Artificial intelligence-based chromosome karyotype image automatic identification method and system

Abstract

The invention provides a chromosome karyotype image automatic identification method and system based on artificial intelligence, and relates to the technical field of image automatic identification, and the method comprises the steps: extracting a plurality of stripe cross section contours which are distributed at equal intervals in the length direction of a chromosome in a chromosome karyotype image; constructing a striated section contour sequence of the chromosome; inputting the striated section contour sequence into a pre-trained recurrent neural network, outputting striated contrast at each section position by the recurrent neural network, and generating a striated degeneration curve of the chromosome according to the striated contrast at the adjacent section position; and extracting an attenuation coefficient of the banding contrast of the banding degeneration curve along the long axis of the chromosome, when the attenuation coefficient is smaller than a preset threshold value, judging that the banding distortion of the chromosome is enzymolysis transition type artificial distortion, and otherwise, judging that the real structure is abnormal. According to the method and the device, accurate identification of the chromosome enzymolysis excessive type striated distortion can be realized, so that artificial distortion and chromosome real abnormality are distinguished.

Description

Technical Field

[0001] This application relates to the field of automatic image recognition technology, and more specifically, to an automatic chromosome karyotype image recognition method and system based on artificial intelligence. Background Technology

[0002] Automatic image recognition technology, as a key technology for intelligent chromosome karyotype analysis, has significant application value in clinical cytogenetic testing. It can replace manual labor in extracting chromosome banding features, identifying morphological abnormalities, and effectively reducing the subjective errors of manual interpretation. It also improves the efficiency and standardization of karyotype analysis and is widely used in genetic disease screening, prenatal diagnosis, and tumor cell testing. It provides automated technical support for the accurate determination of chromosome structural abnormalities and promotes the development of clinical chromosome testing towards intelligence and efficiency.

[0003] In practical applications of chromosome karyotype image recognition, improper control of enzymatic digestion parameters during slide preparation can easily lead to artificial aberrations such as blurred bands, boundary ablation, and band fusion. These aberrations caused by excessive enzymatic digestion are highly similar to the visual characteristics of real structural abnormalities such as natural chromosome curvature and fragment loss. Existing automatic image recognition models struggle to effectively distinguish between the two, easily misjudging artificial aberrations as pathological abnormalities. Such misjudgments directly lead to distorted karyotype analysis results, increase the risk of false positives in clinical diagnosis, and affect the reliability of treatment outcomes. Therefore, how to accurately identify chromosome band aberrations caused by excessive enzymatic digestion and thus distinguish between artificial aberrations and real chromosome abnormalities has become a challenge for the industry. Summary of the Invention

[0004] This application provides an artificial intelligence-based method and system for automatic identification of chromosome karyotype images, which can accurately identify chromosome enzymatically over-digested banding aberrations, thereby distinguishing between artificial aberrations and real chromosome abnormalities.

[0005] In a first aspect, this application provides an automatic chromosome karyotype image recognition method based on artificial intelligence, the method comprising the following steps:

[0006] Obtain the karyotype image of the chromosome to be identified, and extract multiple striped cross-sectional contours that are equally spaced along the length direction from the chromosome in the karyotype image;

[0007] Arrange all the banded cross-sectional outlines of the chromosome in order from centromere to telomere to construct the chromosome banded cross-sectional outline sequence;

[0008] The striped cross-sectional contour sequence is input into a pre-trained recurrent neural network. The recurrent neural network outputs the striped contrast at each cross-sectional position and generates the striped degradation curve of the chromosome based on the striped contrast at adjacent cross-sectional positions.

[0009] The attenuation coefficient of the banding contrast along the long axis of the chromosome is extracted from the banding degradation curve. When the attenuation coefficient is less than a preset threshold, the banding aberration of the chromosome is determined to be an enzymatic over-digestion type artificial aberration; otherwise, it is determined to be a real structural abnormality.

[0010] In this embodiment, extracting multiple equally spaced banded cross-sectional contours along the length direction of the chromosomes in the chromosome karyotype image specifically includes:

[0011] Chromosome instance segmentation is performed on the chromosome karyotype image to obtain a binary mask image of a single chromosome;

[0012] The central line of the chromosome is extracted based on the binary mask image, and the length direction from the centromere to the telomere on the central line is determined.

[0013] Multiple sampling points are set at equal intervals along the center line. At each sampling point, the pixel grayscale value is sampled in a direction perpendicular to the center line to generate the textured cross-sectional profile at that sampling point.

[0014] In this embodiment, arranging all the banded cross-sectional outlines of the chromosome in order from centromere to telomere to construct the chromosome banded cross-sectional outline sequence specifically includes:

[0015] Identify the centromere position of each chromosome in the chromosome karyotype image, and determine the telomere direction of the short arm and the telomere direction of the long arm of the chromosome;

[0016] Starting from the centromere position, the stripe cross-sectional profiles at each sampling point are sorted along the short arm telomere direction and the long arm telomere direction, respectively.

[0017] All sorted banded cross-sectional profiles were sequentially spliced ​​together in the order from the short arm telomere through the centromere to the long arm telomere to construct the banded cross-sectional profile sequence of the chromosome.

[0018] In this embodiment, the striped cross-sectional contour sequence is input into a pre-trained recurrent neural network, and the recurrent neural network outputs the striped contrast at each cross-sectional position, specifically including:

[0019] The length of each textured cross-section contour in the textured cross-section contour sequence is normalized, and each contour is converted into a one-dimensional gray-scale feature vector of fixed length, thus obtaining a feature vector sequence that corresponds one-to-one with the cross-section position.

[0020] The feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-sectional position. The recurrent neural network receives the feature vector of the current cross-sectional position at each time step and updates its internal memory state.

[0021] The hidden state vectors output by the recurrent neural network at each time step are extracted. All hidden state vectors are arranged in order of cross-sectional position to form a hidden state drift field that reflects the evolution of chromosome banding along the long axis.

[0022] Perform a contrast response activation operation on each hidden state vector in the hidden state drift field to map the high-dimensional hidden state to a scalar form of the initial value of the stripe contrast.

[0023] The initial value of the band contrast is normalized to eliminate the contrast baseline drift caused by differences in staining conditions between different chromosomes, thus obtaining the band contrast at each cross-sectional position.

[0024] In this embodiment, generating the banding degradation curve of the chromosome based on the banding contrast at adjacent cross-sectional positions specifically includes:

[0025] Traverse all adjacent cross-section positions in ascending order of cross-section position, calculate the difference in strip contrast between each pair of adjacent cross-sections, and obtain the attenuation amplitude at each cross-section position.

[0026] Arrange all attenuation amplitudes in order of cross-sectional position to construct a band contrast gradient field that reflects the rate of change of band contrast along the long axis of chromosome;

[0027] Starting from the initial section position, the attenuation amplitudes in the band contrast gradient field are sequentially accumulated along the long axis of the chromosome to obtain the cumulative attenuation amplitude sequence of each section position relative to the initial section position;

[0028] An initial stripe degradation curve is generated based on the cumulative attenuation magnitude sequence of each cross-section position relative to the initial cross-section position;

[0029] The initial banding degradation curve is smoothed by moving average to eliminate spikes and fluctuations introduced by image noise or local irregularities in the banding, thus obtaining the final banding degradation curve of the chromosome.

[0030] In this embodiment, extracting the attenuation coefficient of the banding contrast along the long axis of the chromosome from the banding degradation curve specifically includes:

[0031] Perform a decay curve segment truncation operation on the stripe degradation curve, extract the effective curve segment from the starting section position to the ending section position, exclude the abnormal fluctuation area at both ends of the curve caused by the stripe boundary effect, and obtain the decay curve segment to be analyzed.

[0032] Attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of band contrast along the long axis of the chromosome.

