A method for recognizing dicentric chromosome images based on CNN convolutional neural network

Abstract

The present invention discloses a method and system for recognizing dicentric chromosome images based on a CNN convolutional neural network, the method comprising: 1) obtaining a sample chromosome image; 2) binarizing the image; 3) morphologically processing and cutting the chromosomes in the binarized image; 4) constructing a CNN model and using the image obtained in step 3) to train the CNN model, and 5) recognizing and classifying dicentric chromosome images based on the CNN model. The present invention can improve the accuracy of dicentric chromosome recognition, improve work efficiency and accuracy, and contribute to biomedical research and clinical diagnosis.

Description

Technical Field

[0001] The invention belongs to the field of biomedical technology, and in particular relates to a dicentric chromosome image recognition method based on a CNN convolutional neural network. Background Art

[0002] With the rapid development of nuclear energy and nuclear technology applications, the potential dangers of nuclear and radiation accidents are increasing. This has put forward new demands and challenges on how to achieve rapid screening of radiation-damaged casualties. As the gold standard for biological dose estimation, dicentric chromosome analysis currently relies mainly on manual analysis. In order to increase the credibility of dose estimation, a lot of manpower is required. At present, radiation cytogenetics laboratories in some developed countries have been equipped with automatic analysis systems. There are also more mature products on the market, such as the Metafer slide scanning system of Zeiss in Germany and the CytoScan HD cytogenetic detection system of Affy in the United States. However, these foreign products are relatively expensive, costing millions. China started late in this field, and automatic chromosome analysis is still in the research stage. Especially in the field of radiation biology, it is still limited to semi-automatic dicentric analysis to estimate dose. Therefore, facing the needs of automated analysis of biological dose estimation, it is very necessary to carry out research on automatic dicentric recognition algorithms for biological dose estimation and improve the automatic recognition rate of dicentrics.

[0003] In modern biomedical research and clinical diagnosis, dicentric chromosome image recognition plays an important role in genetic research and diagnosis of chromosomal abnormalities. Currently, the recognition technology and software on the market have the following shortcomings:

[0004] First, the degree of localization of software is not high. In the existing market, dicentric chromosome image recognition software is mainly developed and controlled by foreign medical or software companies. Domestic users need to rely on foreign service providers to obtain necessary technical support, and they will encounter problems such as poor communication in language communication and code updates, and the needs cannot be solved in time. In addition, the high price of foreign software makes it unaffordable for many small and medium-sized laboratories, which limits the development of related research.

[0005] Secondly, existing software performs poorly in program optimization and is relatively cumbersome. This problem is particularly prominent when processing large amounts of image data. In today's biomedical research, large-scale, high-throughput experiments have become the norm. The recognition software has cumbersome processing steps and cannot quickly and accurately process large amounts of image data in batches, which will slow down the progress of research.

[0006] Finally, existing software is not flexible enough to cope with different lens hardware, different film settings, chromosome images of different imaging quality, and different usage scenarios. For example, some software can only process chromosomes of a certain background color format, but cannot adapt to other types. Some software can only process high-quality images, and the recognition effect of images with more noise and lower quality is greatly reduced. All these limit the scope of application and effect of the software.

[0007] Therefore, new technologies and methods are needed to at least partially address the deficiencies in the prior art. Summary of the invention

[0008] The main goal of the present invention is to develop an accurate and efficient system that can automatically identify dicentric chromosomes from chromosome images and mark them for display. By using advanced techniques of image morphological processing and deep learning, the system can quickly and accurately perform chromosome cutting and identification, further improving the current manual analysis of chromosome images and the method of morphological processing analysis alone, improving work efficiency and accuracy, while filling the gap in domestic chromosome recognition software development, and contributing to biomedical research and clinical diagnosis.

[0009] More specifically, according to one aspect of the present invention, a method for recognizing dicentric chromosome images based on a CNN convolutional neural network is provided, comprising:

[0010] 1) Obtain a map of the sample chromosome;

[0011] 2) Binarize the image;

[0012] 3) Perform morphological processing and cropping on the chromosomes in the binary image;

[0013] 4) constructing a CNN model and using the images obtained in step 3) to train the CNN model, and

[0014] 5) Recognition and classification of dicentric chromosome images based on CNN model.

