Chromosome image analysis method and system based on progressive segmentation and focused classification

Abstract

The application provides a chromosome image analysis method and system based on progressive segmentation and focused classification, and comprises the following steps: a progressive segmentation step: performing segmentation on a metaphase chromosome image by using progressive segmentation to obtain a single chromosome image set; a focused classification step: identifying the categories of all single chromosomes in the single chromosome image set by using a focused classification algorithm, and finally generating a karyotype analysis graph. The application combines a traditional chromosome processing method and deep learning to form a progressive segmentation method, which gradually and effectively separates the adherent chromosome clusters, has high segmentation precision, and does not involve manual participation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic chromosome analysis, in particular, to a chromosome image analysis method and system based on progressive segmentation and focused classification. In particular, preferably, it relates to a full-automatic chromosome image analysis method based on progressive segmentation and focused classification. BACKGROUND

[0002] The chromosome image analysis method, i.e., the chromosome karyotype analysis algorithm, is a common means for diagnosing whether chromosomes have number and structural abnormalities, such as 21 trisomy syndrome, leukemia, etc. Chromosome karyotype analysis plays a crucial role in the prenatal diagnosis of whether a fetus has serious defects or genetic diseases. A normal human body includes 46 chromosomes: 22 pairs of autosomes and 1 pair of sex chromosomes XX or XY. Early human chromosome karyotype analysis needs to be completed manually by professional medical workers, which requires a high level of knowledge and operation proficiency of the operator and is a time-consuming and labor-intensive work with extremely low efficiency. Therefore, in recent years, many scholars have devoted themselves to the research of full-automatic chromosome analysis methods.

[0003] Chromosome image analysis mainly includes two steps: chromosome segmentation and chromosome classification. The accuracy of chromosome image segmentation has a non-negligible influence on the accuracy of later classification, and the segmentation of touching and overlapping chromosomes has always been an important link affecting the accuracy of segmentation. In addition, the accuracy of chromosome classification determines whether the whole chromosome karyotype analysis method can be used in clinical practice.

[0004] The Chinese invention patent document with the publication number CN113781505A discloses a chromosome segmentation method and device, a chromosome analyzer, and a storage medium. The method includes: performing image processing on an original image to obtain a mask image corresponding to the original image, the original image containing crossing chromosomes; performing image skeleton extraction on the mask image to obtain a skeleton image; determining a target pixel point on the skeleton according to the pixel value of each pixel point in the skeleton image, the target pixel point corresponding to a crossing point of the crossing chromosomes in the original image; determining a plurality of contour key points corresponding to the crossing chromosomes in the original image; determining a segmentation point of the crossing chromosomes in the original image according to the crossing point and the plurality of contour key points; and segmenting the crossing chromosomes in the original image based on the segmentation point.

[0005] For the related art in the above, the inventors believe that the traditional segmentation method is mainly threshold segmentation, and the final result of such segmentation method depends heavily on image quality and manual intervention; and most of the existing deep learning chromosome segmentation methods are specifically for synthetic sticky chromosome data sets, and the effect is not good when they are applied to real chromosome data. In the research of chromosome image classification method, the application of deep learning classification has become the mainstream, but a part of scholars directly use artificially segmented single chromosome data for classification, and another part of scholars have no manual intervention but the classification accuracy is low. SUMMARY

[0006] In view of the defects in the prior art, the purpose of the present application is to provide a chromosome image analysis method and system based on progressive segmentation and focused classification.

[0007] According to the chromosome image analysis method based on progressive segmentation and focused classification provided by the present application, the following steps are included:

[0008] Progressive segmentation step: using progressive segmentation to segment the metaphase chromosome image to obtain a single chromosome image set;

[0009] Focused classification step: using a focused classification algorithm to identify the categories of all single chromosomes in the single chromosome image set, and finally generating a karyotype analysis diagram.

[0010] Preferably, the progressive segmentation step includes the following steps:

[0011] Global threshold segmentation step: the average value of all pixel point values in the entire metaphase chromosome image is calculated as a threshold value, the pixel points below the threshold value are set to a first predetermined value, and the pixel points above the threshold value are set to a second predetermined value, and the entire metaphase chromosome image is divided into foreground and background according to the first predetermined value and the second predetermined value;

[0012] Chromosome cluster classification step: determining the real chromosome cluster in the foreground, and classifying the real chromosome cluster through a chromosome cluster classification network to obtain category information of the chromosome cluster;

[0013] Chromosome instance segmentation step: segmenting the chromosome cluster according to the category information of the chromosome cluster to obtain a single chromosome image set in the metaphase chromosome image.

[0014] Preferably, the chromosome cluster classification step includes the following steps:

[0015] Foreground discrimination step: before classification, the single chromosome and the real chromosome cluster of the foreground object are discriminated by using a convex hull and a predetermined circumscribed figure, wherein the single chromosome is directly output to the single chromosome set, and the real chromosome cluster will be further segmented by the following steps;

[0016] Multi-scale feature map acquisition step: input the real chromosome cluster into the residual network to obtain a multi-scale feature map;

[0017] Multi-scale feature map calibration step: after inputting the multi-scale feature map into the respective squeeze and excitation module, a calibrated multi-scale feature map is output;

[0018] Multi-scale feature fusion step: the calibrated multi-scale feature map is fused by a formula to obtain a multi-scale fusion feature map;

[0019] Final feature map acquisition step: after the multi-scale fusion feature map is subjected to an average pooling operation, feature connection is performed to obtain a final feature map of each chromosome cluster;

[0020] Category information acquisition step: after the final feature map is subjected to a plurality of fully connected layers and a normalized exponential function excitation function, category information of the chromosome cluster is obtained; the category information of the chromosome cluster is a touching chromosome cluster, an overlapping chromosome cluster, and a touching and overlapping chromosome cluster.

