A chromosome image segmentation method

Abstract

The present application provides a chromosome image segmentation method, relates to the technical field of chromosome image processing, first acquires a first image dataset composed of a first metaphase image of a first cell, trains a rotation target detection model constructed by using a training set in the first image dataset, then inputs a collected second metaphase image of a second cell into the trained rotation target detection model for detection, accurately locates and closely covers each chromosome of the second metaphase image of the second cell, further constructs a semantic segmentation model for segmenting a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes, trains the semantic segmentation model by using a training set of the acquired second image dataset, finally jointly trains the trained rotation target detection model and the semantic segmentation model to predict a metaphase image of a cell to be detected, so that the segmentation and extraction of multiple chromosomes are completed, and the segmentation accuracy of chromosomes is effectively improved.

Description

Technical Field

[0001] This invention relates to the technical field of chromosome image processing, and in particular to a chromosome image segmentation method. Background Technology

[0002] In clinical diagnosis, karyotype analysis refers to the process of pairing, numbering, and grouping chromosomes of the cells to be tested according to the inherent morphological characteristics and rules of the organism, and then performing morphological analysis. Karyotype analysis based on chromosome G-banding and microscopic imaging is an important means of diagnosing genetic symptoms. Chromosomes are mainly composed of DNA and proteins, and play a role in storing and transmitting genetic information. The non-rigid nature of chromosomes means that different chromosome instances in cell images are extremely prone to bending, overlapping, and sticking together, forming various structures. This makes the accurate segmentation and extraction of intact chromosomes one of the most complex steps, causing great difficulties for clinicians. Therefore, the invention of an efficient automated chromosome segmentation and analysis method is urgently needed.

[0003] Currently, chromosome segmentation is typically performed manually by clinical professionals to accurately extract each chromosome from disordered raw photomicrographs. The quality of chromosome segmentation depends heavily on the experience and expertise of the professionals, requiring a high level of skill. Furthermore, the segmentation process is susceptible to subjective factors and fatigue, leading to errors. Existing technology discloses a chromosome segmentation method that first uses a primary chromosome segmentation model to extract a mask image containing overlapping chromosomes using directional bounding boxes. Then, a trained secondary overlapping chromosome segmentation model is used to perform secondary segmentation on the overlapping chromosomes in the high-power region image containing overlapping chromosomes, resulting in a mask image containing only individual chromosomes from the overlapping chromosomes. However, chromosomes in metaphase typically appear as flexible strips in different directions. Using directional bounding boxes to extract a mask image containing overlapping chromosomes cannot accurately locate and tightly cover each chromosome in the metaphase cell image. Moreover, the extracted mask image may contain non-target chromosomes or incomplete chromosomes, making it difficult to accurately segment individual chromosomes from images containing non-target chromosomes or incomplete chromosomes, thus reducing the accuracy of chromosome segmentation. Summary of the Invention

[0004] To address the issue of low chromosome segmentation accuracy in current chromosome segmentation methods, this invention proposes a chromosome image segmentation method that can accurately locate and identify each chromosome in a chromosome image, capture chromosome information in any direction within the image, and effectively improve chromosome segmentation accuracy.

[0005] To achieve the above-mentioned technical effects, the technical solution of the present invention is as follows:

[0006] A chromosome image segmentation method includes the following steps:

[0007] S1. Obtain a certain number of first-cell mid-phase images to form the first image dataset;

[0008] S2. Preprocess the first cell mid-phase image in the first image dataset;

[0009] S3. Construct a rotation target detection model for capturing chromosomes in arbitrary orientations in first metaphase images;

[0010] S4. Divide the first image dataset into a training set, a validation set, and a test set. Use the training set to train the constructed rotating object detection model, then use the validation set to evaluate the rotating object detection model during the training process, and use the test set to test the effectiveness of the rotating object detection model to obtain the trained rotating object detection model.

