A method and system for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication.
By employing capability gating and consistency adjudication of multiple anomaly detection methods, the detection blind spots and false alarms in chromosome karyotype analysis have been resolved, achieving efficient and accurate chromosome anomaly detection.
Patent Information
- Application Number
- CN202611286681.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies in chromosome karyotype analysis suffer from several problems, including blind spots in single-image representation detection, false alarms and errors introduced by the direct fusion of different anomaly detection methods, and a lack of global verification constraints based on karyotype biology.
We employ multiple anomaly detection methods with different representational invariances, combine capability gating and consistency adjudication, generate capability weights based on image quality level and anomaly type, perform gating fusion, and utilize karyotype biological constraints for global verification.
It improves the coverage and accuracy of chromosomal abnormality detection, reduces false alarm rates and the burden of manual review, and improves processing efficiency.
Smart Images

Figure CN123118504A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing and computer-aided cytogenetics analysis technology, specifically, it relates to a method and system for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication. Background Technology
[0002] G-banded chromosome karyotype analysis is an important method for chromosome abnormality analysis. Traditional procedures typically rely on professionals to examine metaphase chromosome images one by one, performing chromosome segmentation, classification, pairing, and abnormality identification. This is labor-intensive, and the interpretation results are easily influenced by human experience. With the development of medical image processing and artificial intelligence technologies, existing solutions can automatically perform chromosome segmentation, classification, or abnormality marking, but the following problems still exist.
[0003] First, single image representations have structural detection blind spots. Central axis strip representations can mitigate the effects of curvature, but they easily lose overall information such as ring morphology and bicentromeres; overall contour representations are good for identifying length and arm-to-arm ratio anomalies, but are not sensitive to fine-grained strip rearrangements; strip coding representations are robust to scale and position changes, but weaken absolute length information. Therefore, single representation detection cannot cover all observable anomaly types.
[0004] Secondly, fusion methods such as voting, union, or fixed weights do not fully consider the differences in capabilities among various detection methods. The effective detection capabilities of different anomaly detection methods vary with the type of anomaly, chromosome category, and image quality. Direct merging may introduce unreliable claims outside the scope of their capabilities, increasing false alarms and the burden of manual review.
[0005] Furthermore, existing anomaly detection results lack global verification based on karyotype biology constraints. For example, chromosome deletions are often accompanied by changes in homologous chromosome length, banding patterns, or material content; reciprocal translocations may manifest as associated breakpoints and fragment complementarity. If corresponding verifiable inferences are not generated based on the anomaly type and global verification is not performed using multiple chromosome instances, it becomes difficult to distinguish between genuine anomalies and false positives caused by segmentation errors or image artifacts.
[0006] The patent document "An Automated Chromosome Karyotype Analysis and Abnormality Detection Method" (CN112288706A) discloses a technical solution that combines attention mechanism with convolutional neural network to achieve chromosome segmentation, category recognition and abnormality labeling through detection localization, target segmentation and multi-task learning. However, this solution mainly learns chromosome category and abnormality category simultaneously by the same model. It does not establish conditional capability gating relationship for multiple abnormality detection methods with different representation invariances, nor does it generate karyotype-level verifiable inferences based on abnormality type.
[0007] The patent document "An Automatic Chromosome Segmentation and Classification Method Based on Deep Learning" (CN113658150A) discloses a technical solution that uses Attention U-Net to filter impurities and segment chromosomes, identifies overlapping or adhered chromosomes through voting using support vector machines, random forests and logistic regression models, and outputs chromosome categories and karyotype analysis maps through the stacking of multiple residual networks. However, its model fusion mainly serves the identification of overlap or adhesion and the determination of chromosome categories. It does not determine the capability weights of each anomaly detection method based on the anomaly type, chromosome category and image quality, nor does it perform karyotype biological consistency verification on anomaly candidates.
[0008] The patent document "Method, Apparatus, Device and Computer-Readable Storage Medium for Chromosomal Abnormality Detection" (CN112037174A) discloses a technical solution that inputs chromosome segmentation maps into multiple karyotype discrimination models, forms karyotype pairing results based on all discrimination values of each map set, and generates case detection results based on the number of elements corresponding to each karyotype in multiple map sets and the number threshold. However, its multiple discrimination models are used to determine the discrimination values of the corresponding karyotypes, and it does not perform conditional capability weighting on abnormality detection methods with different characterization invariances based on abnormality type, chromosome category and image quality, nor does it generate multiple types of karyotype biological inferences and calculate consistency scores for structural abnormality declarations.
[0009] Therefore, there is a need for a chromosome karyotype abnormality detection scheme that can utilize the complementary relationships between multiple image representations, selectively accept the outputs of each abnormality detection method according to the specific detection scenario, and then verify the abnormality candidates through karyotype-level biological relationships. Summary of the Invention
[0010] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication.
