolaou staining method can be adopted to obtain a cell image for analysis, and the papanicolaou staining technology is one of the most common staining methods in human exfoliative cytology and is widely applied to the fields of diagnosis of exfoliative cytology of respiratory systems, urinary systems, reproductive systems and the like, identification of microbial infection and the like. Compared with the cell image obtained by Fowler root dyeing, the cell image under the Pasteur dyeing has good cell transparency, bright color and clear cell structure. The DNA ploidy analysis method based on the Papanicolaou staining cell glass image can effectively solve the problems of long time consumption, much manual participation and the like in the staining and analyzing links in the existing DNA ploidy quantitative analysis method based on the cell image of the Fuller's root staining. However, the papanicolaou staining mode stains the cell nucleus and cytoplasm in the cell simultaneously, has a certain staining effect on non-DNA substances in the cell nucleus, and can also dope the non-DNA substances when performing subsequent DNA ploid quantitative analysis, so that how to perform DNA ploid quantitative analysis calculation on a cell image based on the papanicou staining mode avoids the inclusion of the non-DNA substances when calculating, ensures accurate analysis results of the DNA ploid quantitative analysis, and becomes a problem to be solved urgently.

Disclosure of Invention

In view of this, the embodiment of the present invention provides a DNA ploid quantitative analysis method based on a cell image of a papanicolaou staining method, which can remove the influence caused by non-DNA substances under papanicolaou staining and ensure the accuracy of DNA ploid quantitative analysis.

The embodiment of the invention also provides a DNA ploid quantitative analysis system based on the cell image of the Papanicolaou staining mode, which can remove DNA substances when performing the DAN analysis result and ensure the accuracy of the DNA ploid quantitative analysis.

The embodiment of the invention is realized as follows:

a method for carrying out quantitative analysis on DNA ploidy based on a cell image in a Papanicolaou staining mode comprises the following steps:

scanning a cell slide obtained by staining in a Papanicolaou staining mode to obtain a cell image;

processing the single cells in the cell image by adopting a set integral optical density analysis model to obtain an integral optical density value of the single cells, wherein the integral optical density analysis model is obtained by training by adopting an integral optical density value under Fowler root dyeing corresponding to the single cells under the Papanicolaou dyeing as a regression target by adopting a regression neural network of an attention mechanism and a characteristic pyramid mechanism;

after all single cells are detected by the set cell detection model, selecting a part of single cells from all normal single cells as sample cells, calculating the DNA indexes of all cells in the cell slide by taking the counted integrated optical density mean value of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide according to the calculated DNA indexes of all cells in the cell slide.

Preferably, the integrated optical density analysis model adopts an EfficientNet network structure, and the structure includes: a convolutional layer, an attention mechanism layer, a feature gold sub-tower layer and a regression layer.

Preferably, the attention mechanism layer is used for performing attention processing on a cell nucleus region in the input single cell by the integrated optical density analysis model during training;

the characteristic pyramid layer is used for fusing cell characteristics of different scales during processing.

Preferably, the regression layer performs regression by using a Huber loss function, including:

where L is the result of the calculated Huber loss function, y represents the true value of the normalized integrated optical density under the Fowler root staining of the cells, f (x) is the predicted value under the Papanicolaou staining of the cells, x represents the input single cell image, and δ is a predefined threshold.

Preferably, the convolution layer is used for performing convolution processing on the characteristics of single cells in the cell image in the input papanicolaou staining mode;

and the regression layer comprises a convolution layer and a global maximum pooling layer, the activation function adopts a Relu activation function, and the output is an integrated optical density value under the Fowler root staining corresponding to a single cell in the cell image under the Papanicolaou staining.

Preferably, before the training to obtain the integrated optical density analysis model, a training data set is prepared for training:

preparing cell slides according to a set number, wherein the cell slides comprise negative cell slides and positive cell slides;

staining cells in the cell slide by using Papanicolaou staining and scanning the cells to obtain a cell image, then staining the cells by using fern root after fading the cells and scanning the cells to obtain a cell image of the same cell sample under the fern root staining and the Papanicolaou staining respectively;

the integrated optical density values under feulgen staining corresponding to individual cells under papanicolaou staining were obtained.