[0033] In this embodiment, the feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-sectional position order. The recurrent neural network receives the feature vector at the current cross-sectional position at each time step and updates its internal memory state, specifically including:

[0034] A time step alignment mechanism is established between each feature vector in the feature vector sequence and the cross-sectional position, mapping each cross-sectional position from centromere to telomere to a time step input node of a pre-trained recurrent neural network;

[0035] At each current time step, the recurrent neural network reads the historical memory state saved in the previous time step and uses the historical memory state as the contextual background information of the current time step to participate in the processing of the current feature vector;

[0036] The feature vector input at the current time step is fused with the historical memory state of the previous time step by the gated fusion unit inside the recurrent neural network according to the preset gate weights to generate the candidate memory state of the current time step.

[0037] The candidate memory state is written into the memory unit of the recurrent neural network to complete the progressive refresh from the memory state of the previous time step to the memory state of the current time step, and the refreshed memory state is passed to the next time step.

[0038] In this embodiment, performing a contrast response activation operation on each hidden state vector in the hidden state drift field to map the high-dimensional hidden state to a scalar form of initial stripe contrast value specifically includes:

[0039] According to the ascending order of the cross-sectional positions, a point-by-point traversal operation is performed on each hidden state vector in the hidden state drift field, and the hidden state vector corresponding to each cross-sectional position is taken out as the input object of the activation operation in turn.

[0040] For each extracted hidden state vector, a fully connected dimensionality reduction mapping is performed, and all dimensional components of the hidden state vector are linearly aggregated according to the trained weight coefficients to obtain a single-dimensional aggregated response value.

[0041] The aggregated response value is subjected to nonlinear activation compression. The dynamic range of the aggregated response value is limited to a preset stripe contrast response range by a nonlinear activation function to obtain the compressed stripe contrast activation value.

[0042] The stripe contrast activation value is used as the initial stripe contrast value at the cross-sectional position to complete the mapping from the high-dimensional hidden state space to the scalar contrast space.

[0043] In this embodiment, attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of band contrast along the long axis of the chromosome, specifically including:

[0044] The attenuation curve segment is divided into multiple continuous sub-intervals according to the cross-sectional position, and each sub-interval covers a fixed number of continuous cross-sectional positions;

[0045] For each sub-interval, the local attenuation rate of the attenuation curve segment is estimated, and the segmented attenuation rate corresponding to the sub-interval is determined based on the change in the vertical axis between the start and end of the segment.

[0046] Arrange the segmented decay rates corresponding to all sub-intervals in order of cross-sectional position to construct a segmented decay rate sequence that reflects the fluctuation of decay rate along the long axis of chromosome;

[0047] The segmented attenuation rate sequence is subjected to weighted fusion processing, and higher fusion weights are assigned to sub-intervals near the centromere region. The weighted fusion result is used as the attenuation coefficient of band contrast along the long axis of the chromosome.

[0048] Secondly, this application provides an artificial intelligence-based automatic chromosome karyotype image recognition system for executing an artificial intelligence-based automatic chromosome karyotype image recognition method, the automatic chromosome karyotype image recognition system comprising:

[0049] The chromosome karyotype feature extraction module is used to acquire the chromosome karyotype image to be identified and extract multiple striped cross-sectional contours that are equally spaced along the length direction from the chromosome in the chromosome karyotype image.

[0050] The chromosome banding profile recognition module is used to arrange all the banding profiles of the chromosome in order from centromere to telomere to construct a chromosome banding profile sequence.

[0051] The chromosome banding degradation recognition module is used to input the banding cross-sectional contour sequence into a pre-trained recurrent neural network. The recurrent neural network outputs the banding contrast at each cross-sectional position and generates the chromosome banding degradation curve based on the banding contrast at adjacent cross-sectional positions.

[0052] The chromosome banding aberration determination module is used to extract the attenuation coefficient of the banding contrast along the long axis of the chromosome from the banding degradation curve. When the attenuation coefficient is less than a preset threshold, the chromosome banding aberration is determined to be an enzymatic over-digestion type artificial aberration; otherwise, it is determined to be a real structural abnormality.

[0053] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0054] A karyotype image of the chromosome to be identified is acquired. Multiple banding profiles, evenly spaced along the length of the chromosome, are extracted from the karyotype image. All banding profiles are arranged in order from centromere to telomere to construct a sequence of chromosome banding profiles. This sequence is input into a pre-trained recurrent neural network (RNN), which outputs the banding contrast at each profile location and generates a chromosome banding degradation curve based on the banding contrast at adjacent profile locations. The attenuation coefficient of the banding contrast along the chromosome's long axis is extracted. If the attenuation coefficient is less than a preset threshold, the chromosome banding aberration is determined to be an enzymatically excessive artificial aberration; otherwise, it is determined to be a genuine structural abnormality.

[0055] Therefore, in this application, when the attenuation coefficient is less than a preset threshold, the banding aberration of the chromosome can be determined to be an enzymatically over-digested artificial aberration; otherwise, it can be determined to be a true structural abnormality. Specifically, firstly, after acquiring the karyotype image of the chromosome to be identified, a binary mask image of a single chromosome is obtained through chromosome instance segmentation. The center line is extracted, and sampling points are set at equal intervals on the center line to generate the banding cross-sectional contour. This can accurately separate a single chromosome, avoiding interference from overlapping multiple chromosomes. Simultaneously, equal-interval sampling can uniformly capture banding details along the chromosome length direction, completely preserving local features of enzymatically over-digested artificial aberrations such as banding blurring and boundary ablation. This solves the problem of incomplete extraction of local banding features and easy omission of aberrations in existing models. The model addresses several key issues. First, by identifying the centromere position and determining the telomere direction of the short and long arms of the chromosome, the banded cross-sectional contours are pieced together in the order from the short-arm telomere through the centromere to the long-arm telomere, constructing a complete banded cross-sectional contour sequence. This accurately reconstructs the natural distribution pattern of bands along the chromosome's long axis, clearly presenting the continuous spatial correlation characteristics of band degradation caused by excessive enzymatic digestion, overcoming the shortcomings of existing models that only focus on the overall morphology and ignore the spatial evolution of bands. Second, the banded cross-sectional contour sequence is length-normalized, transformed into a fixed-length feature vector sequence, and input into a pre-trained recurrent neural network. The memory mechanism of the recurrent neural network captures the temporal correlation of the banded sequence, extracting latent features. The state drift field is converted into a scalar form of band contrast, and then amplitude normalization correction is used to eliminate baseline drift caused by differences in staining conditions. Subsequently, the difference in band contrast between adjacent sections is calculated, a gradient field is constructed, the attenuation amplitude is accumulated, and smoothing is performed to generate an accurate band degradation curve. This can transform the blurred visual features of excessive enzymatic distortion into quantifiable and analyzable curve features, effectively avoiding the problem of confusion between the local abrupt features of real structural anomalies (such as natural curvature and fragment loss) and the gradual features of artificial distortion. At the same time, the smoothing process eliminates the interference caused by image noise and local irregularities of the band, improving the reliability of the band degradation curve. Finally, by truncating the effective segment of the band degradation curve and excluding... By addressing boundary anomalies, the effective curve segment is divided into sub-intervals, and the segmented attenuation rate is estimated. Weighted fusion of these attenuation rates is then performed, highlighting the attenuation characteristics of the centromere region to obtain the attenuation coefficient. This coefficient is then compared with a preset threshold to accurately distinguish between enzymatically over-digested artificial aberrations and true structural abnormalities. This solves the problem of misjudgment caused by the similarity of their visual features, effectively preventing artificial aberrations from being misdiagnosed as pathological abnormalities, reducing the risk of distorted karyotype analysis results, decreasing the probability of false positives in clinical diagnosis, and ensuring the reliability of diagnostic results in scenarios such as genetic disease screening, prenatal diagnosis, and tumor cell testing. Furthermore, this improves the stability and intelligence of automatic image recognition technology, promoting the development of clinical chromosome testing towards greater accuracy and efficiency.

