TWI856681B - Deep learning-based automatic detection and identification system and method for chromosomes in karyotype diagrams - Google Patents

Abstract

The present invention provides a method and system for automatic detection and identification of chromosomes in a karyotype image based on deep learning, which can establish a chromosome automatic detection and identification model through machine learning, and then directly extract and classify the features of the chromosomes in the original chromosome image, so as to effectively reduce the time spent by clinical personnel on image judgment.

Description

Deep learning-based automatic detection and identification system and method for chromosomes in karyotype diagrams

The present invention relates to a genetic molecular identification system, and in particular to a karyotype chromosome automatic detection and identification system and method based on deep learning.

Chromosomes are genetic aggregates in cells and are the main carriers of genetic information (genes). Normal humans have 46 chromosomes, including 22 pairs of chromosomes and 1 pair of sex chromosomes. Chromosomal abnormalities can cause chromosomal diseases, which include abnormalities in the number of chromosomes or abnormalities in their structure. Chromosomal diseases can lead to mental retardation, developmental malformations, abnormal sex differentiation, cancer, infertility, etc.

According to statistics, about 2% of prenatal fetuses will have chromosomal abnormalities. Therefore, chromosome karyotype analysis has become an important item in prenatal screening. Currently, clinical chromosome karyotype analysis requires doctors or medical examiners to personally interpret chromosome images, and a subject will provide at least 4 chromosome images, which makes chromosome karyotype analysis a burden on medical manpower. Although automatic chromosome analysis systems have been developed to replace manual analysis, the existing automatic chromosome analysis systems are based on image editing software. Not only is the accuracy low, but the analysis process still requires the participation of medical examiners, so it is still impossible to reduce the time cost and labor cost of testing.

The main purpose of the present invention is to provide a method and system for automatic detection and identification of chromosomes in karyotype diagrams based on deep learning, which can generate a chromosome automatic detection and identification model through a machine learning module, thereby automatically detecting and identifying each pair of chromosomes in the chromosome diagram, so as to reduce the detection cost of chromosome karyotype analysis.

Therefore, in order to achieve the above-mentioned purpose, the present invention provides a method for automatic detection and identification of karyotype chromosomes based on deep learning, which is executed by a processor using computer-readable instructions stored in a database and mainly includes the following steps:

Step A: Select a chromosome training atlas containing multiple feature maps corresponding to each pair of chromosomes.

Step B: Output the chromosome training atlas to a machine learning model to perform the first stage of object detection, that is, detect and select each pair of chromosomes in the chromosome training atlas, and analyze the selection results of each pair of chromosomes to generate a first prediction frame to identify each pair of chromosomes.

Step C: Output the first prediction frames of each pair of chromosomes to a second object detection model for analysis. The second object detection model detects each pair of chromosomes in the chromosome training atlas and adjusts the first prediction frame to generate a second prediction frame for identifying each pair of chromosomes.

Step D: Using the second prediction frames corresponding to each pair of chromosomes and their corresponding chromosome numbers as parameters, a chromosome automatic detection and recognition model is generated to automatically identify and classify at least one pair of chromosomes in a chromosome pattern to be tested.

Furthermore, since the chromosome automatic detection and recognition model can be used to quickly and accurately identify each pair of chromosomes in a chromosome pattern to be tested and classify them, the karyotype chromosome automatic detection and recognition method based on deep learning disclosed in the present invention further includes a step E, wherein the step E is to output a chromosome pattern to be tested to the chromosome automatic detection and recognition model, and the chromosome automatic detection and recognition model detects and selects the chromosomes in the chromosome pattern to be tested, and identifies the chromosome numbers of the selected chromosomes to generate an identification result.

In the embodiment of the present invention, the feature images in step A are captured from a chromosome original image. Specifically, a chromosome original image is output to an image marking software, and the image marking software marks each pair of chromosomes in the chromosome original image, and after cutting and classifying the chromosome serial numbers according to the marks, the feature images can be obtained.

