A method and system for intelligent chromosome karyotype segmentation integrating deep learning

Abstract

This application provides a method and system for intelligent chromosome karyotype segmentation that integrates deep learning. The method involves: enhancing chromosome microscopic images; performing region segmentation based on the enhanced chromosome images to obtain multiple chromosome region sub-images; for any chromosome region sub-image, extracting the chromosome pixel area and band count corresponding to that sub-image, and determining the structural coverage index of the sub-image based on these values; comparing the structural coverage index with a preset chromosome banding distribution model to perform anomaly screening and identify candidate chromosome adhesion regions; and segmenting the candidate adhesion regions using a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image. This application can screen and identify candidate adhesion regions based on the chromosome skeleton banding density, improving the accuracy of chromosome image segmentation.

Description

Technical Field

[0001] This application relates to the field of image segmentation technology, and more specifically, to a method and system for intelligent segmentation of chromosome karyotypes that integrates deep learning. Background Technology

[0002] Chromosomal karyotype analysis is one of the important diagnostic tools in clinical cytogenetics, and it has significant application value in the early screening of hereditary diseases, prenatal diagnosis, and detection of abnormal karyotypes. By analyzing the morphology, number, and structure of chromosomes, various chromosomal abnormality-related diseases can be effectively identified. However, chromosomes exhibit a relatively elongated and flexible structure during metaphase of cell division. Even chromosomes with the same serial number may show different curved shapes and spatial configurations in different cell nuclei or at different imaging times, resulting in their morphology exhibiting strong non-rigid variation characteristics.

[0003] Currently, chromosome karyotype analysis and overlapping chromosome segmentation still mainly rely on manual interpretation and segmentation by cytogenetic experts. This method is not only time-consuming and labor-intensive, but also highly dependent on the operator's experience level, which can easily lead to subjective differences and affect the consistency and stability of the analysis results. Although some chromosome segmentation methods based on image processing or automated analysis have been proposed in the existing technology, their segmentation accuracy and robustness are still insufficient when faced with complex situations such as chromosome morphological diversity, blurred boundaries, and chromosome adhesion and overlap. They are difficult to meet the actual clinical needs for high accuracy and high stability. Therefore, there is an urgent need for an automated karyotype segmentation method that can adapt to complex chromosome structural changes and effectively improve the segmentation accuracy of adhered regions. Summary of the Invention

[0004] This application provides a chromosome karyotype intelligent segmentation method and system that integrates deep learning, which can screen and identify candidate regions of adhesion based on the banding density of the chromosome skeleton, thereby improving the accuracy of chromosome image segmentation.

[0005] In a first aspect, this application provides a chromosome karyotype intelligent segmentation method that integrates deep learning. This method can be executed by a network device, or it can be executed by a chip configured in the network device. This application does not limit the execution of this method.

[0006] Specifically, the method includes:

[0007] Acquire chromosome microscopic images of the sample to be analyzed, and perform image enhancement on the chromosome microscopic images to obtain chromosome enhanced images;

[0008] Based on the enhanced chromosome image, region segmentation is performed to obtain multiple chromosome region sub-images;

[0009] For any chromosome region sub-image, extract the chromosome pixel area and band count corresponding to the chromosome region sub-image, and determine the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and band count.

[0010] The structural coverage index is compared with a preset chromosome banding distribution model, and anomaly screening is performed on the chromosome region sub-map to determine candidate chromosome adhesion regions.

[0011] The chromosome adhesion candidate region is segmented based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image.

[0012] In conjunction with the first aspect, in certain implementations of the first aspect, performing region segmentation based on the enhanced chromosome image to obtain multiple chromosome region sub-images specifically includes:

[0013] The chromosome region is separated from the chromosome enhancement image based on the threshold segmentation algorithm to obtain the initial binary segmentation image;

[0014] The initial binary segmentation image is subjected to morphological processing to obtain a connected region image. Each independent connected region in the connected region image is extracted, and multiple independent connected regions are selected as chromosome region sub-images based on morphological features.

[0015] In conjunction with the first aspect, in some implementations of the first aspect, for any chromosome region sub-image, extracting the chromosome pixel area and banding number corresponding to that chromosome region sub-image specifically includes:

[0016] The number of foreground pixels in the chromosome region sub-image is counted to obtain the pixel area of ​​the chromosome region sub-image;

[0017] Chromosomal boundary points of the chromosome region sub-map are extracted, and a chromosome skeleton structure is generated through a regularized connection strategy. The directional feature analysis of the chromosome skeleton structure is performed, and the gray-scale distribution sequence along the main axis of the chromosome skeleton structure is extracted.

[0018] Peak detection is performed based on the grayscale distribution sequence to identify local peak points in grayscale changes, and the display bands are filtered according to the interval and amplitude between adjacent peaks to count the number of effective display bands.

[0019] In conjunction with the first aspect, in some implementations of the first aspect, image enhancement is performed on the chromosome microscopic image to obtain a chromosome enhanced image. Specifically, this includes: after performing background noise filtering and contrast enhancement on the chromosome microscopic image, a multi-scale feature enhancement algorithm is used to visualize the band structure in the chromosome image to obtain a chromosome enhanced image.

[0020] In conjunction with the first aspect, in certain implementations of the first aspect, acquiring chromosome microscopic images of the sample to be analyzed specifically includes:

[0021] Chromosome slides were prepared from the cell samples to be tested and stained with fluorescent dyes. The chromosome slides were then irradiated with an excitation light source based on a preset wavelength range, and a microscopic image of the chromosomes in the sample was acquired.