[0011] A method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication according to the present invention includes: Step S1: Obtain images of chromosomes in metaphase, segment them to obtain multiple chromosome instances, and determine the chromosome category and image quality level of each chromosome instance; Step S2: For the same chromosome instance, at least two anomaly detection methods based on different chromosome image information and with different characterization invariances are used to detect anomalies, and anomaly declarations and detection confidence are obtained. Step S3: Based on the correspondence between anomaly type, chromosome category, image quality level and the detection capabilities of each anomaly detection method, determine the capability weight corresponding to each anomaly declaration, and perform gated fusion on the anomaly declarations based on the capability weight and detection confidence to obtain an anomaly candidate set; Step S4: In the anomaly declaration of the anomaly candidate set, generate a verifiable inference set based on the karyotype biological constraints corresponding to the anomaly type, and use multiple chromosome instances to perform global verification of the verifiable inference set to obtain a consistency score. Step S5: Determine the grading results of candidate abnormal statements based on the gating fusion results and consistency scores, and output chromosomal abnormality triage information.
[0012] Preferably, in step S1, the image processing unit acquires a metaphase chromosome image, segments it to obtain multiple chromosome instances, and determines the chromosome category and image quality level of each chromosome instance; Step S2: For the same chromosome instance, the characterization divergent detection unit uses at least two anomaly detection methods based on different chromosome image information and with different characterization invariances to perform anomaly detection, and obtains anomaly declaration and detection confidence. Step S3: The processor determines the capability weight corresponding to each anomaly declaration based on the correspondence between the anomaly type, chromosome category, image quality level and the detection capability of each anomaly detection method. Based on the capability weight and detection confidence, the anomaly declarations are gated and fused to obtain an anomaly candidate set. Step S4: The processor generates a verifiable inference set based on the karyotype biological constraints corresponding to the anomaly type in the anomaly declaration in the anomaly candidate set. The verifiable inference set is then globally verified using multiple chromosome instances to obtain a consistency score. Step S5: The hierarchical output unit determines the hierarchical results of the candidate abnormality declarations based on the gating fusion results and the consistency score, and outputs the chromosomal abnormality triage information.
[0013] Preferably, the anomaly detection method includes: Anomaly detection is performed based on the central axis band sequence of chromosome instances; anomaly detection is performed based on at least one of the morphological information of the overall outline, relative length, centromere index, arm ratio, and mean curvature of the central axis of chromosome instances; anomaly detection is performed based on the serialization encoding of the light and dark bands of chromosome instances; and anomaly detection is performed based on the pairing difference characterization after homologous chromosome alignment.
[0014] In step S1, the image processing unit acquires a metaphase chromosome image, segments it to obtain multiple chromosome instances, and determines the chromosome category and image quality level of each chromosome instance.
[0015] In step S2, for the same chromosome instance, the characterization divergent detection unit selects at least two abnormality detection methods to retain morphological information, band information or homologous chromosome difference information, so that the corresponding missed detection ranges do not overlap at least partially.
[0016] Preferably, in step S3, the processor obtains a capability matrix containing capability weights based on the correspondence between anomaly type, chromosome category, image quality level, and the detection capabilities of each anomaly detection method, including: Obtain a labeled sample set with manual verification results, and stratify it according to anomaly type, chromosome category, and image quality level.
[0017] The sensitivity, false alarm rate, and sample support of each anomaly detection method under each stratification condition were statistically analyzed, and the raw capability value was calculated based on the sensitivity and false alarm rate.
[0018] Based on the sample support, the original capability values are shrunk or the confidence interval is corrected to determine the capability weights and write them into the capability matrix.
[0019] The capability matrix is periodically calibrated based on the results of manual review.
[0020] Preferably, the gating fusion includes hard gating and soft gating.
[0021] Hard gating disables the anomaly declarations output by the anomaly detection method when the sensitivity of the anomaly detection method is lower than the minimum sensitivity limit, the false alarm rate is higher than the maximum false alarm rate limit, or the sample support is lower than the minimum sample support limit under the target stratification conditions.
[0022] Soft gating merges abnormal claims that pass hard gating and point to the same chromosome instance, the same abnormality type, and the same compatible breakpoint region, and calculates candidate scores based on capability weights and detection confidence.
[0023] When at least one anomaly detection method in the effective capability domain reports an anomaly declaration and the candidate score reaches the candidate threshold, the corresponding anomaly declaration is added to the anomaly candidate set.
[0024] Preferably, the karyotype biological constraints include at least two of the following: chromosome number constraints, homologous pairing integrity constraints, chromosome relative length and arm ratio constraints, chromosome material variation constraints, and abnormal pattern constraints.
[0025] In step S4, the processor generates an inference about the change in the number of instances of the target chromosome category based on the abnormal chromosome number.
[0026] Based on the inferences that there are length asymmetries, banding mismatches in breakpoint regions, or changes in chromosome material content among the abnormally generated homologous chromosomes due to deletions or duplications.
[0027] The inference can be based on the fact that at least two chromosome categories have corresponding breakpoints and complementary fragment origins, generated by reciprocal translocation abnormalities, or on the fact that the length relationship or banding mirror relationship of the two arms of the target chromosome meets a preset abnormality pattern, generated by isochromocytosis abnormalities.
[0028] A consistency score is calculated based on the satisfaction and reliability weights of each inference.
[0029] In step S5, the hierarchical output unit classifies candidate anomaly claims into high-confidence anomalies, anomalies to be reviewed, or undetected anomalies, and calibrates the corresponding confidence thresholds based on the validation dataset to meet the target sensitivity.