In order to improve the recognition performance of the chromosome automatic detection and recognition model disclosed in the present invention in real time, one embodiment of the present invention further includes a step F, which means that when a feedback training model detects that a pair of chromosomes has more than two second prediction frames, and the overlap rate of these second prediction frames is higher than a predetermined IOU threshold, the chromosome pattern to be tested is input into the chromosome training atlas, and steps B to D are executed again.

In the embodiment of the present invention, in order to obtain a better recognition effect, before the chromosome training atlas is output to the first object detection model, at least a portion of the feature maps is subjected to an image pre-processing procedure, such as adjusting the size, brightness or direction, etc.

In order to increase the convenience and applicability of clinical use, in step E, the recognition result is input into an image post-processing model, and the image post-processing model performs post-processing such as background removal, sorting, and position correction on the chromosome image to be tested.

In another embodiment of the present invention, a deep learning-based karyotype chromosome automatic detection and identification system is disclosed, which mainly includes a database and a processor, wherein the database at least stores a plurality of chromosome feature maps; the processor transmits data with the database, and generates a chromosome automatic detection and identification model through a machine learning module, and then identifies and classifies each pair of chromosomes in a chromosome pattern to be tested through the automatic identification module to generate an identification result.

The machine learning module has a machine learning model, and in one embodiment of the present invention, the machine learning model includes a two-stage object detection process.

In another embodiment of the present invention, the machine learning module further includes a feedback training model for judging whether two prediction frames selected for the same pair of chromosomes should be retained through a predetermined IOU threshold, and feeding back the chromosome pattern and its judgment result to the machine learning module to update the chromosome automatic detection and recognition model in real time, thereby achieving the effect of improving the recognition performance of the chromosome automatic detection and recognition model.

In the next embodiment of the present invention, the processor further includes an image pre-processing model, which receives the chromosome feature map from the database and performs an image pre-processing procedure, such as rotation, mirroring, cropping, brightness adjustment, etc.

In another embodiment of the present invention, the processor further includes an image post-processing model, which receives the chromosome pattern to be detected and identified by the chromosome automatic detection and identification model, and performs image post-processing on the identification result according to an image processing instruction.

The present invention discloses a method and system for automatic detection and identification of chromosomes in karyotype images based on deep learning. It is capable of establishing a chromosome automatic detection and identification model through machine learning. Through the chromosome automatic detection and identification model, the features of the chromosomes in the original chromosome image can be directly extracted and classified. This can not only effectively reduce the time spent by clinical personnel on image judgment, but also improve the recognition accuracy of difficult images.

The term "data" can also be understood as "information" or "information", which refers to symbols, files, instructions, text, numbers, etc. that can be identified or processed.

The term "machine learning" refers to using a machine learning model to learn and improve in data, find patterns and relationships, and make decisions and predictions based on its learning and analysis results.

The term "chromosome image" or "chromosome original map" refers to the karyotype diagram obtained from the chromosome specimen. In principle, each karyotype diagram has 22 pairs of somatic chromosomes plus 1 pair of sex chromosomes. Clinically, the chromosomes are roughly divided into 9 groups based on their size: A, B, C, D, E, F, G, X and Y. The large chromosomes in group A are A1, A2, A3, B4 and B5; the smallest groups are G, X and Y.

The chromosome pattern determined as normal in the present invention means that the structure and characteristics of each pair of chromosomes in the chromosome image are normal; the chromosome pattern determined as abnormal in the present invention means that the structure and characteristics of each pair of chromosomes in the chromosome image are pathological, hyperplastic, broken, folded, or cannot be interpreted, such as hyperplasia of the 5th or 21st pair of chromosomes, too many features in the chromosome pattern are blocked by impurities, or heterochromosomal groups cannot be completely excluded during the observation process.

The term "feature map" refers to the result obtained by extracting the features of each pair of chromosomes from a chromosome original map through a graphic feature processing module.

The term "machine learning" refers to the process whereby a processor uses an algorithm to identify patterns in an input data set and build a data model based on the patterns.