[0021] Preferably, the chromosome instance segmentation step comprises the following steps:

[0022] Touching segmentation step: the skeleton graph, terminal points, and contour graph of the touching chromosome cluster are extracted to obtain cutting points of the intersection region;

[0023] Connecting the pair of cutting points to obtain a guide line, performing dilation on the guide line to obtain a connection region, calculating the mean value of all pixel points in the connection region as a threshold, then retaining all pixel points in the connection region that are lower than the threshold, and then using the least binary method to process these pixel points to fit a segmentation line between the touching chromosomes, and using the segmentation line to segment the touching chromosome cluster;

[0024] Overlapping segmentation step: using U-Net to extract the overlapping region of the overlapping chromosome cluster;

[0025] Splicing the images on both sides of the overlapping region with the overlapping region respectively to separate all single chromosomes in the overlapping chromosome cluster;

[0026] Touching and overlapping segmentation step: first, the overlapping chromosome cluster in the touching and overlapping chromosome cluster is segmented using the overlapping segmentation step, and then a set of single chromosomes in the touching and overlapping chromosome cluster and a touching chromosome cluster are obtained, and the convex hull and the predetermined circumscribed figure are used to distinguish the touching chromosome cluster in the touching and overlapping chromosome cluster, wherein the separated single chromosomes are directly output to the single chromosome set, and the touching chromosome cluster will be further segmented by the following steps;

[0027] The touching segmentation step is used to segment the touching chromosome cluster in the touching and overlapping chromosome cluster;

[0028] Single chromosome collection step: after the touching segmentation step, the overlapping segmentation step and the touching overlapping segmentation step, the touching chromosome clusters, the overlapping chromosome clusters and the touching overlapping chromosome clusters are all segmented, and all single chromosome image sets in the metaphase chromosome image are obtained.

[0029] Preferably, the focusing classification step comprises the following steps:

[0030] Focusing step: focusing on the salient region in the input single chromosome image, and extracting the salient region;

[0031] Local network step: inputting the salient region image obtained in the focusing step into the local network to extract the local feature map of the single chromosome;

[0032] Global network step: inputting the single chromosome image into the global network to obtain the global feature map;

[0033] Classification network step: connecting the local feature map and the global feature map, and inputting them into the classification network to output the single chromosome class information;

[0034] Analysis graph generation step: generating the karyotype analysis graph by combining the single chromosome image set and the single chromosome class information.

[0035] Preferably, the focusing step comprises the following steps:

[0036] Basic feature extraction step: inputting the single chromosome image into the basic feature extraction network to extract a plurality of scale single chromosome feature maps;

[0037] Multi-dimensional attention feature generation step: inputting the plurality of scale single chromosome feature maps into the multi-dimensional attention mechanism network to generate a multi-dimensional attention feature map;

[0038] Multi-dimensional attention feature fusion step: inputting the generated multi-dimensional attention feature map into the feature fusion network to generate a final fusion feature map;

[0039] Mask extraction step: inputting the final fusion feature map into the mask generation network to extract the mask of the salient region;

[0040] Mask processing step: performing morphological processing on the generated mask of the salient region to remove noise points and fill in holes, multiplying the input original image to obtain a single chromosome salient region image.

[0041] According to the chromosome image analysis system based on progressive segmentation and focusing classification provided by the application, the following modules are included:

[0042] Progressive segmentation module: using progressive segmentation to segment the metaphase chromosome image to obtain a single chromosome image set;

[0043] Focus classification module: a focus classification algorithm is used to identify the categories of all single chromosomes in a single chromosome image set, and finally a karyotype analysis diagram is generated.

[0044] Preferably, the progressive segmentation module comprises the following modules:

[0045] Global threshold segmentation module: the average value of all pixel point values in the entire metaphase chromosome image is calculated as a threshold, pixel points below the threshold are set to a first predetermined value, and pixel points above the threshold are set to a second predetermined value, and the entire metaphase chromosome image is divided into foreground and background according to the first predetermined value and the second predetermined value;

[0046] Chromosome cluster classification module: determine the real chromosome cluster in the foreground, and classify the real chromosome cluster through the chromosome cluster classification network to obtain the category information of the chromosome cluster;

[0047] Chromosome instance segmentation module: according to the category information of the chromosome cluster, the chromosome cluster is segmented to obtain a single chromosome image set in the metaphase chromosome image.

[0048] Preferably, the chromosome cluster classification module comprises the following modules:

[0049] Foreground discrimination module: before classification, the single chromosome and the real chromosome cluster of the foreground object are discriminated by using the convex hull and the predetermined circumscribed figure, wherein the single chromosome is directly output to the single chromosome set, and the real chromosome cluster will be further segmented by the following modules;

[0050] Multi-scale feature map acquisition module: input the real chromosome cluster into the residual network to obtain a multi-scale feature map;

[0051] Multi-scale feature map calibration module: after inputting the multi-scale feature map into the respective squeeze excitation module, a calibrated multi-scale feature map is output;

[0052] Multi-scale feature fusion module: the calibrated multi-scale feature map is fused by the formula to obtain a multi-scale fusion feature map;

[0053] Final feature map acquisition module: after the multi-scale fusion feature map is subjected to an average pooling operation, a feature connection is performed to obtain a final feature map of each chromosome cluster;

[0054] Category information acquisition module: after the final feature map is subjected to a plurality of fully connected layers and a normalized exponential function excitation function, category information of the chromosome cluster is obtained; the category information of the chromosome cluster is a touching chromosome cluster, an overlapping chromosome cluster, and a touching and overlapping chromosome cluster.

[0055] Preferably, the chromosome instance segmentation module comprises the following modules:

[0056] Touch segmentation module: the skeleton graph, terminal point and contour graph of the touch chromosome cluster are extracted to obtain the cutting point of the intersection region;

[0057] The guide line is obtained by connecting the pair of cutting points, the connection region is obtained by performing inflation processing on the guide line, the mean value of all pixel points in the connection region is calculated as a threshold, then all pixel points in the connection region below the threshold are reserved, and the segmentation line between the touch chromosomes is fitted by processing the pixel points by using the minimum binary method, and the touch chromosome cluster is segmented by using the segmentation line;

[0058] Overlap segmentation module: the overlapping region of the overlapping chromosome cluster is extracted by using U-Net;

[0059] The image on both sides of the overlapping region is spliced with the overlapping region respectively, and all single chromosomes in the overlapping chromosome cluster are separated;

[0060] Touch overlap segmentation module: the overlapping chromosome cluster in the touch overlapping chromosome cluster is segmented by using the overlap segmentation module, and then a plurality of single chromosomes in the touch overlapping chromosome cluster and a set of touch chromosome clusters are obtained, the touch chromosome cluster in the touch overlapping chromosome cluster is distinguished by using the convex hull and the predetermined circumscribed figure, wherein the plurality of single chromosomes separated are directly output to the single chromosome set, and the touch chromosome cluster is continuously segmented by the following module;

[0061] The touch chromosome cluster in the touch overlapping chromosome cluster is segmented by using the touch segmentation module;

[0062] Single chromosome set module: after the touch segmentation module, the overlap segmentation module and the touch overlap segmentation module, the touch chromosome cluster, the overlapping chromosome cluster and the touch overlapping chromosome cluster are all segmented, and a set of all single chromosome images in the metaphase chromosome image is obtained.