[0030] The average number of review candidates per unit sample is limited to within the pre-set manual review burden budget.
[0031] A chromosome karyotype abnormality detection system based on capability gating and consistency adjudication, provided by the present invention, includes: The image processing module acquires images of chromosomes in metaphase, segments them into multiple chromosome instances, and determines the chromosome category and image quality level of each chromosome instance. The characterization divergence detection module uses at least two anomaly detection methods based on different chromosome image information and with different characterization invariances to detect anomalies for the same chromosome instance, and obtains anomaly declarations and detection confidence. The capability-gated fusion module determines the capability weights corresponding to each anomaly declaration based on the correspondence between anomaly type, chromosome category, image quality level and the detection capabilities of each anomaly detection method. Based on the capability weights and detection confidence, the anomaly declarations are gated and fused to obtain an anomaly candidate set. The consistency scoring module generates a set of verifiable inferences based on the karyotype biological constraints corresponding to the anomaly type in the anomaly candidate set. It then uses multiple chromosome instances to perform global verification of the set of verifiable inferences to obtain a consistency score. The hierarchical output module determines the hierarchical results of candidate abnormality declarations based on the gating fusion results and consistency scores, and outputs chromosomal abnormality triage information.
[0032] Preferably, the characterization divergence detection module selects at least two of the following anomaly detection methods for execution: Anomaly detection is performed based on the central axis band sequence of chromosome instances; anomaly detection is performed based on at least one of the morphological information of the overall outline, relative length, centromere index, arm ratio, and mean curvature of the central axis of chromosome instances; anomaly detection is performed based on the serialization encoding of the light and dark bands of chromosome instances; and anomaly detection is performed based on the pairing difference characterization after homologous chromosome alignment.
[0033] At least two abnormality detection methods are selected to preserve morphological information, banding information, or homologous chromosome difference information, so that the corresponding missed detection ranges do not overlap at least partially.
[0034] Preferably, the capability-gated fusion module obtains a capability matrix containing capability weights based on the correspondence between anomaly type, chromosome category, image quality level and the detection capability of each anomaly detection method.
[0035] The capability matrix is established from a set of labeled samples with manually verified results, which is stratified according to anomaly type, chromosome category, and image quality level.
[0036] The capability-gated fusion module statistically analyzes the sensitivity, false alarm rate, and sample support of each anomaly detection method under each stratification condition, and calculates the original capability value based on the sensitivity and false alarm rate.
[0037] Based on the sample support, the original capability values are shrunk or the confidence interval is corrected to determine the capability weights and write them into the capability matrix.
[0038] The capability matrix is periodically calibrated based on the results of manual review.
[0039] Preferably, the capability-gated fusion module uses hard gating to shield anomaly claims output by anomaly detection methods that fail to meet minimum sensitivity, maximum false alarm rate, or minimum sample support.
[0040] Abnormal claims that pass hard gating and point to the same chromosome instance, the same abnormality type, and the same compatibility breakpoint region are merged, and candidate scores are calculated based on capability weights and detection confidence.
[0041] When at least one anomaly detection method in the effective capability domain reports an anomaly declaration and the candidate score reaches the candidate threshold, the corresponding anomaly declaration is added to the anomaly candidate set.
[0042] Preferably, the consistency scoring module generates corresponding verifiable inferences based on chromosome number abnormalities, deletion or duplication abnormalities, reciprocal translocation abnormalities, or isochromatic abnormalities; determines the satisfaction level of each inference using multiple chromosome instances in the same karyotype; and calculates a consistency score based on satisfaction level and reliability weights.
[0043] The hierarchical output module calculates a comprehensive score based on the gating fusion results and consistency score, and outputs high-confidence anomalies, anomalies pending review, or undetected anomalies based on the high-confidence threshold and low-confidence threshold calibrated by the validated dataset; the average number of review candidates per unit sample is limited to the preset manual review burden budget.
[0044] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention employs at least two abnormality detection methods with different characterization invariance for the same chromosome instance, respectively preserving morphological information, banding information, or homologous chromosome difference information, thereby improving the coverage of observable features of different chromosomal abnormalities and reducing the risk of structural missed detections caused by a single characterization.
[0045] 2. This invention establishes a correspondence between abnormality type, chromosome category, image quality, and the detection capabilities of various abnormality detection methods, and performs gated fusion on abnormality declarations based on this correspondence, thereby suppressing unreliable outputs outside the capability range and reducing the number of abnormality candidates and the burden of manual review.
[0046] 3. This invention generates verifiable inferences based on the karyotype biological constraints corresponding to the abnormality type, and performs global verification using multiple chromosome instances in the same karyotype, which improves the consistency between local abnormality declarations and karyotype-level information and reduces the risk of false alarms caused by segmentation gaps, local occlusion, or banding artifacts.
[0047] 4. This invention achieves hierarchical adaptation of abnormal candidates and manual review resources by comprehensively gating fusion results and consistency scores, and calibrates confidence thresholds based on target sensitivity and manual review burden budget.