A system for carrying out quantitative analysis on DNA ploidy based on a cell image in a Papanicolaou staining mode comprises: an acquisition unit, a processing unit and a DNA ploid quantitative analysis unit, wherein,

the acquisition unit is used for scanning the cell slide obtained by staining in a Papanicolaou staining mode to obtain a cell image;

the processing unit is used for processing the single cells in the cell image by adopting a set integral optical density analysis model to obtain an integral optical density value of the single cells, wherein the integral optical density analysis model is obtained by adopting a regression neural network of an attention machine system and a characteristic pyramid mechanism and training an integral optical density value under Fowler root dyeing corresponding to the single cells under the Pasteur dyeing as a regression target;

and the DNA ploid quantitative analysis unit is used for selecting part of single cells from all normal single cells as sample cells after detecting all the single cells by adopting the set cell detection model, calculating the DNA indexes of all the cells in the cell slide by taking the counted integrated optical density mean value of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide.

Preferably, the processing unit is further configured to use an EfficientNet network structure for the integrated optical density analysis model, where the structure includes: a convolutional layer, an attention mechanism layer, a feature gold sub-tower layer and a regression layer.

As can be seen from the above, in the embodiment of the present invention, a cell slide obtained by staining in a papanicolaou staining manner is scanned to obtain a cell image, a single cell in the cell image is processed by using a set integrated optical density analysis model to obtain an integrated optical density value of the single cell, then all the single cells are detected by using a set cell detection model, then a part of the single cells are selected from all the single cells detected as normal cells as sample cells, a counted average value of the integrated optical density of the sample cells is used as a control value, DNA indexes of all the cells in the cell slide are calculated, and a result of quantitative analysis of DNA of the cell slide is obtained. The integrated optical density analysis model is obtained by training a value of integrated optical density under Fowler root staining corresponding to a single cell under Papanicolaou staining as a regression target by adopting a regression neural network of an attention mechanism and a characteristic pyramid mechanism. Therefore, when the DNA ploidy quantitative analysis is performed on the cell image, the invention adopts the integral optical density analysis model, the model is obtained by training the integral optical density value under Fowler root staining corresponding to a single cell under Papanicolaou staining as a regression target, the influence caused by non-DNA substances under Papanicolaou staining can be removed, and an attention mechanism and a characteristic pyramid mechanism are introduced during the construction of the model, so that the network can automatically learn the cell nucleus region needing important attention, fuse the cell characteristic information of different scales, enhance the characterization capability of the provided characteristic, and ensure the accuracy of the DNA ploidy quantitative analysis.

Drawings

FIG. 1 is a flow chart of a method for quantitative analysis of DNA ploidy based on a cell image of Papanicolaou staining mode according to an embodiment of the present invention;

FIG. 2 is a diagram illustrating an example of the DNA ploid quantitative analysis according to the embodiment of the present invention;

fig. 3 is a detailed diagram of a network result of an integrated optical density analysis model according to an embodiment of the present invention;

FIG. 4 is a schematic structural diagram of an FPN layer according to an embodiment of the present invention;

FIG. 5 is a schematic structural diagram of an attention suppressing layer provided in an embodiment of the present invention;

FIG. 6 is a schematic structural diagram of a system for performing quantitative analysis of DNA ploidy based on a cell image obtained by Papanicolaou staining according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings and examples.

As can be seen from the background art, the DNA ploidy analysis method based on the Papanicolaou staining cell can effectively solve the problems of long time consumption, much manual participation and the like in the staining and analyzing links in the existing DNA ploidy quantitative analysis method based on the cell image of the Fuller's staining mode. However, when the papanicolaou staining mode is used for staining cells, the cell nucleus and cytoplasm in the cells are stained at the same time, non-DNA substances in the cell nucleus are also stained to a certain extent, and the non-DNA substances are doped when the subsequent quantitative analysis of the DNA ploid is carried out, so that the quantitative analysis of the DNA ploid is inaccurate. Therefore, in the embodiment of the present invention, a cell slide obtained by staining in a papanicolaou staining manner is scanned to obtain a cell image, a single cell in the cell image is processed by using a set integral optical density analysis model to obtain an integral optical density value of the single cell, a set cell detection model is used to detect all single cells, a part of single cells are selected from all normal single cells as sample cells, a counted average value of the integral optical density of the sample cells is used as a control value, DNA indexes of all cells in the cell slide are calculated, and a DNA quantitative analysis result of the cell slide is obtained according to the calculated DNA indexes of all cells in the cell slide. The integrated optical density analysis model is obtained by training a value of integrated optical density under Fowler root staining corresponding to a single cell under Papanicolaou staining as a regression target by adopting a regression neural network of an attention mechanism and a characteristic pyramid mechanism.