[0022] In conjunction with the first aspect, in certain implementations of the first aspect, the structural coverage index of the chromosome region sub-image is determined based on the chromosome pixel area and the number of bands, specifically including:

[0023] Obtain the chromosome skeleton structure of the chromosome region submap;

[0024] Based on the chromosome skeleton structure, multiple chromosome skeleton segments are determined. For any chromosome skeleton segment, the corresponding banding distribution density is determined according to the chromosome pixel area and the number of bands.

[0025] The banding density corresponding to each chromosome skeleton segment is used as the structural coverage index of the chromosome region subgraph.

[0026] In conjunction with the first aspect, in some implementations of the first aspect, a fluorescence imaging camera is used as the image acquisition device for the chromosome microscopic image.

[0027] Secondly, this application provides a chromosome karyotype intelligent segmentation system integrating deep learning, which includes an image processing unit, the image processing unit comprising:

[0028] The image acquisition module is used to acquire chromosome microscopic images of the sample to be analyzed, and to perform image enhancement on the chromosome microscopic images to obtain chromosome enhanced images;

[0029] The image recognition module is used to perform region segmentation based on the enhanced chromosome image to obtain multiple chromosome region sub-images;

[0030] The image recognition module further extracts the chromosome pixel area and band count corresponding to any chromosome region sub-image, and determines the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and band count.

[0031] The image recognition module is also used to compare the structural coverage index with a preset chromosome banding distribution model, perform anomaly screening on the chromosome region sub-image, and determine candidate chromosome adhesion regions.

[0032] The image segmentation module segments the candidate regions of chromosome adhesion based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image.

[0033] Thirdly, this application provides a computer terminal device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described intelligent chromosome karyotype segmentation method that integrates deep learning.

[0034] Fourthly, this application provides a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to perform the operations described above in a chromosome karyotype intelligent segmentation method incorporating deep learning.

[0035] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0036] This application provides a chromosome karyotype intelligent segmentation method and system integrating deep learning. First, a chromosome microscopic image of the sample to be analyzed is acquired. The chromosome microscopic image is then enhanced to obtain an enhanced chromosome image. Based on the enhanced chromosome image, region segmentation is performed to obtain multiple chromosome region sub-images. For any chromosome region sub-image, the chromosome pixel area and banding number corresponding to that sub-image are extracted, and a structural coverage index for that sub-image is determined based on the chromosome pixel area and banding number. The structural coverage index is compared with a preset chromosome banding distribution model to perform anomaly screening on the chromosome region sub-image, identifying candidate chromosome adhesion regions. Finally, based on a deep learning segmentation model, the candidate chromosome adhesion regions are segmented to obtain the corresponding chromosome karyotype segmentation image.

[0037] Therefore, this application achieves preprocessing and structural separation of microscopic chromosome images through image enhancement and region segmentation. Based on this, a structural coverage index based on pixel area and band count is introduced to quantify the local structural features of chromosomes. This transforms the traditional visual experience-based interpretation process into a computable statistical feature analysis process. By comparing the skeleton band density in the structural coverage index with a pre-defined chromosome band distribution model, abnormal structural regions are screened and located, identifying complex situations such as chromosome entanglement, overlap, or adhesion. This reduces the search range and interference factors in subsequent segmentation tasks. Furthermore, a deep learning segmentation model is introduced only for candidate adhesion regions for refined processing, and the model parameters are intelligently adjusted based on the skeleton band density. This allows the model to focus on learning the feature representation of complex boundary regions, avoiding the error propagation problem caused by global segmentation, thereby improving the segmentation accuracy and boundary separation capability of adhered chromosomes.

[0038] In summary, this application can screen and identify candidate adhesion regions based on the banding density of the chromosome skeleton, thereby improving the accuracy of chromosome image segmentation. Attached Figure Description

[0039] Figure 1 This is an exemplary flowchart illustrating a chromosome karyotype intelligent segmentation method incorporating deep learning, according to some embodiments of this application.

[0040] Figure 2 These are example images of chromosome Q-banding provided in some embodiments of this application;

[0041] Figure 3 This is a schematic diagram of the structure of an image processing unit according to some embodiments of this application;

[0042] Figure 4 This is a schematic diagram of the structure of a computer terminal device that implements a chromosome karyotype intelligent segmentation method incorporating deep learning, according to some embodiments of this application. Detailed Implementation

[0043] This application acquires chromosome microscopic images of the sample to be analyzed, enhances the chromosome microscopic images to obtain enhanced chromosome images, performs region segmentation based on the enhanced chromosome images to obtain multiple chromosome region sub-images, extracts the chromosome pixel area and banding number corresponding to the chromosome region sub-image for any chromosome region sub-image, and determines the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and banding number, compares the structural coverage index with a preset chromosome banding distribution model, performs anomaly screening on the chromosome region sub-image, and identifies chromosome adhesion candidate regions, and performs adhesion region segmentation on the chromosome adhesion candidate regions based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image. This application can screen and identify adhesion candidate regions based on the chromosome skeleton banding density, thus improving the accuracy of chromosome image segmentation.