[0063] Compared with the prior art, the present application has the following beneficial effects:

[0064] 1. The full-automatic chromosome image analysis method based on progressive segmentation and focus classification of the present application combines the traditional chromosome processing method and deep learning to form a progressive segmentation method, which gradually and effectively separates the adherent chromosome cluster, has high segmentation precision, and does not involve manual participation;

[0065] 2. In order to better realize chromosome classification, the present application further extracts the chromosome saliency feature by using the focus network, combines the global feature to generate enhanced chromosome feature representation, and greatly improves the classification precision;

[0066] 3. The application is a kind of full-automatic chromosome image analysis method, which can automatically generate chromosome karyotype analysis graph on the basis of high-precision segmentation and classification, and lay a certain foundation for the development of full-automatic chromosome image analysis system. BRIEF DESCRIPTION OF DRAWINGS

[0067] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments thereof, read in conjunction with the accompanying drawings:

[0068] Figure 1 is the overall structure diagram of the full-automatic chromosome image analysis method based on progressive segmentation and focused classification of the application;

[0069] Figure 2 is a schematic diagram of metaphase G-banded chromosome image in the application;

[0070] Figure 3 is a global threshold segmentation binary graph in the application;

[0071] Figure 4 is a chromosome cluster classification network structure diagram in the application;

[0072] Figure 5 is a chromosome cluster image class label diagram in the application;

[0073] Figure 6 is a touching chromosome segmentation result diagram in the application;

[0074] Figure 7 is an overlapping chromosome cluster segmentation result diagram in the application;

[0075] Figure 8 is a clustering network structure diagram in the application;

[0076] Figure 9 is a pixel point attention mechanism structure diagram in the application;

[0077] Figure 10 is a channel attention mechanism structure diagram in the application;

[0078] Figure 11 is a spatial attention mechanism structure diagram in the application;

[0079] Figure 12 is a generated karyotype analysis graph in the application. DETAILED DESCRIPTION

[0080] The application will be described in detail below with specific examples. The following examples will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the application. These are within the scope of protection of the application.

[0081] Embodiment one of the application discloses a full-automatic chromosome image analysis method based on progressive segmentation and focused classification, as shown in the figure, comprising the following steps: Figure 1

[0082] Progressive segmentation step: using the progressive segmentation method to segment the metaphase chromosome image, so as to obtain a single chromosome image set.

[0083] The progressive segmentation method includes three steps: (1) global threshold segmentation; (2) chromosome cluster classification; (3) chromosome instance segmentation. The specific segmentation process is as follows: that is, the progressive segmentation step comprises the following steps:

[0084] Global threshold segmentation step (first step): by calculating the average value of all pixel point values in the entire metaphase chromosome image as a threshold, the pixel points below the threshold are set to a first predetermined value, and the pixel points above the threshold are set to a second predetermined value, and the entire metaphase chromosome image is divided into foreground and background according to the first predetermined value and the second predetermined value.

[0085] Specifically, the global threshold segmentation is to calculate the average value of all pixel point values in the entire metaphase chromosome image as a threshold, set the pixel points below the threshold to 0, and set the pixel points above the threshold to 1, so as to divide the entire metaphase chromosome image into foreground and background.

[0086] Chromosome cluster classification step (second step): determining the real chromosome cluster in the foreground, and the real chromosome cluster is classified by the chromosome cluster classification network to obtain the category information of the chromosome cluster. That is, the chromosome cluster classification is classified by the chromosome cluster classification network.

[0087] The specific implementation steps of the chromosome cluster classification are: that is, the chromosome cluster classification step comprises the following steps:

[0088] Foreground discrimination step: before classification implementation, the single chromosome and the real chromosome cluster of the foreground object are discriminated by using the convex hull and the predetermined circumscribed graph. Specifically, before classification implementation, the single chromosome and the real chromosome cluster of the foreground object are discriminated by using the convex hull and the minimum circumscribed rectangle, wherein the single chromosome is directly output to the single chromosome set, and the real chromosome cluster will be further segmented by the following steps.

[0089] ​Multi-scale feature map acquisition step: input the real chromosome cluster into the residual network 50 (ResNet50) to obtain multi-scale feature maps {F2, F3, F4, F5}, wherein F2, F3, F4, F5 are the feature images output by the 10th, 22nd, 40th and 49th layers of the ResNet50 network respectively.

[0090] Multi-scale feature map calibration step: after inputting {F2, F3, F4, F5} into the respective squeeze and excitation module, the calibrated multi-scale feature maps {F'2, F'3, F'4, F'5} are output, wherein F'2, F'3, F'4, F'5 represent the respective calibrated multi-scale feature maps obtained after F2, F3, F4, F5 are input into the squeeze and excitation module. Specifically, since the squeeze and excitation (SE) module can change the weight coefficients of each channel of the convolution feature layer, it can learn important feature information in the chromosome image, and the difference features between each chromosome cluster are mainly concentrated in the bottom feature layer of the network. Therefore, by adjusting the excitation parameters in the SE module, the input multi-scale features can be enriched with bottom layer information, that is, after inputting {F2, F3, F4, F5} into the respective SE module, the calibrated multi-scale feature maps {F'2, F'3, F'4, F'5} are output.

[0091] Multi-scale feature fusion step: in order to make each layer of feature representation contain more rich bottom layer feature information, the feature fusion of {F'2, F'3, F'4, F'5} is performed by the following formula to obtain multi-scale fusion feature maps {A2, A3, A4, A5}.

[0092]

[0093]

[0094]

[0095]

[0096] Wherein, {A2, A3, A4, A5} represents the multi-scale fusion feature maps obtained by the above calculation formula, A2 and F'2, A3 and F'3, A4 and F'4, A5 and F'5 each have the same image size; Conv 1×1 () represents one-dimensional convolution calculation; Down() represents down-sampling calculation; Up() represents up-sampling calculation; represents pixel-by-pixel addition.

[0097] Final feature map acquisition step: after the multi-scale fusion feature map is subjected to average pooling operation, it is connected with the feature to obtain the final feature map of each chromosome cluster.

[0098] Category information obtaining step: the final feature map is obtained after passing through multiple fully connected layers and a normalization exponential function (Softmax) activation function, and the category information of the chromosome cluster is obtained; the category information of the chromosome cluster is a touching chromosome cluster, an overlapping chromosome cluster, and a touching and overlapping chromosome cluster. That is, the final feature map is obtained after passing through two fully connected layers and a Softmax activation function, and the category information of the chromosome cluster is obtained.

[0099] After all real chromosome clusters are input into the chromosome cluster classification network, they can be divided into three categories: a touching chromosome cluster, an overlapping chromosome cluster, and a touching and overlapping chromosome cluster.