Therefore, when the DNA ploidy quantitative analysis is performed on the cell image, the invention adopts the integral optical density analysis model, the model is obtained by training the integral optical density numerical value under Fowler root staining corresponding to a single cell under Papanicolaou staining as a regression target, the influence caused by non-DNA substances under Papanicolaou staining can be removed, and an attention mechanism and a characteristic pyramid mechanism are introduced during the construction of the model, so that the network can automatically learn the cell nucleus region needing important attention, fuse the cell characteristic information of different scales, enhance the characterization capability of the provided characteristic, and ensure the accuracy of the DNA quantitative analysis of the ploidy.

Fig. 1 is a flow chart of a method for performing DNA ploid quantitative analysis based on a cell image of papanicolaou staining mode provided by the embodiment of the invention, which comprises the following specific steps:

step 101, scanning a cell slide obtained by staining in a Papanicolaou staining mode to obtain a cell image;

102, processing single cells in the cell image by adopting a set integral optical density analysis model to obtain an integral optical density value of the single cells, wherein the integral optical density analysis model is obtained by training by adopting an attention mechanism and a regression neural network of a characteristic pyramid mechanism and taking an integral optical density value under Fowler root dyeing corresponding to the single cells under the Papanicolaou dyeing as a regression target;

and 103, after detecting all single cells by using the set cell detection model, selecting a part of single cells from all normal single cells as sample cells, calculating the DNA indexes of all cells in the cell slide by using the counted integrated optical density mean value of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide according to the calculated DNA indexes of all cells in the cell slide.

In the method, the integrated optical density analysis model adopts an EfficientNet network structure, and the structure comprises the following steps: a convolution layer, an attention mechanism layer, a characteristic gold sub-tower layer and a regression layer.

The convolutional layer generally has a 3 × 3 structure, and performs convolution processing on the input characteristics of a single cell.

And the attention mechanism layer is used for performing attention processing on the nucleus area in the input single cell by the integrated optical density analysis model during training.

And the characteristic pyramid layer is used for fusing cell characteristics of different scales during processing.

Performing regression in a regression layer using a Huber loss function, comprising:

where L is the result of the calculated Huber loss function, y represents the true value of the normalized integrated optical density under the Fowler root staining of the cells, f (x) is the predicted value under the Papanicolaou staining of the cells, x represents the input single cell image feature, and δ is a predefined threshold.

And the regression layer adopts two 3-by-3 convolution layers and a global maximum pooling layer, the activation function adopts a Relu activation function, and the output is an integrated optical density value under Fowler's root staining corresponding to a single cell in a cell image under the Pasteur staining.

In this method, the integrated optical density analysis model is specifically trained as follows.

The first step is as follows: a training data set is prepared.

Under the step, firstly, cell slide preparation is carried out, wherein the cell slide preparation comprises a negative cell slide and a positive cell slide which are prepared according to a set quantity; secondly, collecting a cell slide to obtain a cell image, specifically, staining cells in the cell slide by using Papanicolaou staining and scanning the cells to obtain a cell image, fading the cell image, staining by using Fuller root and scanning the cell image to obtain a cell image of the same cell sample under the Fuller root staining and the Papanicolaou staining respectively; finally, integrated optical density values under feulgen staining corresponding to single cells under papanicolaou staining were obtained.

The second step is that: an integrated densitometric analysis model is set and trained based on a prepared training data set.

Considering the cellular images obtained by two staining methods under the same cellular sample, the DNA content of the cell nucleus should be consistent. Therefore, the embodiment of the invention constructs an integral spectral density analysis model, and in the model, the mapping relation from the single cell under the Papanicolaou staining to the Fowler root staining integral spectral density value is directly constructed through a deep learning technology so as to further calculate the DNA ploidy value of the single cell under the Papanicolaou staining.

Specifically, first, a network structure is constructed;

in the step, an integrated optical density analysis model is constructed by adopting a regression target, and the integrated optical density value under the Fowler root staining corresponding to the single cell under the Papanicolaou staining is used as the regression target of the corresponding single cell under the Papanicolaou staining. That is, in the constructed integrated optical density analysis model, a single cell image in the cell images under the papanicolaou staining is input, and an integrated optical density value under the feulgen staining corresponding to the cell is output for training.