[0048] 5. Before the processor enters global verification and hierarchical calculation, the capability gating fusion module of the present invention performs hard gating on the anomaly declarations according to the capability matrix, and only passes the anomaly candidate set that has passed the hard gating to the subsequent modules, thereby reducing the number of anomaly declarations and data transmission volume in the subsequent calculations, thereby reducing the computational load and storage resource occupation, and improving the processing efficiency of chromosome karyotype images under the same hardware configuration. Attached Figure Description
[0049] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a schematic diagram of a method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication. Figure 2 This is a schematic diagram of the process of capability gating fusion, consistency verification and hierarchical output in an embodiment of the present invention. Detailed Implementation
[0050] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the protection scope of the present invention.
[0051] This invention provides a method for detecting chromosome karyotype abnormalities based on capability gating and consistency adjudication. This method addresses the problems of structural detection blind spots in single chromosome image representation, the potential for false positives from direct fusion of multiple abnormality detection methods, and the lack of karyotype-level verification for local abnormality declarations. Figure 1 For example, specifically including: Step S1: Image preprocessing.
[0052] Specifically, images of the metaphase chromosomes to be analyzed are obtained. The instance segmentation network is used to obtain One chromosome instance. Each chromosome instance is represented by an instance mask. and its circumscribed region image Indicates. The classification network outputs the first... The probability of each chromosome instance belonging to each chromosome class is calculated, and the chromosome class is determined according to the following formula:
[0053] in, Indicates the first Chromosome category of each chromosome instance; This represents a predefined set of chromosome categories; Represents a set Any candidate category in; The classification network represents the classification based on instance images. Output the category to which this chromosome instance belongs. The probability of. Represents a chromosome instance index. ; Indicates the first Instance mask for each chromosome instance; Indicates the first Image of the outer region of a chromosome instance; This indicates the operation of selecting the independent variable corresponding to the maximum value.
[0054] Quality indicators such as banding clarity, segmentation integrity, overlap, and centromere discernibility were extracted, and after normalization, the following calculations were performed. The overall quality score of each chromosome instance :
[0055] in, This indicates the total number of quality indicators; Indicates the first The first chromosome instance The normalized quality index has a value range of 0 to 1. The larger the value, the better the image quality of the corresponding index. Indicates the first The weights of each quality indicator are calculated, and the sum of the weights of each indicator is 1. Indicates the quality index. .
[0056] According to the preset quality threshold and Determine the image quality level for each chromosome instance:
[0057] in, . Indicates the first Image quality level of each chromosome instance.
[0058] Step S2: Characterization of divergence detection. For each chromosome instance... The central axis stripe representation, overall morphology representation, stripe coding representation, and homologous pairing difference representation are constructed respectively, and the anomaly declaration and detection confidence are output by the anomaly detection method corresponding to each representation.
[0059] Specifically, anomaly detection methods maintain characterization invariance to changes in chromosome pose, size, position, or grayscale, i.e., they possess stable characterization properties. The first anomaly detection method targets the first The abnormal declaration formed by a chromosome instance is denoted as Exception declaration At least include instance identifier Exception types and abnormal location information Anomaly location information is used to indicate the breakpoint interval, the location of the abnormal segment, or the affected chromosome category. It can be empty if no breakpoint is involved.
[0060] The construction process of the central axis strip characterization includes: Extract the first The central axis of each chromosome instance Straighten the chromosomes along the central axis and place them at the normalized position on the central axis. At that point, the sampling neighborhood of the normal line The average of the gray values within the range is used to form a central axis strip sequence:
[0061] in, Indicates the pixel position in the normal sampling neighborhood; Indicates pixel position The grayscale value at that location. Indicates the first Chromosome instances at normalized positions The value of the central axis strip at that location; This indicates the number of pixels contained in the sampling neighborhood of the normal.
[0062] This characterization can reduce the impact of chromosome bending and posture changes on band alignment information.
[0063] The process of constructing the overall morphological representation includes: Construct the first The overall morphological feature vector of a chromosome instance Overall morphological feature vector Based on relative length arm ratio Centromere index Mean curvature of the central axis and instance mask area At least one of the components. Relative length The arm ratio is the ratio of the length of this chromosome instance to the total length of all chromosomes of the same karyotype. It is the ratio of the length of the longer arm to the length of the shorter arm.
[0064] Overall morphological characterization is used to preserve information on chromosome length, centromere position, arm ratio, outline, and curvature.
[0065] The process of constructing the strip-coded representation includes: For the central axis strip sequence Perform grayscale standardization, and according to Each grayscale range is quantized into a length of [length]. strip-coded sequence . No. Encoding of each sampling position Determined according to the following formula:
[0066] in, and They represent the first The lower and upper boundaries of each grayscale range; This represents the strip grayscale sequence after zero-mean or interval normalization. Represents a striped coded sequence The Middle Encoding of each sampling position.
[0067] This representation improves stability to scale and location variations through grayscale normalization and serialization encoding.
[0068] The construction process of the homologous pairing differential representation includes: From category Selecting homologous chromosome instances that pair up with each other from the same chromosome instances. and After aligning the representation sequences of the two, the paired difference representation is calculated:
[0069] in, and These represent the normalized banding or outline representations of two homologous chromosomes, respectively. This represents the positional mapping obtained through dynamic time warping, affine registration, or feature point registration. Represents a chromosome instance and In the Paired differential characterization of each sampling location; Indicates the relationship with the first An index of homologous chromosome instances paired with each chromosome instance.