[0100] Chromosome instance segmentation step (third step): according to the category information of the chromosome cluster, the chromosome cluster is segmented to obtain a set of all single chromosome images in the metaphase chromosome image. That is, the chromosome instance segmentation includes a touching segmentation module, an overlapping segmentation module, and a touching and overlapping segmentation module.

[0101] The specific segmentation steps of the chromosome instance segmentation step include the following steps:

[0102] Touching segmentation step (touching segmentation module): the skeleton graph, terminal point, and contour graph of the touching chromosome cluster are extracted to obtain the cutting points of the intersection region. That is, (1.1) the touching segmentation module is used to segment the touching chromosome cluster, and in order to quickly and accurately implement the segmentation, the skeleton graph, terminal point, and contour graph are extracted to obtain the cutting points of the intersection region.

[0103] The guide line is obtained by connecting the above-mentioned pair of cutting points, the connection region is obtained by performing inflation processing on the guide line, the mean value of all pixel points in the connection region is calculated as a threshold, and then all pixel points in the connection region that are lower than the threshold are retained. Then, the minimum binary method is used to process these pixel points to fit the segmentation line between the touching chromosomes, and the touching chromosome cluster is segmented by using the segmentation line. That is, (1.2) the pair of cutting points and the basic image operation are used to calculate the segmentation line between the touching chromosomes, and the segmentation line can completely segment the touching chromosome cluster.

[0104] Overlapping segmentation step (overlapping segmentation module): a classic U-shaped network (U-Net) is used to extract the overlapping region of the overlapping chromosome cluster. Specifically, (2.1) the overlapping segmentation module is used to segment the overlapping chromosome cluster, and since the overlapping chromosome image has a small size and a complex structure, a simplified U-Net is used to extract the overlapping region. Here, the structure of the simplified U-Net is a network structure composed of only two times of down-sampling and corresponding up-sampling.

[0105] The images on both sides of the overlapping region are spliced with the overlapping region respectively to separate all single chromosomes in the overlapping chromosome cluster. Specifically, (2.2) the overlapping region can be approximated as a rectangle, and then the images on both sides of the parallel edges of the rectangle are spliced with the overlapping region to obtain all single separated chromosomes.

[0106] Touch overlap segmentation step (touch and overlap segmentation module): first, the overlapping chromosome cluster in the touch overlap chromosome cluster is segmented by the overlap segmentation step, and then a set of single chromosomes in the touch overlap chromosome cluster and the touch chromosome cluster are obtained. The touch chromosome cluster in the touch overlap chromosome cluster is distinguished by the convex hull and the predetermined circumscribed graph. The single chromosomes that have been separated are directly output to the single chromosome set without processing, and the touch chromosome cluster will be further segmented by the following steps. Specifically, (3.1) the touch and overlap segmentation module is used to segment the chromosome cluster with both touch and overlap. First, the overlapping chromosome cluster is segmented by the overlap segmentation module. (3.2) After the operation of (3.1), the overlapping chromosome cluster is completely segmented to obtain a set of single chromosomes and touch chromosome clusters, and then the touch chromosome cluster is distinguished by the convex hull and the minimum circumscribed rectangle.

[0107] The touch segmentation step is used to segment the touch chromosome cluster in the touch overlap chromosome cluster. That is, (3.3) the touch chromosome cluster is segmented by the touch segmentation module.

[0108] Single chromosome set step: after the touch segmentation step, the overlap segmentation step and the touch overlap segmentation step, the touch chromosome cluster, the overlapping chromosome cluster and the touch overlap chromosome cluster are all segmented to obtain a set of all single chromosome images in the metaphase chromosome image. After the above three segmentation steps, all chromosome clusters can be segmented to obtain a set of all single chromosomes in the metaphase chromosome image.

[0109] Focus classification step: a focus classification algorithm is used to identify the categories of all single chromosomes in the single chromosome image set, and finally a karyotype analysis graph is generated. That is, a focus classification algorithm is used to identify the categories of all single chromosomes, and finally a karyotype analysis graph is generated. The focus classification network (Focus-Net) includes four modules: focus module, local network module, global network module and classification module.

[0110] The specific classification process of the focus classification step includes the following steps:

[0111] Focus step: focusing on the salient region in the input single chromosome image and extracting the salient region. Specifically, the focus module is used to focus on the salient region in the input single chromosome image and extract it. The focus module is composed of a feature extraction network, a multi-dimensional attention mechanism network, a feature fusion network and a mask generation network.

[0112] The salient region specific generation process is as follows: the focusing step includes the following steps:

[0113] The basic feature extraction step: a single chromosome image is input into a basic feature extraction network to extract a plurality of scale single chromosome feature maps. Specifically, (4.1) a joint residual network 50, a feature pyramid network (ResNet50-FPN) is used as the basic feature extraction network, a single chromosome image is input, and five scale single chromosome feature maps can be extracted.

[0114] The multi-dimensional attention feature generation step: the plurality of scale single chromosome feature maps are input into a multi-dimensional attention mechanism network to generate multi-dimensional attention feature maps. Specifically, (4.2) the multi-scale feature maps are input into the multi-dimensional attention mechanism network to generate multi-dimensional attention feature maps, and the multi-dimensional attention mechanism network is composed of pixel point, channel, and space attention mechanisms.

[0115] The multi-dimensional attention feature fusion step: the generated multi-dimensional attention feature maps are input into a feature fusion network to generate a final fusion feature map. Specifically, (4.3) the generated multi-dimensional attention feature maps are input into the feature fusion network, the purpose is to effectively combine the attention feature maps through a reasonable fusion method, first, the multi-scale pixel point attention feature maps are processed through upsampling to have the same resolution as the input image, the processed pixel point attention feature maps are fused through addition to obtain a pixel point attention fusion feature map, the channel and space attention feature maps are processed in the same way to obtain a channel attention fusion feature map and a space attention fusion feature map, and finally, the three types of fusion feature maps are fused again through weighted addition to generate a final feature map, and the weight coefficients are all set to 1.

[0116] The mask extraction step: the final fusion feature map is input into a mask generation network to extract a salient region mask. Specifically, (4.4) the feature map generated in (4.3) is input into the mask generation network, the purpose is to extract a salient region mask, and the mask generation network is composed of Conv1x1 and a Softmax activation function.

[0117] The mask processing step: the generated salient region mask is morphologically processed to remove noise points and fill in small holes, and is multiplied with the input original image to obtain a single chromosome salient region image. Specifically, (4.5) the generated salient region mask is morphologically processed to remove noise points and fill in small holes, and is multiplied with the input original image to obtain a single chromosome salient region image.

