Medium-term division cell optimization method and system for chromosome aberration detection
By constructing a graded dataset to train a primary screening and grading model and combining multi-dimensional morphological features and an AI confidence weighting mechanism, the problems of low screening efficiency and unstable image quality of mid-stage dividing cells were solved, achieving efficient and reliable image screening and meeting the needs of chromosome aberration analysis.
Patent Information
- Application Number
- CN202511242692.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-02
- Publication Date
- 2025-12-26
AI Technical Summary
Existing technologies for screening metaphase cells are inefficient and highly subjective. Automated methods rely on single feature evaluation, leading to unstable image quality and an inability to fully reflect key indicators such as cell morphology integrity and chromosome overlap, thus affecting the accuracy of chromosome aberration analysis.
By constructing a graded dataset to train a preliminary screening and grading model, multi-dimensional morphological features are extracted. Combined with an AI confidence weighting mechanism, a scoring sequence is generated and the target image is output, realizing multi-feature fusion evaluation and adaptive optimization.
It significantly improves the efficiency, objectivity, and stability, adaptability, and accuracy of metaphase cell screening, and provides high-quality image evidence for subsequent chromosomal aberration analysis.
Smart Images

Figure CN121213467A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of information processing, in particular, to a metaphase cell optimization method and system for chromosome aberration detection. BACKGROUND
[0002] In the field of chromosome aberration detection, the chromosome image of metaphase cell is the key basis for analyzing chromosome structural abnormalities and biological dose estimation. In the prior art, the conventional process relies on manual screening or simple automation: manual screening requires operators to select well-shaped cells from a large number of images, which has problems such as low efficiency, strong subjectivity, poor consistency of results, and is difficult to meet the detection needs of large quantities of samples; although some automatic methods introduce image recognition technology to assist in screening, they mostly rely on a single feature (such as sharpness or dispersion) to evaluate image quality, which cannot fully reflect key indicators such as cell morphological integrity and chromosome overlap, resulting in unstable image quality after screening, which is easily affected by sample differences and imaging condition fluctuations, and interferes with the accuracy of subsequent aberration analysis. Therefore, how to efficiently and objectively screen high-quality metaphase cells from a large number of images is still a technical problem to be solved in this field.
[0003] The above information disclosed in the BACKGROUND section is only for the purpose of enhancing the understanding of the background of the present application, and therefore it can include information that does not constitute the prior art known to those of ordinary skill in the art. SUMMARY
[0004] The purpose of the present application is to overcome the technical barriers in the prior art that the screening of metaphase cells for chromosome aberration detection has low efficiency, strong subjectivity, and the automatic method relies on a single feature evaluation, resulting in unstable image quality. A metaphase cell optimization method and system for chromosome aberration detection is proposed, which first constructs a hierarchical dataset to train a preliminary screening hierarchical model to realize real-time preliminary hierarchical screening of chromosome images, then extracts multi-dimensional morphological features of the selected images to construct a basic scoring function, combines an AI confidence weighting mechanism to generate a scoring sequence and outputs the target image according to the priority, overcomes the problems of low screening efficiency, poor consistency of results and evaluation dimension in the prior art, improves the stability and reliability of the screening results, and provides high-quality image basis for subsequent chromosome aberration analysis.
[0005] In the first aspect, a technical solution provided in an embodiment of the present application is: a metaphase cell optimization method for chromosome aberration detection, comprising the following steps: S1, constructing an image hierarchical dataset by hierarchically labeling the acquired chromosome images of metaphase; S2, training an artificial intelligence model through the image hierarchical dataset to obtain a preliminary screening hierarchical model; S3, screening the real-time chromosome image data through the preliminary screening grading model to obtain a selected image data set; S4, extracting morphological features of each chromosome image in the selected image data set and corresponding weight parameters to construct a basic scoring function; S5, the basic scoring function combines an AI confidence weighting mechanism to determine a scoring sequence of each level of image; the scoring sequence is arranged in descending order and the target image is output in turn.
[0006] As a preferred, the hierarchical labeling of the obtained metaphase chromosome image to construct the image hierarchical data set comprises the following steps: Extracting hierarchical feature indicators of the metaphase chromosome image and comparing and marking with the hierarchical indicator reference library to determine the hierarchical label of each chromosome image, and classifying and storing each chromosome image according to the hierarchical label to construct the image hierarchical data set.
[0007] As a preferred, the hierarchical feature indicators include but are not limited to chromosome distribution uniformity, focusing quality, overlapping degree, background interference and morphological definition.