[0015] According to the implementation scheme of the present invention, in step 2), the grayscale threshold is determined according to the pixel value of the image for global binarization, including calculating the background color and contrast by taking the mode and extreme value difference of the pixel values. When the background color is greater than 244, the background color and the chromosome color difference are large, and direct global binarization is selected; if the background color does not exceed 244 and is close to the chromosome color value, the grayscale threshold interval [50, 100] and [100, 150] determined based on the mode is enabled for weighted binarization operation. When the extreme value difference is large (for example, >60), the median in the grayscale threshold interval is selected as the threshold for global binarization; when the extreme value difference is small (for example, ≤60), the grayscale threshold is determined according to the following formula:

[0016] (median within the grayscale threshold range × weight - 55) / 255,

[0017] The weights are selected according to the extreme value difference and the corresponding weight values ​​are [1.75, 1.80, 1.85, 1.90, 1.95, 2.00].

[0018] According to an embodiment of the present invention, step 3) comprises:

[0019] Use median filtering to filter and delete noise or tiny impurity pixels in binary images;

[0020] The image is expanded and eroded so that the 1-value area in the binary image completely covers the chromosome position to prevent the inability to completely extract a single chromosome image and further enhance the visual characteristics of the chromosome in the binary image.

[0021] Performing connected domain search on the binary graph to derive a large dataset of 1-valued pixel position matrices covering the chromosome region; and

[0022] Individual images are cut out one by one from the original image corresponding to each chromosome position matrix in the data set and saved.

[0023] According to an embodiment of the present invention, an angle checking algorithm is set in the saving process, including: extracting the upper, lower, left and right boundary pixel positions of a single chromosome, using the trigonometric function formula angle a=tan(y / x) to calculate the approximate angle and then rotating the screenshot so that the chromosomes in the image are all at an angle of 45° to ensure the quality of the capture and the uniformity of the morphology, which is convenient for subsequent identification.

[0024] According to the implementation scheme of the present invention, in step 4), the CNN convolutional neural network architecture includes an input layer, a convolution layer, a pooling layer and an output layer, wherein the activation function is a Relu function; during model training, 50% of the total sample data is randomly selected as a training set, 30% as a validation set, and the last 20% is used as a test set to obtain the recognition accuracy of the model.

[0025] According to an embodiment of the present invention, model training includes:

[0026] ① Read all sample grayscale images as numerical matrix groups;

[0027] ② Randomly select 50% of the image numerical matrix and input it into the fully connected layer neural network for training;

[0028] ③After a single training is completed, 30% of the image numerical matrix is ​​used as a validation set to verify the training effect and calculate the loss function. If the loss function converges, the training is completed; if the loss function does not converge, the eigenvalue parameters in the training are recorded and back-propagated to the next training;

[0029] ④ When the training reaches 50 times, the training stops and the accuracy is output; and

[0030] ⑤Save the training model and corresponding parameters.

[0031] According to an embodiment of the present invention, step 5) includes selecting the model with the highest accuracy in the training set for recognition and classification, and identifying the dicentric chromosomes.

[0032] According to an embodiment of the present invention, the method for identifying dicentric chromosome images based on a CNN convolutional neural network further comprises step 6), marking the identified dicentric chromosomes.

[0033] According to another aspect of the present invention, there is also provided a dicentric chromosome image recognition device based on a CNN convolutional neural network, comprising:

[0034] 1) An acquisition module for acquiring a sample chromosome image; 2) A binarization processing module for binarizing an image; 3) A morphological processing and cropping module for morphologically processing and cropping chromosomes in a binarized image; 4) A CNN model construction and training module for constructing and training a CNN model; and 5) An identification and classification module for identifying and classifying dicentric chromosome images.

[0035] According to another aspect of the present invention, there is also provided an electronic device, comprising: a memory and one or more processors; the memory is used to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in the present invention.

[0036] The present invention can effectively fill the domestic technical gap in the field of chromosome identification and analysis. Globally, automatic chromosome identification and analysis are widely needed in many fields, and my country is also in urgent need of independent identification systems / software in this regard. The present invention is not only of great value to my country's scientific research and medical fields, but also provides new ideas and methods for related research in other countries and regions.

[0037] The present invention uses advanced convolutional neural networks (CNN) and morphological preprocessing methods, which makes it extremely accurate in chromosome image recognition and abnormality detection. In particular, by increasing the training material and the number of training times, the accuracy of the algorithm can be further improved. This is of great value to the fields of medical diagnosis, genetic research, etc., because in these fields, accuracy is crucial.