[0118] Local network step: input the salient region image obtained in the focusing step into the local network to extract the local feature map of the single chromosome. Specifically, the local network module is used to generate the local feature map, and inputting the salient region image obtained in the focusing step into the local network module can extract the local feature map of the single chromosome, and here the local network module is composed of a Resnet50 network.

[0119] Global network step: input the single chromosome image into the global network to obtain the global feature map. Specifically, the global network module is used to generate the global feature map, and inputting the single chromosome image into the global network module can obtain the global feature map, and here the global network module is also composed of a Resnet50 network.

[0120] Classification network step: connect the local feature map and the global feature map, input into the classification network, and output the single chromosome class information. Specifically, the classification network module is used to identify the class of the single chromosome, first connect the local and global feature maps obtained in the local network step and the global network step, and then input into the classification network module to output the chromosome class, and here the classification network module is composed of 2 fully connected layers and 1 Softmax excitation function.

[0121] Analysis diagram generation step: generate a karyotype analysis diagram combining the single chromosome image set and the single chromosome class information. That is, finally, the single chromosome set and the class information are combined to generate the final karyotype analysis diagram.

[0122] The embodiment one of the present application also discloses a full-automatic chromosome image analysis system based on progressive segmentation and focused classification, as shown in the figure, comprising the following modules: Figure 1

[0123] Progressive segmentation module: the progressive segmentation is used for segmenting the metaphase chromosome image to obtain a single chromosome image set.

[0124] The progressive segmentation module comprises the following modules:

[0125] Global threshold segmentation module: the average value of all pixel point values in the whole metaphase chromosome image is calculated as a threshold, the pixel points lower than the threshold are set as a first predetermined value, the pixel points higher than the threshold are set as a second predetermined value, and the whole metaphase chromosome image is divided into foreground and background according to the first predetermined value and the second predetermined value.

[0126] Chromosome cluster classification module: the real chromosome cluster in the foreground is determined, the real chromosome cluster is classified by the chromosome cluster classification network to obtain the class information of the chromosome cluster.

[0127] The chromosome cluster classification module comprises the following modules:

[0128] ​The foreground discrimination module: before classification, the single chromosomes and real chromosome clusters of the foreground objects are discriminated by using the convex hull and the predetermined circumscribed graph, wherein the single chromosomes are directly output to the single chromosome set, and the real chromosome clusters will continue to be segmented through the following steps.

[0129] The multi-scale feature map acquisition module: the real chromosome clusters are input into the residual network 50 (ResNet50) to obtain multi-scale feature maps {F2, F3, F4, F5}, wherein F2, F3, F4, and F5 are the feature images output by the 10th, 22nd, 40th, and 49th layers of the ResNet50 network, respectively.

[0130] The multi-scale feature map calibration module: after the input {F2, F3, F4, F5} is input into the respective squeeze-and-excitation module, the calibrated multi-scale feature maps {F'2, F'3, F'4, F'5} are output.

[0131] The multi-scale feature fusion module: the feature fusion of {F'2, F'3, F'4, F'5} is performed through the formula to obtain multi-scale fusion feature maps {A2, A3, A4, A5}.

[0132]

[0133]

[0134]

[0135]

[0136] Wherein, {A2, A3, A4, A5} represents the multi-scale fusion feature maps obtained by the above calculation formula, A2 and F'2, A3 and F'3, A4 and F'4, and A5 and F'5 each have the same image scale; Conv 1×1 () represents one-dimensional convolution calculation; Down() represents down-sampling calculation; Up() represents up-sampling calculation; represents pixel-by-pixel addition.

[0137] The final feature map acquisition module: after the multi-scale fusion feature maps are subjected to the average pooling operation, the feature connection is performed to obtain the final feature maps of each chromosome cluster.

[0138] The category information acquisition module: after the final feature maps are subjected to the multi-layer fully connected layer and the normalization exponential function (Softmax) excitation function, the category information of the chromosome clusters is obtained; the category information of the chromosome clusters is the touching chromosome cluster, the overlapping chromosome cluster, and the touching and overlapping chromosome cluster.

[0139] Chromosome instance segmentation module: according to the category information of the chromosome cluster, the chromosome cluster is segmented to obtain a single chromosome image set in the metaphase chromosome image.

[0140] The chromosome instance segmentation module comprises the following modules:

[0141] Touch segmentation module: the skeleton graph, terminal point and contour graph of the touch chromosome cluster are extracted to obtain the cutting point of the intersection region.

[0142] The guide line is obtained by connecting the above-mentioned pair of cutting points, the connection region is obtained by performing inflation processing on the guide line, the mean value of all pixel points in the connection region is calculated as a threshold, and then all pixel points in the connection region below the threshold are retained, and then the pixel points are processed by using the minimum binary method to fit the segmentation line between the touch chromosomes, and the touch chromosome cluster is segmented by using the segmentation line.

[0143] Overlap segmentation module: the overlapping region of the overlapping chromosome cluster is extracted by using a classic U-shaped network (U-Net).

[0144] The images on both sides of the overlapping region are spliced with the overlapping region respectively to separate all single chromosomes in the overlapping chromosome cluster.

[0145] Touch overlap segmentation module: first, the overlapping chromosome cluster in the touch overlapping chromosome cluster is segmented by using the overlap segmentation module, and then a set of single chromosomes in the touch overlapping chromosome cluster and the touch chromosome cluster are obtained, and the touch chromosome cluster in the touch overlapping chromosome cluster is distinguished by using the convex hull and the predetermined circumscribed figure, wherein the separated single chromosomes are directly output to the single chromosome set without processing, and the touch chromosome cluster will be further segmented by the following steps.

[0146] The touch overlap segmentation module is segmented by using the touch segmentation module.

[0147] Single chromosome set module: after the touch segmentation module, the overlap segmentation module and the touch overlap segmentation module, the touch chromosome cluster, the overlapping chromosome cluster and the touch overlapping chromosome cluster are all segmented to obtain a single chromosome image set in the metaphase chromosome image.

[0148] Focus classification module: the categories of all single chromosomes in the single chromosome image set are identified by using a focus classification algorithm, and finally a karyotype analysis graph is generated.

[0149] The focus classification module comprises the following modules:

[0150] Focus module: focusing on the significant region in the input single chromosome image, and extracting the significant region.

[0151] The focus module comprises the following modules:

[0152] Basic feature extraction module: adopt a basic feature extraction network, input a single chromosome image, and extract a plurality of scale single chromosome feature maps.