[0008] As a preferred, the training of the artificial intelligence model through the image hierarchical data set to obtain the preliminary screening grading model comprises the following steps: S21, extracting morphological features of the chromosome image from the image hierarchical data set, the morphological features include but are not limited to contour number, skeleton length, intersection number, distribution uniformity, roundness index and background brightness feature; normalizing the extracted morphological features to form an input feature vector; S22, configuring the network structure of the artificial intelligence model, taking the normalized input feature vector as the input and the hierarchical label of the corresponding image as the output, and training the artificial intelligence model by using the cross-validation method; S23, optimizing the classification accuracy by adjusting the model hyperparameters, evaluating the model performance using the validation set, and forming a preliminary screening grading model for quality level classification of the chromosome image.
[0009] As a preferred, the artificial intelligence module includes but is not limited to any one of a convolutional neural network model, a support vector machine model and a random forest model.
[0010] As a preferred, the screening of the real-time chromosome image data through the preliminary screening grading model to obtain the selected image data set comprises the following steps: Performing noise reduction, contrast enhancement and background segmentation processing on the real-time chromosome image data to determine the chromosome region; extracting morphological features of the chromosome region images, the morphological features including but not limited to contour number, skeleton length, intersection number, distribution uniformity, roundness index and background brightness feature, mapping the extracted features into a feature vector consistent with the training stage; inputting the feature vector into the preliminary screening grading model to obtain a predicted result of image quality grade of the model output; setting a quality grade threshold according to a grading index reference library, screening out images with a predicted grade higher than or equal to the threshold to form a selected image data set.
[0011] As preferred, morphological features of each chromosome image in the selected image data set are extracted and combined with corresponding weight parameters to construct a basic score function, including the following steps: S41, extracting morphological features of each chromosome image, the morphological features including but not limited to contour number, skeleton length, intersection number, distribution uniformity, roundness index and background brightness feature; S42, normalizing each morphological feature and obtaining a weight parameter of each morphological feature through analytic hierarchy process; S43, obtaining the basic score function through weighted summation of the morphological features and corresponding weight parameters.
[0012] As preferred, the basic score function is combined with an AI confidence weighting mechanism to determine a score sequence of each level of image; the score sequence is arranged in descending order and the target image is output in sequence, including the following steps: S51, obtaining a classification confidence corresponding to each grading label, dynamically adjusting the weight parameter of the basic score function according to the classification confidence, and constructing an adaptive weight vector; S52, calculating the adaptive weight vector and the normalized morphological feature vector to obtain a final score of the fusion AI confidence; S53, performing S51-S52 in turn to obtain a final score of each level of image to construct a score sequence; S54, arranging the images in descending order to generate a priority queue; outputting the images at the front end of each image queue in sequence as target images.
[0013] As preferred, the process of obtaining the classification confidence is as follows: if the model is a support vector machine model, the output value of the decision function is extracted and mapped into a score in the interval [0, 1] through a Sigmoid function as the classification confidence; if the model is a convolutional neural network model, the class probability value output by the Softmax layer is obtained as the classification confidence.
[0014] In a second aspect, the application also provides a technical solution of a metaphase cell optimization system, which is suitable for the metaphase cell optimization method for chromosome aberration detection. The data acquisition module: hierarchical labeling is performed on the acquired metaphase chromosome image to construct an image hierarchical dataset. The model construction module: an artificial intelligence model is trained by using the image hierarchical dataset to obtain a preliminary screening hierarchical model. The data screening module: the real-time chromosome image data are screened by using the preliminary screening hierarchical model to obtain a candidate image dataset. The data operation module: morphological features of each chromosome image in the candidate image dataset are extracted, and corresponding weight parameters are used to construct a basic scoring function. The data output module: the basic scoring function is combined with an AI confidence weighting mechanism to determine a scoring sequence of each level of image; the scoring sequence is arranged in descending order, and target images are output in sequence.