[0011] S5. Acquire images of the second metaphase of the cell, input the images of the second metaphase of the cell into the trained rotating target detection model, locate and cover each chromosome in the images of the second metaphase of the cell, output semantic segmentation images, and form the second image dataset;

[0012] S6. Preprocess the semantic segmentation images in the second image dataset;

[0013] S7. Construct a semantic segmentation model for segmenting a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes;

[0014] S8. Divide the second image dataset into a training set, a validation set, and a test set. Use the training set to train the constructed semantic segmentation model, then use the validation set to evaluate the semantic segmentation model during the training process, and use the test set to test the effectiveness of the semantic segmentation model to obtain the trained semantic segmentation model.

[0015] S9. Acquire the metaphase image of the cell to be detected, input the metaphase image of the cell to be detected into the trained rotating target detection model, output the fourth metaphase image of the cell, input the fourth metaphase image of the cell into the semantic segmentation model, and output the segmentation result of a single chromosome in the metaphase image of the cell to be detected.

[0016] In this technical solution, a first image dataset consisting of first metaphase images is first acquired. A rotational target detection model is trained using the training set in the first image dataset. Then, the acquired second metaphase images are input into the trained rotational target detection model for detection. Each chromosome in the second metaphase image is then precisely located and closely covered, capturing chromosomes in any direction in the first metaphase image. At the same time, it can effectively locate and identify each chromosome in a cluster containing two or more overlapping chromosomes. Furthermore, a semantic segmentation model is constructed to segment a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes. The semantic segmentation model is trained using the training set of the acquired second image dataset. Finally, the trained rotational target detection model and semantic segmentation model are combined to predict the metaphase image to be detected, thereby completing the segmentation and extraction of multiple chromosomes and effectively improving the chromosome segmentation accuracy.

[0017] Preferably, in step S2, the specific process of preprocessing the first mid-cell image in the first image dataset is as follows:

[0018] S21. Manually label the first metaphase image in the first image dataset using directed bounding boxes, and tightly cover each chromosome in the first metaphase image to obtain the label image of the first metaphase image;

[0019] S22. Let the initial angle θ of the label image of the first mid-cell image in S21 be θ. Starting from the initial angle θ, rotate the label image in the first mid-cell image several times in a counterclockwise or clockwise direction with ω as the rotation angle change, until the angle of the label image of the first mid-cell image after rotation is 180°.

[0020] Preferably, in step S3, the rotating target detection model adopts the Oriented RepPoints network, which uses an adaptive point learning method to detect rotating target images.

[0021] Preferably, during the training process of the rotating target detection model described in S4, the network parameters and weights of the rotating target detection model are updated using stochastic gradient descent.

[0022] Preferably, the content of the image within the directed bounding box is predicted and saved using a trained rotating target detection model.

[0023] Preferably, in step S6, the specific steps for preprocessing the semantic segmentation images in the second image dataset are as follows:

[0024] S61. Sample the semantic segmentation images to the same size;

[0025] S62. Use LabelMe software to manually annotate semantic segmentation images of uniform size;

[0026] S63. Unify the format of the semantic segmentation images annotated in S62 to JSON format;

[0027] S64. Convert the JSON format image into the corresponding mask image.

[0028] Preferably, in step S62, the chromosome tag in the semantic segmentation image is marked as 1, and the background portion of the image other than the chromosome in the semantic segmentation image is marked as 0.

[0029] Preferably, in step S7, the semantic segmentation model adopts a multi-scale NestedUNet network based on a dual attention mechanism, and the multi-scale NestedUNet network based on a dual attention mechanism is an encoder-decoder structure.

[0030] Preferably, a pre-trained ResNet50 network is selected as the backbone network of the encoder.

[0031] Preferably, the multi-scale NestedUNet network includes a first residual block, a second residual block, a third residual block, a first ASPP module, and a second ASPP module. The first residual block, the second residual block, the third residual block, the first ASPP module, and the second ASPP module are connected sequentially. A CBAM module is provided between the first residual block and the second residual block, between the second residual block and the third residual block, between the third residual block and the first ASPP module, and between the first ASPP module and the second ASPP module. An ECA module is connected to the first residual block, the second residual block, the third residual block, the first ASPP module, and the second ASPP module.