[0153] Multi-dimensional attention feature generation module: input the single chromosome feature map of multiple scales into the multi-dimensional attention mechanism network to generate a multi-dimensional attention feature map.

[0154] Multi-dimensional attention feature fusion module: the generated multi-dimensional attention feature map is input into the feature fusion network to generate a final fusion feature map.

[0155] Mask extraction module: input the final fusion feature map into the mask generation network to extract the mask of the significant region.

[0156] Mask processing module: the mask of the significant region is generated for morphological processing to remove noise points and fill holes, and is multiplied with the input original image to obtain a single chromosome significant region image.

[0157] Local network module: input the significant region image obtained by the focusing module into the local network to extract a local feature map of a single chromosome.

[0158] Global network module: input a single chromosome image into a global network to obtain a global feature map.

[0159] Classification network module: connect the local feature map and the global feature map, input into a classification network, and output single chromosome class information.

[0160] Analysis graph generation module: generate a karyotype analysis graph by combining a single chromosome image set and single chromosome class information.

[0161] Embodiment two of the present application also provides a chromosome image analysis method based on progressive segmentation and focused classification, which first uses a progressive segmentation method to segment a metaphase chromosome image, thereby obtaining a single chromosome set, and then uses a focused classification algorithm to identify the class of all single chromosomes.

[0162] 1、The metaphase G-band chromosome image shown in Figure 2 is subjected to global threshold processing, the average value of all pixel point values in the entire metaphase chromosome image is calculated as a threshold, pixel points lower than the threshold are set to 0, and pixel points higher than the threshold are set to 1, thereby dividing the entire metaphase chromosome image into foreground and background, and the processing result image is as shown in Figure 3 .

[0163] 2、The foreground includes single chromosomes and chromosome clusters, and before classifying the chromosome clusters, a convex hull and a minimum circumscribed rectangle are used to distinguish the real chromosome clusters.

[0164] 3. Train the chromosome cluster classification network using the chromosome cluster image dataset, and obtain a chromosome cluster classification network structure diagram as shown in Figure 4 .

[0165] 4. Classify the chromosome cluster images obtained in step 2 using the trained chromosome cluster classification network. The specific classification steps are as follows:

[0166] (1) Input the real chromosome clusters in step 2 into ResNet50 to obtain multi-scale feature maps {F2, F3, F4, F5} (from low to high).

[0167] (2) After inputting {F2, F3, F4, F5} into the respective SE modules, obtain calibrated multi-scale feature maps {F'2, F'3, F'4, F'5}.

[0168] (3) In order to make each layer of features represent more rich bottom feature information, the following formula is used to fuse {F'2, F'3, F'4, F'5}:

[0169]

[0170]

[0171]

[0172]

[0173] Where {A2, A3, A4, A5} represents the multi-scale fusion feature maps obtained by the above calculation formula, A2 and F'2, A3 and F'3, A4 and F'4, A5 and F'5 each have the same image size, Conv 1×1 () represents one-dimensional convolution calculation, Down() and Up() represent down-sampling and up-sampling calculation, represents pixel-wise addition.

[0174] (4) After the multi-scale fusion feature maps are subjected to average pooling operation, they are connected by feature connection to obtain the final feature maps of each chromosome cluster.

[0175] (5) After the final feature maps are subjected to two fully connected layers and a Softmax activation function, the class information of the chromosome clusters is obtained.

[0176] (6) After all real chromosome clusters in step 2 are input into the chromosome cluster classification network, they can be divided into three categories: touching chromosome clusters, overlapping chromosome clusters, and touching and overlapping chromosome clusters, as shown in Figure 5 .

[0177] 5、Chromosome instance segmentation is to segment three types of chromosome clusters respectively, and the segmentation process is as follows:

[0178] (1) Segmentation of touching chromosome clusters. In order to quickly and accurately implement segmentation, the skeleton graph, terminal point and contour graph are extracted first to obtain the cutting points of the intersection region; then the guide line is obtained by connecting the above-mentioned pairs of cutting points, and the connection region is obtained by performing dilation on the guide line; the mean value of all pixel points in the connection region is calculated as a threshold, and all pixel points in the connection region below the threshold are retained; then the touching chromosome segmentation line is fitted by using the minimum binary method to process these pixel points, and the touching chromosome cluster can be completely segmented by the segmentation line, as shown in Figure 5 .

[0179] (2) Segmentation of overlapping chromosome clusters. Since the image size of overlapping chromosomes is small and the structure is complex, the simplified U-Net is used to extract the overlapping region first. The structure of the simplified U-Net is a network structure composed of only twice down-sampling and corresponding up-sampling; secondly, the overlapping region can be approximated as a rectangle, and the single separated chromosome can be obtained by splicing the image on both sides of the parallel edges of the rectangle and the overlapping region, as shown in Figure 7 .

[0180] (3) For touching and overlapping chromosome segmentation, the overlapping region is extracted by using the simplified U-Net, and the overlapping chromosome cluster is separated by splicing. After this operation, the overlapping chromosome cluster is completely segmented, and a set of single chromosomes and touching chromosome clusters are obtained. The convex hull and the minimum circumscribed rectangle are used to distinguish the touching chromosome cluster, wherein the separated single chromosomes are directly output to the single chromosome set without processing, and the touching chromosome cluster will be further segmented by the following steps; the separation of touching chromosomes is as described in step (1) above.

[0181] After the above-mentioned 3 segmentation steps, all chromosome clusters can be separated, and a set of all single chromosomes in the metaphase chromosome image is obtained.

[0182] 6、Before performing chromosome classification and recognition, the focus network—Focus-Net is trained using the single chromosome data set. The network structure diagram of the trained Focus-Net is shown in Figure 8 , which includes four modules: clustering module, local network module, global network module and classification module. PAB, CAB and SAB represent pixel attention module, channel attention module and spatial attention module.

[0183] 7、Focus on the significant area in the input single chromosome image and extract it by using the focus module. The focus module is composed of a feature extraction network, a multi-dimensional attention mechanism network, a feature fusion network and a mask generation network. The specific generation process of the significant area is as follows:

[0184] (1) Using ResNet50-FPN as the basic feature extraction network, a single chromosome image can be input and a single chromosome feature map at 5 scales can be extracted.

[0185] (2) Input the multi-scale feature map into the multi-dimensional attention mechanism network to generate a multi-dimensional attention feature map. The multi-dimensional attention mechanism network here consists of three attention mechanisms: pixel attention, channel attention, and spatial attention. The pixel attention mechanism is as follows: Figure 9 As shown, the channel attention mechanism is as follows Figure 10 As shown, the spatial attention mechanism is as follows Figure 11 As shown.