[0056] In summary, the technical solution adopted in this application can achieve accurate identification of chromosome enzymatically over-digested banding aberrations, thereby distinguishing between artificial aberrations and real chromosomal abnormalities. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0058] Figure 1 This is an exemplary flowchart of an automatic chromosome karyotype image recognition method based on artificial intelligence provided in this application;

[0059] Figure 2 This is a schematic diagram of chromosome banding cross-sectional contour extraction provided in this application;

[0060] Figure 3 This is a module structure diagram of an artificial intelligence-based automatic chromosome karyotype image recognition system provided in this application;

[0061] Figure 4 This is a schematic diagram illustrating the application scenario of the AI-based automatic chromosome karyotype image recognition system provided in this application. Detailed Implementation

[0062] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0063] This application provides an artificial intelligence-based method and system for automatic chromosome karyotype image recognition. The core of this method involves acquiring a chromosome karyotype image to be identified, extracting multiple banded cross-sectional contours evenly spaced along the length of the chromosome from the karyotype image, arranging all the banded cross-sectional contours in order from centromere to telomere to construct a chromosome banded cross-sectional contour sequence, inputting the banded cross-sectional contour sequence into a pre-trained recurrent neural network, which outputs the banded contrast at each cross-sectional position and generates a chromosome banded degradation curve based on the banded contrast at adjacent cross-sectional positions, extracting the attenuation coefficient of the banded contrast along the chromosome's long axis, and determining that the chromosome banding aberration is an enzymatic over-ablation type of artificial aberration when the attenuation coefficient is less than a preset threshold; otherwise, it is determined to be a true structural abnormality.

[0064] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown in the figure, this is an exemplary flowchart of an automatic chromosome karyotype image recognition method based on artificial intelligence according to this embodiment of the present application. The automatic chromosome karyotype image recognition method includes the following steps:

[0065] In step S1, an image of the chromosome karyotype to be identified is obtained, and multiple striped cross-sectional contours with equal spacing along the length direction are extracted from the chromosome in the chromosome karyotype image.

[0066] It should be noted that the chromosome karyotype image to be identified in this application refers to a grayscale microscopic image containing the complete chromosome banding structure obtained after microscopic imaging of a chromosome banding specimen.

[0067] In practice, firstly, the chromosome specimens that have completed banding are optically acquired using a dedicated microscopic imager for chromosome karyotype analysis. During the acquisition process, the magnification and exposure time parameters are fixed. After acquiring the color microscopic image, the color channel information is removed by grayscale conversion, and only the pixel brightness information is retained. The processed grayscale image is then used as the chromosome karyotype image to be identified.

[0068] In this embodiment, the extraction of multiple equally spaced banded cross-sectional contours along the length direction of the chromosomes in the chromosome karyotype image can be achieved using the following steps:

[0069] Chromosome instance segmentation is performed on the chromosome karyotype image to obtain a binary mask image of a single chromosome;

[0070] The central line of the chromosome is extracted based on the binary mask image, and the length direction from the centromere to the telomere on the central line is determined.

[0071] Multiple sampling points are set at equal intervals along the center line. At each sampling point, the pixel gray value is sampled along the length direction perpendicular to the center line to generate the textured cross-sectional profile at that sampling point.

[0072] It should be noted that, in this application, chromosome instance segmentation refers to an image segmentation operation that accurately distinguishes a single chromosome region from the background region and other chromosome regions in a chromosome karyotype image; the binary mask image refers to a black-and-white binary image containing only two pixel values ​​that can clearly highlight the outline of a single chromosome; the chromosome centerline refers to a continuous curve that extends along the length of a single chromosome and is located at the geometric center of the chromosome, reflecting the chromosome's orientation; the banded cross-sectional contour refers to a sequence formed by arranging the pixel gray values ​​collected at the sampling point along the direction perpendicular to the chromosome centerline, which can reflect the banded distribution characteristics of the chromosome at the sampling point.

[0073] It should also be noted that in this application, references Figure 2 As shown in the figure, this figure is a schematic diagram of chromosome banding cross-sectional contour extraction provided in the embodiment of this application. The figure sequentially shows the process of obtaining a single chromosome binary mask image by segmenting the chromosome karyotype image, extracting the chromosome center line and determining the centromere and length direction, setting equally spaced sampling points on the center line, and sampling pixel gray values ​​at each sampling point along the direction perpendicular to the center line to generate the banding cross-sectional contour.

[0074] In specific implementation, firstly, the acquired chromosome karyotype image to be identified is binarized using the Otsu thresholding algorithm. Then, a morphological opening operation is performed using a 3×3 rectangular structuring element (the size of the structuring element is set according to the minimum width of the chromosome to remove small noise points in the image). The processed black and white image containing only the foreground and background of a single chromosome is used as a binary mask image for that single chromosome. Secondly, the obtained binary mask image is processed using the Zhang-Suen thinning algorithm to extract the chromosome centerline. This algorithm iteratively deletes non-central pixels on the chromosome edge, retains central pixels, and connects all the retained central pixels in sequence to form a continuous curve as the chromosome centerline. Simultaneously, the chromosome centerline is extracted by searching for chromosome... The constriction position with the lowest gray value on the chromosome centerline is taken as the centromere. The direction from the centromere to the ends of the chromosome is taken as the length direction from the centromere to the telomeres. Finally, on the extracted chromosome centerline, multiple sampling points are set with a fixed step size of 5 pixels (the step size is determined according to the pixel resolution of the chromosome karyotype image and the actual length of the chromosome, ensuring that the number of sampling points is 10-20, which can fully cover the length direction of the chromosome). At each sampling point, the pixel gray value is read point by point along the length direction perpendicular to the centerline from one edge of the chromosome to the other edge. The read gray values ​​are arranged in the traversal order, and the arranged gray value sequence is taken as the band cross-sectional contour at that sampling point.

[0075] In step S2, all the banded cross-sectional outlines of the chromosome are arranged in order from centromere to telomere to construct the chromosome banded cross-sectional outline sequence.

[0076] In this embodiment, the sequence of chromosome banding cross-sectional outlines is constructed by arranging all the banding cross-sectional outlines of the chromosome in order from centromere to telomere, which can be achieved by the following steps:

[0077] Identify the centromere position of each chromosome in the chromosome karyotype image, and determine the telomere direction of the short arm and the telomere direction of the long arm of the chromosome;

[0078] Starting from the centromere position, the stripe cross-sectional profiles at each sampling point are sorted along the short arm telomere direction and the long arm telomere direction, respectively.

[0079] All sorted banded cross-sectional profiles were sequentially spliced ​​together in the order from the short arm telomere through the centromere to the long arm telomere to construct the banded cross-sectional profile sequence of the chromosome.

[0080] It should be noted that, in this application, the centromere position refers to a characteristic position on the chromosome where the constricted and thinned gray value is significantly lower than that of the surrounding area and can distinguish the short arm and long arm of the chromosome; the short arm telomere direction refers to the extension direction from the centromere to the end of the shorter arm of the chromosome; the long arm telomere direction refers to the extension direction from the centromere to the end of the longer arm of the chromosome; and the banded cross-sectional contour sequence refers to a continuous contour sequence reflecting the distribution of bands along the long axis of the chromosome.

[0081] In specific implementation, firstly, for the binary mask image of the chromosome and its centerline, a gray-level minimum value detection algorithm is used to traverse the centerline and surrounding pixels. By setting a gray-level threshold (which is determined based on the average gray-level value of the chromosome karyotype image), regions with gray-level values ​​below the threshold and exhibiting a constricted shape are found within the centerline. The center pixel of this region is taken as the centromere position of the chromosome. Subsequently, the lengths of the centerlines on both sides of the centromere are measured, and the extension direction of the shorter side is taken as the short arm telomere direction, and the extension direction of the longer side is taken as the long arm telomere direction. Secondly, starting from the sampling point corresponding to the centromere position, the pixel distance between each sampling point and the centromere sampling point is counted point by point along the short arm telomere direction. The distance is then calculated from... The stripe cross-sectional contours corresponding to each sampling point are sorted in ascending order to form a short-arm contour sequence. At the same time, the same method is used along the long-arm telomere direction, sorting the sampling points in ascending order according to the pixel distance between the sampling points and the centromere sampling points to form a long-arm contour sequence. These two sorted contour sequences are used as the sorting results of the stripe cross-sectional contours corresponding to the short arm and long arm, respectively. Finally, the short-arm contour sequence is arranged in the sorting order, followed by splicing the stripe cross-sectional contours corresponding to the centromere position, and then splicing the long-arm contour sequence. The complete spliced ​​contour sequence is used as the stripe cross-sectional contour sequence of the chromosome, ensuring that the contour sequence is consistent with the actual spatial distribution of the chromosome from the short-arm telomere through the centromere to the long-arm telomere.