The integrated optical density analysis model is characterized in that a main network of the integrated optical density analysis model adopts an EfficientNet network structure, the EfficientNet network structure amplifies the network from three dimensions of depth, width and resolution ratio through simple and efficient composite coefficients, a characteristic pyramid layer can be inserted into the network structure, the characteristic extraction capability is enhanced by fusing cell characteristic information of different scales, and the integrated optical density analysis model is suitable for the problem of large difference of morphological sizes of different cells; an attention mechanism layer can be inserted into the network structure, so that the network can automatically learn the nucleus area in the single cell which needs to pay attention, and the relationship between the single cell in the cell image under the Papanicolaou staining and the corresponding nucleus integrated optical density under the Fowler root staining can be better determined.

The regression layer in the integrated densitometric model employs a Huber loss function:

where L is the result of the calculated Huber loss function, y represents the true value of the normalized integrated optical density under the Fowler root staining of the cells, f (x) is the predicted value under the Papanicolaou staining of the cells, x represents the input single cell image, and δ is a predefined threshold.

The third step, DNA ploidy value calculation

After the integral optical density of the single cell is obtained, detecting all single cells by using a set cell detection model, selecting a part of single cells from all normal single cells as sample cells, calculating the DNA indexes of all cells in the cell slide by using the counted average value of the integral optical density of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide.

The method for calculating the DNA ploidy value provided by the embodiment of the invention is exemplified in detail.

The specific example of the quantitative analysis of DNA ploidy provided by the embodiment of the invention shown in FIG. 2 is schematically illustrated.

The first step is as follows: a training data set is prepared.

Under the step, firstly, cell slide preparation is carried out, wherein the cell slide preparation comprises a negative cell slide and a positive cell slide which are prepared according to a set quantity; secondly, collecting a cell slide to obtain a cell image, specifically, staining cells in the cell slide by using Papanicolaou staining and scanning the cells to obtain a cell image, fading the cell image, staining by using Fuller root and scanning the cell image to obtain a cell image of the same cell sample under the Fuller root staining and the Papanicolaou staining respectively; finally, integrated optical density values under feulgen staining corresponding to single cells under papanicolaou staining were obtained.

Here, the procedure for obtaining integrated optical density values under feulgen staining corresponding to individual cells under papanicolaou staining is:

1) the nuclear integrated optical density value in the cell image under the Fowler root staining is calculated by using the existing integrated optical density calculation method. The cell sample after the feulgen staining can be scanned by a full-automatic cell image analysis system of a MotiCytomer (Miaodi medical diagnosis System Co., Ltd.), cell nuclei of about 8000 are collected, more than 100 parameter characteristic values are obtained for each cell nucleus, the cell nuclei can be automatically classified through a parameter characteristic value system, and DNA ploidy values of the cell nuclei are determined;

2) carrying out cell detection on the cell image under the Papanicolaou staining by using the set target detection model to obtain a single cell; the set target detection model is trained and used for detecting single cells in the cell image under the papanicolaou staining.

3) And (3) carrying out coordinate alignment on the cell image of the Papanicolaou staining and the cell image of the Fuller root staining, and obtaining the integral optical density value of the cell nucleus under the Papanicolaou staining and the corresponding Fuller root staining.

And a second step of setting an integrated optical density analysis model and training based on the prepared training data set.

The Fowler root staining method only stains the cell nucleus in the cell, while the Papanicolaou staining method has staining effects on both the cell nucleus and cytoplasm, and has staining effects on various substances such as DNA, protein and the like in the cell nucleus. If the integrated optical density is directly calculated, the obtained DNA ploidy value of the Papanicolaou staining cell has certain difference with the DNA ploidy value of the Fuller root staining cell. Considering the cell images obtained by two staining methods under the same cell sample, the DNA content of the cell nucleus is kept consistent. Therefore, the embodiment of the invention constructs an integrated optical density analysis model, and in the model, the mapping relation between the single cell under the papanicolaou staining and the integrated optical density value under the feulgen staining is directly constructed through a deep learning technology so as to further calculate the DNA ploidy value of the cell under the papanicolaou staining.

Specifically, as shown in fig. 3, fig. 3 is a detailed schematic diagram of a network result of the integrated optical density analysis model provided in the embodiment of the present invention. Firstly, constructing a network structure;

in the step, an integrated optical density analysis model is constructed by adopting a regression target, and the integrated optical density value under the Fowler root staining corresponding to the single cell under the Papanicolaou staining is used as the regression target of the corresponding single cell under the Papanicolaou staining. That is, in the constructed integrated optical density analysis model, the characteristics of a single cell image in a cell image under papanicolaou staining are input, and the integrated optical density value under the feulgen staining corresponding to the cell is output for training.

The integrated optical density analysis model has a main network adopting an EfficientNet network structure, and the EfficientNet network structure amplifies the network from three dimensions of depth, width and resolution ratio through simple and efficient compound coefficients, so that the feature extraction capability of the main network is enhanced.