[0032] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0033] This invention proposes a chromosome image segmentation method. First, a first image dataset consisting of first metaphase images is acquired. A rotational target detection model is trained using the training set in the first image dataset. Then, the acquired second metaphase images are input into the trained rotational target detection model for detection. Next, each chromosome in the second metaphase image is precisely located and closely overlapped, capturing chromosomes in any direction in the first metaphase image. It can also effectively locate and identify each chromosome in a cluster containing two or more overlapping chromosomes. Furthermore, a semantic segmentation model is constructed to segment a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes. The semantic segmentation model is trained using the training set of the acquired second image dataset. Finally, the trained rotational target detection model and semantic segmentation model are combined to predict the metaphase image to be detected, thereby completing the segmentation and extraction of multiple chromosomes and effectively improving the chromosome segmentation accuracy. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a chromosome image segmentation method proposed in an embodiment of the present invention;

[0035] Figure 2 This represents a first mid-cell image as presented in an embodiment of the present invention;

[0036] Figure 3 A labeled image representing the first mid-cell image proposed in this embodiment of the invention;

[0037] Figure 4 This represents the semantic segmentation image proposed in the embodiments of the present invention;

[0038] Figure 5 This represents the mask image proposed in the embodiments of the present invention;

[0039] Figure 6 This diagram illustrates the structure of the multi-scale NestedUNet network proposed in this embodiment of the invention. Detailed Implementation

[0040] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.

[0041] To better illustrate this embodiment, some parts of the accompanying drawings may be omitted, enlarged, or reduced, and do not represent actual dimensions. The descriptions of directions such as "up" and "down" are not intended to limit this patent.

[0042] It is understandable to those skilled in the art that some well-known details may be omitted from the accompanying drawings;

[0043] The positional relationships depicted in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.

[0044] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0045] Example 1

[0046] like Figure 1 As shown, this embodiment proposes a chromosome image segmentation method, including the following steps:

[0047] S1. Obtain a certain number of first-cell mid-phase images to form the first image dataset;

[0048] In step S1, 4000 images of the first cell mid-phase are acquired.

[0049] S2. Preprocess the first cell mid-phase image in the first image dataset;

[0050] In step S2, the specific process of preprocessing the first mid-cell image in the first image dataset is as follows:

[0051] S21. See also Figure 2 and Figure 3 The first metaphase image of the first image dataset is manually labeled using directed bounding boxes, and each chromosome in the first metaphase image is tightly covered to obtain the label image of the first metaphase image;

[0052] In step S21, using the roLabelImg software, each chromosome in the first metaphase image is tightly covered with a directed bounding box, labeled with 1, and saved as an XML file.

[0053] S22. Let the initial angle θ of the label image of the first mid-cell image in S21 be θ. Starting from the initial angle θ, rotate the label image in the first mid-cell image several times in a counterclockwise or clockwise direction with ω as the rotation angle change, until the angle of the label image of the first mid-cell image after rotation is 180°.

[0054] In step S22, data augmentation is used to rotate the labeled image of the first cell mid-phase image every 30° between 0° and 180°, resulting in a total of 28,000 images, which expands the number of images in the first dataset.

[0055] S3. Construct a rotation target detection model for capturing chromosomes in arbitrary orientations in first metaphase images;

[0056] S4. Divide the first image dataset into a training set, a validation set, and a test set. Use the training set to train the constructed rotating object detection model, then use the validation set to evaluate the rotating object detection model during the training process, and use the test set to test the effectiveness of the rotating object detection model to obtain the trained rotating object detection model.

[0057] In step S4, the augmented first image dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set consists of 19,600 images, the validation set consists of 5,600 images, and the test set consists of 4,900 images.

[0058] S5. Acquire images of the second metaphase of the cell, input the images of the second metaphase of the cell into the trained rotating target detection model, locate and cover each chromosome in the images of the second metaphase of the cell, output semantic segmentation images, and form the second image dataset;

[0059] In step S5, the number of semantic segmentation images output is 1700, and the second image dataset consists of 1700 semantic segmentation images.

[0060] S6. Preprocess the semantic segmentation images in the second image dataset;

[0061] S7. Construct a semantic segmentation model for segmenting a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes;

[0062] S8. Divide the second image dataset into a training set, a validation set, and a test set. Use the training set to train the constructed semantic segmentation model, then use the validation set to evaluate the semantic segmentation model during the training process, and use the test set to test the effectiveness of the semantic segmentation model to obtain the trained semantic segmentation model.