[0044] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific implementation methods. (Reference) Figure 1 The figure is an exemplary flowchart of a chromosome karyotype intelligent segmentation method incorporating deep learning, according to some embodiments of this application. This chromosome karyotype intelligent segmentation method 100 incorporating deep learning mainly includes the following steps:

[0045] In step S101, a chromosome microscopic image of the sample to be analyzed is acquired, and the chromosome microscopic image is enhanced to obtain an enhanced chromosome image.

[0046] Preferably, in some embodiments, acquiring chromosome microscopic images of the sample to be analyzed specifically includes:

[0047] Chromosome slides were prepared from the cell samples to be tested and stained with fluorescent dyes. The chromosome slides were then irradiated with an excitation light source based on a preset wavelength range, and a microscopic image of the chromosomes in the sample was acquired.

[0048] In practice, a microscopic imaging system is used to optically magnify and image the excited chromosome slides, and an image acquisition device acquires the imaging results to obtain chromosome microscopic images. During the image acquisition process, the exposure time, gain parameters, and focal length can be adjusted, and automatic focusing is performed based on image sharpness evaluation indicators to improve the contrast of the chromosome microscopic images. Finally, the acquired image signals are converted into digital image data and output for subsequent chromosome segmentation and karyotype analysis.

[0049] Optionally, in some embodiments, during the staining of chromosomal DNA with fluorescent dyes, this application preferably uses quinoline fluorescent dyes that have selective binding ability to AT base pairs, and uses a high-sensitivity fluorescence imaging camera as the image acquisition device for the chromosome microscopic image. In some other embodiments, other devices or equipment capable of acquiring microscopic images may also be used, and this application does not limit them.

[0050] In some specific embodiments of this application, chromosome slides can be scanned and acquired using an automated scanning platform, and the chromosome distribution image can be obtained by stitching together multiple field images.

[0051] Optionally, in some embodiments, image enhancement is performed on the chromosome micrograph to obtain an enhanced chromosome image. Specifically, this includes: after background noise filtering and contrast enhancement of the chromosome micrograph, a multi-scale feature enhancement algorithm is used to visualize the band structure in the chromosome image to obtain the enhanced chromosome image.

[0052] It should be noted that during chromosome microscopy, after chromosomal DNA is stained with fluorescent dyes (such as Q-banding dyes), as... Figure 2 As shown, Figure 2 These are example images of chromosome Q banding provided in some embodiments of this application. The differences in base composition in the chromosome karyotype will appear as alternating bright and dark bands under ultraviolet excitation. Among them, the AT-rich regions have a stronger binding ability with the dye and show a brighter fluorescence signal, while the regions with higher GC base content show a relatively darker signal, thus forming a banding distribution with periodic variation characteristics.

[0053] In its specific implementation, this application employs median filtering to smooth the chromosome microscopic image during background noise reduction processing to suppress high-frequency random noise generated during fluorescence imaging. For example, in some embodiments, a 3×3 to 5×5 median filtering window can be used to suppress salt-and-pepper noise. Furthermore, background estimation is performed on the image based on morphological opening operations of structuring elements. Circular / elliptical structuring elements (adaptive size of 15~50 pixels) are selected to extract low-frequency background, and non-uniform background correction is achieved through difference operations between the original image and the background image, thereby obtaining a pre-processed image with a uniform background.

[0054] Next, the background-corrected image undergoes contrast enhancement processing. An adaptive histogram equalization method is used to divide the image into blocks. Specifically, the image is divided into 16×16 sub-regions, and gray-level redistribution is performed within each sub-region. A contrast limit threshold is set to prevent excessive noise amplification, thereby improving the gray-level distinction between the chromosome banding region and the background, resulting in a contrast-enhanced image. A multi-scale feature enhancement mechanism is then used to make the banding structure in the chromosome image more explicit. This includes: constructing a multi-scale Gaussian spatial representation; convolving the image with Gaussian kernels of different standard deviations (1, 2, 4) to obtain smooth images at multiple scale levels; and performing differential Gaussian operations on images at adjacent scale levels to extract the corresponding scale range. The image extracts detailed change information and enhances the edge and texture response of the banding region. Based on the detailed feature extraction, a direction-sensitive filter is introduced to perform directional response analysis on the image. For example, Gabor filter banks are used to perform convolution processing in the 0°, 45°, 90° and 135° directions. By matching the structural features of the periodic distribution of chromosome banding along the length direction, the band response in specific directions is enhanced to highlight the periodic structure of the banding. Furthermore, the feature information extracted at different scales is weighted and fused, where the weight coefficients are adaptively determined according to the local contrast at each scale. For example, the weight parameters can be calculated through local variance to enhance the overall continuity of the banding region and suppress background noise interference, resulting in a banding structure enhanced image.

[0055] In step S102, region segmentation is performed based on the enhanced chromosome image to obtain multiple chromosome region sub-images;

[0056] Preferably, in some embodiments, performing region segmentation based on the enhanced chromosome image to obtain multiple chromosome region sub-images specifically includes:

[0057] The chromosome region is separated from the chromosome enhancement image based on the threshold segmentation algorithm to obtain the initial binary segmentation image;

[0058] The initial binary segmentation image is subjected to morphological processing to obtain a connected region image. Each independent connected region in the connected region image is extracted, and multiple independent connected regions are selected as chromosome region sub-images based on morphological features.