[0015] The application has the following beneficial effects: (1) In view of the problems of low efficiency, strong subjectivity and poor consistency of results in the prior art of manual screening of metaphase cells, the application proposes a scheme of training a preliminary screening hierarchical model based on a hierarchical labeling dataset and combining multi-feature scoring. The four-level hierarchical dataset is constructed by professional personnel according to multiple indexes such as chromosome distribution uniformity, so that the artificial intelligence model learns objective hierarchical standards to realize real-time image rapid preliminary screening. Multi-dimensional morphological features such as contour number and skeleton length are extracted to construct a basic scoring function, and the weight is dynamically adjusted in combination with AI confidence to perform comprehensive sorting. This process converts artificial experience into quantifiable feature indexes and model parameters. Through machine self-learning and multi-feature fusion evaluation, the traditional artificial subjective judgment is replaced, and the screening efficiency and objectivity and consistency of the results are significantly improved. (2) In view of the problem that the prior art of automatic screening method relies on single feature evaluation of image quality, and cannot comprehensively reflect key indexes such as cell morphological integrity and chromosome overlapping degree, resulting in unstable screening quality, the application proposes a multi-dimensional morphological feature fusion evaluation scheme. Multiple complementary features such as contour number (reflecting fragmentation degree), skeleton intersection point number (reflecting overlapping degree) and distribution uniformity (reflecting spatial dispersion) are extracted, and a basic scoring function is constructed by using the analytic hierarchy process to give each feature a reasonable weight. The quality of the metaphase cell image is comprehensively quantitatively evaluated, the one-sidedness of single feature evaluation is overcome, the screened image is more in line with the requirements of “morphological integrity, moderate dispersion and less overlapping” for subsequent aberration analysis, and the stability and reliability of the screening result are significantly improved. (3) In view of the problem that the existing screening method is susceptible to sample differences and imaging condition fluctuations and has poor adaptability to images of different quality levels, the application proposes a dynamic evaluation scheme fusing AI confidence weighted, the classification confidence of the image (such as the Sigmoid mapping score of SVM or the Softmax probability of CNN) of the preliminary screening grading model is obtained, the feature weights of the basic scoring function are dynamically adjusted, the key feature weights of high-quality level images are enhanced, and a penalty coefficient is introduced for low-quality level images, so that the final score not only retains the comprehensive evaluation advantage of multiple morphological features, but also can be adaptively optimized in combination with the confidence judgment of the image quality by the model, effectively offsetting the interference of sample differences and imaging condition fluctuations, and significantly improving the adaptability of the screening method to complex scenes and the accuracy of the screening result.
[0016] The above summary of the application is only a summary of the technical solutions of the application. In order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the specific embodiments of the application are described below. BRIEF DESCRIPTION OF DRAWINGS
[0017] Other features, objects and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings. The drawings are for the purpose of illustrating preferred embodiments only and are not to be construed as limiting the application. Moreover, in the accompanying drawings, like reference numerals designate like parts throughout the several views.
[0018] Figure 1 A preferred method flowchart for metaphase cells for chromosome aberration detection of the application.
[0019] Figure 2 A visualization diagram of pre-sorted metaphase cells of the application.
[0020] Figure 3 A visualization diagram of post-sorted metaphase cells of the application.
[0021] Figure 4 A preferred system framework diagram of metaphase cells of the application. DETAILED DESCRIPTION
[0022] In order to make the purposes, technical solutions and advantages of the application more clear and obvious, the application is further described in detail below with reference to the drawings and examples. It should be understood that the specific embodiments described herein are only the best mode of the application, which are used to explain the application and do not limit the protection scope of the application. All other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.
[0023] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.
[0024] Example: Figure 1 As shown, a preferred method for metaphase cells used in chromosome aberration detection includes the following steps: S1. Perform hierarchical annotation on the obtained chromosome images of metaphase division to construct an image hierarchical dataset; S2. Train the artificial intelligence model using the image grading dataset to obtain the initial screening and grading model; S3. The real-time chromosome image data is graded and screened using a preliminary screening and grading model to obtain a candidate image dataset. S4. Extract the morphological features and corresponding weight parameters of each chromosome image in the candidate image dataset to construct a basic scoring function; S5. The basic scoring function, combined with the AI confidence weighting mechanism, determines the scoring sequence of images at each level; the scoring sequence is arranged in descending order and the target images are output sequentially.
[0025] Understandably, this embodiment employs a technique of first hierarchically labeling images of chromosomes undergoing metaphase division to construct an image hierarchical dataset, then training an artificial intelligence model using this dataset to obtain a preliminary screening and hierarchical model. This model is then used to hierarchically screen real-time chromosome image data to obtain a candidate image dataset. Next, morphological features and corresponding weight parameters of the candidate images are extracted to construct a basic scoring function. Finally, the basic scoring function is combined with an AI confidence weighting mechanism to determine the scoring sequence and output the target images in descending order. This technique achieves the transformation of human experience into quantifiable model parameters and feature indicators, leveraging the autonomous learning ability and multi-feature fusion evaluation of the artificial intelligence model to replace the subjective judgment of traditional manual screening, overcoming the one-sidedness of single feature evaluation. At the same time, the weights are dynamically adjusted through AI confidence to adapt to images of different quality levels, thereby significantly improving the efficiency, objectivity, consistency, and result stability and reliability of metaphase cell screening, providing high-quality image evidence for subsequent chromosome aberration analysis.