[0063] In step S8, the second image dataset is divided into a training set, a validation set, and a test set in a ratio of 7:2:1. The training set consists of 1190 semantic segmentation images, the validation set consists of 340 semantic segmentation images, and the test set consists of 170 semantic segmentation images.

[0064] S9. Acquire images of the cell metaphase to be detected, input the images of the cell metaphase to be detected into the trained semantic segmentation model, output the fourth cell metaphase image, input the fourth cell metaphase image into the semantic segmentation model, and output the segmentation result of a single chromosome in the cell metaphase image to be detected.

[0065] In this embodiment, a first image dataset consisting of first metaphase images is first acquired. A rotational target detection model is trained using the training set in the first image dataset. Then, the acquired second metaphase images are input into the trained rotational target detection model for detection. Each chromosome in the second metaphase image is then precisely located and closely covered, capturing chromosomes in any direction in the first metaphase image. At the same time, it can effectively locate and identify each chromosome in a cluster containing two or more overlapping chromosomes. Furthermore, a semantic segmentation model is constructed to segment a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes. The semantic segmentation model is trained using the training set of the acquired second image dataset. Finally, the trained rotational target detection model and semantic segmentation model are combined to predict the metaphase image to be detected, thereby completing the segmentation and extraction of multiple chromosomes and effectively improving the chromosome segmentation accuracy.

[0066] Example 2

[0067] See Figure 1 In step S3, the rotating target detection model includes the Oriented RepPoints network, which is built on MMRotate, an open-source toolkit for rotating object detection based on PyTorch. MMRotate features support for multiple angle representations, modular design, and powerful benchmark models and state-of-the-art performance. The Oriented RepPoints network uses an adaptive point learning method to detect rotating target images. This method can capture the geometric information of targets in any direction, which is of great significance for meeting the requirements of karyotype analysis to extract each chromosome. In object detection, horizontal rectangular bounding boxes or axisymmetric bounding boxes are typically used to detect target regions. However, in chromosome segmentation tasks, chromosomes in metaphase of cell division often appear as flexible strips in different directions due to their non-rigid nature. For densely distributed chromosomes in an image, this means that the horizontal rectangular bounding boxes or axisymmetric bounding boxes will not only contain the target chromosomes but also contain a large number of complete or incomplete chromosome instances, greatly complicating the subsequent chromosome instance extraction. Therefore, the directed bounding boxes in the rotation object detection model are used to accurately locate and tightly cover each chromosome in the second metaphase image, capturing chromosomes in any direction in the first metaphase image. At the same time, it can also effectively locate and identify each chromosome in a cluster containing more than two overlapping chromosomes. During the training process of the S4 rotation object detection model, the network parameters and weights of the rotation object detection model are updated using stochastic gradient descent, where the warm-up strategy is set to linear and the learning rate decay strategy is set to step.

[0068] See Figure 4 In step S5, the trained rotating target detection model is used to predict and save the content of the image within the directed bounding box, capture chromosomes in any direction in the first metaphase image of the cell, and effectively locate and identify each chromosome in a cluster of two or more overlapping chromosomes, outputting a semantic segmentation image that is further segmented by the semantic segmentation model.

[0069] Example 3

[0070] See Figure 1 In step S6, the specific steps for preprocessing the semantic segmentation images in the second image dataset are as follows:

[0071] S61. Sample the semantic segmentation images to the same size;

[0072] In step S61, see Figure 4 The semantic segmentation images output by the rotating object detection model are in JPG format. All JPG semantic segmentation images are saved in a uniform size of 256*256.

[0073] S62. Use LabelMe software to manually annotate semantic segmentation images of uniform size;

[0074] In step S62, the chromosome label in the semantic segmentation image is marked as 1, and the background portion of the image other than the chromosome in the semantic segmentation image is marked as 0.

[0075] S63. Unify the format of the semantic segmentation images annotated in S62 to JSON format;

[0076] S64. Convert the JSON format image into the corresponding mask image.

[0077] In step S64, see Figure 5 The data processing method converts the saved JSON file into a mask image as a label file for semantic segmentation data, and maps it one-to-one with the source file.