[0059] In specific implementation, a threshold segmentation algorithm is used to separate the chromosome region from the chromosome enhancement image to obtain an initial binary segmentation image. The threshold segmentation algorithm is preferably a locally adaptive thresholding method commonly used in image segmentation, i.e., calculating an adaptive threshold based on a local window (e.g., 15×15 or 31×31) to adapt to the uneven brightness in the chromosome microscopic image, thereby achieving effective separation of the chromosome foreground and background regions. Morphological processing is then performed on the initial binary segmentation image to obtain a connected region image. This morphological processing includes: using opening operations to remove isolated noise points (e.g., small regions with an area less than 50-200 pixels), using closing operations to fill in internal broken structures within the chromosome region, and smoothing the chromosome boundaries through a combination of dilation and erosion operations. In some embodiments, structural elements with a size of 3×3 to 7×7 (e.g., circles or ellipses) can be used to perform the morphological operations.

[0060] Further, connected component labeling processing is performed on the connected region image. An eight-neighbor connectivity analysis method is used to extract each independent connected region in the image, and region feature parameters are calculated for each connected region, including area, bounding rectangle size, aspect ratio, perimeter, and compactness. Based on these region feature parameters, each independent connected region is screened, with regions that meet chromosome morphological characteristics preferably retained as candidate regions. The screening conditions include: region area within a preset range (e.g., 1000~10000 pixels), aspect ratio greater than a preset threshold (e.g., greater than 2), and region contrast and texture structure features, used to remove background noise, small fragments, and non-chromosomal structures from the connected region image. The selected candidate connected regions are then cropped from the chromosome enhancement image to obtain multiple independent chromosome region sub-images.

[0061] In step S103, for any chromosome region sub-image, the chromosome pixel area and banding number corresponding to the chromosome region sub-image are extracted, and the structural coverage index of the chromosome region sub-image is determined based on the chromosome pixel area and banding number.

[0062] Preferably, in some embodiments, for any chromosome region sub-image, extracting the chromosome pixel area and banding number corresponding to that chromosome region sub-image specifically includes:

[0063] The number of foreground pixels in the chromosome region sub-image is counted to obtain the pixel area of ​​the chromosome region sub-image;

[0064] Chromosomal boundary points of the chromosome region sub-map are extracted, and a chromosome skeleton structure is generated through a regularized connection strategy. The directional feature analysis of the chromosome skeleton structure is performed, and the gray-scale distribution sequence along the main axis of the chromosome skeleton structure is extracted.

[0065] Peak detection is performed based on the grayscale distribution sequence to identify local peak points in grayscale changes, and the display bands are filtered according to the interval and amplitude between adjacent peaks to count the number of effective display bands.

[0066] In specific implementation, the chromosome region sub-image is binarized, and the foreground region is corrected for connectivity. The pixel area of ​​the chromosome region sub-image is obtained by counting the number of pixels in the binary image whose pixel value is the foreground value. In some embodiments, hole filling operation can be performed on the binary image before pixel counting to improve the accuracy of area calculation. This application will not elaborate on this. Then, the boundary points of the chromosome region sub-image are extracted. Specifically, edge detection operators (such as the Canny operator) can be used to obtain the chromosome outline, and the region outline is constructed based on the set of boundary points. On this basis, a regular connection strategy is used to smooth and complete the boundary, for example, by eliminating boundary jaggedness through polygon approximation.

[0067] Furthermore, a skeleton extraction process is performed based on the chromosome region, and a single-pixel-width chromosome skeleton structure is generated using a thinning algorithm (such as the Zhang-Suen thinning algorithm or the distance transform skeletonization method). In some other embodiments of this application, the chromosome region sub-image can also be input into a pre-trained multi-scale convolutional neural network model to extract multi-scale features of the chromosome and generate a corresponding skeleton keypoint map, wherein the keypoints include endpoints, bends, and intersections. Based on the skeleton keypoint map, response intensity is screened for various keypoints, and keypoints with confidence scores higher than a preset threshold are selected as candidate keypoints, and those with spatial distribution relationships are eliminated. After removing redundant or outliers, a stable set of key points is obtained. Based on this, a regularized connection strategy is used to topologically connect the key points to construct a preliminary skeleton structure. Specifically: for endpoints, the connection is extended along the local direction to the nearest bend or intersection; for bends, the connection is made to endpoints, bends, or intersections in their neighborhood; for intersections, at least three edges are connected to maintain topological connectivity, thereby forming an initial skeleton network that satisfies the characteristics of chromosome structure. Furthermore, curve fitting is performed on the polyline segments in the initial skeleton, for example, by using spline curves to smoothly connect the key points, in order to reduce the discontinuity of the skeleton polylines and obtain the chromosome skeleton structure.

[0068] Optionally, in some embodiments, for branched chromosome skeleton structures, short branches (e.g., branches less than 5-15 pixels in length) are removed using branch point detection and pruning strategies, retaining the main trunk structure to obtain a stable chromosome skeleton. After obtaining the chromosome skeleton structure, directional feature analysis is performed on it. Specifically, principal component analysis is used to determine the main axis direction of the skeleton, and grayscale projection processing is performed on the chromosome region sub-image along the direction perpendicular to the main axis to convert the two-dimensional image into a one-dimensional grayscale distribution sequence. In some embodiments, samples can be taken at equal intervals along the skeleton path (e.g., once every 2-5 pixels), and the average grayscale value can be calculated in the local neighborhood (e.g., a 3×3 or 5×5 window) to enhance the stability of the grayscale sequence. Qualitative analysis is performed on the grayscale distribution sequence. Firstly, the grayscale sequence is smoothed, for example, by using a moving average filter. Then, local peaks in the grayscale changes are identified through local maximum search, and these peaks are used as candidate locations for banding. Next, the candidate locations are filtered based on the interval between adjacent peaks and the peak amplitude. A peak amplitude threshold is set to eliminate low-contrast pseudo-peaks, and a peak spacing threshold (e.g., 5-20 pixels) is set to avoid duplicate counting of the same band. In some implementations, peak significance indicators (such as the difference between peak height and the neighborhood mean) can be used for further banding filtering. The effective peaks after filtering are counted to obtain the number of bands corresponding to the chromosome region sub-image.