[0026] As a preferred embodiment, the step of hierarchically labeling the acquired metaphase chromosome images to construct an image hierarchical dataset includes the following steps: The grading feature indexes of the metaphase chromosome images are extracted, and the grading feature indexes are compared with a grading index reference library to mark and determine the grading labels of the chromosome images, and the chromosome images are classified and stored according to the grading labels to construct an image grading data set.
[0027] It can be understood that, under the standard lymphocyte culture and processing flow, a large number of cells in the metaphase cell division state are induced, and a high-throughput digital scanning platform is used for chromosome image acquisition. In the acquisition process, a 10x objective is used for low-magnification rapid scanning to obtain wide-field images for preliminary screening; a manual expert with experience in chromosome analysis is invited to mark the 10x images, which are divided into four grades: grade 1 (high-quality images), grade 2, grade 3, and grade 4 (quality decreases in turn). In the marking process, factors such as chromosome distribution, focusing quality, overlapping degree, and background interference are considered. The constructed marking image set is used to train a chromosome screening grading model to realize the initial screening marking classification under a 10x field of view. In this embodiment, the grading feature indexes (including chromosome distribution uniformity, focusing quality, etc.) of the metaphase chromosome images are extracted, compared with the grading index reference library to determine the grading labels of the chromosome images, and then the images are classified and stored according to the grading labels to construct an image grading data set. The technical means realizes the conversion of the professional's evaluation standard of the chromosome image quality into a quantifiable and comparable feature index system, provides standardized and normalized sample data for subsequent training of the initial screening grading model, ensures that the basic data for model learning has a unified quality evaluation benchmark, and thus lays a data foundation for the model to accurately learn the image quality grade division rule.
[0028] As a preferred embodiment, the grading feature indexes include but are not limited to chromosome distribution uniformity, focusing quality, overlapping degree, background interference, and morphological clarity.
[0029] It can be understood that, in this embodiment, the chromosome distribution uniformity, focusing quality, overlapping degree, background interference, and morphological clarity are used as grading feature indexes, and the technical means of comprehensive evaluation of the metaphase chromosome image quality in multiple dimensions is realized to comprehensively depict the image quality, avoid the limitations of single feature evaluation, make the grading labels accurately reflect the differences in key dimensions such as distribution rationality, imaging clarity, and interference degree, provide a scientific feature basis for constructing a standardized image grading data set, and thus ensure that the subsequent initial screening grading model can learn a comprehensive and accurate quality grading rule.
[0030] As a preferred embodiment, the training of the artificial intelligence model through the image grading data set to obtain the initial screening grading model includes the following steps: S21, morphological features of the chromosome image are extracted from the image grading data set, and the morphological features include but are not limited to contour number, skeleton length, intersection point number, distribution uniformity, roundness index and background brightness feature; the extracted morphological features are normalized to form an input feature vector; S22, the network structure of the artificial intelligence model is configured, the normalized input feature vector is taken as the input, the grading label corresponding to the image is taken as the output, and the artificial intelligence model is trained by using the cross-validation method; S23, the classification accuracy is optimized by adjusting the model hyperparameters, the model performance is evaluated by using the verification set, and the preliminary screening grading model for classifying the quality level of the chromosome image is formed.
[0031] It can be understood that the embodiment adopts the technical means of extracting multi-dimensional morphological features such as contour number and skeleton length from the image grading data set, performing normalization processing to form an input feature vector, configuring the network structure of the artificial intelligence model, taking the normalized feature vector and the corresponding grading label as input and output, training the model by using the cross-validation method, adjusting the hyperparameters to optimize the classification accuracy, and evaluating the verification set to form the preliminary screening grading model, realizes the conversion of the visual features of the image into quantized feature vectors that can be learned by the model, makes the model accurately learn the mapping relationship between the grading features and the quality level through a standardized training process, improves the accuracy and generalization ability of the model for classifying the quality level of the chromosome image, and provides reliable model support for the efficient grading and screening of subsequent real-time images.
