Data augmentation technique optimized for stable and unstable extrachromosomal multi-object images

Abstract

The present invention relates to a data augmentation technique optimized for stable and unstable data and overlapping objects of various other types of images, the data augmentation technique comprising the steps of: augmenting data of an image for insufficient learning data; determining whether the quality of the image is low; and performing preprocessing on the image having low quality. According to one embodiment, quantitative and qualitative improvement of data is sought by using commonly used techniques such as rotation and folding, as well as a mosaic method applied for the first time in the world, and mapping technology is applied to address difficulties in applying data augmentation techniques to overlapping images. Accordingly, it is possible to increase the accuracy of learning and improve the reliability of a program by augmenting data even with a small amount of images.

Description

Data augmentation techniques optimized for stable and unstable non-chromosomal multi-object images. The present invention relates to a data augmentation technique optimized for data of stable and unstable chromosomal extra-multi-object images, and in particular, to improve the quantity and quality of data by using the mosaic method, which was applied for the first time in the world, along with commonly used rotation and folding, and above all, to an image mapping method optimized for stable data and overlapping objects of various images by applying mapping technology to solve the problem of difficulty in individually recognizing (detecting) multiple objects in overlapping images. In addition, the present invention relates to a data augmentation technique optimized for data of stable and unstable chromosomal extra-multi-object images, and in particular, to an image mapping method optimized for stable and unstable data and overlapping objects of various images, by using a mosaic method together with methods such as horizontal flip, vertical flip, rotation, random brightness contrast, and downscale to improve the quantity and quality of data, and applying a mapping technique to solve the problem of difficulty in individually recognizing (detecting) multiple objects in overlapping images. Typically, counting chromosomes and identifying abnormal chromosomes is a labor-intensive task that is performed manually from start to finish. Furthermore, even when performed by experienced examiners, chromosome alignment and analysis require significant time. Furthermore, there are limitations that can hinder the efficiency of communication related to the results. Since each institution has different interpretation standards and different people interpret the test, the results may vary. Also, even if the same person interprets the test, different interpretation results may be obtained several months later, meaning that errors in interpretation results always occur. If an error occurs during the experimental process or during the interpretation of the results, the experiment must be restarted from the beginning, which requires significant costs, manpower, and time. If this occurs during a radiation emergency, it could delay appropriate medical intervention, putting lives at risk. Furthermore, arbitrarily manipulating a chromosome image can result in the original image being altered or distorted, resulting in a different image. Therefore, choosing a data augmentation technique optimized for your data is crucial. This is because if the image is altered or distorted, the accuracy of deep learning will significantly decrease. Meanwhile, Automated Dicentric Chromosome Identification