A Feature Pyramid (FPN) layer is inserted into the network structure, as shown in fig. 4, fig. 4 is a schematic structural diagram of the FPN layer provided in the embodiment of the present invention, and the capability of extracting features is enhanced by fusing cell feature information of different scales, so as to adapt to the problem of large difference in morphology and size of different cells.

An attention mechanism (attention) layer may also be inserted into the network structure, as shown in fig. 5, fig. 5 is a schematic structural diagram of the attention mechanism layer provided in the embodiment of the present invention, so that the network automatically learns the cell nucleus region in a single cell that needs to be focused, and better determines the relationship between the single cell in the cell image under the papanicolaou staining and the corresponding cell nuclear integrated optical density under the feulgen staining.

The regression layer in the integrated optical density analysis model adopts two 3-by-3 convolution layers and a global maximum pooling layer, the activation function adopts a Relu activation function, and the output is the integrated optical density value under the Fowler root staining corresponding to a single cell in a cell image under the Pasteur staining.

In an integrated optical density analysis model, common loss functions of a regression layer comprise a square loss function, an absolute value loss function or a Huber loss function and the like, the square loss function is most common, and the defect that a large penalty is given to an abnormal point, so that the robustness is not enough; the absolute value loss function has the characteristic of resisting abnormal point interference, but discontinuous conduction at y-f (x) is difficult to optimize. The regression target of the regression layer in the integrated optical density analysis model of the embodiment of the invention is to analyze the corresponding cell under the Papanicolaou staining by taking the integrated optical density value obtained from the cell nucleus in the single cell in the Fuller staining mode as the target, and to facilitate the calculation of the loss function, the integrated optical density value of the regression layer is normalized, and the adopted loss function is the Huber loss function:

where L is the result of the calculated Huber loss function, y represents the true value of the normalized integrated optical density under the Fowler root staining of the cells, f (x) is the predicted value under the Papanicolaou staining of the cells, x represents the input single cell image, and δ is a predefined threshold.

The third step, calculating the DNA ploid quantitative value

After the integral optical density of the single cell is obtained, detecting all single cells by using a set cell detection model, selecting a part of single cells from all normal single cells as sample cells, calculating the DNA indexes of all cells in the cell slide by using the counted average value of the integral optical density of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide.

Fig. 6 is a schematic structural diagram of a system for performing DNA ploid quantitative analysis based on a cell image in papanicolaou staining mode according to an embodiment of the present invention, including: an acquisition unit, a processing unit and a DNA ploid quantitative analysis unit, wherein,

the acquisition unit is used for scanning the cell slide obtained by staining in a Papanicolaou staining mode to obtain a cell image;

the processing unit is used for processing the single cells in the cell image by adopting a set integral optical density analysis model to obtain an integral optical density value of the single cells, wherein the integral optical density analysis model is obtained by adopting a regression neural network of an attention machine system and a characteristic pyramid mechanism and training an integral optical density value under Fowler root dyeing corresponding to the single cells under the Pasteur dyeing as a regression target;

and the DNA ploid quantitative analysis unit is used for selecting part of single cells from all normal single cells as sample cells after detecting all the single cells by adopting the set cell detection model, calculating the DNA indexes of all the cells in the cell slide by taking the counted integrated optical density mean value of the sample cells as a control value, and obtaining the DNA quantitative analysis result of the cell slide.

In this system, the processing unit is further configured to adopt an EfficientNet network structure for the integrated optical density analysis model, where the structure includes: a convolution layer, an attention mechanism layer, a characteristic gold sub-tower layer and a regression layer.

The embodiment of the invention simplifies the DNA ploid analysis process based on cell images under Papanicolaou staining and improves the precision of DNA ploid quantitative analysis, in particular to the following steps: the embodiment of the invention provides an integral optical density analysis model, which can realize the automatic calculation of the integral optical density value of a single cell of a cell image under the Papanicolaou staining; the integrated optical density analysis model provided by the embodiment of the invention uses an EfficientNet network as a main network, the network is amplified from three dimensions of depth, width and resolution by adopting a simple and efficient composite coefficient based on a neural structure search technology, and a characteristic pyramid layer and an attention mechanism layer are introduced, so that cell information of different scales can be fused, the capability of extracting characteristics is enhanced, the integrated optical density analysis model is suitable for practical problems of multiple difference of various scanners, size difference of different types of cells and the like, and has higher robustness.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.