[0032] The interpretation of each morphological feature is as follows: a) Contour number N_contour: The number of effective connected domains in the center region of the image is counted to judge the integrity of the number of chromosomes; b) Skeleton total length L_skeleton: The skeleton algorithm is used to extract the center axis of the chromosome, and the sum of the skeleton structures with a length greater than 30 pixels is calculated to exclude the interference of artifacts; c) Skeleton intersection point number N_cross: A fixed sliding window combined with a multi-template matching technique is used to detect the intersection structure, the template ignores the flag bit 2, and only matches the structures with values of 0 and 1 to evaluate the degree of chromosome stacking; d) Spatial distribution index D_spatial: The Euclidean distance mean and standard deviation of the centers of gravity of all chromosomes to the center of the image are calculated, and the density of the distribution of the chromosomes in the divided regions of the image is combined to evaluate the spatial uniformity of the chromosomes in the image; e) Roundness index C_shape: The equivalent roundness index of each connected region is counted to identify circular or elliptical artifacts; f) Background brightness average L bg: By evaluating the average brightness of the background area, the uniformity of the image background is judged to avoid the interference of bright background or stray light on the recognition result.
[0033] As a preferred embodiment, the artificial intelligence module includes but is not limited to any one of a convolutional neural network model, a support vector machine model, and a random forest model.
[0034] It can be understood that the embodiment adopts the technical means of taking the convolutional neural network model, the support vector machine model, and the random forest model as the preliminary screening grading model, realizes the adaptation of different types of artificial intelligence models to the chromosome image grading requirements, the convolutional neural network can deeply extract the deep features of the image, the support vector machine is good at small sample high-dimensional feature classification, and the random forest can process nonlinear feature relationship, thereby providing diversified technical paths for image grading, ensuring stable quality level classification in different sample size and feature complexity scenarios, and improving the adaptability and accuracy of the preliminary screening grading model for chromosome image quality evaluation.
[0035] As a preferred embodiment, the preliminary screening grading model includes the following steps: The real-time chromosome image data is subjected to noise reduction, contrast enhancement, and background segmentation processing to determine the chromosome region; The morphological features of the chromosome region image are extracted, the morphological features include but are not limited to contour number, skeleton length, intersection point number, distribution uniformity, roundness index, and background brightness feature, and the extracted features are mapped into feature vectors consistent with the training stage; The feature vectors are input into the preliminary screening grading model to obtain the image quality level prediction result output by the model; the quality level threshold is set according to the grading index reference library, and the images with a prediction level higher than or equal to the threshold are screened out to form the selected image data set.
[0036] It can be understood that the embodiment adopts the technical means of taking the convolutional neural network model, the support vector machine model, and the random forest model as the preliminary screening grading model, realizes the adaptation of different models to the chromosome image grading scene, the convolutional neural network can capture the deep morphological features of the chromosomes in the image through the multi-layer network structure, the support vector machine can effectively process high-dimensional morphological features and realize accurate classification under small sample conditions, and the random forest can process the nonlinear relationship between features through multi-decision tree ensemble learning, thereby providing a flexible and optional technical solution for image grading, ensuring that accurate quality level classification results can be stably output in different scenarios of sample size and feature complexity, and improving the universality and reliability of the preliminary screening grading model for chromosome image quality evaluation.
[0037] As a preferred embodiment, the morphological features of each chromosome image in the selected image data set are extracted to construct a basic score function with corresponding weight parameters, including the following steps: S41, extract the morphological features of each chromosome image, including but not limited to the number of contours, skeleton length, number of intersection points, distribution uniformity, roundness index and background brightness characteristics; S42, normalize each morphological feature, and obtain the weight parameter of each morphological feature by analytic hierarchy process; S43, weighted sum of morphological features and corresponding weight parameters to obtain the basic score function.
[0038] It can be understood that the embodiment adopts the technical means of extracting the number of contours, skeleton length and other multi-dimensional morphological features of chromosomes in the selected image, normalizing these features to eliminate dimensional differences, determining the weight parameters of each feature by analytic hierarchy process, and then weighting and summing the morphological features and corresponding weight parameters to construct the basic score function, which realizes the conversion of the quality characteristics of chromosome images into quantifiable numerical evaluation, gives different features reasonable weight to reflect their influence on image quality, and makes the basic score function fully and objectively reflect the comprehensive quality of chromosome images, lays a quantitative foundation for subsequent precise sorting combined with AI confidence weighting mechanism, and improves the scientificity and accuracy of image quality evaluation.