[0069] Preferably, in some embodiments, the structural coverage index of the chromosome region sub-image is determined based on the chromosome pixel area and the number of bands, specifically including:

[0070] Obtain the chromosome skeleton structure of the chromosome region submap;

[0071] Based on the chromosome skeleton structure, multiple chromosome skeleton segments are determined. For any chromosome skeleton segment, the corresponding banding distribution density is determined according to the chromosome pixel area and the number of bands.

[0072] The banding density corresponding to each chromosome skeleton segment is used as the structural coverage index of the chromosome region subgraph.

[0073] In specific implementation, the chromosome skeleton structure is segmented to determine multiple chromosome skeleton segments. The segmentation method can be adaptive, using key skeleton points (such as endpoints, bends, or intersections) as boundaries, or equidistant segmentation based on skeleton path length; this application does not limit this approach. In some embodiments, the skeleton can be divided into several continuous segments with a fixed step size (e.g., every 10-30 pixels). For any chromosome skeleton segment, its corresponding local chromosome region is determined. Specifically, a cross-sectional window perpendicular to the skeleton direction can be constructed at each sampling point of the skeleton segment. (For example, a width of 5-15 pixels), and merge the pixel areas covered by each cross-sectional window to obtain the local region corresponding to the skeleton segment; within the local region, count the number of foreground pixels as the local pixel area of ​​the skeleton segment, and count the number of visible bands within the skeleton segment based on the grayscale distribution sequence or peak detection method; based on the local pixel area and the number of visible bands, calculate the visible band distribution density corresponding to the skeleton segment, which can be characterized by the number of visible bands per unit skeleton length; finally, label and arrange the visible band distribution densities corresponding to each chromosome skeleton segment, and use them as the structural coverage index of the chromosome region sub-image.

[0074] In step S104, the structural coverage index is compared with the preset chromosome banding distribution model, and anomaly screening is performed on the chromosome region sub-map to determine candidate chromosome adhesion regions.

[0075] It should be noted that, under normal circumstances, a single chromosome has a relatively independent skeletal structure in space. Its banding distribution exhibits continuous and regular variations along the chromosome's main axis; that is, the number, spacing, and grayscale changes of the bands have stable statistical characteristics within a local area. Therefore, under normal conditions without structural interference, the banding density of a chromosome is relatively smooth within a local region, maintaining a consistent correspondence with its length and structural morphology. However, when two or more chromosomes become entangled, overlapped, or adhered, a pixel-level structural superposition effect occurs in the microscopic imaging plane, causing the originally independent banding structures to mix spatially. This mixing effect leads to an increase in the frequency of grayscale changes within a local region, blurred banding boundaries, and abnormal banding statistics, resulting in a significant deviation in the banding density per unit skeletal length or unit area.

[0076] Specifically, in overlapping or adhered regions, due to the superposition of banding structures from multiple chromosomes, local areas may exhibit an abnormal increase in the number of bands or banding intervals that do not conform to the statistical regularity of a single chromosome. In crossover regions, structural directional conflicts can lead to non-periodic mutations in the grayscale sequence, thereby disrupting the original continuity of the banding distribution. Therefore, banding density can statistically reflect whether there is spatial superposition or structural interference in chromosome structures. Based on the above mechanism, this application, by modeling and analyzing the banding density along the chromosome skeleton direction and comparing it with a reference distribution model of normal chromosomes, can effectively identify regions where abnormal shifts in banding density occur, thereby enabling the location and screening of risk areas for chromosome entanglement, overlap, or adhesion.

[0077] Preferably, in some embodiments, the structural coverage index is compared with a preset chromosome banding distribution model, and anomaly screening is performed on each chromosome region sub-map to determine candidate chromosome adhesion regions, specifically including:

[0078] For any chromosome region subgraph, obtain the multiple banding distribution densities obtained by segmenting it along the skeleton, and set corresponding sequence labels for each skeleton segment according to the skeleton length direction;

[0079] Based on the sequence labels corresponding to each skeleton segment, the banding distribution density sequence is compared with the preset chromosome banding distribution model segment by segment, and the density deviation value corresponding to each skeleton segment is calculated. The deviation value includes the relative deviation ratio between the banding density of the current skeleton segment and the reference density.

[0080] When the banding density of a certain skeleton segment exceeds a preset threshold, the skeleton segment is marked as an abnormal skeleton segment, and continuous abnormal skeleton segments are merged through continuity analysis.

[0081] The abnormal skeleton segment is mapped back to its spatial location in the original chromosome region subgraph, and a candidate region mask is generated based on the local region range corresponding to the skeleton segment and used as a candidate region for chromosome adhesion.