[0082] In step S3, the striped cross-sectional contour sequence is input into a pre-trained recurrent neural network. The recurrent neural network outputs the striped contrast at each cross-sectional position and generates the striped degradation curve of the chromosome based on the striped contrast at adjacent cross-sectional positions.

[0083] In this embodiment, the striped cross-sectional contour sequence is input into a pre-trained recurrent neural network, and the recurrent neural network outputs the striped contrast at each cross-sectional position, which can be achieved by the following steps:

[0084] The length of each textured cross-section contour in the textured cross-section contour sequence is normalized, and each contour is converted into a one-dimensional gray-scale feature vector of fixed length, thus obtaining a feature vector sequence that corresponds one-to-one with the cross-section position.

[0085] The feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-sectional position. The recurrent neural network receives the feature vector of the current cross-sectional position at each time step and updates its internal memory state.

[0086] The hidden state vectors output by the recurrent neural network at each time step are extracted. All hidden state vectors are arranged in order of cross-sectional position to form a hidden state drift field that reflects the evolution of chromosome banding along the long axis.

[0087] Perform a contrast response activation operation on each hidden state vector in the hidden state drift field to map the high-dimensional hidden state to a scalar form of the initial value of the stripe contrast.

[0088] The initial value of the band contrast is normalized to eliminate the contrast baseline drift caused by differences in staining conditions between different chromosomes, thus obtaining the band contrast at each cross-sectional position.

[0089] It should be noted that, in this application, length normalization refers to a regularization operation that uniformly adjusts the grayscale sequences of banded cross-section contours of different pixel lengths to the same dimension; the internal memory state of the recurrent neural network represents the historical banded evolution information retained by the network when processing temporal features; the hidden state vector represents the high-dimensional feature vector encoding the local features of the banded texture output by the recurrent neural network at a single time step; the hidden state drift field represents the set of all hidden state vectors arranged in order of cross-section position, used to characterize the continuous evolution features of the banded texture along the chromosome long axis; the contrast response activation operation represents a feature transformation operation that maps the high-dimensional hidden state vector to a single-value banded texture discrimination; the initial value of the banded texture contrast represents the original banded texture clarity value without correction for staining differences; the amplitude normalization correction represents a numerical regularization operation that eliminates the numerical baseline shift caused by staining depth and exposure differences; and the banded texture contrast represents a standard value that, after correction, can objectively reflect the clarity of the banded texture at each cross-section position.

[0090] In specific implementation, firstly, the grayscale sequence of each group in the banded cross-section contour sequence is length-normalized using a linear interpolation algorithm, and all contour grayscale sequences are uniformly interpolated to a fixed length of 64 dimensions (the dimension is set according to the conventional grayscale distribution characteristics of the chromosome banded cross-section, taking into account both feature integrity and computational efficiency). The normalized 64-dimensional grayscale sequence is used as a one-dimensional grayscale feature vector for the corresponding cross-section position. The one-dimensional grayscale feature vectors of all cross-section positions are combined in spatial order to obtain a feature vector sequence that corresponds one-to-one with the cross-section position. Secondly, the feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-section position order. The recurrent neural network receives the feature vector of the current cross-section position at each time step and updates its internal memory state. Next, the 128-dimensional hidden state vectors output at each time step of the recurrent neural network are extracted, and all hidden state vectors are arranged sequentially according to the cross-sectional position. The overall high-dimensional feature set after arrangement is used as the hidden state drift field reflecting the evolution law of chromosome banding along the long axis. Subsequently, a contrast response activation operation is performed on each hidden state vector in the hidden state drift field to map the high-dimensional hidden state into a scalar form of initial banding contrast value. Finally, the initial banding contrast values ​​at all cross-sectional positions are traversed, and outliers exceeding 3 times the average gray level difference are removed. The global mean is then calculated, and baseline compensation is completed by subtracting the global mean from each initial banding contrast value. The compensated integer value is used as the banding contrast at each cross-sectional position after eliminating the difference in staining conditions.

[0091] In this embodiment, the feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-sectional position. The recurrent neural network receives the feature vector of the current cross-sectional position at each time step and updates its internal memory state, which can be achieved by the following steps:

[0092] A time step alignment mechanism is established between each feature vector in the feature vector sequence and the cross-sectional position, mapping each cross-sectional position from centromere to telomere to a time step input node of a pre-trained recurrent neural network;

[0093] At each current time step, the recurrent neural network reads the historical memory state saved in the previous time step and uses the historical memory state as the contextual background information of the current time step to participate in the processing of the current feature vector;

[0094] The feature vector input at the current time step is fused with the historical memory state of the previous time step by the gated fusion unit inside the recurrent neural network according to the preset gate weights to generate the candidate memory state of the current time step.

[0095] The candidate memory state is written into the memory unit of the recurrent neural network to complete the progressive refresh from the memory state of the previous time step to the memory state of the current time step, and the refreshed memory state is passed to the next time step.

[0096] It should be noted that the pre-trained recurrent neural network described in this application is specifically a pre-trained long short-term memory (LSTM) recurrent neural network. This means that after parameter iterative optimization based on a large number of labeled chromosome karyotype samples, it has the ability to encode the temporal evolution features of chromosome banding and can stably output the contrast features of banding. This network retains the banding correlation features of continuous sections on the long axis of chromosomes through memory units, avoids the loss of temporal features, and adapts to the continuous evolution characteristics of chromosome banding along the length direction. The training process of this pre-trained LSTM recurrent neural network adopts a supervised learning approach. First, a chromosome karyotype training sample set is constructed, selecting 12,000 labeled chromosome samples, including 4,000 normal chromosome samples, 4,000 samples of artificially aberrant chromosomes due to excessive enzymatic digestion, and 4,000 samples of real structural abnormalities. All samples undergo chromosome instance segmentation, band contour extraction, and length normalization to obtain a standardized 64-dimensional grayscale feature vector sequence as the network input sample. Second, training annotation data is created. Chromosome karyotype analysis professionals manually annotate each cross-sectional position of each sample, quantifying the clarity of the labeled bands into a scalar value between 0 and 1. This value is used as the true annotation value for the band contrast at the corresponding cross-sectional position. Finally, the LSTM network training structure is built. The network input dimension was set to 64, the hidden state vector dimension to 128, and the output layer to be a single fully connected layer with the ReLU activation function. The Adam optimizer was used for network parameter iteration, with an initial learning rate of 0.001 and a batch size of 32. Subsequently, the initial value of the band contrast of the network output and the manually labeled ground truth value were both normalized to 0-1. The mean squared error between the two was used as the loss function. The forward and backward propagation were iteratively performed on batches of training samples to update the internal control weights and the weight coefficients of the fully connected layer. Finally, when the network loss function value decreased by less than 0.0001 for 50 consecutive iterations, the network training was considered to have converged. The iteration was stopped, and all trained parameter weights were saved. The LSTM network with saved weights was used as the pre-trained recurrent neural network in this application.

[0097] It should also be noted that the time step alignment mechanism described in this application refers to a matching mechanism that establishes a one-to-one correspondence between the chromosome cross-section position and the time step input node of the recurrent neural network; the time step input node refers to the input unit of the recurrent neural network that receives a single set of feature vectors; the historical memory state refers to the feature data that records the band evolution information retained after the recurrent neural network has completed the processing of the previous time step; the contextual background information refers to the reference information reflecting the correlation of band features; the gated fusion unit refers to the core module inside the recurrent neural network used to fuse the current feature and the historical memory, specifically a joint module composed of the forget gate and the input gate of the LSTM recurrent neural network; the gate weight refers to the weight parameters used to adjust the fusion ratio of the current feature and the historical memory; the candidate memory state refers to the current time step feature fusion result to be written into the memory unit after gated fusion; the memory unit refers to the core component in the recurrent neural network used to store the memory state; the progressive refresh refers to the operation of iteratively updating the memory information by overwriting the previous time step memory state with the current candidate memory state.