[0039] Exemplarily, taking the steps of constructing the basic score function as an example, the technician first extracts the morphological features of the contours, skeleton length and intersection points of the chromosomes in the selected image, such as counting the number of independent connected domains of the chromosomes by edge detection algorithm, extracting the skeleton axis of the chromosomes by Zhang-Suen thinning algorithm and calculating the total length, detecting the skeleton intersection structure by 5x5 neighborhood template matching; then normalizing these features, such as mapping the skeleton length to the interval [0, 1] (the longer the length, the higher the normalized value), and mapping the number of intersection points to the interval [0, 1] (the fewer the number, the higher the normalized value); then by analytic hierarchy process, invite chromosome analysis experts to score the importance of each feature, and determine the weight parameters of skeleton length as 0.3 and the weight parameters of intersection point number as 0.25 after consistency test; finally, construct the basic score function: Score_base=0.3x normalized skeleton length + 0.25x(1-normalized intersection point number) + other feature weighting value, realize the quantitative evaluation of image quality.
[0040] The total score of the image is calculated by using the following score function to measure the image quality: Further, the formula of the basic score function is: ; wherein R(·) represents a normalization function, are the weighting coefficients corresponding to each morphological feature, which are determined according to actual task requirements and experience parameters.
[0041] As a preferred embodiment, the basic scoring function combines an AI confidence weighting mechanism to determine the scoring sequence of each level of image; the scoring sequence is arranged in descending order and the target image is output in turn, including the following steps: S51, acquiring the classification confidence corresponding to each hierarchical label, dynamically adjusting the weight parameters of the basic scoring function according to the classification confidence, and constructing an adaptive weight vector; S52, calculating the final score of the fusion AI confidence by using the adaptive weight vector and the normalized morphological feature vector; S53, performing S51-S52 in rounds to obtain the final score of each level of image and construct a scoring sequence; S54, arranging the images in descending order to generate a priority queue; outputting the images at the front end of each level of image queue in turn as the target image.
[0042] It can be understood that the embodiment adopts the technical means of first acquiring the classification confidence corresponding to each hierarchical label, dynamically adjusting the weight parameters of the basic scoring function according to the confidence to construct an adaptive weight vector, then calculating the final score of the fusion AI confidence by using the adaptive weight vector and the normalized morphological feature vector, performing the above steps in rounds to obtain the final score of each level of image and construct a scoring sequence, and finally arranging the images in descending order to generate a priority queue and outputting the images at the front end of each level of image queue in turn as the target image, which realizes dynamic optimization of feature weights through AI confidence, makes the score retain the comprehensive evaluation advantage of multiple morphological features, and can also adaptively adjust the confidence judgment of the model on the image quality, ensures that high-quality images are more prominent in the sorting, and at the same time, the reasonable screening of images at each level is taken into account, thereby improving the accuracy and pertinence of the output of the target image and better meeting the needs of subsequent chromosome aberration analysis for high-quality images.
[0043] Further, dynamically adjusting the weight parameters of the basic scoring function according to the classification confidence to construct an adaptive weight vector includes the following steps: a) for high-quality level images (such as images predicted to be level 1), increasing the morphological feature weight (such as w1-w6), and introducing a confidence adjustment coefficient a1 (a1>1) to enhance the influence of AI judgment; b) for low-quality level images (such as images predicted to be level 3), reducing the morphological feature weight, and introducing a confidence penalty coefficient a2 (0 c) the weight adjustment formula is: w i' = w i × (1 + a i × Confidence), where w_i is the original weight corresponding to each morphological feature, i e (1, 6), a i is the adjustment coefficient of the corresponding level image, and Confidence is the confidence value.
[0044] The adjusted weight vector is combined with the normalized morphological feature vector to calculate the final score of the fusion AI confidence: Score_final = Σ(w i' × R i ) × b × Confidence; where R i is the normalized morphological feature value, and b is the confidence direct influence coefficient, calibrated by experimental data (such as b = 0.2).