[0082] In practice, the skeleton structure of each chromosome region subgraph is extracted and traversed along the topological path of the skeleton. During the traversal, the pixels are numbered according to the connection order of the skeleton, and the skeleton is segmented at fixed pixel intervals. The chromosome skeleton is divided into multiple skeletons with sequential relationships, and each skeleton segment corresponds to a relative position in the chromosome structure.

[0083] For each skeleton segment, a local analysis region is constructed at its centerline position. That is, at each sampling point of the skeleton segment, a cross-sectional window of fixed width is established with the normal direction of that point as the reference. The pixel area covered by the window is taken as the local region of the skeleton segment. Based on the number of local bands and the length of the skeleton segment, the band distribution density of the skeleton segment is obtained. The band distribution densities of all skeleton segments are arranged in order on the skeleton to form a complete band density sequence.

[0084] Based on this, a pre-defined chromosome banding distribution model is introduced. This model is derived from the statistical results of historical normal chromosome samples, including the normal banding density range of skeletal segments at different locations. For the current chromosome region sub-image, the banding density of each skeletal segment is compared one by one with the reference density at the corresponding position in the model, and the degree of deviation between the two is calculated. For example, it is measured using a relative deviation ratio. When the banding density of some skeletal segments is significantly higher or lower than the allowable range of the model, these skeletal segments are marked as abnormal skeletal segments. Then, a continuity analysis is performed on these abnormal skeletal segments. If multiple abnormal skeletal segments appear consecutively in the skeletal sequence, the continuous region is considered to have the overall characteristic of structural abnormality. Finally, the continuous abnormal skeletal segments are mapped back to their spatial locations in the original chromosome image. Specifically, based on the set of pixels corresponding to the skeletal segments, they are extended to the corresponding local cross-sectional region to generate a mask region in two-dimensional space. This mask region is the risk area in the chromosome region sub-image where there are entanglements, overlaps, or adhesions.

[0085] In step S105, the chromosome adhesion candidate region is segmented based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image.

[0086] Preferably, in some embodiments, segmenting the candidate chromosome adhesion regions based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image specifically includes:

[0087] Obtain the banding distribution density corresponding to each backbone segment in the candidate region of chromosome adhesion, and generate a bias regulation factor based on the banding distribution density corresponding to each backbone segment.

[0088] Based on the bias regulation factor of the candidate chromosome adhesion region, interval mapping is performed to generate the corresponding segmentation probability threshold and loss function weight;

[0089] The segmentation probability threshold and loss function weights are input into a deep learning segmentation model to adaptively segment and regulate the chromosome adhesion candidate regions, thereby obtaining the corresponding chromosome karyotype segmentation image.

[0090] In practice, the skeletal structure within the candidate region of chromosome adhesion is segmented to obtain multiple sequentially related skeletal segments. The banding distribution density for each segment is calculated, and this density is then compared with a pre-defined chromosome banding distribution model to obtain the density deviation value for each segment. Based on this, the density deviation values ​​of all segments are statistically fused to generate a deviation control factor. Specifically, the average density deviation value, the weighted deviation value based on the segment length, and the deviation variance are calculated, and a weighted combination is used to obtain the deviation control factor. This deviation control factor characterizes the degree of structural abnormality in the candidate adhesion region.

[0091] Furthermore, the bias control factor is introduced into the pre-trained deep learning segmentation model to dynamically adjust the key segmentation parameters of the model, wherein the key segmentation parameters include the segmentation probability threshold, the loss function weight, and the intensity of post-processing morphological operations.

[0092] Specifically, regarding the adjustment of the segmentation probability threshold, a linear mapping is performed between the base threshold and the bias control factor to obtain a dynamic segmentation threshold. As the bias control factor increases, the segmentation threshold is raised, thereby enhancing the ability to distinguish the boundaries of adhered regions. Regarding the adjustment of the loss function, the weight coefficients of the boundary constraint loss are adjusted according to the bias control factor, strengthening boundary learning constraints when the adhesion level is high. Regarding post-processing adjustment, the size of the structural elements in morphological operations is dynamically adjusted according to the bias control factor, enhancing region separation when adhesion is severe. The adhesion level is graded based on the magnitude of the bias control factor. When the bias control factor exceeds a preset threshold range, a significant structural abnormality is determined in the chromosome region, triggering a strong constraint segmentation strategy. When the bias control factor is in the intermediate range, a medium constraint strategy is used for segmentation optimization. When the bias control factor is low, a standard segmentation strategy is used. Through these control mechanisms, the deep learning segmentation model can adaptively adjust the segmentation strategy according to the degree of structural abnormality in the candidate chromosome adhesion regions, thereby achieving stable karyotype segmentation under different adhesion complexity conditions and improving the accuracy and structural consistency of the segmentation results.

