Chromosome karyotype analysis method and system based on deep learning
Through the chromosome karyotyping analysis method based on mask rcnn, the module is removed and the loss function is optimized, and the problems of artificial dependence and low accuracy in traditional methods are solved, achieving efficient and accurate chromosome classification.
Patent Information
- Application Number
- CN202311357036.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-19
- Publication Date
- 2025-08-12
- Estimated Expiration
- 2043-10-19
AI Technical Summary
Traditional chromosomal karyotyping methods rely on artificial experience, are time-consuming and laborious, and have low accuracy. The two-stage and single-stage methods based on deep learning have problems such as information loss and error propagation or not being optimized for chromosomal features.
Using deep learning-based chromosome karyotyping analysis method, using mask rcnn as the benchmark network, the chromosome classification effect is enhanced through public feature removal modules and optimization of loss functions, including image preprocessing, feature extraction, suggestion box cutting, classification and mask supervision, and the central loss function and penalty term are used to improve the differences between classes.
It improves the accuracy and efficiency of chromosome classification, reduces artificial dependence, shortens analysis time, and enhances the discriminantity and accuracy of the classifier.
Smart Images
Figure CN117197587B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of chromosome karyotype analysis and relates to a chromosome karyotype analysis method and system based on deep learning. Background Art
[0002] Human chromosomes consist of 23 pairs of chromosomes, 22 of which are autosomes and one pair is the sex chromosome (XY for males and XX for females). Chromosomes can become abnormal for a variety of reasons, and when these abnormalities occur, they can lead to disease. For example, an extra chromosome 21 can cause Down syndrome, with patients typically experiencing intellectual disability, unusual facial features, and heart problems. A partial deletion of chromosome 5 can lead to clonal syndrome.
[0003] Chromosome karyotype analysis is an important screening method for chromosomal abnormalities. Human chromosome karyotype analysis is usually performed by professional clinical cytogeneticists and includes the following steps:
[0004] ① Sample preparation: Collect biological tissue samples from patients, such as peripheral blood or amniotic fluid, culture the cells in the sample until they enter metaphase, and then use some techniques to stain the metaphase cells (such as Giemsa staining) to enhance their visibility;
[0005] ② Metaphase image acquisition: Place the sample under a microscope for observation to obtain the metaphase chromosome image that meets the requirements;
[0006] ③ Karyotype analysis: extract each chromosome in the image, classify it, and then number it according to the classification results and arrange them in order to form a karyotype map;
[0007] Abnormality Detection: Count the chromosomes in the karyotype to determine if there are any abnormalities. Observe the structure of each chromosome in the karyotype to determine if there are any structural abnormalities. Step 3 is the most important step in the entire analysis process, as it determines the reliability of the final chromosome abnormality detection results.
[0008] Traditional chromosome karyotype analysis requires experienced clinical cytogeneticists to observe microscopic images of metaphase cells to identify suitable regions for karyotype analysis. Morphologically, they then observe and record each chromosome's size, long-arm ratio, centromere position, and banding characteristics. These characteristics are then compared with standard chromosome features to determine the chromosome's type. The chromosomes are then segmented, numbered, and arranged sequentially to form a karyotype map. This method is complex and requires a high level of technical experience.
[0009] The traditional manual karyotyping method has the following disadvantages:
[0010] 1. High reliance on experts. Different human chromosomes appear highly similar in appearance and can only be distinguished by a few features, such as the ratio of long and short arms, centromere position, and banding distribution. Only experienced experts can make this distinction, and the learning cost for ordinary people is high and requires years of experience accumulation.
[0011] 2. Time-consuming and labor-intensive. Traditional karyotyping methods involve complex procedures, requiring experts to first identify each chromosome, then observe its size, long-arm ratio, centromere position, and banding distribution. These are then compared one-to-one with standard chromosome markers. Each chromosome is then numbered and arranged into a karyotype map. The entire process is time-consuming and labor-intensive, typically taking about a month from patient specimen collection to the production of a karyotype analysis report.
[0012] 3. Low accuracy. Traditional methods rely on the human eye to distinguish chromosome features, but chromosomes are highly similar in appearance, which makes classification errors very likely.