[0045] Illustratively, the skilled person first obtains the classification confidence from the preliminary screening grading model: if the model is a convolutional neural network, the class probability output by the Softmax layer is directly extracted (such as the confidence of a certain 1-level image is 0.92, and the confidence of a certain 3-level image is 0.65); then the weight parameters of the basic score function are dynamically adjusted according to the confidence, for 1-level high-confidence images, the skeleton length weight is increased from 0.3 to 0.35 (w_i' = 0.3 x (1 + 0.5 x 0.92)), for 3-level low-confidence images, the cross point number weight is reduced from 0.25 to 0.2 (w_i' = 0.25 x (1 + 0.3 x 0.65)), and an adaptive weight vector is constructed; then the final score is calculated, such as the Score_final of the 1-level image = 0.35 x 0.9 (normalized skeleton length) + 0.2 x 0.8 (normalized distribution uniformity) + 0.2 x 0.92 (confidence) = 0.315 + 0.16 + 0.184 = 0.659; after calculating the final scores of all selected images in the round, the priority queue is generated in descending order, and the top N images with the highest scores in the queue are output as target images for subsequent chromosome aberration analysis, such as Figure 2 and Figure 3 are the visualized images before and after preliminary screening and sorting of the staining images by the embodiment.
[0046] As a preferred embodiment, the process of obtaining the classification confidence is: If the model is a support vector machine model, the output value of the decision function is extracted and mapped to the score in the interval [0, 1] by the Sigmoid function as the classification confidence; If the model is a convolutional neural network model, the class probability value of the Softmax layer output is obtained as the classification confidence.
[0047] In the embodiment of the present application, a technical solution is also provided, which is a metaphase cell optimization system, which is applicable to the metaphase cell optimization method for chromosome aberration detection. Figure 4 As shown in the figure, the system comprises: The data acquisition module is configured to perform hierarchical labeling on the metaphase chromosome images to construct an image hierarchical dataset. The model construction module is configured to train an artificial intelligence model by using the image hierarchical dataset to obtain a preliminary screening hierarchical model. The data screening module is configured to perform hierarchical screening on real-time chromosome image data by using the preliminary screening hierarchical model to obtain a candidate image dataset. The data operation module is configured to extract morphological features of each chromosome image in the candidate image dataset and corresponding weight parameters to construct a basic scoring function. The data output module is configured to determine a scoring sequence of each level of image by using the basic scoring function in combination with an AI confidence weighting mechanism, and to output target images in descending order of the scoring sequence.
[0048] It can be understood that the embodiment uses the data acquisition module to perform hierarchical labeling on the metaphase chromosome images to construct an image hierarchical dataset, the model construction module uses the dataset to train an artificial intelligence model to obtain a preliminary screening hierarchical model, the data screening module performs hierarchical screening on real-time images by using the preliminary screening hierarchical model to obtain a candidate image dataset, the data operation module extracts morphological features of the candidate images and corresponding weight parameters to construct a basic scoring function, and the data output module determines a scoring sequence by using the basic scoring function in combination with an AI confidence weighting mechanism and outputs target images in descending order. The technical means realizes the cooperation of each module, converts the whole process of metaphase cell optimization into a systematic and automated processing process, ensures efficient linkage from data preparation to target output through the technical connection between modules, and thus stably and reliably completes the optimization of metaphase cells and provides high-quality image support for chromosome aberration detection.
[0049] The above-described specific embodiments are preferred embodiments of the present application, and the specific implementation range of the present application is not limited thereto. The scope of the present application includes but is not limited to the specific embodiments, and equivalent changes made in accordance with the shape and structure of the present application are within the scope of protection of the present application.
Claims
1. A method for metaphase split cell for chromosome aberration detection, preferably characterized in that, The method comprises the following steps: S1, grading labeling of the obtained metaphase split chromosome images to construct an image grading dataset; S2, training an artificial intelligence model through the image grading dataset to obtain a preliminary screening grading model; S3, grading screening of real-time chromosome image data through the preliminary screening grading model to obtain a selected image dataset; S4, extracting morphological features of each chromosome image in the selected image dataset and corresponding weight parameters to construct a basic scoring function; S5, the basic scoring function combines with an AI confidence weighting mechanism to determine a scoring sequence of each level of image; the scoring sequence is arranged in descending order and the target image is output in sequence.
2. A method for metaphase split cell analysis for chromosome aberration detection according to claim 1, characterized in that, The method of grading labeling of the obtained metaphase split chromosome images to construct an image grading dataset comprises the following steps: Extracting grading feature indexes of the metaphase split chromosome images and comparing and marking with a grading index reference library to determine grading labels of each chromosome image, and storing each chromosome image according to the grading labels to construct an image grading dataset.