The term "module" refers to a specific functional component composed of one or more basic functional components. Generally speaking, a module includes a module input interface, a model, and its code description. After inputting data or information into the model through the module input interface, the model is executed to obtain a result.

The term "model" can also be understood as code, which is a computer-readable instruction that is executed on a computer or processor.

The term "sequence number" or "chromosome sequence number" refers to the number of the chromosome, for example, the first pair of chromosomes is number 1, and so on.

The term "IOU (intersection over union)" refers to the intersection of the predicted object area and the real object area. In the present invention, IOU is calculated by the following formula: IOU = the area of the overlap of two frames / the sum of the areas of two frames.

The term "correction" or "position correction" refers to the ability to identify the centromere of a chromosome and adjust the position of the short arm of the chromosome to be above the centromere and the long arm to be below the centromere.

One embodiment of the present invention provides a method for automatic detection and identification of chromosomes in a karyotype diagram based on deep learning, which is executed by a system, and the system mainly includes a database and a processor, wherein: the database is used to store data and/or data, including multiple chromosome original images, multiple chromosome feature maps, multiple computer-readable instructions, etc.

The processor can transmit data and/or data in the database wirelessly or wiredly to execute the machine learning model and perform automatic detection and recognition of chromosome karyotype. Specifically, the processor includes at least an input/output module, a machine learning module and an automatic recognition module, wherein: the input/output module is used to transmit data and/or data with the database, and can input it into the machine learning module or the automatic recognition module.

The machine learning module has a machine learning model, and the machine learning model receives a part of the chromosome feature graphs of the database, performs object detection procedures in the chromosome feature graphs, extracts the sequence number of each pair of chromosomes and its predicted frame, and uses the predicted frame of each pair of chromosomes and its sequence number as parameters to generate a chromosome automatic detection and recognition model.

The automatic identification module receives a chromosome pattern to be tested and the chromosome automatic detection identification model, and identifies each pair of chromosomes in the chromosome pattern to be tested by inputting the chromosome pattern to be tested into the chromosome automatic detection identification model to generate an identification result.

In one embodiment of the present invention, the machine learning model includes two-stage object detection, that is, firstly, the first-stage object detection is performed on the chromosome feature graphs, and then the second-stage object detection is performed on the first-stage object detection results to obtain the sequence number of each pair of chromosomes and its prediction frame.

In another embodiment of the present invention, the machine learning module further includes a feedback training model, which determines which of the multiple prediction frames of the same pair of chromosomes should be retained through a predetermined IOU threshold, and feeds the chromosome pattern and its judgment result back to the machine learning module as a training data set for retraining to generate a new chromosome automatic detection and recognition model.

In one embodiment of the present invention, the processor further includes an image post-processing model, which receives the chromosome pattern to be detected and identified by the chromosome automatic detection and identification model, and performs image processing on the identification result according to an image processing instruction, such as capturing chromosomes with specific serial numbers or images containing them, removing the background of the selected image or specific area, adjusting the color tone, correcting the position of the chromosome, and sorting all chromosomes in the image according to the identified chromosome serial numbers.

In another embodiment of the present invention, the processor further includes an image pre-processing model, which performs an image pre-processing procedure such as image augmentation (Efficient Det Resize Crop) on at least one of the chromosome feature maps before the chromosome feature maps are input into the machine learning model.

Furthermore, the method for automatic detection and identification of chromosomes in karyotype diagrams based on deep learning disclosed in the embodiments of the present invention includes the following steps:

Step 101: Select a chromosome training atlas, which contains multiple feature maps corresponding to each pair of chromosomes, wherein the feature map of each pair of chromosomes is obtained by outputting a chromosome original image to an image labeling software, such as imagelaber software, and the image labeling software labels each pair of chromosomes in the chromosome original image, and then performs cropping and chromosome number classification according to the labels.

Step 102: Before the chromosome training atlas enters the machine learning module, the feature maps are first input into the image pre-processing module for image augmentation processing. Specifically, when the feature map exceeds the set size, the excess part is first cut off, and then rotation, mirroring, brightness adjustment and other processing are performed; when the feature map does not exceed the set size, there is no need to perform cutting and subsequent image pre-processing procedures.