[2014] (ADCI) selects high-quality data by giving priority to various data and tracks centromeres by using gradient vector field-based active contour and centerline extraction methods during contour extraction. However, it takes a long time to use active contour-based region segmentation techniques and soft vector machines for classifying the source chromosome, and only classifies the source chromosome, and cannot analyze translocations of stable chromosomes. In addition, the Ikaros-karyotyping platform (Metasystem) is a karyotype analysis program that uses deep learning to search for metaphase chromosome clusters and classifies karyotypes from metaphase chromosome clusters. However, there is no explanation that deep learning or a series of image processing is used for searching for individual chromosomes, and since it does not use CNN, which shows the highest performance in image classification tasks, the accuracy is very low, and only the mobile origin among the abnormal chromosomes is classified. On the other hand, to create automated analysis programs using big data and deep learning, data quality and quantity must be high above all else. In other words, the more high-quality data you have, the more powerful the program can be. However, in the past, there was little data or a lot of low-quality data, which made learning difficult and resulted in significantly lower accuracy. In particular, the stable chromosomes were stained using fluorescent dyes, with chromosomes 1, 2, and 4 stained with red fluorescence, and the remaining chromosomes stained with blue fluorescence, and the chromosomes appear in fluorescent color on a black background. As this is a task that requires reading a total of 46 chromosomes, if even one of them is wrong, the image becomes unusable, so even if you have a large amount of data, the number of images that can be applied to actual learning is greatly reduced. Additionally, since chromosome images are not readily available, it is very difficult to secure the images required for learning-verification-testing. Therefore, data augmentation techniques are used because there is a lack of data for learning (small amount of images) due to overlapping chromosomes, chromosomes that are too small to distinguish (small object size), and images with mismatched chromosome numbers (class imbalance). There are various data augmentation techniques available, and after applying various techniques, I use the optimized method that fits the data I have. Previously, a mosaic method was used to synthesize a single image by taking a portion of each of the four images along with methods such as horizontal flip, vertical flip, rotation, random brightness contrast, and downscale, but this method lacked accuracy for overlapping chromosomes. On the other hand, developing an automated chromosome analysis program for deep learning requires many steps. First, image segmentation is required, followed by image classification, and finally object detection. In the image segmentation stage, there are thresholding-based, watershed-based, geometric-based, and deep-learning-based methods. In the image classification stage, there are supportive vector machines, Bayes classifiers, and deep-learning-based methods. And in object detection, there are deep-learning-based methods. Referring to Figure 1, a deep-learning based model was selected for object detection. In the case of stable type, chromosomes appear in red and blue fluorescent series on a black background, and numerical and structural chromosomes must be read. Numerically, 46±1 should be counted, and structurally, translocation, dicentric, ring, and fragment should be distinguished from normal chromosomes. The problem with object detection in stable mode is that the quality of chromosome images is low, for example, low contrast, saturated chromosomes, and variable intensity. Additionally, problems arose in that the number and shape of the chromosome images were difficult to determine due to overlapping images, the size was too small to distinguish the number of chromosomes, and there were too few chromosomes on one slide. Meanwhile, when performing object detection in unstable environments, problems arise, such as the low quality of chromosome images making it difficult to clearly distinguish chromosomes, the inability to determine the number or shape due to overlapping chromosome images, the inability to distinguish the number of chromosomes due to the size being too small, the number of chromosomes on a slide being too small, and the chromosomes having a variety of shapes and sizes. The technical task of the present invention is to improve the conventional problem, and to provide a data augmentation technique optimized for stable and unstable data as well as various types of images, i.e., multi-objects, which can increase the accuracy of learning and improve the reliability of the program by augmenting data even with a small amount of images. In addition, the technical task to be achieved by the present invention is to improve the conventional problem, and to provide a data augmentation technique optimized for stable and unstable data as well as overlapping objects of various images, which significantly reduces the time, effort, and cost required for selecting and counting chromosome images by automatically determining whether there are numerical or structural abnormalities in chromosomes by analyzing chromosome images in the metaphase state of cell division. In addition, the technical task to be achieved by the present invention is to improve the conventional problem, in the case of images analyzed by each individual, errors occurred when the same person re-analyzed them or compared them with images analyzed by other people, whereas the automated chromosome analysis provides the reading results in a form that can be identified at a glance, so that fast and accurate reading results can be obtained, and provides a data augmentation technique optimized for stable and unstable data and overlapping objects of various images. In order to solve these problems, in accordance with the characteristics of the present invention, a data augmentation technique optimized for overlapping