3. The method according to claim 2, wherein the grading feature indexes include but are not limited to chromosome distribution uniformity, focusing quality, overlapping degree, background interference and morphological definition. The method of training an artificial intelligence model through the image grading dataset to obtain a preliminary screening grading model comprises the following steps:
4. A method for metaphase split cell analysis for chromosome aberration detection according to claim 1, characterized in that, S21, extracting morphological features of the chromosome images from the image grading dataset, wherein the morphological features include but are not limited to contour number, skeleton length, intersection point number, distribution uniformity, roundness index and background brightness feature; normalizing the extracted morphological features to form an input feature vector; S22, configuring a network structure of the artificial intelligence model, taking the normalized input feature vector as input and the grading label of the corresponding image as output, and training the artificial intelligence model by using a cross-validation method; S23, optimizing the classification accuracy by adjusting model hyperparameters, evaluating the model performance by using a validation set, and forming a preliminary screening grading model for quality level classification of the chromosome images.
5. The method according to claim 4, wherein the artificial intelligence model includes but is not limited to any one of a convolutional neural network model, a support vector machine model and a random forest model.
6. The method according to claim 2, wherein the method of grading screening of real-time chromosome image data through the preliminary screening grading model to obtain a selected image dataset comprises the following steps: performing noise reduction, contrast enhancement and background segmentation processing on the real-time chromosome image data to determine a chromosome region; extracting morphological features of the chromosome region image, wherein the morphological features include but are not limited to contour number, skeleton length, intersection point number, distribution uniformity, roundness index and background brightness feature, and mapping the extracted features to a feature vector consistent with the training stage; The feature vector is input into the preliminary screening classification model to obtain a model output image quality level prediction result; a quality level threshold is set according to a classification index reference library, and images with a predicted level higher than or equal to the threshold are screened out to form a selected image data set.
7. A method for metaphase split cell analysis for chromosome aberration detection according to claim 1 or 6, characterized in that, The morphological features of each chromosome image in the selected image data set are extracted, and corresponding weight parameters are used to construct a basic scoring function, including the following steps: S41, morphological features of each chromosome image are extracted, including but not limited to contour number, skeleton length, intersection number, distribution uniformity, circularity index and background brightness feature; S42, each morphological feature is normalized, and the weight parameters of each morphological feature are obtained by the analytic hierarchy process; S43, the basic scoring function is obtained by weighted summation of the morphological features and the corresponding weight parameters.
8. The preferred method for metaphase cell detection of chromosome aberration according to claim 1, characterized in that, The basic scoring function determines the scoring sequence of each level image by combining the AI confidence weighting mechanism; the scoring sequence is arranged in descending order, and the target image is output in turn, including the following steps: S51, the classification confidence corresponding to each classification label is obtained, the weight parameters of the basic scoring function are dynamically adjusted according to the classification confidence, and an adaptive weight vector is constructed; S52, the adaptive weight vector and the normalized morphological feature vector are calculated to obtain the final score of the fusion AI confidence; S53, S51~S52 are executed in turn to obtain the final score of each level image to construct a scoring sequence; S54, the images are arranged in descending order to generate a priority queue; the images at the front end of each image queue are output in turn as target images.
9. The preferred method for metaphase cell detection of chromosome aberration according to claim 8, characterized in that, The process of obtaining the classification confidence is as follows: If the model is a support vector machine model, the output value of the decision function is extracted and mapped to the score in the interval [0, 1] by the Sigmoid function as the classification confidence; If the model is a convolutional neural network model, the class probability value output by the Softmax layer is obtained as the classification confidence.
10. A metaphase cell selection system suitable for use in a metaphase cell selection method for detecting chromosomal aberrations according to any one of claims 1 to 9, characterized in that It includes: Data acquisition module: classifying and annotating the obtained metaphase chromosome images to construct an image classification data set; Model construction module: training an artificial intelligence model by using the image classification data set to obtain a preliminary screening classification model; Data screening module: screening real-time chromosome image data by using the preliminary screening classification model to obtain a selected image data set; Data operation module: extracting morphological features of each chromosome image in the selected image data set and corresponding weight parameters to construct a basic scoring function; Data output module: the basic scoring function determines the scoring sequence of each level image by combining the AI confidence weighting mechanism; the scoring sequence is arranged in descending order, and the target image is output in turn.
Citation Information
Patent Citations
Chromosome division phase image quality evaluation method based on deep learning
CN111325711A
Chromosome metaphase image scoring method and device, electronic equipment and storage medium
CN112785566A
Chromosome division phase positioning and sorting method based on multi-scale feature fusion
CN113807259A
Chromosome image quality evaluation method and device, computer equipment and storage medium
CN117437633A
Centromere detector and method for determining radiation exposure from chromosome abnormalities
WO2012061669A2