[0186] (3) The generated multi-dimensional attention feature map is input into the feature fusion network. The purpose is to effectively combine the attention feature maps through a reasonable fusion method. First, the multi-scale pixel attention feature map can be upsampled to make its resolution the same as the input image. The processed pixel attention feature maps are fused by addition to obtain the pixel attention fusion feature map. The channel and spatial attention feature maps are processed in the same way to obtain the channel attention fusion feature map and the spatial attention fusion feature map. Finally, the three types of fusion feature maps are fused again by weighted addition to generate the final feature map. Here, the weight coefficients are all set to 1.

[0187] (4) Input the feature map generated in (3) into the mask generation network in order to extract the mask of the salient region. The mask generation network consists of Conv 1×1 It consists of the Softmax activation function.

[0188] (5) The mask for generating salient regions is first morphologically processed to remove noise points and fill small holes, and then multiplied with the original input image to obtain a single chromosome salient region image.

[0189] 8. Use the local network module to generate local feature maps. Input the salient region image obtained in step 7 into the local network module to extract the local feature map of a single chromosome. Here, the local network module is composed of a ResNet50 network.

[0190] 9. Simultaneously, a global feature map is generated using the global network module. A global feature map can be obtained by inputting a single chromosome image into the global network module, which is also composed of a ResNet50 network.

[0191] 10. Next, use the classification network module to identify the category of a single chromosome. First, connect the local and global feature maps obtained in steps 8 and 9, and then input them into the classification network module to output the chromosome category. The classification network module here consists of two fully connected layers and one Softmax activation function.

[0192] 11. Finally, the extracted individual chromosomes and their corresponding categories are combined to generate... Figure 2 Karyotype analysis diagram of chromosome images, such as Figure 12 As shown.

[0193] The fully automated chromosome image analysis method based on progressive segmentation and focused classification of the present invention combines traditional segmentation methods and deep learning methods to gradually and accurately segment touching and overlapping chromosome clusters in metaphase chromosome images. At the same time, it automatically focuses on the salient regions of individual chromosomes and extracts effective local features, which, combined with global features, greatly improves classification accuracy.

[0194] This invention presents a fully automated chromosome image analysis method based on progressive segmentation and focused classification, comprising two stages: First, a progressive segmentation stage consisting of global threshold segmentation, a chromosome cluster classification network, and chromosome instance segmentation. Global threshold segmentation is first used to extract foreground targets from human metaphase chromosome images. Then, the chromosome cluster classification network classifies adhered chromosome clusters in the foreground into three categories. Finally, chromosome instance segmentation is performed based on the characteristics of each chromosome cluster. Second, a focused classification stage utilizes a focused network (Focus-Net) composed of a focusing module, a global network module, a local network module, and a classification module. This network automatically focuses on salient regions, extracts global and local features, and identifies chromosome categories, ultimately generating a chromosome karyotype analysis map. Through this method, this invention progressively and effectively achieves complete segmentation of touching and overlapping chromosomes, while automatically focusing on salient regions of individual chromosomes, resulting in richer extracted feature information and significantly improved segmentation and classification accuracy. The entire chromosome analysis process is performed without human intervention.

[0195] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0196] The specific embodiments of the present application have been described. It is to be understood that the application is not limited to particular details described herein and that various modifications can be made therein without departing from the scope of the claimed application. Embodiments and features disclosed in this document, including in the examples, can be combined with each other, unless specifically contradicted by or inconsistent with each other.

Claims

1. A chromosome image analysis method based on progressive segmentation and focus classification, characterized in that, Includes the following steps: Progressive segmentation steps: The metaphase chromosome image is segmented using progressive segmentation to obtain a set of individual chromosome images; Focused classification steps: The focused classification algorithm is used to identify the category of all individual chromosomes in a single chromosome image set, and finally a karyotype analysis map is generated; The progressive segmentation step includes the following steps: Global threshold segmentation step: The average value of all pixels in the entire metaphase chromosome image is calculated as the threshold. Pixels below the threshold are set to the first predetermined value, and pixels above the threshold are set to the second predetermined value. The entire metaphase chromosome image is divided into foreground and background according to the first and second predetermined values. Chromosome cluster classification steps: Identify the real chromosome clusters in the foreground, and classify the real chromosome clusters through the chromosome cluster classification network to obtain the category information of the chromosome clusters; Chromosome instance segmentation steps: Segment the chromosome clusters according to their category information to obtain a set of all individual chromosome images in the metaphase chromosome image; The focused classification step includes the following steps: Focusing step: Focus on the salient region in the input single chromosome image and extract the salient region; Local network step: Input the salient region image obtained in the focusing step into the local network to extract the local feature map of a single chromosome; Global network steps: Input a single chromosome image into the global network to obtain a global feature map; The classification network process involves concatenating local and global feature maps, inputting them into the classification network, and outputting single chromosome category information. Analysis map generation steps: Combine a collection of individual chromosome images and individual chromosome category information to generate a karyotype analysis map.

2. The chromosome image analysis method based on progressive segmentation and focus classification according to claim 1, characterized in that, The chromosome cluster classification process includes the following steps: Foreground discrimination step: Before classification, the individual chromosomes and real chromosome clusters of the foreground object are identified using convex hull and a predetermined bounding graph. Individual chromosomes are directly output to a single chromosome set, while real chromosome clusters will be further segmented through the following steps. Multi-scale feature map acquisition steps: Input the real chromosome clusters into the residual network to obtain multi-scale feature maps; Multi-scale feature map calibration steps: After inputting the multi-scale feature maps into their respective extrusion excitation modules, the calibrated multi-scale feature maps are output. Multi-scale feature fusion steps: The calibration multi-scale feature map is fused using a formula to obtain a multi-scale fused feature map; Final feature map acquisition steps: After the multi-scale fused feature map is subjected to average pooling, feature concatenation is performed to obtain the final feature map of each chromosome cluster; Category information acquisition steps: After the final feature map passes through multiple fully connected layers and a normalized exponential function activation function, the category information of the chromosome clusters is obtained; the category information of the chromosome clusters is touching chromosome clusters, overlapping chromosome clusters, and touching-overlapping chromosome clusters.