objects of various images as well as stable data is provided. The steps of using the above image data augmentation technique are: It is used to solve problems of insufficient data, such as when the chromosome images on the slide are too small to distinguish their sizes, when the number of abnormal chromosomes is small to insufficient images to learn from, when the number of chromosomes on a single slide is insufficient, and when the chromosome images overlap each other to make it difficult to distinguish between them. The steps of using the above image data augmentation technique are: Random normal, horizontal flip, vertical flip, rotation (90℃, 180℃, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform were used. The steps of using the above image data augmentation technique are: Mapping randomly selects four images, detects each chromosome image, and randomly samples K points. The above K is randomly set to 40 to 60. The distance between the above K points is set to a maximum of 50. A score map is applied to the above overlapping chromosome images. The step of performing the above preprocessing is: When learning about multi-object images as well as chromosomes, there are images that hinder learning. These include overlapping images, images where individual images are small and difficult to distinguish, such as centromeres, images with an abnormal number of objects such as chromosomes in one slide, images where the staining is not balanced and difficult to distinguish, images where the background is dark and difficult to distinguish individual objects, and images where the object color is too blurry and cannot be distinguished from the background. These images require a preprocessing step for learning. The step of performing the above preprocessing automatically removes noise to obtain a clear multi-object image without manually removing noise, so that various images can be read. In order to solve these problems, in accordance with the characteristics of the present invention, a data augmentation technique optimized for unstable data as well as overlapping objects of various images is provided. The steps of using the above image data augmentation technique are: It is used to solve problems of insufficient data, such as when the chromosome images on the slide are too small to distinguish their sizes, when the number of abnormal chromosomes is small to not have enough images to learn, when the number of chromosomes on a slide is insufficient, and when the chromosome images are overlapping and difficult to distinguish, etc. The steps of using the above image data augmentation technique are: Random normal, horizontal flip, vertical flip, rotation (90℃, 180℃, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform were used. The steps of using the above image data augmentation technique are: Mapping randomly selects four images, detects each chromosome image, and randomly samples K points. The above K is randomly set to 40 to 60. The distance between the above K points is set to a maximum of 50. A score map is applied to the above overlapping chromosome images. The step of performing the above preprocessing is: When learning about multi-object images as well as chromosomes, there are images that hinder learning. These include overlapping images, images where individual images are small and difficult to distinguish, such as centromeres, images with an abnormal number of objects such as chromosomes in one slide, images where the staining is not balanced and difficult to distinguish, images where the background is dark and difficult to distinguish individual objects, and images where the object color is too blurry and cannot be distinguished from the background. These images require a preprocessing step for learning. The step of performing the above preprocessing automatically removes noise to obtain a clear multi-object image without manually removing noise, so that various images can be read. According to one embodiment, a data augmentation technique optimized for stable and unstable data can be provided, which can increase the accuracy of learning and improve the reliability of a program by augmenting data even with a small amount of images. In addition, according to one embodiment, a data augmentation technique optimized for objects in various images as well as stable and unstable data can be provided, which analyzes a multi-object image of a cell division metaphase state, automatically determines whether there are numerical or structural abnormalities in the multi-object, and significantly reduces the time, effort, and cost required for selecting and counting cell images. In addition, according to one embodiment, in the case of images analyzed by each individual, errors occurred when the same person re-analyzed them or compared them with images analyzed by other people. In contrast, the present invention provides a data augmentation technique optimized for stable and unstable data as well as objects in various images, which can obtain fast and accurate interpretation results by providing interpretation results in a form that can be identified at a glance. Figures 1 and 2 are drawings showing examples of typical target aberrations. FIG. 3 is an overview of the development of an automated chromosome analysis in a data augmentation technique optimized for overlapping chromosomes of unstable data according to the first embodiment of the present invention. FIG. 4 is a diagram illustrating the application of various data augmentation techniques to an unstable chromosome image according to the first embodiment of the present invention. FIG. 5 is a diagram showing an example of mapping in a data augmentation technique optimized for overlapping chromosomes of unstable data according to the first embodiment of the present invention. FIG. 6 is a diagram showing an example of the result of applying image mapping, which is a data augmentation technique optimized for overlapping chromosomes of unstable data according to the first embodiment of the present invention. FIG. 7 is a diagram showing an example of applying score mapping to an overlapping chromosome in a data augmentation technique optimized for overlapping chromosomes