[0078] In step S7, the semantic segmentation model employs a multi-scale NestedUNet network based on a dual attention mechanism. This multi-scale NestedUNet network has an encoder-decoder structure, composed of UNets of varying depths. It further enhances the combination of shallow and deep features and reduces the number of parameters through deep supervision. UNet networks are known for their height invariance, and their encoder-decoder structure determines the stability and accuracy of their segmentation results, especially in medical image segmentation. Because medical images have blurred boundaries and complex gradients, requiring a large amount of high-resolution information, and because medical data is difficult to acquire and highly specialized, the target region often needs to be accurately determined in conjunction with its surrounding environment. Therefore, both high-level semantic information and low-level features are particularly important. The encoder-decoder structure includes a skip connection substructure, which combines high-level semantic information and low-level features to ultimately obtain an accurate segmentation result. A pre-trained ResNet50 network was selected as the backbone network of the encoder. The ResNet50 network contains 49 convolutional layers and one fully connected layer. The ResNet50 network structure is divided into seven parts. The first part does not contain residual blocks and mainly performs convolution, regularization, activation function, and max pooling calculations on the input. The second, third, fourth, and fifth parts all contain residual blocks. In the ResNet50 network structure, each residual block has three convolutional layers. The network has a total of 49 convolutional layers, plus the final fully connected layer, for a total of 50 layers.

[0079] See Figure 6 The multi-scale NestedUNet network has a first residual block X. 0-0 Second residual block X 1-0 The third residual block X 2-0 First ASPP module X 3-0 Second ASPP module X 4-0 The first residual block X 0-0 Second residual block X 1-0 The third residual block X 2-0 The first ASPP module X 3-0 Second ASPP module X 4-0 Connect sequentially, the first residual block X 0-0 With the second residual block X 1-0 Between, the second residual block X 1-0 With the third residual block X 2-0 Between, the third residual block X 2-0 With the first ASPPX 3-0 Between modules, the first ASPP module X 3-0 With the second ASPP module X4-0 Each of them has a CBAM module, the first residual block X 0-0 Second residual block X 1-0 The third residual block X 2-0 First ASPP module X 3-0 Second ASPP module X 4-0 Each module is connected to an ECA module. The Convolutional Attention Module (CBAM) is a simple yet effective attention module for feedforward convolutional neural networks. Given an intermediate feature map as input, the CBAM module infers attention maps sequentially along both channel and spatial dimensions, then multiplies the attention maps with the input feature map for adaptive feature optimization. The effective channel attention module (ECA) adds only a few parameters but achieves significant performance gains. It employs a non-dimensionality-reduction local cross-channel interaction strategy, which can be effectively implemented using one-dimensional convolution. Furthermore, an adaptive method for selecting the one-dimensional convolution kernel size is used to determine the coverage of the local cross-channel interaction.

[0080] In this embodiment, a pre-trained ResNet50 network is first selected as the encoder backbone, which helps the semantic segmentation model converge quickly. Then, the CBAM module is set in the first residual block X. 0-0 Second residual block X 1-0 The third residual block X 2 -0 First ASPP module X 3-0 Second ASPP module X 4-0 Subsequently, the CBAM module greatly enriches the high-level chromosome semantic information through spatial and channel dimensions while also effectively preserving the detailed features of the chromosome outline; the first residual block X 0-0 Second residual block X 1-0 The third residual block X 2-0 First ASPP module X 3-0 Second ASPP module X 4-0 Each connection includes an ECA module. The purpose of introducing an ECA module at the very beginning of the skip connection is to avoid dimensionality reduction and effectively capture information from cross-channel interactions, thus improving segmentation accuracy while reducing the number of parameters. Finally, the first ASPP module X is introduced in the last two layers. 3-0 Second ASPP module X 4-0 The goal is to replace the original residual blocks set in the last two layers, thereby expanding the receptive field through parallel sampling of dilated convolutions of different sizes and effectively capturing global contextual information. This avoids the problem of segmentation errors caused by getting stuck in local chromosome features, and ultimately obtains more refined segmentation results.