3. The chromosome image analysis method based on progressive segmentation and focus classification according to claim 2, characterized in that, The chromosome instance segmentation step includes the following steps: Touch segmentation steps: Extract the skeleton map, terminal points and contour map of the touch chromosome clusters to obtain the cutting points of the intersection region; By connecting each pair of cutting points, guide lines are obtained. The guide lines are then expanded to obtain the connected regions. The mean value of all pixels in the connected regions is calculated as a threshold. All pixels in the connected regions that are below the threshold are retained. These pixels are then processed using the least binary method to fit the dividing lines between the touching chromosomes. The dividing lines are then used to segment the clusters of touching chromosomes. Overlap segmentation step: Use U-Net to extract the overlapping regions of overlapping chromosome clusters; The images on both sides of the overlapping region are stitched together with the overlapping region to separate all individual chromosomes in the overlapping chromosome cluster; Touch overlap segmentation steps: First, the overlap segmentation step is used to segment the overlapping chromosome clusters in the touch overlap chromosome clusters, and then the set of several individual chromosomes and touch chromosome clusters in the touch overlap chromosome clusters are obtained. The touch chromosome clusters in the touch overlap chromosome clusters are identified by using the convex hull and the predetermined circumscribed graph. The several individual chromosomes that have been separated are directly output to the set of individual chromosomes, while the touch chromosome clusters will be further segmented through the following steps. The touch segmentation step is used to segment touch chromosome clusters within touch overlapping chromosome clusters; Single chromosome set step: After the touch segmentation step, the overlap segmentation step, and the touch-overlap segmentation step, the touch chromosome clusters, overlap chromosome clusters, and touch-overlap chromosome clusters are all segmented, resulting in a set of all single chromosome images in the metaphase chromosome image.

4. The chromosome image analysis method based on progressive segmentation and focus classification according to claim 1, characterized in that, The focusing step includes the following steps: Basic feature extraction steps: Using a basic feature extraction network, a single chromosome image is input, and single chromosome feature maps at multiple scales are extracted; Multidimensional attention feature generation steps: Input single chromosome feature maps at multiple scales into a multidimensional attention mechanism network to generate multidimensional attention feature maps; Multidimensional attention feature fusion step: The generated multidimensional attention feature map is input into the feature fusion network to generate the final fused feature map; Mask extraction steps: Input the final fused feature map into the mask generation network to extract the mask of the salient region; Masking steps: The mask of the salient region is generated, morphological processing is performed to remove noise points and fill holes, and it is multiplied with the original input image to obtain the image of the salient region of a single chromosome.

5. A chromosome image analysis system based on progressive segmentation and focus classification, characterized in that, Includes the following modules: Progressive segmentation module: Uses progressive segmentation to segment metaphase chromosome images to obtain a set of individual chromosome images; Focused classification module: The focused classification algorithm is used to identify the category of all individual chromosomes in a single chromosome image set, and finally generate a karyotype analysis map; The progressive segmentation module includes the following modules: Global threshold segmentation module: By calculating the average value of all pixels in the entire metaphase chromosome image as a threshold, pixels below the threshold are set to a first predetermined value, and pixels above the threshold are set to a second predetermined value. The entire metaphase chromosome image is divided into foreground and background based on the first and second predetermined values. Chromosome cluster classification module: Identifies real chromosome clusters in the foreground. Real chromosome clusters are classified through a chromosome cluster classification network to obtain chromosome cluster category information; Chromosome instance segmentation module: Segments chromosome clusters based on their category information to obtain a set of all individual chromosome images in the metaphase chromosome image; The focused classification module includes the following modules: Focusing module: Focuses on the salient region in the input single chromosome image and extracts the salient region; Local network module: Input the salient region image obtained by the focusing module into the local network to extract the local feature map of a single chromosome; Global network module: Inputs a single chromosome image into the global network to obtain a global feature map; The classification network module connects local and global feature maps, inputs them into the classification network, and outputs single chromosome category information. Analysis diagram generation module: Combines a collection of single chromosome images and single chromosome category information to generate a karyotype analysis diagram.

6. The chromosome image analysis system based on progressive segmentation and focus classification according to claim 5, characterized in that, The chromosome cluster classification module includes the following modules: Foreground discrimination module: Before classification, the convex hull and the predetermined outer graph are used to distinguish the individual chromosomes and real chromosome clusters of the foreground object. The individual chromosomes are directly output to the single chromosome set, and the real chromosome clusters will be further segmented through the following module. Multi-scale feature map acquisition module: Input the real chromosome clusters into the residual network to obtain multi-scale feature maps; Multi-scale feature map calibration module: After inputting multi-scale feature maps into their respective extrusion excitation modules, it outputs calibrated multi-scale feature maps; Multi-scale feature fusion module: Performs feature fusion on the calibration multi-scale feature map using formulas to obtain a multi-scale fused feature map; Final feature map acquisition module: After the multi-scale fused feature map is subjected to average pooling, feature concatenation is performed to obtain the final feature map of each chromosome cluster; Category information acquisition module: After the final feature map passes through multiple fully connected layers and a normalized exponential function activation function, the category information of the chromosome clusters is obtained; the category information of the chromosome clusters is touching chromosome clusters, overlapping chromosome clusters, and touching-overlapping chromosome clusters.

7. The chromosome image analysis system based on progressive segmentation and focus classification according to claim 6, characterized in that, The chromosome instance segmentation module includes the following modules: Touch segmentation module: Extracts the skeleton map, terminal points and contour map of the touch chromosome clusters to obtain the cutting points of the intersection region; By connecting each pair of cutting points, guide lines are obtained. The guide lines are then expanded to obtain the connected regions. The mean value of all pixels in the connected regions is calculated as a threshold. All pixels in the connected regions that are below the threshold are retained. These pixels are then processed using the least binary method to fit the dividing lines between the touching chromosomes. The dividing lines are then used to segment the clusters of touching chromosomes. Overlap segmentation module: Uses U-Net to extract the overlapping regions of overlapping chromosome clusters; The images on both sides of the overlapping region are stitched together with the overlapping region to separate all individual chromosomes in the overlapping chromosome cluster; Touch overlap segmentation module: First, the overlap segmentation module is used to segment the overlapping chromosome clusters in the touch overlap chromosome clusters, and then the set of several individual chromosomes and touch chromosome clusters in the touch overlap chromosome clusters are obtained. The touch chromosome clusters in the touch overlap chromosome clusters are identified by using the convex hull and the predetermined outer graph. The several individual chromosomes that have been separated are directly output to the single chromosome set, while the touch chromosome clusters will be further segmented by the following module. The touch segmentation module is used to segment touch chromosome clusters within overlapping touch chromosome clusters. Single chromosome set module: After the touch segmentation module, overlap segmentation module and touch overlap segmentation module, the touch chromosome cluster, overlap chromosome cluster and touch overlap chromosome cluster are all segmented to obtain the set of all single chromosome images in the metaphase chromosome image.

Images