[0098] In specific implementation, firstly, an index mapping method is used to establish a time step alignment mechanism. Each cross-sectional position from centromere to telomere is sequentially and continuously numbered (starting from 1 and increasing sequentially). Each numbered cross-sectional position is directly mapped to a time step input node of the pre-trained LSTM recurrent neural network, ensuring that the one-dimensional grayscale feature vector of each cross-sectional position corresponds to a unique time step input node. Secondly, at each current time step, the LSTM recurrent neural network retrieves the historical memory state saved in the previous time step through its internal memory retrieval module. The initial historical memory state is set as a 128-dimensional all-zero vector, with the same dimension as the hidden state vector. This historical memory state is directly used as the contextual background information of the current time step and input into the network processing module along with the one-dimensional grayscale feature vector of the current time step for encoding the current feature vector. Then, through... The fusion operation is performed through the gated fusion unit (composed of a forget gate and an input gate) inside the LSTM recurrent neural network. The weight of the forget gate is set to 0.3 and the weight of the input gate is set to 0.7 (the weight parameters are determined by pre-training based on the correlation of the temporal features of chromosome banding). The feature vector input at the current time step is weighted and summed with the historical memory state at the previous time step according to the above gate weights. The summed feature data is used as the candidate memory state at the current time step. Finally, the generated candidate memory state is directly written into the memory unit of the LSTM recurrent neural network through the memory writing module, overwriting the historical memory state retained at the previous time step. This completes the progressive refresh from the memory state at the previous time step to the memory state at the current time step. At the same time, the refreshed memory state at the current time step is passed to the next time step, providing contextual background information for the feature processing at the next time step.

[0099] In this embodiment, performing a contrast response activation operation on each hidden state vector in the hidden state drift field, and mapping the high-dimensional hidden state to a scalar form of initial stripe contrast value, can be achieved through the following steps:

[0100] According to the ascending order of the cross-sectional positions, a point-by-point traversal operation is performed on each hidden state vector in the hidden state drift field, and the hidden state vector corresponding to each cross-sectional position is taken out as the input object of the activation operation in turn.

[0101] For each extracted hidden state vector, a fully connected dimensionality reduction mapping is performed, and all dimensional components of the hidden state vector are linearly aggregated according to the trained weight coefficients to obtain a single-dimensional aggregated response value.

[0102] The aggregated response value is subjected to nonlinear activation compression. The dynamic range of the aggregated response value is limited to a preset stripe contrast response range by a nonlinear activation function to obtain the compressed stripe contrast activation value.

[0103] The stripe contrast activation value is used as the initial stripe contrast value at the cross-sectional position to complete the mapping from the high-dimensional hidden state space to the scalar contrast space.

[0104] It should be noted that, in this application, the input object of the activation operation represents a single hidden state vector selected for contrast response activation processing; the fully connected dimensionality reduction mapping represents a feature mapping operation that converts a high-dimensional hidden state vector into a single-dimensional value through a fully connected layer; the aggregated response value represents a single-dimensional feature value obtained after fully connected dimensionality reduction mapping, reflecting the comprehensive characteristics of the hidden state vector; the nonlinear activation compression represents a regularization operation that restricts the numerical range of the aggregated response value to a preset interval through a nonlinear function; the stripe contrast response interval represents a numerical range reflecting the clarity of the stripe; the stripe contrast activation value represents an intermediate value that conforms to the preset interval after nonlinear compression; and the scalar contrast space represents a feature space that contains only single-dimensional stripe contrast values.

[0105] In specific implementation, firstly, following the ascending order of chromosome cross-section positions from centromere to telomere, an index-based traversal is used to perform a point-by-point traversal of each hidden state vector in the hidden state drift field. The hidden state vectors are extracted sequentially starting from 1 according to their corresponding cross-section positions. Each extracted hidden state vector is used as the input object for the current contrast response activation operation. Secondly, for each extracted 128-dimensional hidden state vector (with dimensions consistent with the hidden state vectors output by the recurrent neural network), a pre-defined single-layer fully connected dimensionality reduction layer is input. The weight coefficients of this fully connected layer are obtained through iterative optimization using sample labeled data during pre-training (the weight coefficients are initialized using the Xavier initialization method, set according to the hidden state dimension and scalar output requirements). The weights are then processed through the fully connected layer. The next layer performs a linear weighted summation of all 128 components of the hidden state vector according to their corresponding weight coefficients, and uses the single-dimensional value obtained by the summation as the aggregate response value corresponding to the hidden state vector. Then, a nonlinear activation compression operation is performed on the obtained aggregate response value. The ReLU nonlinear activation function is selected to limit the dynamic range of the aggregate response value to a preset stripe contrast response interval of 0 to 1. The compressed value within this interval is used as the stripe contrast activation value. Finally, the obtained stripe contrast activation value is directly used as the initial value of the stripe contrast at the corresponding cross-sectional position, completing the mapping from the high-dimensional hidden state space to the scalar contrast space, ensuring that the hidden state vector at each cross-sectional position can correspond to a unique scalar form initial value of stripe contrast.

[0106] In this embodiment, the striped cross-sectional contour sequence is input into a pre-trained recurrent neural network, and the recurrent neural network outputs the striped contrast at each cross-sectional position, which can be achieved by the following steps:

[0107] Traverse all adjacent cross-section positions in ascending order of cross-section position, calculate the difference in strip contrast between each pair of adjacent cross-sections, and obtain the attenuation amplitude at each cross-section position.

[0108] Arrange all attenuation amplitudes in order of cross-sectional position to construct a band contrast gradient field that reflects the rate of change of band contrast along the long axis of chromosome;

[0109] Starting from the initial section position, the attenuation amplitudes in the band contrast gradient field are sequentially accumulated along the long axis of the chromosome to obtain the cumulative attenuation amplitude sequence of each section position relative to the initial section position;

[0110] An initial stripe degradation curve is generated based on the cumulative attenuation magnitude sequence of each cross-section position relative to the initial cross-section position;

[0111] The initial banding degradation curve is smoothed by moving average to eliminate spikes and fluctuations introduced by image noise or local irregularities in the banding, thus obtaining the final banding degradation curve of the chromosome.

[0112] It should be noted that, in this application, the attenuation amplitude represents the change in the clarity of banding between adjacent sections; the banding contrast gradient field represents the gradient field reflecting the rate of change of banding contrast along the chromosome's long axis; the cumulative attenuation amplitude sequence represents the total attenuation of banding contrast relative to the initial section; the initial banding degradation curve represents the original curve generated solely based on the cumulative attenuation amplitude sequence without noise processing, which can initially reflect the banding degradation trend; the moving average smoothing process represents the conventional data processing operation of numerically smoothing the initial curve through a sliding window to eliminate local fluctuations; the spur fluctuations represent sharp fluctuations in the curve caused by image noise and local irregularities in the banding; and the banding degradation curve represents the curve reflecting the overall degradation trend of the banding along the chromosome's long axis.

[0113] In practice, firstly, following the ascending order of cross-sectional positions from centromere to telomere, each pair of adjacent cross-sectional positions is selected sequentially using an index traversal method. The band contrast value corresponding to each pair of adjacent cross-sections is read, and the band contrast value of the next cross-section is subtracted from the band contrast value of the previous cross-section. The calculated difference is used as the attenuation amplitude at the next cross-sectional position. If the difference is negative, it is determined that there is no attenuation of band contrast, and the attenuation amplitude at that position is set to 0, ensuring that each cross-sectional position (except for the initial cross-section) corresponds to a unique attenuation amplitude. Secondly, all the calculated attenuation amplitudes are arranged in ascending order according to their corresponding cross-sectional positions. This ordered set of attenuation amplitudes is used as the band contrast gradient field, reflecting the rate of change of band contrast along the chromosome long axis. Then, starting from the initial cross-sectional position (the first cross-section on the centromere side), the cumulative attenuation amplitude of this cross-section is calculated. The initial value is set to 0. Then, along the long axis of the chromosome, the attenuation amplitude of the current section is added to the cumulative attenuation amplitude of the previous section. The value obtained from each addition is taken as the cumulative attenuation amplitude of the current section position relative to the starting section position. All cumulative attenuation amplitudes are arranged in order of section position to obtain a cumulative attenuation amplitude sequence. Then, a two-dimensional coordinate correspondence is established with the section position number as the horizontal axis and the corresponding cumulative attenuation amplitude as the vertical axis. All coordinate points are connected in order of section position to form a continuous curve as the initial banding degradation curve. Finally, the initial banding degradation curve is smoothed using a 3-point moving average algorithm. The average value of the current point and the two adjacent points is calculated point by point. This average value is used to replace the original value of the current point. The smoothing process is completed by traversing all coordinate points. The smoothed continuous curve is taken as the final chromosome banding degradation curve.