Step 103: Output the chromosome training atlas to a machine learning model to perform the first stage object detection, that is, detect and select the chromosomes in each feature graph, and analyze the selection results to generate the first prediction frame corresponding to each pair of chromosomes, generate the sequence number of each pair of chromosomes and a first prediction frame for selecting each pair of chromosomes.

Step 104: Output the first prediction frames to a second object detection model for analysis, adjust the position and/or size of the first prediction frame according to the analysis results, generate the serial number of each pair of chromosomes and a second prediction frame for selecting each pair of chromosomes.

Step 105: Using the second prediction frames and their corresponding chromosome numbers as parameters, a chromosome automatic detection and recognition model is generated to automatically identify and classify at least one pair of chromosomes in a chromosome pattern to be tested.

Step 106: Output a chromosome pattern to be tested to the chromosome automatic detection and recognition model. The chromosome automatic detection and recognition model detects and selects each pair of chromosomes in the chromosome pattern to be tested, and marks the chromosome numbers of the selected chromosomes to generate an identification result.

Step 107: According to a user instruction, the recognition result is input into an image post-processing model to perform post-processing of the recognition result. Specifically, the image containing the chromosome pair with a specific serial number is extracted from the recognition result, and other objects except the chromosome with the specific serial number are removed; or the position of each pair of chromosomes is corrected; or the chromosomes are sorted according to the size of their serial numbers.

Step 108: When any pair of chromosomes in a recognition result has more than 2 second prediction frames, the recognition result will be input into the feedback training model to compare the IOU of each second prediction frame and the true frame with a predetermined IOU threshold, and determine the second prediction frame to be retained based on the comparison result. When the number of retained second prediction frames is greater than or equal to 2, multiple second prediction frames are merged into a single second prediction frame, and the chromosome pattern and judgment result are input into the chromosome training atlas, and steps 102 to 105 are executed again to generate a new chromosome automatic detection and recognition model.

The predetermined IOU threshold is usually set according to the size of the chromosome. For example, the 1st to 5th pair of chromosomes and the 13th to 16th pair of chromosomes are large and medium-sized chromosomes, respectively, and their predetermined IOU threshold is set to 0.7; while other small chromosomes, such as the 19th to 22nd chromosomes, have their predetermined IOU thresholds set to 0.5 or 0.6.

Furthermore, in step 108, when the second prediction frames do not intersect each other or have the same size and the IOU is zero, the feedback training model will incorporate the judgment parameters based on the distance between the center points of the second prediction frames, the length and width of each second prediction frame, etc.

Among them, the deep learning-based automatic detection and identification method of karyotype chromosomes disclosed in the present invention generates a feature map in the backbone network composed of Res2Net101 and BiFPN, and then obtains the prediction frame of each pair of chromosomes and its corresponding prediction score through a two-stage object detection framework composed of one stage and two stage.

In order to verify the effects that can be achieved by the technical features of the present invention, the following examples are given with detailed explanations.

Example 1: Data processing

The original chromosome image files used in this example come from the Genetic Laboratory of the Department of Women's Medicine, Taichung Veterans General Hospital; the chromosome samples come from the amniotic fluid of the fetus of the pregnant woman, a total of 5,000 images, 229,852 chromosomes, among which each original chromosome image file has a corresponding karyotype map.

Before machine training, each chromosome original image file must be marked and processed as follows to become a chromosome image: a person who has received medical training and acceptance will identify each chromosome based on its characteristics and check it against the standard karyotype map to ensure accuracy; at the end of each chromosome image, check whether the number of chromosomes is accurate; check the annotation of each chromosome through another marker and confirm whether the total number of chromosomes is correct; the tool for annotating chromosome images is the imagelaber software in Matlab, and the storage format is gTruth or xml file, and xml is converted from gTruth.