[0093] Optionally, in some embodiments, in this embodiment, the deep learning segmentation model adopts the U-Net network model commonly used in existing technologies for chromosome karyotype segmentation. The U-Net model includes an encoding feature extraction module, a skip connection module, and a decoding reconstruction module, used to achieve pixel-level chromosome region segmentation. During model inference, candidate chromosome adhesion regions are input into the U-Net model. The encoder extracts multi-scale chromosome texture features and banding structure features, and the decoder restores the spatial resolution to generate a probability map of each pixel belonging to different chromosome instances. During adaptive adjustment, the segmentation probability threshold is applied to the probability map output by the model to binarize the pixel-level classification results, that is, when the probability of a pixel belonging to a certain chromosome category is greater than the segmentation probability threshold... If the value is positive, the pixel is determined to belong to the chromosome region; otherwise, it is determined to be background or an adhered boundary region. Simultaneously, the weights of the loss function are introduced into the boundary constraint loss term during model training or fine-tuning. This boundary constraint loss term is weighted under DiceLoss or BoundaryLoss constraints, enhancing boundary learning ability when adhesion is high, thereby improving the model's ability to separate adjacent chromosome boundary regions. In some implementations, a morphology-based post-processing step can be introduced after model output. The size of the structuring element is dynamically adjusted according to the bias control factor F, and erosion and dilation operations are performed on the segmentation results to further separate adhered regions and smooth segmentation boundaries. Finally, a chromosome karyotype segmentation image is output, achieving adaptive and fine segmentation of adhered chromosome regions.

[0094] Furthermore, in another aspect of this application, in some embodiments, this application provides a chromosome karyotype intelligent segmentation system incorporating deep learning, the system including an image processing unit, referenced... Figure 3 The figure is a schematic diagram of the exemplary hardware and / or software structure of an image processing unit according to some embodiments of this application. The image processing unit 200 includes: an image acquisition module 201, an image recognition module 202, and an image segmentation module 203, which are described below:

[0095] The image acquisition module 201 is used to acquire chromosome microscopic images of the sample to be analyzed, and to perform image enhancement on the chromosome microscopic images to obtain chromosome enhanced images;

[0096] Image recognition module 202 is used to perform region segmentation based on the enhanced chromosome image to obtain multiple chromosome region sub-images;

[0097] The image recognition module 202 further extracts the chromosome pixel area and band count corresponding to any chromosome region sub-image, and determines the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and band count.

[0098] The image recognition module 202 is also used to compare the structural coverage index with a preset chromosome banding distribution model, perform anomaly screening on the chromosome region sub-image, and determine candidate chromosome adhesion regions.

[0099] The image segmentation module 203 segments the chromosome adhesion candidate regions based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image.

[0100] The foregoing detailed an example of a chromosome karyotype intelligent segmentation method and system integrating deep learning provided in the embodiments of this application. It is understood that the corresponding device includes hardware structures and / or software modules for performing each function in order to achieve the above functions.

[0101] Those skilled in the art should readily recognize that, based on the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a certain function in the application is executed in a manner that drives hardware or computer software depends on the specific application and design constraints of the technical solution. Therefore, those skilled in the art can use different methods to implement the described function for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0102] In addition, this application also provides a computer terminal device, which includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute the above-described intelligent chromosome karyotype segmentation method that integrates deep learning.

[0103] In some embodiments, reference Figure 4 The figure is a schematic diagram of the structure of a computer terminal device implementing a chromosome karyotype intelligent segmentation method incorporating deep learning, according to some embodiments of this application. The chromosome karyotype intelligent segmentation method incorporating deep learning in the above embodiments can... Figure 4 The computer terminal device 300 shown is used to implement this, and the computer terminal device 300 includes at least one communication bus 301, communication interface 302, processor 303 and memory 304.

[0104] The processor 303 can be a general-purpose central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more devices used to control the execution of a chromosome karyotype intelligent segmentation method that incorporates deep learning as described in this application.

[0105] The communication bus 301 may include a path for transmitting information between the aforementioned components.

[0106] Memory 304 may be read-only memory (ROM) or other types of static storage devices capable of storing static information and instructions, random access memory (RAM) or other types of dynamic storage devices capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 304 may exist independently and be connected to processor 303 via communication bus 301. Memory 304 may also be integrated with processor 303.

[0107] The memory 304 stores program code for executing the scheme of this application, and its execution is controlled by the processor 303. The processor 303 executes the program code stored in the memory 304. The program code may include one or more software modules. In the above embodiments, the determination of candidate regions for chromosome adhesion can be achieved by the processor 303 and one or more software modules in the program code in the memory 304.

[0108] Communication interface 302 uses any transceiver-like device for communicating with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area network (WLAN), etc.

[0109] Optionally, the computer terminal device 300 may also include a power supply 305 for providing power to various devices or circuits in the real-time computer terminal device.

[0110] In a specific implementation, as one example, a computer terminal device may include multiple processors, each of which may be a single-core (single-CPU) processor or a multi-core (multi-CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0111] The aforementioned computer terminal device can be a general-purpose computer terminal device or a dedicated computer terminal device. In specific implementations, the computer terminal device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer terminal device.

[0112] In addition, other aspects of this application provide a computer-readable storage medium storing at least one computer program, which is loaded and executed by a processor to perform the operations described above in a chromosome karyotype intelligent segmentation method incorporating deep learning.

[0113] In summary, the chromosome karyotype intelligent segmentation method and system integrating deep learning disclosed in this application first acquires a chromosome microscopic image of the sample to be analyzed, performs image enhancement on the chromosome microscopic image to obtain an enhanced chromosome image; performs region segmentation based on the enhanced chromosome image to obtain multiple chromosome region sub-images; for any chromosome region sub-image, extracts the chromosome pixel area and banding number corresponding to the chromosome region sub-image, and determines the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and banding number; compares the structural coverage index with a preset chromosome banding distribution model, performs anomaly screening on the chromosome region sub-image, and determines chromosome adhesion candidate regions; performs adhesion region segmentation on the chromosome adhesion candidate regions based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image. This application can screen and identify adhesion candidate regions based on the chromosome skeleton banding density, improving the accuracy of chromosome image segmentation.