[0038] Another advantage of the present invention is its versatility. Specifically, the method can be used not only to identify and detect dicentric chromosomes, but also to identify and detect other abnormal morphologies of chromosomes, such as rings or broken chromosomes, by adding new classifications. This allows the invention to cope with a variety of complex and ever-changing identification and detection tasks, and its application range and value far exceed methods based on pure morphological analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flowchart of dicentric chromosome image recognition based on a CNN convolutional neural network according to an embodiment of the present invention;

[0040] Figure 2 A comparison diagram of the effects of a binarization method and other binarization methods in a dicentric chromosome image recognition method based on a CNN convolutional neural network according to an embodiment of the present invention;

[0041] Figure 3 It is a diagram showing the effect of adaptive binarization in the dicentric chromosome image recognition method based on CNN convolutional neural network according to an embodiment of the present invention;

[0042] Figure 4 A schematic diagram of the morphological chromosome interception step in the dicentric chromosome image recognition method based on a CNN convolutional neural network according to an embodiment of the present invention;

[0043] Figure 5 A CNN convolutional neural network architecture diagram of a morphological chromosome interception step in a dicentric chromosome image recognition method based on a CNN convolutional neural network according to an embodiment of the present invention;

[0044] Figure 6 A schematic diagram of a dicentric chromosome image recognition device based on a CNN convolutional neural network according to an embodiment of the present invention;

[0045] Figure 7 is a schematic diagram of a principle of an electronic device according to an embodiment of the present invention; and

[0046] Figure 8 This is a comparison diagram of the effects of the dicentric chromosome image recognition method based on the CNN convolutional neural network according to an embodiment of the present invention and other recognition methods. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] Figure 1 Flow chart of the dicentric chromosome image recognition method based on CNN convolutional neural network according to the embodiment of the present invention. As shown in the figure, the dicentric chromosome image recognition method based on CNN convolutional neural network according to the embodiment can be designed as a software program / module and implemented by computer.

[0049] First, the sample chromosome images are obtained and batch input. In the preliminary stage, the program quickly inputs large quantities of chromosome images according to the chromosome image number to ensure that all image data can be accurately and efficiently read and stored.

[0050] These images are usually obtained in the form of grayscale images, but the brightness and contrast of the images vary greatly. The method of the present invention can calculate the background color and contrast based on the pixel values ​​of the grayscale image by taking the mode and the extreme value difference of the values, and then combine the two values: the mode determines the grayscale threshold interval [50,100] and [100,150], and the extreme value difference is large, such as >60 (the specific value can be determined after multiple experiments based on the material image), then the median in the grayscale threshold interval is selected as the threshold for global binarization and filling in the chromosome crossover If the extreme value difference is small, for example ≦60 (the specific value can be determined after multiple experiments based on the material image), the weight is selected according to its value (the weight is: [1.75, 1.80, 1.85, 1.90, 1.95, 2.00], and the weight parameter of the corresponding size is selected according to the extreme value difference). The specific grayscale threshold is determined by weighted calculation in the grayscale interval (the specific formula is: (interval median × weight-55) / 255), and then global binarization is performed to fill the color "holes" caused by chromosome crossing. This method realizes the adaptive binarization of the original grayscale image, the black and white image is more in line with the chromosome position, and the black and white domain segmentation edge is finer, which provides support for subsequent morphological analysis and properly solves the problem of blurred chromosome boundaries after single threshold or Otsu global threshold binarization.

[0051] Figure 2 FIG. 1 is a comparison diagram of the effects of the binarization method and other binarization methods in the dicentric chromosome image recognition method based on the CNN convolutional neural network according to the embodiment of the present invention. It can be clearly seen from the comparison diagram that Figure 2Among the three methods (adaptive grayscale binarization, Otsu binarization and single fixed threshold binarization of the present invention), the adaptive grayscale binarization illustration effect of the present invention has less edge blurring burrs and background noise, and the chromosome boundary is clearer.

[0052] Figure 3 This is a diagram of the adaptive binarization effect in the dicentric chromosome image recognition method based on the CNN convolutional neural network according to the embodiment of the present invention. In the embodiment, original images with different brightness, color and background grayscale are selected for adaptive grayscale binarization arrangement and display. It can be seen that the binarization effect of the present invention is stable and universal.

[0053] Morphological processing and cropping can then be performed.