[0070] Homologous pairing differential characterization can suppress common banding patterns of the same chromosome class and highlight local differences between homologous chromosomes.
[0071] No. Anomaly detection methods will characterize the corresponding input. Input the classifier or anomaly scorer to obtain the anomaly type. Detection confidence:
[0072] in, Indicates the first The first anomaly detection method for the first Each chromosome instance belongs to the abnormal type. The detection confidence level, Indicates the first The first anomaly detection method targets the first The input representation used for each chromosome instance; Indicates the first Anomaly detection methods target anomaly types A function that outputs unnormalized fractions; This represents the Sigmoid (S-shaped activation) function or the Softmax (normalization exponent) function.
[0073] An anomaly declaration is generated when the detection confidence level reaches the declaration threshold of the corresponding anomaly detection method.
[0074] Step S3: Capability Gating Fusion.
[0075] Specifically, first obtain a labeled sample set with manual verification results, and then categorize them according to anomaly type. Chromosome categories and image quality level Perform layering and construct layered units.
[0076] For the An anomaly detection method is used to count the number of true positive samples in each stratified unit. False negative sample count Number of false positive samples and the number of true negative samples And calculate sensitivity False alarm rate and sample support :
[0077]
[0078]
[0079] in, This represents a positive number used to prevent the denominator from being zero and to smooth out small sample statistics.
[0080] Calculate the first [unit] within the hierarchical unit based on sensitivity and false alarm rate. The original capability values of the anomaly detection methods:
[0081] in, Indicates the first Anomaly detection methods in anomaly types Chromosome categories and image quality level The original capability value in the corresponding hierarchical unit.
[0082] The original capability values are then shrunk according to the sample support to obtain the capability weights to be written into the capability matrix:
[0083]
[0084] in, Indicates the first Anomaly detection methods target anomaly types Overall capability priors across all chromosome categories and image quality levels; Indicates the coefficient of shrinkage; A positive number representing the control of shrinkage strength. This represents the capability weights for writing into the capability matrix.
[0085] When the sample support is low, the shrinkage coefficient decreases, and the ability weights shrink towards the prior of the overall ability.
[0086] In more optimized examples, the lower bound of the confidence interval can be used instead of the original capability value to reduce the risk of overestimation of capability due to small sample size. The results of manual review will be used as new labeled samples to periodically update the above statistics and capability matrix.
[0087] In more preferred examples, gating fusion includes both hard gating and soft gating.
[0088] For the first to be detected Chromosome instances and candidate abnormality types According to chromosome category and image quality level Implement hard gating:
[0089] in, This represents the result of hard gating, with a value of 1 indicating the first... The anomaly declarations output by the anomaly detection method are valid in the current capability domain, and a value of 0 indicates that the corresponding anomaly declaration is masked; Indicates the minimum sensitivity; Indicates the maximum false alarm rate; This indicates the minimum level of sample support. This indicates an indicator function; the value is 1 if all conditions within the square brackets are met, and 0 otherwise.
[0090] Abnormal claims that are hard-gated and point to the same chromosome instance, the same anomalous type, and a compatible breakpoint region are merged, and candidate scores are calculated according to soft gating:
[0091] in, This indicates the ability to detect anomaly types. A collection of anomaly detection methods. Indicates the first Each chromosome instance corresponds to an abnormality type. Candidate scores; Represents the set of anomaly detection methods Multiply the terms in the table together.
[0092] When at least one anomaly detection method within the effective capability domain reports an anomaly declaration and candidate scores Reaching the candidate threshold When this happens, the corresponding exception declaration is added to the exception candidate set. Consistent claims from multiple reliable anomaly detection methods thus reinforce each other, while limiting the impact of low-capacity weighting methods on candidate scores.
[0093] Step S4: Calculate the consistency score.
[0094] Specifically, such as Figure 2 As shown, for the anomaly candidate set Each candidate exception declaration in Generate an exception based on the exception type. A set of verifiable inferences for verifiable inferences. Calculate the first verifiable inference using multiple chromosome instances of the same karyotype. Satisfaction of verifiable inferences The satisfaction level ranges from 0 to 1, and a consistency score is calculated:
[0095] in, Indicates candidate exception declaration Consistency score. Indicates the first The reliability weight of a verifiable inference. The reliability weight is determined based on the measurement stability of the corresponding feature or the statistical results of the validation set.
[0096] In more preferred examples, for chromosome number abnormalities, the target chromosome category is set as follows: The actual number of instances is The expected number of instances is The minimum quantity difference required to identify a quantity anomaly is Satisfaction of the inference of quantity change This can be calculated as the ratio of the difference between the actual number of instances and the expected number of instances to the minimum difference, with the maximum value limited to 1. (The superscript is not included in this calculation.) Used to identify the expected value.