[0114] The above descriptions are merely embodiments of this application, and common knowledge such as specific technical solutions or characteristics in the solutions are not described in detail here. It should be noted that those skilled in the art can make several modifications and improvements without departing from the technical solutions of this application, and these should also be considered within the scope of protection of this application, without affecting the effectiveness of the implementation of this application or the practicality of the patent.

[0115] The scope of protection claimed in this application shall be determined by the content of its claims. The specific embodiments described in the specification can be used to interpret the content of the claims. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of the invention. Therefore, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A chromosome karyotype intelligent segmentation method integrating deep learning, characterized in that, include: Acquire chromosome microscopic images of the sample to be analyzed, and perform image enhancement on the chromosome microscopic images to obtain chromosome enhanced images; Based on the enhanced chromosome image, region segmentation is performed to obtain multiple chromosome region sub-images; For any chromosome region sub-image, extract the chromosome pixel area and band count corresponding to the chromosome region sub-image, and determine the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and band count. The structural coverage index is compared with a preset chromosome banding distribution model, and anomaly screening is performed on the chromosome region sub-map to determine candidate chromosome adhesion regions. The chromosome adhesion candidate region is segmented based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image.

2. The method as described in claim 1, characterized in that, Based on the enhanced chromosome image, region segmentation is performed to obtain multiple chromosome region sub-images, specifically including: The chromosome region is separated from the chromosome enhancement image based on the threshold segmentation algorithm to obtain the initial binary segmentation image; The initial binary segmentation image is subjected to morphological processing to obtain a connected region image. Each independent connected region in the connected region image is extracted, and multiple independent connected regions are selected as chromosome region sub-images based on morphological features.

3. The method as described in claim 1, characterized in that, For any chromosome region subimage, extracting the chromosome pixel area and banding number corresponding to that chromosome region subimage specifically includes: The number of foreground pixels in the chromosome region sub-image is counted to obtain the pixel area of ​​the chromosome region sub-image; Chromosomal boundary points of the chromosome region sub-map are extracted, and a chromosome skeleton structure is generated through a regularized connection strategy. The directional feature analysis of the chromosome skeleton structure is performed, and the gray-scale distribution sequence along the main axis of the chromosome skeleton structure is extracted. Peak detection is performed based on the grayscale distribution sequence to identify local peak points in grayscale changes, and the display bands are filtered according to the interval and amplitude between adjacent peaks to count the number of effective display bands.

4. The method as described in claim 1, characterized in that, Image enhancement of the chromosome microscopic image to obtain an enhanced chromosome image specifically includes: after background noise filtering and contrast enhancement of the chromosome microscopic image, a multi-scale feature enhancement algorithm is used to visualize the band structure in the chromosome image to obtain the enhanced chromosome image.

5. The method as described in claim 1, characterized in that, Acquiring chromosome microscopic images of the sample to be analyzed specifically includes: Chromosome slides were prepared from the cell samples to be tested and stained with fluorescent dyes. The chromosome slides were then irradiated with an excitation light source based on a preset wavelength range, and a microscopic image of the chromosomes in the sample was acquired.

6. The method as described in claim 1, characterized in that, The structural coverage index of the chromosome region sub-image is determined based on the chromosome pixel area and the number of bands, specifically including: Obtain the chromosome skeleton structure of the chromosome region submap; Based on the chromosome skeleton structure, multiple chromosome skeleton segments are determined. For any chromosome skeleton segment, the corresponding banding distribution density is determined according to the chromosome pixel area and the number of bands. The banding density corresponding to each chromosome skeleton segment is used as the structural coverage index of the chromosome region subgraph.

7. The method as described in claim 1, characterized in that, A fluorescence imaging camera was used as the image acquisition device for the chromosome microscopic images.

8. A chromosome karyotype intelligent segmentation system integrating deep learning, comprising an image processing unit, wherein the image processing unit is used to execute the chromosome karyotype intelligent segmentation method integrating deep learning as described in any one of claims 1 to 7, characterized in that, The image processing unit includes: The image acquisition module is used to acquire chromosome microscopic images of the sample to be analyzed, and to perform image enhancement on the chromosome microscopic images to obtain chromosome enhanced images; The image recognition module is used to perform region segmentation based on the enhanced chromosome image to obtain multiple chromosome region sub-images; The image recognition module further extracts the chromosome pixel area and band count corresponding to any chromosome region sub-image, and determines the structural coverage index of the chromosome region sub-image based on the chromosome pixel area and band count. The image recognition module is also used to compare the structural coverage index with a preset chromosome banding distribution model, perform anomaly screening on the chromosome region sub-image, and determine candidate chromosome adhesion regions. The image segmentation module segments the candidate regions of chromosome adhesion based on a deep learning segmentation model to obtain the corresponding chromosome karyotype segmentation image.

9. A computer terminal device, characterized in that, The computer terminal device includes a memory and a processor. The memory stores code, and the processor is configured to acquire the code and execute a chromosome karyotype intelligent segmentation method incorporating deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing at least one computer program, characterized in that, The computer program is loaded and executed by a processor to perform the operations described in any one of claims 1 to 7 of the chromosome karyotype intelligent segmentation method incorporating deep learning.

Images