[0081] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A chromosome image segmentation method, characterized in that, Includes the following steps: S1. Obtain a certain number of first-cell mid-phase images to form the first image dataset; S2. Preprocess the first cell mid-phase image in the first image dataset; S3. Construct a rotation target detection model for capturing chromosomes in arbitrary directions in the preprocessed first metaphase image; S4. Divide the first image dataset into a training set, a validation set, and a test set. Use the training set to train the constructed rotating object detection model, then use the validation set to evaluate the rotating object detection model during the training process, and use the test set to test the effectiveness of the rotating object detection model to obtain the trained rotating object detection model. S5. Acquire images of the second metaphase of the cell, input the images of the second metaphase of the cell into the trained rotating target detection model, locate and cover each chromosome in the images of the second metaphase of the cell, output semantic segmentation images, and form the second image dataset; S6. Preprocess the semantic segmentation images in the second image dataset; S7. Construct a semantic segmentation model for segmenting a single chromosome from a semantic segmentation image containing multiple overlapping chromosomes; S8. Divide the preprocessed second image dataset into a training set, a validation set, and a test set. Use the training set to train the constructed semantic segmentation model, then use the validation set to evaluate the semantic segmentation model during the training process, and use the test set to test the effectiveness of the semantic segmentation model to obtain the trained semantic segmentation model. S9. Acquire the metaphase image of the cell to be detected, input the metaphase image of the cell to be detected into the trained rotating target detection model, output the fourth metaphase image of the cell, input the fourth metaphase image of the cell into the semantic segmentation model, and output the segmentation result of a single chromosome in the metaphase image of the cell to be detected.

2. The chromosome image segmentation method according to claim 1, characterized in that, In step S2, the specific process of preprocessing the first mid-cell images in the first image dataset is as follows: S21. Manually label the first metaphase image in the first image dataset using directed bounding boxes, and tightly cover each chromosome in the first metaphase image to obtain the label image of the first metaphase image; S22. Let the initial angle of the label image of the first cell mid-phase image in S21 be... The angle of the label image in the first mid-cell image is changed from the initial angle. Begin by moving counter-clockwise or clockwise. The rotation angle is changed by performing several rotations until the angle of the label image of the first mid-cell image after rotation is 180°.

3. The chromosome image segmentation method according to claim 2, characterized in that, In step S3, the rotating target detection model uses an Oriented RepPoints network, which employs an adaptive point learning method to detect rotating target images.

4. The chromosome image segmentation method according to claim 3, characterized in that, During the training of the rotating target detection model described in S4, the network parameters and weights of the rotating target detection model are updated using stochastic gradient descent.

5. The chromosome image segmentation method according to claim 4, characterized in that, The trained rotating object detection model is used to predict and save the content of the image within the directed bounding box.

6. The chromosome image segmentation method according to claim 1, characterized in that, In step S6, the specific steps for preprocessing the semantic segmentation images in the second image dataset are as follows: S61. Sample the semantic segmentation images to the same size; S62. Use LabelMe software to manually annotate semantic segmentation images of uniform size; S63. Unify the format of the semantic segmentation images annotated in S62 to JSON format; S64. Convert the JSON format image into the corresponding mask image.

7. The chromosome image segmentation method according to claim 6, characterized in that, In step S62, the chromosome label in the semantic segmentation image is marked as 1, and the background portion of the image other than the chromosome in the semantic segmentation image is marked as 0.

8. The chromosome image segmentation method according to claim 7, characterized in that, In step S7, the semantic segmentation model adopts a multi-scale NestedUNet network based on a dual attention mechanism, which is an encoder-decoder structure.

9. The chromosome image segmentation method according to claim 8, characterized in that, A pre-trained ResNet50 network was selected as the backbone network of the encoder.

10. The chromosome image segmentation method according to claim 8, characterized in that, The multi-scale NestedUNet network includes a first residual block, a second residual block, a third residual block, a first ASPP module, and a second ASPP module. The first residual block, the second residual block, the third residual block, the first ASPP module, and the second ASPP module are connected sequentially. A CBAM module is provided between the first residual block and the second residual block, between the second residual block and the third residual block, between the third residual block and the first ASPP module, and between the first ASPP module and the second ASPP module. An ECA module is connected to the first residual block, the second residual block, the third residual block, the first ASPP module, and the second ASPP module.

Images