[0097] In more preferred examples, for deletion or duplication anomalies, the changes in the length asymmetry and material content of homologous chromosomes are calculated. The material content characterization value of each chromosome instance is determined according to the following formula:
[0098] in, Indicates the first The material content characterization value of each chromosome instance. Indicates the local background grayscale; This represents the normalization factor used to eliminate differences in imaging exposure or scale. The material content characterization value is a surrogate quantity at the image level, not a direct measurement of DNA (deoxyribonucleic acid) quality.
[0099] For homologous chromosome examples and The length asymmetry is calculated based on the relative lengths of the two. The degree of material asymmetry was calculated based on the material content characterization values of both. Then, the asymmetry is mapped to the inference satisfaction level in the range of 0 to 1 using a piecewise linear function or a sigmoid (S-shaped activation) function. Wherein, the subscript... Used to indicate length; subscript Used to identify the content characterization value of a substance.
[0100] In more preferred examples, for reciprocal translocation abnormalities, the two chromosome categories are considered. and The candidate breakpoints are verified to ensure their positions correspond and that the band patterns of the swapped segments are complementary. Indicates the relationship with chromosome category Another chromosome category involved in mutual translocation verification.
[0101] In more preferred examples, for isochromosomal abnormalities, the deviation of the ratio of the lengths of the two arms from 1 and the mirror similarity of the band patterns of the two arms are calculated, and the corresponding inference satisfaction is formed. Thus, the local abnormality declaration can be globally verified based on karyotype-level information such as chromosome number, homologous pairing, relative length, arm ratio, material changes, and abnormal patterns.
[0102] Step S5: Graded diagnostic output.
[0103] Specifically, such as Figure 2 As shown, for candidate exception declarations The candidate scores obtained by gating fusion With consistency score The combination is the overall score:
[0104] in, This represents the combined weight of the gating fusion score, and its value ranges from 0 to 1. Indicates candidate exception declaration The overall score.
[0105] In more preferred examples, the combination type of anomaly detection method is input as an additional calibration variable into the logistic regression model to obtain a comprehensive score.
[0106] when When, mark the candidate anomaly declaration as a high-confidence anomaly; when When, mark the candidate exception declaration as pending review; when At that time, no abnormality was detected in the output. Indicates a high confidence threshold; This indicates a low confidence threshold. No abnormality detected is only used as supplementary triage information and is not used for automatically excluding abnormalities or automatically generating negative diagnostic reports.
[0107] On the validation dataset, the high-confidence and low-confidence thresholds are calibrated based on target sensitivity and the budget for human review:
[0108]
[0109] in, and These represent the calibrated high-confidence threshold and low-confidence threshold, respectively, with superscript indicating the threshold values. Used to identify the target value obtained through calibration; This indicates the specificity of the validation set under the corresponding threshold combination; This indicates the validation set sensitivity under the corresponding threshold combination; Indicates the preset target sensitivity; Indicates the number of validation samples; Indicates the first The number of candidate anomalies that need to be submitted for manual review for each verification sample at the corresponding threshold; This represents the average budget for manual review per unit sample.
[0110] The above calibration limits the number of abnormal candidates per unit sample while meeting the target sensitivity.
[0111] The final output of chromosomal abnormality triage information includes at least the candidate abnormality type, affected chromosomal instances and categories, candidate breakpoints or fragment locations, abnormality detection methods involved in the declaration, detection confidence and capability weights of each abnormality detection method, gating status, candidate score, consistency score, comprehensive score, grading results, and verifiable inferences that support or refute the candidate abnormalities, for relevant personnel to review.
[0112] The present invention also provides a chromosomal karyotype abnormality detection system based on capability gating and conformity adjudication. The chromosomal karyotype abnormality detection system based on capability gating and conformity adjudication can be implemented by executing the process steps of the chromosomal karyotype abnormality detection method based on capability gating and conformity adjudication. That is, those skilled in the art can understand the chromosomal karyotype abnormality detection method based on capability gating and conformity adjudication as a preferred embodiment of the chromosomal karyotype abnormality detection system based on capability gating and conformity adjudication.
[0113] The present invention provides a chromosome karyotype abnormality detection system based on capability gating and consistency adjudication, comprising: an image processing module, a characterization divergence detection module, a capability gating fusion module, a consistency scoring module, and a hierarchical output module, wherein each module forms a complete processing chain around the same metaphase chromosome image.
[0114] Specifically, the image processing module acquires images of the metaphase chromosomes to be analyzed, performs instance segmentation on the metaphase chromosome images, and obtains multiple chromosome instances. For each chromosome instance, the chromosome category is determined, and the image quality level is determined based on banding clarity, segmentation completeness, overlap, and centromere discernibility.
[0115] The image processing module transmits chromosome instances, chromosome categories, and image quality levels to the characterization divergence detection module, and transmits karyotype-level data formed by multiple chromosome instances to the consistency scoring module.
[0116] The characterization divergence detection module constructs at least two input characterizations with different characterization invariances for the same chromosome instance, and performs anomaly detection on each, obtaining anomaly declarations and detection confidence scores.
[0117] In more preferred embodiments, the input representation is selected from central axis strip representation, overall morphological representation, strip coding representation, and homologous pairing difference representation. The construction of each representation and the determination of the detection confidence ensure that the selected anomaly detection method retains morphological information, strip information, or homologous chromosome difference information, respectively.