[0013] Currently, deep learning-based karyotyping methods are primarily divided into two types: two-stage karyotyping methods and single-stage karyotyping methods. Two-stage karyotyping methods first use semantic segmentation or instance segmentation to segment individual chromosomes from metaphase images, and then use image classification methods to classify the segmented chromosomes. For segmentation, the most commonly used models are the UNet, a segmentation network designed specifically for medical image segmentation, or improved versions of UNet. For classification, the most popular approach is to improve convolutional neural networks to achieve higher accuracy, such as incorporating RNNs or LSTMs to capture chromosomal sequence features, or fusing global and local information to capture multi-dimensional information. Single-stage karyotyping methods typically incorporate existing, established object detection or instance segmentation frameworks into the karyotyping task. For example, the YOLO series for object detection can perform chromosome detection and classification, while the Mask RCNN series for instance segmentation can perform chromosome detection, segmentation, and classification.
[0014] However, the two-stage karyotyping method based on deep learning has the following main disadvantages:
[0015] 1. Information loss: During the first stage of semantic segmentation, chromosomes are cut out and then classified separately. This may cause some global information to be lost, affecting the subsequent classification accuracy.
[0016] 2. Error propagation: If there are errors in the segmentation step of the first stage, these errors will propagate to the classification step of the second stage, which may affect the accuracy of the final classification.
[0017] A drawback of single-stage karyotyping methods based on deep learning is that they fail to account for the unique characteristics of chromosomes. Currently, most single-stage karyotyping methods simply apply established object detection or instance segmentation frameworks to the task of chromosome karyotyping without specifically optimizing for chromosome characteristics, resulting in poor clinical effectiveness. Summary of the Invention
[0018] The purpose of the present invention is to provide a chromosome karyotype analysis method and system based on deep learning to complete chromosome karyotype analysis, and to optimize the effect of chromosome classification according to the characteristics of chromosomes, thereby improving overall performance.
[0019] In order to achieve the above objectives, the basic scheme of the present invention is: a chromosome karyotype analysis method based on deep learning, comprising the following steps:
[0020] S1, acquire metaphase images of chromosomes and perform preprocessing;
[0021] S2, uses the backbone network to extract the image feature F;
[0022] S3, using multiple suggestion boxes to trim feature F and obtain features F1~Fn;
[0023] S4, remove the common features of features F1~Fn to obtain modified features f1~fn, use the modified features f1~fn to perform classification to obtain the predicted classification results, and use the loss functions IG Loss and CE Loss to jointly supervise the predicted classification results;
[0024] S5, fine-tune the proposed box and use CE Loss to supervise the fine-tuning results;
[0025] S6, according to the predicted classification results, takes out the mask of the corresponding category, uses binary cross loss BCE Loss to supervise it, and obtains the chromosome karyotype analysis results.
[0026] The working principle and beneficial effects of this basic solution are as follows: Since the differences between different categories of chromosomes are small and there are a large number of similar features, the classifier cannot accurately extract discriminative features. This technical solution removes the common features between different types of chromosomes, leaving only the most discriminative features of each type of chromosome, thereby increasing the inter-class differences between different types of chromosomes and better distinguishing different types of chromosomes. This technical solution sets a loss function that uses the center loss function as a benchmark to reduce the intra-class variance; on the basis of the center loss, it adds a term that penalizes the size of the difference between categories belonging to the same group, which is used to increase the inter-class differences between similar categories, facilitate classification, and improve classification performance.
[0027] Furthermore, the method for preprocessing the metaphase image of the chromosome is to scale the image and randomly flip it.
[0028] Collect the required images and pre-process them for subsequent use.
[0029] Furthermore, the method for obtaining feature F in step S2 is: inputting the preprocessed image into a convolutional neural network, extracting multi-scale features of the image, and using a feature pyramid network to fuse the multi-scale features to obtain feature F.
[0030] Extract multi-scale features of the image and then fuse them to obtain fused features, collect rich feature information of the image, and facilitate classification.
[0031] Furthermore, the method for obtaining features F1~Fn in step S3 is: inputting feature F into the RPN region proposal network to obtain n suggestion boxes, cropping feature F according to the suggestion boxes using a bilinear interpolation method, and uniformly scaling it to a first preset size to obtain features F1~Fn.