[0114] In step S4, the attenuation coefficient of the banding contrast along the long axis of the chromosome is extracted from the banding degradation curve. When the attenuation coefficient is less than a preset threshold, the banding aberration of the chromosome is determined to be an enzymatic over-digestion type artificial aberration; otherwise, it is determined to be a real structural abnormality.

[0115] In this embodiment, the attenuation coefficient of the banding contrast along the long axis of the chromosome can be extracted from the banding degradation curve using the following steps:

[0116] Perform a decay curve segment truncation operation on the stripe degradation curve, extract the effective curve segment from the starting section position to the ending section position, exclude the abnormal fluctuation area at both ends of the curve caused by the stripe boundary effect, and obtain the decay curve segment to be analyzed.

[0117] Attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of band contrast along the long axis of the chromosome.

[0118] It should be noted that the attenuation curve segment truncation operation described in this application refers to the operation of removing abnormal fluctuation areas from the complete band degradation curve and retaining the effective part that can be used for attenuation coefficient extraction; the effective curve segment refers to the part of the band degradation curve that has no obvious boundary effect and can truly reflect the attenuation law of band contrast along the long axis; the boundary effect refers to the local distortion phenomenon of the curve caused by abnormal gray-level fluctuations due to incomplete bands at both ends of the chromosome; the abnormal fluctuation area refers to the area of ​​sharp fluctuation and numerical change at both ends of the band degradation curve due to the boundary effect; the attenuation curve segment to be analyzed refers to the curve segment that reflects the overall attenuation trend of the band; the attenuation coefficient represents the comprehensive value that reflects the overall attenuation degree of band contrast along the long axis of the chromosome, and is used for subsequent aberration type determination.

[0119] In practice, firstly, the attenuation curve segment is truncated on the final band degradation curve. A fixed-number truncation method is used, and based on the total number of chromosome cross-section sampling points, two curve segments corresponding to the start and end of the curve are removed (the number of truncations is determined based on the normal influence range of the boundary effect to ensure that the abnormal fluctuation regions at both ends can be completely excluded), and the remaining curve part in the middle is retained. This retained curve part is used as the attenuation curve segment to be analyzed. Secondly, attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of band contrast along the long axis of the chromosome.

[0120] In this embodiment, attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of the band contrast along the long axis of the chromosome, which can be achieved by the following steps:

[0121] The attenuation curve segment is divided into multiple continuous sub-intervals according to the cross-sectional position, and each sub-interval covers a fixed number of continuous cross-sectional positions;

[0122] For each sub-interval, the local attenuation rate of the attenuation curve segment is estimated, and the segmented attenuation rate corresponding to the sub-interval is determined based on the change in the vertical axis between the start and end of the segment.

[0123] Arrange the segmented decay rates corresponding to all sub-intervals in order of cross-sectional position to construct a segmented decay rate sequence that reflects the fluctuation of decay rate along the long axis of chromosome;

[0124] The segmented attenuation rate sequence is subjected to weighted fusion processing, and higher fusion weights are assigned to sub-intervals near the centromere region. The weighted fusion result is used as the attenuation coefficient of band contrast along the long axis of the chromosome.

[0125] It should be noted that, in this application, the sub-interval refers to a continuous curve segment that is uniformly divided according to the cross-sectional position of the attenuation curve segment to be analyzed. Each sub-interval covers a fixed number of continuous cross-sectional positions, which is used to achieve accurate estimation of the local attenuation rate. The local attenuation rate estimation refers to the operation of calculating the band contrast attenuation rate within the sub-interval by using the start and end features of the curve segment within the sub-interval. The vertical axis change represents the cumulative attenuation amplitude difference between the start and end of the attenuation curve segment within the sub-interval, which is used to quantify the attenuation degree of the sub-interval. The segmented attenuation rate represents the local attenuation rate of the band contrast along the chromosome long axis within a single sub-interval, reflecting the speed of band degradation within the sub-interval. The segmented attenuation rate sequence represents the set of all segmented attenuation rates arranged in order of the cross-sectional positions corresponding to the sub-intervals, which can reflect the variation and fluctuation of the attenuation rate along the chromosome long axis. The fusion weight represents a parameter used to adjust the degree of influence of each sub-interval on the final attenuation coefficient, and its allocation is based on the band stability of the chromosome region corresponding to each sub-interval. The attenuation coefficient represents a comprehensive value reflecting the overall attenuation degree of band contrast along the chromosome long axis, which is the core basis for subsequent determination of the chromosome aberration type.

[0126] In practice, firstly, the attenuation curve segment to be analyzed is uniformly divided according to the increasing order of cross-sectional positions. Based on the total number of cross-sections contained in the curve segment to be analyzed, it is divided into three continuous sub-intervals. Each sub-interval covers a fixed number of continuous cross-sectional positions. If the total number of cross-sections is not divisible by 3, the remaining cross-sections are assigned to the sub-interval closest to the telomeres. Each continuous curve segment after division is considered as a sub-interval. Secondly, the local attenuation rate of the attenuation curve segment within each sub-interval is estimated using the difference method. The cumulative attenuation amplitude corresponding to the starting cross-section and the cumulative attenuation amplitude corresponding to the ending cross-section of each sub-interval are read. The cumulative attenuation amplitude at the ending cross-section is subtracted from the cumulative attenuation amplitude at the starting cross-section to obtain the vertical axis change of the sub-interval. This vertical axis change is then divided by the number of cross-sections contained in the sub-interval, and the calculated value is taken as the segmental attenuation rate corresponding to the sub-interval. Next, the segmented attenuation rates corresponding to all sub-intervals are arranged in ascending order according to their corresponding cross-sectional positions, that is, in the order of sub-intervals near the centromere, intermediate sub-intervals, and sub-intervals near the telomere. This ordered sequence of segmented attenuation rates is used as the segmented attenuation rate sequence reflecting the fluctuation of attenuation rate along the chromosome long axis. Finally, the segmented attenuation rate sequence is weighted and fused. Based on the characteristics that the banding in the centromere region of the chromosome is more stable and the attenuation pattern is more representative, a fusion weight of 0.4 is assigned to the sub-intervals near the centromere region, a fusion weight of 0.3 is assigned to the sub-intervals in the intermediate region, and a fusion weight of 0.3 is assigned to the sub-intervals near the telomere region. The segmented attenuation rate of each sub-interval is multiplied by the corresponding fusion weight, and all the product results are summed. The summed value is used as the attenuation coefficient of banding contrast along the chromosome long axis.

[0127] It should be noted that the preset threshold mentioned in this application represents a critical value used to distinguish between enzymatically over-enzymatic artificial aberrations and real structural abnormalities. Its value is determined based on the statistical results of the attenuation coefficient of a large number of normal chromosome samples with enzymatically over-enzymatic artificial aberrations, which can accurately classify the two types of aberrations. The enzymatically over-enzymatic artificial aberration refers to the abnormal attenuation of band contrast caused by excessively long enzymatic digestion time and excessively high enzymatic concentration during chromosome banding treatment, which belongs to artificial manipulation abnormalities that are not part of the chromosome's own structure. The real structural abnormalities refer to structural defects such as deletions, duplications, and inversions that exist in the chromosome itself, which are manifested as a natural and stable abnormal attenuation of band contrast, and belong to the structural abnormalities of the chromosome itself.