of unstable data according to the first embodiment of the present invention. FIG. 8 is a diagram showing an example before and after application of score mapping for overlapping chromosomes in a data augmentation technique optimized for overlapping chromosomes of unstable data according to the first embodiment of the present invention. FIG. 9 is a diagram showing an example of an analysis result for individually recognizing overlapping chromosomes after applying score mapping of overlapping chromosomes in a data augmentation technique optimized for overlapping chromosomes of unstable data according to the first embodiment of the present invention. FIG. 10 is a drawing showing an example of removing noise and the like to better distinguish chromosome images from unstable data according to the first embodiment of the present invention. FIG. 11 is an overview of the development of an automated chromosome analysis in a data augmentation technique optimized for overlapping chromosomes of unstable data according to the second embodiment of the present invention. FIG. 12 is a diagram illustrating the application of various data augmentation techniques to an unstable chromosome image according to the second embodiment of the present invention. FIG. 13 is a diagram showing an example of mapping in a data augmentation technique optimized for overlapping chromosomes of unstable data according to a second embodiment of the present invention. FIG. 14 is a diagram showing an example of the result of applying image mapping, which is a data augmentation technique optimized for overlapping chromosomes of unstable data according to the second embodiment of the present invention. FIG. 15 is a diagram showing an example of applying score mapping to an overlapping chromosome in a data augmentation technique optimized for overlapping chromosomes of unstable data according to a second embodiment of the present invention. FIG. 16 is a diagram showing an example before and after application of score mapping for overlapping chromosomes in a data augmentation technique optimized for overlapping chromosomes of unstable data according to a second embodiment of the present invention. FIG. 17 is a diagram showing an example of an analysis result for individually recognizing overlapping chromosomes after applying score mapping of overlapping chromosomes in a data augmentation technique optimized for overlapping chromosomes of unstable data according to a second embodiment of the present invention. FIG. 18 is a diagram showing an example of removing noise, etc. to better distinguish a chromosome image from unstable data according to the second embodiment of the present invention. Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings so that those skilled in the art can easily implement the present invention. The present invention may be implemented in various different forms and is not limited to the embodiments described herein. Additionally, throughout the specification, whenever a part is said to "include" a component, this does not mean that it excludes other components, but rather that it may include other components, unless otherwise specifically stated. FIG. 3 is an overview of the development of a stable chromosome automated analysis according to the first embodiment of the present invention, FIG. 4 is a diagram showing the application of various data augmentation techniques to a stable chromosome image according to the first embodiment of the present invention, FIG. 5 is a diagram showing an example of mapping in a data augmentation technique optimized for an overlapping chromosome of stable data according to the first embodiment of the present invention, FIG. 6 is a diagram showing an example of the result of applying image mapping of a data augmentation technique optimized for an overlapping chromosome of stable data according to the first embodiment of the present invention, FIG. 7 is a diagram showing an example of applying score mapping in the result of applying image mapping optimized for an overlapping chromosome of stable data according to the first embodiment of the present invention, FIG. 8 is a diagram showing an example of before and after applying score mapping for an overlapping chromosome in a data augmentation technique optimized for an overlapping chromosome of stable data according to the first embodiment of the present invention, and FIG. 9 is a diagram showing an example of the result of analysis for individually recognizing an overlapping chromosome after applying score mapping of an overlapping chromosome in a data augmentation technique optimized for an overlapping chromosome of stable data according to the first embodiment of the present invention, and FIG. 10 is a drawing showing an example of removing noise, etc. to better distinguish chromosome images from stable data according to the first embodiment of the present invention. In the present invention, multi-objects include chromosomes, cell images in microscope images, microscope images showing repeated structures of materials, etc., and as an example of the embodiment, a chromosome image will be described below. Referring to FIG. 3, a data augmentation technique optimized for overlapping chromosomes of stable data according to the features of the present invention is Step (210) of selecting a chromosome image, applying various data augmentation techniques, and applying a mapping technique to overlapping chromosomes; Step of performing preprocessing on low quality images (S220); Step of performing an object detection model after preprocessing (S230); It includes a step (S240) of performing post-processing. The data augmentation technique optimized for overlapping chromosomes of stable data according to the features of the present invention having such a configuration is described in detail as follows. First, randomly select an image to increase data (S210). We apply various data augmentation techniques, especially for overlapping chromosomes. Apply the mapping technique. The steps of using the above image data augmentation technique are: Data augmentation techniques are used because there is insufficient data for learning (small amount of images) due to