[0032] The instance segmentation framework is used to complete chromosome segmentation in the chromosome karyotype analysis task end-to-end, which is beneficial for subsequent targeted classification.
[0033] Furthermore, in step S4, the method for removing the common features of features F1 to Fn is as follows: the features F1 to Fn are passed through the CFR common feature removal module to remove the common features of the chromosomes, thereby increasing the inter-class variance between each category and obtaining the modified features f1 to fn;
[0034] The method of removing the common features of chromosomes through the CFR common feature removal module is as follows:
[0035] The mth channel of features F1~Fn is taken out in sequence and spliced into a 7×7×n feature. Then, the element at each position is averaged in the channel direction to obtain a channel representative feature Fchannel_m of size 7×7×1. The Fchannel_m obtained from each channel is then spliced in the channel direction to obtain a common feature Fcommon of size 7×7×256. Finally, the common feature Fcommon is subtracted from the features F1~Fn in sequence to obtain the corrected features f1~fn.
[0036] This technical solution uses a common feature removal module to remove similar features between chromosomes of different categories, leaving only the most discriminative features, thereby improving the inter-class differences between chromosomes of different categories.
[0037] Furthermore, in step S4, the method for classifying using the modified features f1 to fn to obtain the predicted classification result is as follows: flatten f1 to fn, pass through three fully connected layers in sequence, map the feature dimension to 24, and obtain the predicted classification result.
[0038] Using the modified features f1~fn for classification makes the comparison clearer and facilitates classification.
[0039] Furthermore, in step S5 , the method for fine-tuning the suggestion box is: passing the features F1 to Fn through two fully connected layers in sequence to obtain adjustment parameters corresponding to the features, and then fine-tuning the suggestion box.
[0040] Fine-tune the suggestion box, optimize the suggestion box, and improve the segmentation effect.
[0041] Furthermore, in step S6, the bilinear interpolation method is used again to crop the feature F and scale it to a second preset size. The scaled feature is input into the convolution layer for 5 identical convolution operations without changing the size of the feature. The feature is doubled by deconvolution, and its number of channels is adjusted to the number of categories to obtain the mask of the feature in each category.
[0042] By scaling the characteristic differences of chromosomes, the inter-class differences of chromosomes of different categories are improved.
[0043] Furthermore, the IG Loss loss function L IG The specific calculation formula is as follows:
[0044]
[0045]
[0046] Among them, x i Represents the feature vector of the i-th sample before entering the last fully connected layer, y i Indicates the category label of the i-th sample, cy i Indicates that all category labels are y i The center point of the sample, b represents the size of the mini-batch, G k represents the set of categories belonging to the same group, c i and c j represents the center points of two different categories belonging to the kth group, and λ represents the parameter for balancing the two items;
[0047] The CE Loss loss function is:
[0048]
[0049] Among them, y i Represents the label value, represents the model prediction value;
[0050] The binary cross loss function BCE Loss is:
[0051]
[0052] Among them, y represents the label value, Represents the model prediction value.
[0053] The loss function of the present invention uses the center loss function as a benchmark to reduce the intra-class variance; on the basis of the center loss, an item that penalizes the difference between categories belonging to the same group is added to increase the inter-class difference between similar categories.
[0054] The present invention also provides a chromosome karyotype analysis system based on deep learning, including a data acquisition module and a processing module. The data acquisition module is used to obtain the metaphase image of the chromosome. The output end of the data acquisition module is connected to the processing module. The processing module executes the method of the present invention to perform chromosome karyotype analysis.
[0055] This system is used to solve the problems of chromosome detection, segmentation and classification in chromosome karyotype analysis, and to realize chromosome karyotype analysis. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] Figure 1 1 is a flowchart of the chromosome karyotype analysis method based on deep learning of the present invention;
[0057] Figure 2 It is a structural schematic diagram of the CFR common feature removal module of the chromosome karyotype analysis method based on deep learning of the present invention. DETAILED DESCRIPTION
[0058] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.
[0059] In the description of the present invention, it should be understood that the terms "longitudinal", "transverse", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention.
[0060] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal communication between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.