The chromosome images were randomly divided into training data set and test data according to the ratio of 3:1. The chromosome images in the aa set were divided into difficult and simple, as shown in Table 2. The judgment criteria for difficult images are: (1) chromosome overlap, especially overlap of more than two chromosomes; (2) dull color and unclear features; (3) chromosomes are too thin and long.

Example 2: Training parameters and performance test results

In the parameter setting during training, the iterative optimization method is Adaptive Moment Estimation (ADAM); the batch size is 1; the initial learning rate is equal to the best learning rate of warm up; in addition, to prevent overfitting, as long as the current validation total loss value exceeds the previous validation total loss value by 5 times, the training is stopped. In the focal loss part, the weight of positive samples is set to 0.6 and the weight of negative samples is set to 0.4.

Using 1250 chromosome images as the test data set and different backbone architectures as the feature extraction network, the performance of the karyotype chromosome automatic detection and identification method and system disclosed in the present invention based on deep learning is tested. The results are shown in Tables 3 and 4.

In Table 3, FPN is the abbreviation of FEATURE PYRAMID NETWORK; BiFPN is the abbreviation of Weighted Bi-directional Feature Pyramid Network; MULTI-IOU is multi-scale IOU, which takes different IOU thresholds for chromosomes of different sizes; DIOU is the abbreviation of Distance-IOU loss; GIOU is the abbreviation of GIOU_Loss.

The verification method used in Table 3 adopts the mAP of COCOAPI, and verifies and calculates the accuracy with reference to the real value of the test set, that is, the performance evaluation indicators used include mAP (IOU=0.5), capture rate (Recall), precision (Precision) and F1 index (F1 score). Among them, mAP is planned based on the mAP50 of COCOAPI, as shown in the following formulas (1) and (2), and the IOU between the measured frame and the real frame should be 0.5 (including 0.5) as the judgment standard to confirm that the predicted frame is true (True); Recall is calculated by the following formula (3); Precision is calculated by the following formula (4); and TP means that both the predicted frame and the real frame are true; TN means that both the predicted frame and the real frame are false; FP means that the predicted frame is true but the real frame is false; FN means that the predicted frame is false but the real frame is true.

AP. 50= AP ( IOUTH = Multi )...(1)

From the results in Table 3, it can be seen that the method and system for automatic detection and recognition of karyotype chromosomes based on deep learning disclosed in the present invention has good recognition performance regardless of the backbone network, with an accuracy rate of over 98%. Among them, the backbone network using Res2Net101+BiFPN+MULTI-IOU+DIOU has a higher capture rate, which means that the recognition performance for difficult images is better.

From the results in Table 4, we can see that A1, A2, A3, B4 and B5 classified as large chromosomes can obtain higher accuracy; while small chromosomes, such as G, X, and Y groups, are more likely to be covered by large chromosomes in the chromosome pattern, so the accuracy is lower.

Please refer to Figure 1 and Figure 2. When the number of training data sets is 800, the accuracy begins to gradually increase; when it is 2000 images, the accuracy reaches 98.87%; when it is 3200 images, the accuracy reaches 98.87%; when it is 5000 images, the accuracy reaches 98.91%. When the amount of data and accuracy are stable, set 150 epochs, verify every 10 epochs, record the accuracy and total loss each time, and find that the best performance can be obtained at epoch 100, with a performance level of 98.91%.

Example 3: Actual test

In this example, simple, difficult and highly difficult chromosome images are selected, as shown in Figures 3 to 5. Figure 3 is a simple chromosome image, with very little chromosome overlap and adhesion, and the chromosome size is moderate and the features are clear; Figure 4 is a difficult chromosome image, with an increase in the number of overlapping and adhesion chromosomes, and the chromosome bending direction is more varied; Figure 5 is a highly difficult chromosome image, with a significant increase in the number of overlapping and adhesion chromosomes, more variable chromosomes, and even chromosomes overlapping impurities.

The three chromosome images were identified by the deep learning-based karyotype chromosome automatic detection and identification method and system disclosed in the present invention, and the accuracy of the identification results was evaluated by a medical examiner. From the evaluation results, it can be seen that the deep learning-based karyotype chromosome automatic detection and identification method and system disclosed in the present invention has 100% accuracy for simple chromosome images; 98% accuracy for difficult chromosome images; and 90% accuracy for highly difficult chromosome images.