overlapping chromosomes, chromosomes that are too small to distinguish (small object size), and images with mismatched chromosome numbers (class imbalance). The steps of using the above image data augmentation technique are: Augment image data using random normal, horizontal flip, vertical flip, rotation (90℃, 180℃, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform. The steps of using the above image data augmentation technique are: Referring to Figure 4, Mapping randomly selects four chromosome images, detects chromosomes, and randomly samples K points. Random sampling and mapping processes are performed as shown in Fig. 5, and K is randomly set to 40 to 60. The distance between the above K points is set to a maximum of 50. The results of mapping augmentation are shown in Fig. 6. A score map such as Fig. 7 or Fig. 8 is applied to the above overlapping chromosomes. Next, preprocessing is performed (S220). The step of performing the above preprocessing is: It is applied when there is a lack of trainable images due to overlapping chromosomes, small size, or images with incorrect chromosome numbers, and when image distinction is not accurate after using data augmentation techniques. The step of performing the above preprocessing automatically removes noise and enables accurate chromosome reading without manually removing noise, even when the original image is used as is. The step of performing the above preprocessing automatically adjusts the intensity using chromosome image deep learning to create an image that is easy to learn, with the background being white, the dyed chromosomes (1, 2, 4) being red, and the remaining chromosomes being gray. Since the stable chromosome is a color image in RGB color with FISH staining, a model was developed to convert it into a Red-Gray image by applying the residual method in the basic U-Net, and color and saturation were enhanced. The step of performing the above preprocessing automatically processes the chromosome image using deep learning, thereby increasing the number of chromosome counts and valid images of abnormal chromosomes. The preprocessing process improves contrast and changes the image of fluorescent (red, blue) chromosomes on a black background into red-gray chromosomes on a white background. Referring to Figure 10, the forward diffusion process is fixed, and the denoising and color adjustment processes are performed in reverse. Next, the object detection model is performed after preprocessing (S230). In object detection, a deep-learning-based method was used. In object detection, various deep-learning-based transformer-based DETA model, state of the arts in object detection, DNN-based YOLOv8x model, and state of the arts in real-time object detection models were selected. In the case of the stable type, the chromosomes appear in a red-gray series on a white background, and numerical and structural chromosomal abnormalities must be identified. Next, post-processing is performed (S240). The post-processing step is well known in the field, so a detailed description is omitted. In the first embodiment of the present invention, when the quality of the chromosome image is low, a preprocessing step is performed. When there is a lack of learnable images due to overlapping chromosomes, small size, images with incorrect chromosome numbers, etc., a data augmentation technique is used to apply geometric changes or morphological transformations to the images, thereby improving the learning efficiency of the model. In addition, in the first embodiment of the present invention, chromosome reading is possible by automatically removing noise through a preprocessing process even when the original image is used as is without removing noise. In addition, in the first embodiment of the present invention, a data augmentation technique was used to solve problems such as overlapping chromosomes, small size, and images with mismatched chromosome numbers. Additionally, in the first embodiment of the present invention, data was augmented using random normal, horizontal flip, vertical flip, rotation (90℃, 180℃, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform. In one embodiment, data augmentation can be performed with a small number of images to increase learning accuracy and improve program reliability. In addition, according to the first embodiment, by analyzing the chromosome image in the metaphase state of cell division and automatically determining whether there are numerical or structural abnormalities in the chromosomes, the time, effort, and cost required for selecting and counting the chromosome image are also significantly reduced. FIG. 11 is an overview of the development of an automated analysis of unstable chromosomes according to the second embodiment of the present invention, FIG. 12 is a diagram of application of various data augmentation techniques to an image of an unstable chromosome according to the second embodiment of the present invention, FIG. 13 is a diagram showing an example of mapping in a data augmentation technique optimized for an overlapping chromosome of unstable data according to the second embodiment of the present invention, FIG. 14 is a diagram showing an example of the result of application of image mapping of a data augmentation technique optimized for an overlapping chromosome of unstable data according to the second embodiment of the present invention, FIG. 15 is a diagram showing an example of applying score mapping in the result of applying image mapping optimized for an overlapping chromosome of unstable data according to the second embodiment of the present invention, FIG. 16 is a diagram showing an example of before and after application of score mapping for an overlapping chromosome in a data augmentation technique optimized for an overlapping chromosome of unstable data according to the second embodiment of the present invention, and FIG. 17 is a diagram showing an analysis result of individually recognizing an overlapping chromosome after applying score mapping of an overlapping chromosome in a data augmentation technique optimized for an overlapping chromosome of unstable data according to the second embodiment of the present invention. This