[0061] Chromosome classification in existing technologies currently faces two main technical problems:
[0062] ① Small inter-class differences, that is, the differences between chromosomes of different classes are minimal. The Denver System of International Cytogenetic Nomenclature for Humans stipulates that human chromosomes can be divided into seven groups based on size and centromere position, represented by ABCDEFG. This rule explains that the chromosomes within each group are very similar in appearance. For example, chromosomes 1 and 2, both belonging to group A, are similar in size and centromere position, differing only in the distribution of banding on the chromosomes. This problem can easily lead to misclassification of chromosomes within the same group.
[0063] ② Large intra-class variability, meaning that chromosomes within the same class vary significantly. This variability is primarily due to subtle differences in the cell division cycle of cells in metaphase images from different batches, resulting in varying degrees of chromosome shortening. This, coupled with differences in film preparation by technicians and in the lighting used during image capture, results in significant variability between chromosomes within the same class from different batches, making it difficult for the classifier to accurately extract the characteristics of each chromosome class.
[0064] This paper discloses a deep learning-based chromosome karyotyping method. Using Mask R-CNN as a baseline network, the classification branch is improved to enhance chromosome classification. Deep learning techniques and the instance segmentation framework Mask R-CNN (Mask R-CNN is an instance segmentation algorithm) are used to perform chromosome detection, segmentation, and classification in an end-to-end karyotyping task. The chromosome classification effect of Mask R-CNN is optimized based on the characteristics of the chromosomes, improving overall performance.
[0065] like Figure 1 As shown, the chromosome karyotype analysis method includes the following steps:
[0066] S1, obtaining a metaphase image of a chromosome and performing preprocessing; preferably, the method for preprocessing the metaphase image of a chromosome is: scaling the image to a size of 448×448×3 and performing random flipping.
[0067] S2, using the backbone network to extract features F of the image; preferably, the method for obtaining features F in step S2 is: inputting the preprocessed image into a convolutional neural network (such as a resnet50 network, which is a convolutional neural network with a depth of 50 layers), extracting multi-scale features of the image, and using a featurized image pyramid (FPN) network to fuse the multi-scale features to obtain features F. Features F include multi-scale, semantic, and positional information extracted from the input image during training, which is obtained by training and optimizing the ResNet and FPN networks for use in subsequent instance segmentation tasks.
[0068] S3, using multiple suggestion boxes to crop feature F to obtain features F1~Fn; preferably, the method for obtaining features F1~Fn in step S3 is: input feature F into the RPN (Region Proposal Network) region proposal network to obtain n suggestion boxes, and according to the suggestion boxes, use the bilinear interpolation method (i.e., ROI Align method) to crop feature F, and uniformly scale it to a first preset size (the first preset size is 7×7×256) to obtain features F1~Fn.
[0069] S4, remove the common features of features F1-Fn to obtain modified features f1-fn, use the modified features f1-fn for classification to obtain a predicted classification result, and use the loss functions IG Loss and CE Loss to jointly supervise the predicted classification result, reducing the intra-class variance of each category while increasing the inter-class variance between similar categories; preferably, the method for using the modified features f1-fn for classification in step S4 to obtain the predicted classification result is: flatten f1-fn (flattening is generally used between convolutional layers and fully connected layers (FC). After feature extraction such as convolutional neural networks, they will pass through several FC layers. The input object of the FC layer can only be one-dimensional. The specific method is to splice each row of the two-dimensional matrix in sequence), pass through three fully connected layers in sequence, map the feature dimension to 24 (chromosome classification is a 24-classification problem (22+X+Y), generally n categories are finally mapped to n dimensions, and the value of each position represents the probability that the target belongs to this category), and obtain the predicted classification result. CE Loss is combined with IG Loss for classification, and this loss improves the classification effect.
[0070] In step S5, in order to make the final proposed box closer to the true box, it is necessary to predict a set of proposed box adjustment parameters, fine-tune the proposed box, and use CE Loss to supervise the fine-tuning results; preferably, the method for fine-tuning the proposed box in step S5 is: pass the features F1~Fn through two fully connected layers in sequence, obtain the adjustment parameters corresponding to the features, and fine-tune the proposed box.