[0054] The input image data is morphologically processed and cropped, for example, it may include steps such as expansion, erosion, and connected domain retrieval. Specific steps may include: first, use median filtering (window selection 5×5) to filter and delete the noise or tiny impurity pixels that appear in the binary image; then expand and erode the image so that the "1" value area in the binary image completely covers the chromosome position to prevent the inability to completely extract a single chromosome image, and further enhance the visual characteristics of the chromosome in the binary image. After that, the binary image is searched for a connected domain, and a large data set of the 1-value pixel position matrix covering the chromosome area is derived. The above-mentioned median filtering denoising, expansion erosion and connected domain retrieval are known in the art and will not be described in detail. Finally, a single image is cut out from the original image corresponding to each chromosome position matrix in the data set and saved. During the saving process, an angle check algorithm can be set to extract the upper, lower, left, and right boundary pixel positions of a single chromosome. y = left and right pixel position difference, x = upper and lower pixel position difference, and the trigonometric function formula angle a = tan (y / x) is used to calculate the approximate angle and rotate the screenshot so that the chromosomes in the image are all at an oblique angle of 45° to ensure the quality of the capture and uniform morphology, which is convenient for later machine learning program recognition.

[0055] Therefore, the present invention uses median filtering denoising, expansion corrosion, connected domain retrieval and rotation algorithm procedures to efficiently lock the chromosome region pixels, cut out the chromosome and accurately cut it, avoiding the loss of important information in the process. The specific process is as follows Figure 4 shown. Figure 4 The figure is a schematic diagram of the steps of morphologically intercepting chromosomes in the dicentric chromosome image recognition method based on a CNN convolutional neural network according to an embodiment of the present invention.

[0056] Then, the CNN convolutional neural network model is built using the deep learning framework keras platform, and classification, training and recognition are performed based on the CNN model.

[0057] The Keras is a highly modular neural network library based on Theano, which can be used to build neural network models. More specifically, a convolutional neural network model is constructed in the present invention, including: first, initialize the network structure: input layer: determine the size of the input data: the picture pixels are 80×80. Next, set the network parameters: convolution layer: select a 5×5 convolution kernel; activation function: select the Relu function as the activation function; pooling layer: set the maximum pooling layer size: 2×2; set the number of repeated convolution and pooling iterations: 50; set the fully connected layer to output the final result. Then train and verify the network: select 70% of the input material package as the training set and 20 as the verification set to start CNN deep learning training. The CNN convolutional neural network architecture of the implementation scheme of the present invention is as follows Figure 5 shown.

[0058] Then, the cropped chromosome image is input into the CNN convolutional neural network model for chromosome classification. This stage can be divided into two parts:

[0059] The first part is the training of the CNN convolutional neural network model. The present invention can select pictures of various brightness, contrast and file formats for chromosome extraction. The samples are divided into three categories: 0 is image impurities, 1 is positive dicentric chromosome images, and 2 is negative non-dicentric chromosome images.

[0060] When training the model, 50% of the total sample data is randomly selected as the training set, 30% as the validation set, and the last 20% as the test set to obtain the recognition accuracy of the model.

[0061] The specific training process is as follows:

[0062] ① Read all sample grayscale images as numerical matrix groups;

[0063] ② Randomly select 50% of the image numerical matrix and input it into the fully connected layer neural network for training;

[0064] ③After a single training, 30% of the image matrix is ​​used as a validation set to verify the training effect and calculate the loss function. If the loss function converges, the training is completed; if the loss function does not converge, the eigenvalue parameters in the training are recorded and back-propagated to the next training;

[0065] ④ When the training reaches 50 times, the training stops and the accuracy is output; and

[0066] ⑤Save the training model and all corresponding parameters.

[0067] The second part is that after the model is trained multiple times and saved, the model with the highest accuracy in the training set is selected for recognition and classification. The accuracy selection method of the present invention can be: the model will obtain a comprehensive accuracy according to the verification set during training. For the test set, the present invention records the actual type of the test sample and the model feedback type data to obtain the number of true positives (TP: True Positive), false positives (FP: False Positive), true negatives (True Negative), and false negatives (False Negative) in the model training results. The accuracy calculation of the present invention includes positive accuracy: TP / (TP+FP), negative accuracy: TN / (TN+FN), and comprehensive accuracy: (TP+TN) / (TP+FP+TN+FN). The implementation plan can select the model with the highest accuracy among all training models to ensure the accuracy and applicability of the recognition effect, and then batch read the pictures to be recognized, intercept a single chromosome in the picture according to the steps of step 2, and input it into the model for recognition feedback to determine the dicentric chromosomes in the picture.