[0128] In specific implementation, determining that the banding aberration of the chromosome is an enzymatically over-digested artificial aberration when the attenuation coefficient is less than a preset threshold, and otherwise determining it as a true structural abnormality, can be achieved in the following way: First, the preset threshold is determined using a statistical calibration method. 1000 normal chromosome samples, 800 samples of enzymatically over-digested artificial aberrations, and 800 samples of true structural abnormalities are selected. The attenuation coefficient of each sample is extracted, and the maximum value of the attenuation coefficient for the enzymatically over-digested samples and the minimum value of the attenuation coefficient for the true structural abnormality samples are calculated. The midpoint between these two values ​​is used as the preset threshold. Then, the attenuation coefficient of the banding contrast along the long axis of the chromosome is directly compared with this preset threshold. When the attenuation coefficient is less than the preset threshold, the banding aberration of the current chromosome is determined to be an enzymatically over-digested artificial aberration; when the attenuation coefficient is not less than the preset threshold, the banding aberration of the current chromosome is determined to be a true structural abnormality. This completes the accurate determination of the chromosome banding aberration type, ensuring that the determination result can be directly used for subsequent processing in chromosome karyotype analysis.

[0129] Therefore, in this application, when the attenuation coefficient is less than a preset threshold, the banding aberration of the chromosome can be determined to be an enzymatically over-digested artificial aberration; otherwise, it can be determined to be a true structural abnormality. Specifically, firstly, after acquiring the karyotype image of the chromosome to be identified, a binary mask image of a single chromosome is obtained through chromosome instance segmentation. The center line is extracted, and sampling points are set at equal intervals on the center line to generate the banding cross-sectional contour. This can accurately separate a single chromosome, avoiding interference from overlapping multiple chromosomes. Simultaneously, equal-interval sampling can uniformly capture banding details along the chromosome length direction, completely preserving local features of enzymatically over-digested artificial aberrations such as banding blurring and boundary ablation. This solves the problem of incomplete extraction of local banding features and easy omission of aberrations in existing models. The model addresses several key issues. First, by identifying the centromere position and determining the telomere direction of the short and long arms of the chromosome, the banded cross-sectional contours are pieced together in the order from the short-arm telomere through the centromere to the long-arm telomere, constructing a complete banded cross-sectional contour sequence. This accurately reconstructs the natural distribution pattern of bands along the chromosome's long axis, clearly presenting the continuous spatial correlation characteristics of band degradation caused by excessive enzymatic digestion, overcoming the shortcomings of existing models that only focus on the overall morphology and ignore the spatial evolution of bands. Second, the banded cross-sectional contour sequence is length-normalized, transformed into a fixed-length feature vector sequence, and input into a pre-trained recurrent neural network. The memory mechanism of the recurrent neural network captures the temporal correlation of the banded sequence, extracting latent features. The state drift field is converted into a scalar form of band contrast, and then amplitude normalization correction is used to eliminate baseline drift caused by differences in staining conditions. Subsequently, the difference in band contrast between adjacent sections is calculated, a gradient field is constructed, the attenuation amplitude is accumulated, and smoothing is performed to generate an accurate band degradation curve. This can transform the blurred visual features of excessive enzymatic distortion into quantifiable and analyzable curve features, effectively avoiding the problem of confusion between the local abrupt features of real structural anomalies (such as natural curvature and fragment loss) and the gradual features of artificial distortion. At the same time, the smoothing process eliminates the interference caused by image noise and local irregularities of the band, improving the reliability of the band degradation curve. Finally, by truncating the effective segment of the band degradation curve and excluding... By addressing boundary anomalies, the effective curve segment is divided into sub-intervals, and the segmented attenuation rate is estimated. Weighted fusion of these attenuation rates is then performed, highlighting the attenuation characteristics of the centromere region to obtain the attenuation coefficient. This coefficient is then compared with a preset threshold to accurately distinguish between enzymatically over-digested artificial aberrations and true structural abnormalities. This solves the problem of misjudgment caused by the similarity of their visual features, effectively preventing artificial aberrations from being misdiagnosed as pathological abnormalities, reducing the risk of distorted karyotype analysis results, decreasing the probability of false positives in clinical diagnosis, and ensuring the reliability of diagnostic results in scenarios such as genetic disease screening, prenatal diagnosis, and tumor cell testing. Furthermore, this improves the stability and intelligence of automatic image recognition technology, promoting the development of clinical chromosome testing towards greater accuracy and efficiency.

[0130] In summary, the technical solution adopted in this application can achieve accurate identification of chromosome enzymatically over-digested banding aberrations, thereby distinguishing between artificial aberrations and real chromosomal abnormalities.

[0131] Example 2: This application provides an automatic chromosome karyotype image recognition system based on artificial intelligence, referring to... Figure 3 As shown in the figure, this is a block structure diagram of an artificial intelligence-based automatic chromosome karyotype image recognition system according to this embodiment of the present application. The automatic chromosome karyotype image recognition system includes:

[0132] The chromosome karyotype feature extraction module 100 is used to acquire the chromosome karyotype image to be identified and extract multiple striped cross-sectional contours that are equally spaced along the length direction from the chromosome in the chromosome karyotype image.

[0133] The chromosome banding cross-sectional contour recognition module 200 is used to arrange all the banding cross-sectional contours of the chromosome in the order from centromere to telomere to construct the chromosome banding cross-sectional contour sequence.

[0134] The chromosome banding degradation recognition module 300 is used to input the banding cross-sectional contour sequence into a pre-trained recurrent neural network. The recurrent neural network outputs the banding contrast at each cross-sectional position and generates the chromosome banding degradation curve based on the banding contrast at adjacent cross-sectional positions.

[0135] The chromosome banding aberration determination module 400 is used to extract the attenuation coefficient of the banding contrast along the long axis of the chromosome from the banding degradation curve. When the attenuation coefficient is less than a preset threshold, the chromosome banding aberration is determined to be an enzymatic over-digestion type artificial aberration; otherwise, it is determined to be a real structural abnormality.

[0136] It should be noted that in this application, references Figure 4 As shown in the figure, this is a schematic diagram of the application scenario of the AI-based automatic chromosome karyotype image recognition system provided in the embodiment of this application. The figure horizontally shows the complete application process from chromosome specimen preparation, karyotype image acquisition, processing by the AI-based automatic chromosome karyotype image recognition system to output the aberration type determination result, and finally providing a reference for clinical diagnosis. The core of the system can accurately distinguish between artificial aberrations and real chromosomal structural abnormalities. The three core clinical application scenarios of the system are marked below, including genetic disease screening, prenatal diagnosis, and tumor cell testing, clearly demonstrating the application of this invention in the field of clinical cytogenetic testing.

[0137] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0139] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. An automatic chromosome karyotype image recognition method based on artificial intelligence, characterized in that, The automatic chromosome karyotype image recognition method includes the following steps: Obtain the karyotype image of the chromosome to be identified, and extract multiple striped cross-sectional contours that are equally spaced along the length direction from the chromosome in the karyotype image; Arrange all the banded cross-sectional outlines of the chromosome in order from centromere to telomere to construct the chromosome banded cross-sectional outline sequence; The striped cross-sectional contour sequence is input into a pre-trained recurrent neural network. The recurrent neural network outputs the striped contrast at each cross-sectional position and generates the striped degradation curve of the chromosome based on the striped contrast at adjacent cross-sectional positions. The attenuation coefficient of the banding contrast along the long axis of the chromosome is extracted from the banding degradation curve. When the attenuation coefficient is less than a preset threshold, the banding aberration of the chromosome is determined to be an enzymatic over-digestion type artificial aberration; otherwise, it is determined to be a real structural abnormality.