[0118] The characterization divergence detection module transmits anomaly declarations and detection confidence scores to the capability gating fusion module.
[0119] The capability-gated fusion module saves or calls the capability matrix. The capability matrix is divided into hierarchical units according to anomaly type, chromosome category, and image quality level, and records the sensitivity, false alarm rate, sample support, and capability weight of each anomaly detection method in the corresponding hierarchical unit.
[0120] The capability-gated fusion module determines the corresponding capability weights from the capability matrix based on the chromosome category, image quality level, and anomaly type corresponding to the anomaly declaration output by the image processing module. It uses hard gating to shield invalid anomaly declarations in the capability domain and soft gating to merge anomaly declarations pointing to the same chromosome instance, the same anomaly type, and compatible breakpoint regions, thus obtaining candidate scores and anomaly candidate sets.
[0121] The capability matrix is statistically analyzed, shrunk, and calibrated using periodic parameters based on the results of manual review. The capability-gated fusion module transmits the abnormal candidate set and candidate scores to the consistency scoring module and the hierarchical output module.
[0122] The consistency scoring module determines the corresponding karyotype biological constraints based on the anomaly type of each candidate anomaly declaration in the anomaly candidate set, from chromosome number constraints, homologous pairing integrity constraints, chromosome relative length and arm ratio constraints, chromosome material change constraints, and anomaly pattern constraints, and generates a set of verifiable inferences.
[0123] The satisfaction level of each verifiable inference is determined using multiple chromosome instances output by the image processing module, and a consistency score is calculated based on the satisfaction level and reliability weights. Verifiable inferences and satisfaction levels are calculated for chromosome number abnormalities, deletion or duplication abnormalities, reciprocal translocation abnormalities, and isochromatic abnormalities.
[0124] The consistency scoring module transmits the consistency score and verifiable inferences that support or refute candidate anomalies to the hierarchical output module.
[0125] The tiered output module calculates a comprehensive score based on candidate scores and consistency scores. It then marks candidate abnormality declarations as high-confidence abnormalities, pending review, or undetected abnormalities according to high-confidence and low-confidence thresholds, and outputs chromosomal abnormality triage information. High-confidence and low-confidence thresholds are calibrated to meet target sensitivity and limit the average number of review candidates per sample.
[0126] The undetected abnormalities output by the triage output module are only used as auxiliary triage information and are not used to automatically exclude abnormalities or automatically generate negative diagnostic reports.
[0127] Therefore, the chromosome instance, chromosome category, and image quality level output by the image processing module are used to determine the input and capability weights of the anomaly detection method, and also for karyotype-level global verification. The candidate score output by the capability gating fusion module and the consistency score output by the consistency scoring module are used together as the input to the hierarchical output module, so that multi-characteristic anomaly detection, conditional capability gating, and karyotype-level consistency adjudication form a closed loop of interconnected system processing.
[0128] Those skilled in the art will understand that, besides implementing the system and its various devices, modules, and units provided by this invention in the form of purely computer-readable program code, the same functions can be achieved entirely through logical programming of the method steps, making the system and its various devices, modules, and units of this invention function as logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system and its various devices, modules, and units provided by this invention can be considered as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; alternatively, the devices, modules, and units for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.
[0129] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.
Claims
1. A method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication, characterized in that, include: Step S1: Obtain images of chromosomes in metaphase, segment them to obtain multiple chromosome instances, and determine the chromosome category and image quality level of each chromosome instance; Step S2: For the same chromosome instance, at least two anomaly detection methods based on different chromosome image information and with different characterization invariances are used to detect anomalies, and anomaly declarations and detection confidence are obtained. Step S3: Based on the correspondence between anomaly type, chromosome category, image quality level and the detection capabilities of each anomaly detection method, determine the capability weight corresponding to each anomaly declaration, and perform gated fusion on the anomaly declarations based on the capability weight and detection confidence to obtain an anomaly candidate set; Step S4: In the anomaly declaration of the anomaly candidate set, generate a verifiable inference set based on the karyotype biological constraints corresponding to the anomaly type, and use multiple chromosome instances to perform global verification of the verifiable inference set to obtain a consistency score. Step S5: Determine the grading results of candidate abnormal statements based on the gating fusion results and consistency scores, and output chromosomal abnormality triage information.
2. The method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication according to claim 1, characterized in that, The anomaly detection method includes: Anomaly detection based on central axis band sequences of chromosome instances; Anomaly detection is performed based on at least one of the morphological information of chromosome instances, including overall outline, relative length, centromere index, arm ratio, and mean curvature of the central axis. Anomaly detection based on the serialization and encoding of light and dark bands from chromosome instances; Anomaly detection based on pairing difference characterization after homologous chromosome alignment; In step S1, the processor acquires a metaphase chromosome image, segments it to obtain multiple chromosome instances, and determines the chromosome category and image quality level of each chromosome instance. In step S2, for the same chromosome instance, the processor selects at least two abnormality detection methods to retain morphological information, band information, or homologous chromosome difference information, so that the corresponding missed detection ranges do not overlap at least partially.