Example 4: Comparison with other models

For other object detection models: Faster-RCNN, YOLOv4, Retinanet, Swin-transformer, YOLOR and Centernet++, the data set disclosed in Example 1 is used for training and testing respectively, and the results are compared with the chromosome automatic detection and recognition model disclosed in the present invention to perform recognition performance. The results are shown in Table 5.

From the results in Table 5, it can be seen that compared with other models, the method and system for automatic detection and identification of karyotype chromosomes based on deep learning disclosed in the present invention performs better in all four indicators, which means that it can effectively distinguish chromosomes from the background and can perform correct classification, so as to achieve the advantage of having a higher accuracy.

In addition, the method and system for automatic detection and identification of chromosomes in karyotype images based on deep learning disclosed in the present invention are combined with another model: DeepACEV2 (H.Bai, T.Zhang, C.Lu, W.Chen, F.Xu, and Z.-B.Han, "Chromosome extraction based on U-Net and YOLOv3," IEEE Access, vol.8, pp.178563-178569, 2020.) for performance comparison, the training methods used were the same, and the results are shown in Table 6; and the accuracy of the karyotype chromosome automatic detection and recognition method and system based on deep learning disclosed in the present invention and DeepACEV2 in identifying difficult images is compared, and the results are shown in Table 7. Among them, DeepACEV2 regards overlapping chromosomes as difficult images, finds overlapping chromosomes based on the IOU of objects and calculates their accuracy, while the difficult images defined by the present invention adopt the standards of clinical medical examiners, so compared with DeepACEV2, the difficulty of the difficult images judged by the present invention is higher.

From the results in Table 6 and Table 7, it can be seen that the accuracy of the present invention is 0.07% higher than that of deepACEv2, and the amount of test data is more than 3 times that of DeepACv2. In the case of more difficult images, the present invention can still obtain a high accuracy of 98.78. This result proves that the method and system for automatic detection and identification of karyotype chromosomes based on deep learning disclosed in the present invention has excellent stability and robustness.

Figure 1 is a graph showing the total data volume (number of images) and accuracy (%).

Figure 2 is a graph showing epoch (number of training times) and accuracy (%).

Figure 3 is a simple chromosome diagram.

Figure 4 is a difficult chromosome diagram.

Figure 5 is a highly difficult chromosome pattern.

Claims (11)