2. The method for automatic chromosome karyotype image recognition based on artificial intelligence as described in claim 1, characterized in that, Extracting multiple equally spaced banded cross-sectional contours along the length direction of chromosomes from the chromosome karyotype image specifically includes: Chromosome instance segmentation is performed on the chromosome karyotype image to obtain a binary mask image of a single chromosome; The central line of the chromosome is extracted based on the binary mask image, and the length direction from the centromere to the telomere on the central line is determined. Multiple sampling points are set at equal intervals along the center line. At each sampling point, the pixel grayscale value is sampled in a direction perpendicular to the center line to generate the textured cross-sectional profile at that sampling point.

3. The method for automatic chromosome karyotype image recognition based on artificial intelligence as described in claim 1, characterized in that, The process of arranging all the banded cross-sectional outlines of the chromosome in order from centromere to telomere to construct the chromosome banded cross-sectional outline sequence specifically includes: Identify the centromere position of each chromosome in the chromosome karyotype image, and determine the telomere direction of the short arm and the telomere direction of the long arm of the chromosome; Starting from the centromere position, the stripe cross-sectional profiles at each sampling point are sorted along the short arm telomere direction and the long arm telomere direction, respectively. All sorted banded cross-sectional profiles were sequentially spliced ​​together in the order from the short arm telomere through the centromere to the long arm telomere to construct the banded cross-sectional profile sequence of the chromosome.

4. The method for automatic identification of chromosome karyotype images based on artificial intelligence as described in claim 1, characterized in that, The striped cross-sectional contour sequence is input into a pre-trained recurrent neural network, and the recurrent neural network outputs the striped contrast at each cross-sectional position, specifically including: The length of each textured cross-section contour in the textured cross-section contour sequence is normalized, and each contour is converted into a one-dimensional gray-scale feature vector of fixed length, thus obtaining a feature vector sequence that corresponds one-to-one with the cross-section position. The feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-sectional position. The recurrent neural network receives the feature vector of the current cross-sectional position at each time step and updates its internal memory state. The hidden state vectors output by the recurrent neural network at each time step are extracted. All hidden state vectors are arranged in order of cross-sectional position to form a hidden state drift field that reflects the evolution of chromosome banding along the long axis. Perform a contrast response activation operation on each hidden state vector in the hidden state drift field to map the high-dimensional hidden state to a scalar form of the initial value of the stripe contrast. The initial value of the band contrast is normalized to eliminate the contrast baseline drift caused by differences in staining conditions between different chromosomes, thus obtaining the band contrast at each cross-sectional position.

5. The method for automatic identification of chromosome karyotype images based on artificial intelligence as described in claim 1, characterized in that, The specific methods for generating chromosome banding degradation curves based on banding contrast at adjacent cross-sectional locations include: Traverse all adjacent cross-section positions in ascending order of cross-section position, calculate the difference in strip contrast between each pair of adjacent cross-sections, and obtain the attenuation amplitude at each cross-section position. Arrange all attenuation amplitudes in order of cross-sectional position to construct a band contrast gradient field that reflects the rate of change of band contrast along the long axis of chromosome; Starting from the initial section position, the attenuation amplitudes in the band contrast gradient field are sequentially accumulated along the long axis of the chromosome to obtain the cumulative attenuation amplitude sequence of each section position relative to the initial section position; An initial stripe degradation curve is generated based on the cumulative attenuation magnitude sequence of each cross-section position relative to the initial cross-section position; The initial banding degradation curve is smoothed by moving average to eliminate spikes and fluctuations introduced by image noise or local irregularities in the banding, thus obtaining the final banding degradation curve of the chromosome.

6. The method for automatic chromosome karyotype image recognition based on artificial intelligence as described in claim 1, characterized in that, The attenuation coefficient of the banding contrast along the long axis of the chromosome, extracted from the banding degradation curve, specifically includes: Perform a decay curve segment truncation operation on the stripe degradation curve, extract the effective curve segment from the starting section position to the ending section position, exclude the abnormal fluctuation area at both ends of the curve caused by the stripe boundary effect, and obtain the decay curve segment to be analyzed. Attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of band contrast along the long axis of the chromosome.

7. The method for automatic identification of chromosome karyotype images based on artificial intelligence as described in claim 4, characterized in that, The feature vector sequence is sequentially input into a pre-trained recurrent neural network according to the cross-sectional position. The recurrent neural network receives the feature vector at the current cross-sectional position at each time step and updates its internal memory state, specifically including: A time step alignment mechanism is established between each feature vector in the feature vector sequence and the cross-sectional position, mapping each cross-sectional position from centromere to telomere to a time step input node of a pre-trained recurrent neural network; At each current time step, the recurrent neural network reads the historical memory state saved in the previous time step and uses the historical memory state as the contextual background information of the current time step to participate in the processing of the current feature vector; The feature vector input at the current time step is fused with the historical memory state of the previous time step by the gated fusion unit inside the recurrent neural network according to the preset gate weights to generate the candidate memory state of the current time step. The candidate memory state is written into the memory unit of the recurrent neural network to complete the progressive refresh from the memory state of the previous time step to the memory state of the current time step, and the refreshed memory state is passed to the next time step.

8. The method for automatic identification of chromosome karyotype images based on artificial intelligence as described in claim 4, characterized in that, Performing a contrast response activation operation on each hidden state vector in the hidden state drift field, mapping the high-dimensional hidden state to a scalar form of stripe contrast initial value specifically includes: According to the ascending order of the cross-sectional positions, a point-by-point traversal operation is performed on each hidden state vector in the hidden state drift field, and the hidden state vector corresponding to each cross-sectional position is taken out as the input object of the activation operation in turn. For each extracted hidden state vector, a fully connected dimensionality reduction mapping is performed, and all dimensional components of the hidden state vector are linearly aggregated according to the trained weight coefficients to obtain a single-dimensional aggregated response value. The aggregated response value is subjected to nonlinear activation compression. The dynamic range of the aggregated response value is limited to a preset stripe contrast response range by a nonlinear activation function to obtain the compressed stripe contrast activation value. The stripe contrast activation value is used as the initial stripe contrast value at the cross-sectional position to complete the mapping from the high-dimensional hidden state space to the scalar contrast space.

9. The method for automatic chromosome karyotype image recognition based on artificial intelligence as described in claim 6, characterized in that, Attenuation registration is performed on the attenuation curve segment to obtain the attenuation coefficient of band contrast along the long axis of the chromosome, specifically including: The attenuation curve segment is divided into multiple continuous sub-intervals according to the cross-sectional position, and each sub-interval covers a fixed number of continuous cross-sectional positions; For each sub-interval, the local attenuation rate of the attenuation curve segment is estimated, and the segmented attenuation rate corresponding to the sub-interval is determined based on the change in the vertical axis between the start and end of the segment. Arrange the segmented decay rates corresponding to all sub-intervals in order of cross-sectional position to construct a segmented decay rate sequence that reflects the fluctuation of decay rate along the long axis of chromosome; The segmented attenuation rate sequence is subjected to weighted fusion processing, and higher fusion weights are assigned to sub-intervals near the centromere region. The weighted fusion result is used as the attenuation coefficient of band contrast along the long axis of the chromosome.

10. An artificial intelligence-based automatic chromosome karyotype image recognition system, used to execute the artificial intelligence-based automatic chromosome karyotype image recognition method as described in any one of claims 1 to 9, characterized in that, The automatic chromosome karyotype image recognition system includes: The chromosome karyotype feature extraction module is used to acquire the chromosome karyotype image to be identified and extract multiple striped cross-sectional contours that are equally spaced along the length direction from the chromosome in the chromosome karyotype image. The chromosome banding profile recognition module is used to arrange all the banding profiles of the chromosome in order from centromere to telomere to construct a chromosome banding profile sequence. The chromosome banding degradation recognition module is used to input the banding cross-sectional contour sequence into a pre-trained recurrent neural network. The recurrent neural network outputs the banding contrast at each cross-sectional position and generates the chromosome banding degradation curve based on the banding contrast at adjacent cross-sectional positions. The chromosome banding aberration determination module is used to extract the attenuation coefficient of the banding contrast along the long axis of the chromosome from the banding degradation curve. When the attenuation coefficient is less than a preset threshold, the chromosome banding aberration is determined to be an enzymatic over-digestion type artificial aberration; otherwise, it is determined to be a real structural abnormality.

Images