is a drawing showing an example, and FIG. 18 is a drawing showing an example of removing noise, etc. to better distinguish a chromosome image from unstable data according to the second embodiment of the present invention. In the present invention, multi-objects include chromosomes, cell images in microscope images, microscope images showing repeated structures of materials, etc., and as an example of the embodiment, a chromosome image will be described below. Referring to FIG. 11, a data augmentation technique optimized for overlapping chromosomes of unstable data according to the second embodiment of the present invention is Step (310) of selecting a chromosome image, applying various data augmentation techniques, and applying a mapping technique to overlapping chromosomes; Step of performing preprocessing on low quality images (S320); Step of performing an object detection model after preprocessing (S330); It includes a step (S340) of performing post-processing. A detailed description of a data augmentation technique optimized for overlapping chromosomes of unstable data according to the features of the present invention having such a configuration is as follows. First, randomly select an image to increase data (S310). We apply various data augmentation techniques, especially for overlapping chromosomes. Apply the mapping technique. The steps of using the above image data augmentation technique are: Data augmentation techniques are used because there is insufficient data for learning (small amount of images) due to overlapping chromosomes, chromosomes that are too small (small object size) to be distinguished, and images with mismatched chromosome numbers (class imbalance). The steps of using the above image data augmentation technique are: Augment image data using random normal, horizontal flip, vertical flip, rotation (90℃, 180℃, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform. The steps of using the above image data augmentation technique are: Referring to Figure 12, Mapping randomly selects four chromosome images to detect chromosomes and randomly samples K points. Random sampling and mapping processes are performed as shown in Fig. 13, and K is randomly set to 40 to 60. The distance between the above K points is set to a maximum of 50. The results of mapping augmentation are shown in Fig. 14. A score map such as Fig. 15 or Fig. 16 is applied to the above overlapping chromosomes. Next, preprocessing is performed (S220). The step of performing the preprocessing is as follows: It is applied when there is a lack of trainable images due to overlapping chromosomes, small size, or images with incorrect chromosome numbers, and when image distinction is not accurate after using data augmentation techniques. The step of performing the above preprocessing automatically removes noise and enables accurate chromosome reading without manually removing noise, even when the original image is used as is. The step of performing the above preprocessing automatically processes the chromosome image using deep learning, thereby increasing the number of chromosome counts and valid images of abnormal chromosomes. Referring to Figure 18, the forward diffusion process is fixed, and the reverse denoising process is performed in reverse. Next, the object detection model is performed after preprocessing (S330). In object detection, a deep-learning-based method was used. In object detection, various deep-learning-based transformer-based DETA model, state of the arts in object detection, DNN-based YOLOv8x model, and state of the arts in real-time object detection models were selected. In the case of unstable type, chromosomes appear in gray series on a white background, and numerical and structural chromosomal abnormalities must be identified. Next, post-processing is performed (S340). The post-processing step is well known in the field, so a detailed description is omitted. In the second embodiment of the present invention, when the quality of the chromosome image is low, a preprocessing step is performed. When there is a lack of learnable images due to overlapping chromosomes, small size, images with incorrect chromosome numbers, etc., a data augmentation technique is used to apply geometric changes or morphological transformations to the images, thereby improving the learning efficiency of the model. In addition, in the second embodiment of the present invention, chromosome reading is possible by automatically removing noise through a preprocessing process even when the original image is used as is without removing noise. In addition, in the second embodiment of the present invention, a data augmentation technique was used to solve problems such as overlapping chromosomes, small size, and images with mismatched chromosome numbers. In addition, in the second embodiment of the present invention, data was augmented using random normal, horizontal flip, vertical flip, rotation (90℃, 180℃, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform. In addition, according to one embodiment, errors occurred not only in individual reading results but also when readings by the same person were reread later. In contrast, by providing reading results in a form that can be identified at a glance, it is possible to obtain fast and accurate reading results. Furthermore, according to one embodiment, data augmentation can be achieved even with a small amount of images, thereby increasing learning accuracy and enhancing program reliability. Specifically, by analyzing chromosome images in the metaphase of cell division and automatically determining whether there are numerical or structural abnormalities in the chromosomes, the time, effort, and cost required to select and count cell images can be significantly reduced. Furthermore, while errors occurred in images analyzed individually by individuals, this technology provides the reading results in a readily identifiable format, enabling fast and accurate reading results. Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements made by those skilled in the art using the basic concept of the present invention defined in the following claims also fall within the scope of the present invention.