[0071] In step S6, based on the predicted classification result, the corresponding class mask is extracted and supervised using binary cross-validation loss (BCE Loss) to ensure that the predicted mask and label are closer together, resulting in the chromosome karyotype analysis result. The network outputs three results: a rectangular box representing the detected chromosome, the chromosome class within the box, and a segmentation mask representing the specific shape of the chromosome within the box. These results can be used to isolate the detected chromosome for karyotype analysis.
[0072] In a preferred embodiment of the present invention, Figure 2 As shown, the method of removing the common features of features F1 to Fn in step S4 is: passing features F1 to Fn through the CFR (Common Feature Removing Module) to remove the common features of chromosomes, increase the inter-class variance between each category, and obtain modified features f1 to fn;
[0073] The method of removing the common features of chromosomes through the CFR common feature removal module is as follows:
[0074] The mth channel of features F1~Fn is taken out in sequence and spliced into a 7×7×n feature. Then, the element at each position is averaged in the channel direction to obtain a channel representative feature Fchannel_m of size 7×7×1. The Fchannel_m obtained from each channel is then spliced in the channel direction to obtain a common feature Fcommon of size 7×7×256. Finally, the common feature Fcommon is subtracted from the features F1~Fn in sequence to obtain the corrected features f1~fn.
[0075] Because chromosomes of different categories differ little and share many similar features, classifiers cannot accurately extract discriminative features. Removing the common features between chromosomes, leaving only the most discriminative features, can reduce intra-class differences within chromosomes of the same category while increasing inter-class differences. This allows the classifier to better distinguish between chromosomes of different categories.
[0076] In a preferred embodiment of the present invention, mask prediction: in step S6, the bilinear interpolation method is used again to crop the feature F and scale it to a second preset size (the second preset size is 14×14×256). The scaled feature is input into the convolution layer for 5 identical convolution operations without changing the size of the feature. The feature is doubled by deconvolution, and its number of channels is adjusted to the number of categories to obtain the mask of the feature in each category.
[0077] In a preferred embodiment of the present invention, the most commonly used classification loss function is the cross entropy loss function CELoss. Due to the small inter-class differences of chromosomes, this loss function cannot accurately distinguish the characteristics of each category. This technical solution designs a new loss function, which uses the center loss function as a benchmark to reduce the intra-class variance; on the basis of the center loss, a term is added to penalize the difference between categories belonging to the same group, which is used to increase the inter-class difference between similar categories. IG Loss function L IG The specific calculation formula is as follows:
[0078]
[0079]
[0080] By using the Center Loss function L c Narrowing the gap between sample features of the same category and increasing the distance between category centers of similar categories increases the inter-class variance and reduces the intra-class variance. i Represents the feature vector of the i-th sample before entering the last fully connected layer, y i Indicates the category label of the i-th sample, cy i Indicates that all category labels are y i The center point of the sample, b represents the size of the mini-batch, G k represents the set of categories belonging to the same group, c i and c j represents the center points of two different categories belonging to the kth group, λ represents the parameter for balancing the two items, and its value range is 0 to 1;
[0081] The CE Loss loss function is:
[0082]
[0083] Among them, y i Represents the label value, represents the model prediction value;
[0084] The binary cross loss function BCE Loss is:
[0085]
[0086] Among them, y represents the label value, Represents the model prediction value.
[0087] For example, chromosomes 1 and 2 look very similar and belong to group A. In this case, when calculating the loss, the difference between features 1 and 2 should be maximized. The specific approach is to push the feature centers of features 1 and 2 apart to enhance the classifier's ability to distinguish between chromosomes 1 and 2. Chromosomes 1 and 8 do not look similar and do not belong to the same group, so this item will not be calculated.
[0088] The present invention also provides a chromosome karyotype analysis system based on deep learning, including a data acquisition module and a processing module. The data acquisition module is used to obtain the metaphase image of the chromosome. The output end of the data acquisition module is connected to the processing module. The processing module executes the method of the present invention to perform chromosome karyotype analysis.