[0068] Finally, the recognition and classification results are presented in the form of images, and the dicentric chromosomes are marked. This process involves designing a user-friendly interface so that non-professional users can also use the software easily.

[0069] The present invention also provides a dicentric chromosome image recognition device based on a CNN convolutional neural network. Figure 6 As shown, the device includes: an acquisition module 610, used to acquire a sample chromosome image; a binarization processing module 620, used to perform binarization processing on the image; a morphological processing and cropping module 630, used to perform morphological processing and cropping on the chromosomes in the binarized image; a CNN model construction and training module 650, used to construct and train the CNN model; and an identification and classification module 650, used to identify and classify dicentric chromosome images.

[0070] Figure 7 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention is shown in FIG. Figure 7 As shown, the electronic device includes a processor 710, a memory 720, an input device 730, and an output device 740; the number of the processor 710 in the electronic device can be one or more. Figure 7 A processor 710 is taken as an example; the processor 710, the memory 720, the input device 730 and the output device 740 in the electronic device can be connected via a bus or other means. Figure 7 The example of connecting through bus is taken in the following.

[0071] The memory 720, as a computer-readable storage medium, can be used to store software programs, computer executable programs and modules, such as program instructions / modules corresponding to the dicentric chromosome image recognition method based on the CNN convolutional neural network in the embodiment of the present invention (for example, the acquisition module 610, the binarization processing module 620, the morphological processing and cropping module 630, the CNN model construction and training module 640, and the recognition and classification module 650). The processor 710 executes various functional applications and data processing of the electronic device by running the software programs, instructions and modules stored in the memory 720, that is, realizing the above-mentioned dicentric chromosome image recognition method based on the CNN convolutional neural network.

[0072] The memory 720 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the terminal, etc. In addition, the memory 720 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other non-volatile solid-state storage device. In some instances, the memory 720 may further include a memory remotely arranged relative to the processor 710, and these remote memories may be connected to the electronic device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The input device 730 may be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the electronic device. The output device 740 may include a display device such as a display screen.

[0073] In order to verify the actual effect of the dicentric chromosome image recognition method based on CNN convolutional neural network proposed in the present invention, the above method was applied to test 19,943 chromosome images in the total sample. The samples are divided into 3 categories, 0 is image impurity, 1 is positive dicentric chromosome image, and 2 is negative non-dicentric chromosome image. The sample contains 13,753 positive chromosome images, 599 impurity images, 5,580 negative chromosome images, and 11 rings (negative chromosomes are counted in model training). The model with the highest accuracy among all training models is selected: the comprehensive accuracy of the validation set: 93.67%, positive accuracy: 90.21%, negative accuracy: 88.5%, and comprehensive accuracy: 89.37%. In addition, for comparison, experiments were also carried out using the morphological recognition method. Some recognition results are shown as follows. Figure 8 As shown, the results show that compared with the morphological identification method, the method of the present invention has higher accuracy and achieves a recognition rate of more than 87%. That is, the method of the present invention has good accuracy and stability in the identification of dicentric chromosomes, and provides a reliable and effective method for the identification of dicentric chromosomes.

[0074] The above description of the embodiments is to facilitate the understanding and application of the present invention by those skilled in the art. It is obvious that those skilled in the art can easily make various modifications to these embodiments and apply the general principles described herein to other embodiments without creative work. Therefore, the present invention is not limited to the embodiments herein, and improvements and modifications made by those skilled in the art based on the disclosure of the present invention without departing from the scope of the present invention should be within the scope of protection of the present invention.

Claims

1. A dicentric chromosome image recognition method based on CNN convolutional neural network, comprising: 1) Obtain a map of the sample chromosome; 2) Binarize the image; 3) Perform morphological processing and cropping on the chromosomes in the binary image; 4) constructing a CNN model and using the images obtained in step 3) to train the CNN model, and 5) Recognition and classification of dicentric chromosome images based on CNN model; Among them, in step 2), the grayscale threshold is determined according to the pixel value of the image for global binarization, including calculating the background color and contrast by taking the mode and extreme value difference of the pixel value. When the background color is greater than 244, the background color and the chromosome color difference are large, and direct global binarization is selected; if the background color does not exceed 244 and is close to the chromosome color value, the grayscale threshold interval [50, 100] and [100, 150] determined based on the mode is enabled for weighted binarization operation. When the extreme value difference is large, the median in the grayscale threshold interval is selected as the threshold for global binarization; when the extreme value difference is small, the grayscale threshold is determined according to the following formula: (median within the grayscale threshold range × weight - 55) / 255, Among them, the weights are selected according to the size of the extreme value difference and the corresponding weight values ​​are [1.75, 1.80, 1.85, 1.90, 1.95, 2.00].