A method for automatic detection and identification of chromosomes in a karyotype diagram based on deep learning comprises the following steps: Step A: selecting a chromosome training atlas, which includes a plurality of feature maps corresponding to each pair of chromosomes; Step B: outputting the chromosome training atlas to a machine learning model, performing a first-stage object detection, and generating a first prediction frame for identifying each pair of chromosomes; Step C: outputting the first prediction frames of each pair of chromosomes to a second object detection model for analysis, and generating a first prediction frame for identifying each pair of chromosomes. The second prediction frame; Step D: using the second prediction frames corresponding to each pair of chromosomes and their corresponding chromosome numbers as parameters, a chromosome automatic detection and recognition model is generated to automatically identify and classify at least one pair of chromosomes in a chromosome pattern to be tested; Step E: outputting a chromosome pattern to be tested to the chromosome automatic detection and recognition model, the chromosome automatic detection and recognition model detects and selects the chromosomes in the chromosome pattern to be tested, and identifies the chromosome numbers of the selected chromosomes to generate an identification result. Step F: When any pair of chromosomes in the recognition result has more than 2 second prediction frames, the recognition result is input into a feedback training model to perform IOU analysis on the second prediction frames, and retain the second prediction frames with IOU higher than a predetermined IOU threshold, and the chromosome pattern to be tested and its updated recognition result are input into the chromosome training atlas, and steps B to D are executed again to generate a new chromosome automatic detection and recognition model; wherein, selection, generation, reception, output, detection, input, recognition and classification are executed by a processor using computer-readable instructions stored in a database. As described in claim 1, the method for automatic detection and identification of chromosomes in karyotype images based on deep learning, wherein the feature images in step A are captured from an original chromosome image. As described in claim 2, the method for automatic detection and identification of chromosomes in a karyotype image based on deep learning, wherein the feature image corresponding to each pair of chromosomes is obtained by outputting a chromosome original image to an image labeling software, the image labeling software labels each pair of chromosomes in the chromosome original image, and cutting and classifying the chromosome sequence numbers according to the labels. As described in claim 1, the method for automatic detection and identification of chromosomes in karyotype images based on deep learning, wherein, in step B, before the chromosome training atlas is output to the first object detection model, an image pre-processing procedure is performed on at least a portion of the feature maps, and the image pre-processing procedure includes size adjustment, brightness adjustment or direction adjustment. As described in claim 1, the method for automatic detection and identification of chromosomes in a karyotype image based on deep learning, wherein in step B, the first object detection model detects and selects each pair of chromosomes in the chromosome training atlas, and analyzes the selection results of each pair of chromosomes to generate the first prediction frame corresponding to each pair of chromosomes. As described in claim 1, the method for automatic detection and identification of karyotype chromosomes based on deep learning, wherein in step C, the second object detection model detects each pair of chromosomes in the chromosome training atlas and adjusts the first prediction frame to generate the second prediction frame corresponding to each pair of chromosomes. As described in claim 1, the method for automatic detection and identification of chromosomes in a karyotype image based on deep learning, wherein in step E, the identification result is input into an image post-processing model, and the image post-processing model extracts a picture containing a predetermined chromosome pair from the chromosome pattern to be detected, and performs background processing to generate an image of the predetermined chromosome pair. As described in claim 1, the method for automatic detection and identification of chromosomes in a karyotype diagram based on deep learning, wherein in step E, the identification result is input into an image post-processing model, and the image post-processing model performs position correction processing on each pair of chromosomes selected. A deep learning-based karyotype chromosome automatic detection and recognition system comprises: A database storing at least a plurality of chromosome feature maps, wherein each pair of chromosomes in each chromosome feature map has a classification label; a processor having: an input/output module for transmitting data to and from the database; a machine learning module having a machine learning model for receiving at least one pair of chromosomes from the database; A chromosome feature map is obtained, and the sequence number and predicted frame of each pair of chromosomes are detected and extracted from the chromosome feature maps, and the sequence number and predicted frame of each pair of chromosomes are used as parameters to generate a chromosome automatic detection and recognition model; an automatic recognition module receives a chromosome pattern to be detected, and inputs the chromosome pattern to be detected into the chromosome automatic detection and recognition model to detect each pair of chromosomes in the chromosome pattern to be detected. chromosomes, and generates a recognition result; wherein, its characteristics are: the machine learning model includes a two-stage object detection procedure and a feedback training model; the first stage of the two-stage object detection procedure is to first detect each pair of chromosomes in the chromosome training atlas and select them, and analyze the selection results of each pair of chromosomes to generate the first prediction frame corresponding to each pair of chromosomes; the second stage is to detect the chromosome training Each pair of chromosomes in the atlas is adjusted to generate the second prediction frame corresponding to each pair of chromosomes; the feedback training model is used to judge whether the two prediction frames selected in the same pair of chromosomes should be retained through a predetermined IOU threshold, and the chromosome image and its judgment results are fed back to the machine learning module as a training data set for retraining to generate a new chromosome automatic detection and recognition model. As described in claim 9, the karyotype chromosome automatic detection and recognition system based on deep learning, wherein the processor further includes an image post-processing model, which receives the chromosome pattern to be tested that has been detected and recognized by the chromosome automatic detection and recognition model, and performs image processing on the recognition result according to an image processing instruction. As described in claim 9, the karyotype chromosome automatic detection and recognition system based on deep learning, wherein the processor further includes an image pre-processing model, which receives the chromosome feature map from the database and performs an image pre-processing procedure.