Claims

1. Step to determine whether the quality of the multi-object image is low; If the quality of the multi-object image is low, it includes a step of performing a preprocessing process. The above preprocessing process is, A data augmentation technique optimized for overlapping multi-object images of stable and heterogeneous images, performing diffusion and denoising processing to address poor quality images (low contrast, saturated chromosomes, variable intensity).

2. In paragraph 1, Data augmentation is used due to insufficient data for learning (small amount of images) due to overlapping images, images that are too small (small object size) to be selected, and images with mismatched numbers (class imbalance), and random normal and horizontal flip, vertical flip, rotation (90, 180, 270℃), random brightness contrast, downscale, mosaic, mapping with transform techniques. It is a data augmentation technique optimized for stable and other types of overlapping multi-object images.

3. In paragraph 2, The steps of using the above image data augmentation technique are: Mapping is a data augmentation technique optimized for stable and overlapping object images of various types of images, which randomly selects four images to detect objects and randomly samples K points.

4. In paragraph 3, The above K is a data augmentation technique optimized for stable and overlapping objects of various types of images, where the number is randomly set to 40 to 60.

5. In paragraph 4, A data augmentation technique optimized for stable and overlapping objects of various types of images, where the distance between the above K points is selected to be at most 50.

6. In paragraph 5, A data augmentation technique optimized for overlapping objects of various types and stable images that applies a score map to the overlapping objects above.

7. In paragraph 1, A data augmentation technique optimized for stable and diverse image data that adjusts diffusion and denoising processes for poor quality images (low contrast, saturated chromosomes, variable intensity).

8. Step to determine whether the quality of the multi-object image is low; If the quality of the multi-object image is low, it includes a step of performing a preprocessing process. The above preprocessing process is, A data augmentation technique optimized for overlapping multi-object images of unstable and heterogeneous types, performing diffusion and denoising processing to address poor quality images (low contrast, saturated chromosomes, variable intensity).

9. In paragraph 8, Data augmentation is used due to insufficient data for learning (small amount of images) due to overlapping images, images that are too small (small object size) to be selected, and images with mismatched numbers (class imbalance). It is a data augmentation technique optimized for unstable and other types of overlapping multi-object images that randomly use normal, horizontal flip, vertical flip, rotation (90, 180, 270℃), random brightness contrast, downscale, mosaic, and mapping with transform techniques.

10. In paragraph 9, The steps of using the above image data augmentation technique are: Mapping is a data augmentation technique optimized for unstable and overlapping objects in various images, which randomly selects four images to detect objects and randomly samples K points.

11. In paragraph 10, The above K is a data augmentation technique optimized for overlapping objects of various types of unstable and different images, where the number is randomly set to 40 to 60. A data augmentation technique optimized for overlapping objects.

12. In paragraph 11, A data augmentation technique optimized for unstable and overlapping objects of various types of images, where the distance between the above K points is selected to be at most 50.

13. In paragraph 12, A data augmentation technique optimized for overlapping objects of various types and unstable images by applying a score map to the overlapping objects above.

14. In paragraph 13, A data augmentation technique optimized for unstable and other types of diverse image data that adjusts diffusion and denoising processes for poor quality images (low contrast, saturated chromosomes, variable intensity).

Images