2. The method for recognizing dicentric chromosome images based on a CNN convolutional neural network as claimed in claim 1, characterized in that: Step 3) includes: Use median filtering to filter and delete noise or tiny impurity pixels in binary images; The image is expanded and eroded so that the 1-value area in the binary image completely covers the chromosome position to prevent the inability to completely extract a single chromosome image and further enhance the visual characteristics of the chromosome in the binary image. Performing connected domain search on the binary graph to derive a large dataset of 1-valued pixel position matrices covering the chromosome region; and Individual images are cut out one by one from the original image corresponding to each chromosome position matrix in the data set and saved.

3. The method for recognizing dicentric chromosome images based on a CNN convolutional neural network as claimed in claim 2, characterized in that: An angle checking algorithm is set in the saving process, including: extracting the upper, lower, left and right boundary pixel positions of a single chromosome, taking y=left and right pixel position difference, x=upper and lower pixel position difference, using the trigonometric function formula angle a=tan(y / x) to calculate the approximate angle and then rotate the screenshot so that the chromosomes in the image are all at an oblique angle of 45° to ensure the quality of the capture and the uniformity of the morphology, which is convenient for subsequent identification.

4. The method for recognizing dicentric chromosome images based on a CNN convolutional neural network as claimed in claim 1, characterized in that: In step 4), the CNN convolutional neural network architecture includes an input layer, a convolution layer, a pooling layer, and an output layer, wherein the activation function is a Relu function; during model training, 50% of the total sample data is randomly selected as a training set, 30% as a validation set, and the last 20% as a test set to obtain the recognition accuracy of the model.

5. The method for recognizing dicentric chromosome images based on CNN convolutional neural network as claimed in claim 4, characterized in that: Model training includes: ① Read all sample grayscale images as numerical matrix groups; ② Randomly select 50% of the image numerical matrix and input it into the fully connected layer neural network for training; ③After a single training is completed, 30% of the image numerical matrix is ​​used as a validation set to verify the training effect and calculate the loss function. If the loss function converges, the training is completed; if the loss function does not converge, the eigenvalue parameters in the training are recorded and back-propagated to the next training; ④ When the training reaches 50 times, the training stops and the accuracy is output; and ⑤Save the training model and corresponding parameters.

6. The method for recognizing dicentric chromosome images based on a CNN convolutional neural network as claimed in claim 4, characterized in that: Step 5) includes selecting the model with the highest accuracy in the training set for recognition and classification, and identifying the dicentric chromosomes.

7. The method for recognizing dicentric chromosome images based on a CNN convolutional neural network as claimed in claim 1, characterized in that: The method further comprises step 6), marking the identified dicentric chromosomes.

8. A dicentric chromosome image recognition device based on CNN convolutional neural network, characterized in that: include: 1) An acquisition module, used to obtain a map of sample chromosomes; 2) Binarization processing module, used for performing binarization processing on the image; 3) A morphological processing and cropping module, used for performing morphological processing and cropping on chromosomes in the binary image; 4) CNN model construction and training module, used to build and train CNN models, and 5) Recognition and classification module, used for recognition and classification of dicentric chromosome images; In the binarization processing module, the grayscale threshold is determined according to the pixel value of the image for global binarization, including the background color and contrast calculated by taking the mode and extreme value difference of the pixel value. When the background color is greater than 244, the background color and the chromosome color difference are large, and direct global binarization is selected; if the background color does not exceed 244 and is close to the chromosome color value, the grayscale threshold interval [50, 100] and [100, 150] determined based on the mode is enabled for weighted binarization. When the extreme value difference is large, the median in the grayscale threshold interval is selected as the threshold for global binarization; when the extreme value difference is small, the grayscale threshold is determined according to the following formula: (median within the grayscale threshold range × weight - 55) / 255, Among them, the weights are selected according to the size of the extreme value difference and the corresponding weight values ​​are [1.75, 1.80, 1.85, 1.90, 1.95, 2.00].

9. An electronic device, characterized in that: include: memory and one or more processors; The memory is used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 7.

Images