[0089] This technical solution uses instance segmentation to solve the chromosome detection, segmentation, and classification problems in karyotype analysis end-to-end. Based on the challenges faced by chromosome classification in Mask RCNN, the corresponding modules are optimized: the classification branch of Mask RCNN is optimized, using a common feature removal module to increase inter-class variability; and the classification loss function of Mask RCNN is optimized to increase inter-class variability of similar classes while reducing intra-class variability within the same class.
[0090] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
[0091] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.
Claims
1. A chromosome karyotype analysis method based on deep learning, characterized in that: The steps include: S1, acquire metaphase images of chromosomes and perform preprocessing; S2, uses the backbone network to extract the image feature F; S3, using multiple suggestion boxes to trim feature F and obtain features F1~Fn; S4, remove the common features of features F1~Fn to obtain modified features f1~fn, use the modified features f1~fn to perform classification to obtain the predicted classification results, and use the loss functions IG Loss and CE Loss to jointly supervise the predicted classification results; IG Loss loss function L IG The specific calculation formula is as follows: Among them, λ represents the parameter of balancing the two terms; x i Represents the feature vector of the i-th sample before entering the last fully connected layer, y i represents the category label of the i-th sample, c yi Indicates that all category labels are y i The center point of the sample, b represents the size of the mini-batch, G k represents the set of categories belonging to the same group, c i and c j Represents the center points of two different categories belonging to the kth group; S5, fine-tune the proposed box and use CE Loss to supervise the fine-tuning results; S6, according to the predicted classification results, takes out the mask of the corresponding category, uses binary cross loss BCE Loss to supervise it, and obtains the chromosome karyotype analysis results.
2. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that The method for preprocessing the metaphase image of chromosomes is to scale the image and randomly flip it.
3. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: The method for obtaining feature F in step S2 is: inputting the preprocessed image into a convolutional neural network, extracting multi-scale features of the image, and using a feature pyramid network to fuse the multi-scale features to obtain feature F.
4. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: The method for obtaining features F1~Fn in step S3 is: input feature F into the RPN region proposal network to obtain n suggestion boxes, and according to the suggestion boxes, use the bilinear interpolation method to crop feature F, and uniformly scale it to a first preset size to obtain features F1~Fn.
5. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: Step S4 is a method for removing the common features of features F1 to Fn: passing features F1 to Fn through a CFR common feature removal module to remove the common features of chromosomes, thereby increasing the inter-class variance between each category and obtaining modified features f1 to fn; The method of removing the common features of chromosomes through the CFR common feature removal module is as follows: The mth channel of features F1~Fn is taken out in sequence and spliced into a 7×7×n feature. Then, the element at each position is averaged in the channel direction to obtain a channel representative feature Fchannel_m of size 7×7×1. The Fchannel_m obtained from each channel is then spliced in the channel direction to obtain a common feature Fcommon of size 7×7×256. Finally, the common feature Fcommon is subtracted from the features F1~Fn in sequence to obtain the corrected features f1~fn.
6. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: In step S4, the method for classifying using the modified features f1 to fn to obtain the predicted classification result is as follows: flatten f1 to fn, pass through three fully connected layers in sequence, map the feature dimension to 24, and obtain the predicted classification result.
7. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: In step S5 , the method for fine-tuning the suggestion box is as follows: passing the features F1 to Fn through two fully connected layers in sequence to obtain the adjustment parameters corresponding to the features, and then fine-tuning the suggestion box.
8. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: In step S6, the bilinear interpolation method is used again to crop the feature F and scale it to a second preset size. The scaled feature is input into the convolution layer for 5 identical convolution operations without changing the size of the feature. The feature is doubled by deconvolution, and its number of channels is adjusted to the number of categories to obtain the mask of the feature in each category.
9. The chromosome karyotype analysis method based on deep learning according to claim 1, characterized in that: CE Loss function L CE for: Among them, y i Represents the tag value, represents the model prediction value; The binary cross loss function BCE Loss is: Among them, y represents the label value, Represents the model prediction value.
10. A chromosome karyotype analysis system based on deep learning, characterized in that: The method comprises a data acquisition module and a processing module, wherein the data acquisition module is used to obtain a metaphase image of a chromosome, an output end of the data acquisition module is connected to the processing module, and the processing module executes the method according to any one of claims 1 to 9 to perform chromosome karyotype analysis.