3. The method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication according to claim 1, characterized in that, In step S3, the processor obtains a capability matrix containing capability weights based on the correspondence between anomaly type, chromosome category, image quality level, and the detection capabilities of each anomaly detection method. This matrix includes: Obtain a labeled sample set with manual verification results, and stratify it according to anomaly type, chromosome category, and image quality level; The sensitivity, false alarm rate, and sample support of each anomaly detection method under each stratification condition were statistically analyzed, and the raw capability value was calculated based on the sensitivity and false alarm rate. Based on the sample support, the original capability values are shrunk or the confidence interval is corrected to determine the capability weights and write them into the capability matrix. The capability matrix is periodically calibrated based on the results of manual review.
4. The method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication according to claim 3, characterized in that, The gating fusion includes both hardware gating and software gating; Hard gating disables the anomaly declarations output by the anomaly detection method when the sensitivity of the anomaly detection method is lower than the minimum sensitivity limit, the false alarm rate is higher than the maximum false alarm rate limit, or the sample support is lower than the minimum sample support limit under the target stratification condition. Soft gating merges abnormal claims that pass hard gating and point to the same chromosome instance, the same abnormality type, and the same compatibility breakpoint region, and calculates candidate scores based on capability weights and detection confidence. When at least one anomaly detection method in the effective capability domain reports an anomaly declaration and the candidate score reaches the candidate threshold, the corresponding anomaly declaration is added to the anomaly candidate set.
5. The method for detecting chromosomal karyotype abnormalities based on capability gating and consistency adjudication according to claim 1, characterized in that, The karyotype biological constraints include at least two of the following: chromosome number constraints, homologous pairing integrity constraints, chromosome relative length and arm ratio constraints, chromosome material variation constraints, and abnormal pattern constraints. In step S4, the processor generates an inference about the change in the number of target chromosome category instances based on the abnormal chromosome number. Based on the inference that there are length asymmetries, banding mismatches in the breakpoint regions, or changes in chromosome material content among the abnormally generated homologous chromosomes due to deletion or duplication; Based on the inference that at least two chromosome categories have corresponding breakpoints and complementary fragment sources, generated by reciprocal translocation abnormalities, or based on the inference that the length relationship or banding mirror relationship of the two arms of the target chromosome meets the preset abnormality pattern, generated by isochromocytosis abnormalities. A consistency score is calculated based on the satisfaction and reliability weights of each inference; In step S5, the processor classifies candidate anomaly claims into high-confidence anomalies, anomalies to be reviewed, or undetected anomalies, and calibrates the corresponding confidence thresholds based on the verification dataset to meet the target sensitivity. The average number of review candidates per unit sample is limited to within the pre-set manual review burden budget.
6. A chromosomal karyotype abnormality detection system based on capability gating and consistency adjudication, characterized in that, include: The image processing module acquires images of chromosomes in metaphase, segments them into multiple chromosome instances, and determines the chromosome category and image quality level of each chromosome instance. The characterization divergence detection module uses at least two anomaly detection methods based on different chromosome image information and with different characterization invariances to detect anomalies for the same chromosome instance, and obtains anomaly declarations and detection confidence. The capability-gated fusion module determines the capability weights corresponding to each anomaly declaration based on the correspondence between anomaly type, chromosome category, image quality level and the detection capabilities of each anomaly detection method. Based on the capability weights and detection confidence, the anomaly declarations are gated and fused to obtain an anomaly candidate set. The consistency scoring module generates a set of verifiable inferences based on the karyotype biological constraints corresponding to the anomaly type in the anomaly candidate set. It then uses multiple chromosome instances to perform global verification of the set of verifiable inferences to obtain a consistency score. The hierarchical output module determines the hierarchical results of candidate abnormality declarations based on the gating fusion results and consistency scores, and outputs chromosomal abnormality triage information.
7. The chromosome karyotype abnormality detection system based on capability gating and consistency adjudication according to claim 6, characterized in that, The characterization divergence detection module selects at least two of the following anomaly detection methods for execution: Anomaly detection based on central axis band sequences of chromosome instances; Anomaly detection is performed based on at least one of the morphological information of chromosome instances, including overall outline, relative length, centromere index, arm ratio, and mean curvature of the central axis. Anomaly detection based on the serialization and encoding of light and dark bands from chromosome instances; Anomaly detection based on pairing difference characterization after homologous chromosome alignment; At least two abnormality detection methods are selected to preserve morphological information, banding information, or homologous chromosome difference information, so that the corresponding missed detection ranges do not overlap at least partially.
8. The chromosome karyotype abnormality detection system based on capability gating and consistency adjudication according to claim 6, characterized in that, The capability-gated fusion module obtains a capability matrix containing capability weights based on the correspondence between anomaly type, chromosome category, image quality level and the detection capabilities of each anomaly detection method. The capability matrix is established from a set of labeled samples with manually verified results, and the labeled sample set is stratified according to anomaly type, chromosome category and image quality level; The capability-gated fusion module statistically analyzes the sensitivity, false alarm rate, and sample support of each anomaly detection method under each stratification condition, and calculates the original capability value based on the sensitivity and false alarm rate. Based on the sample support, the original capability values are shrunk or the confidence interval is corrected to determine the capability weights and write them into the capability matrix. The capability matrix is periodically calibrated based on the results of manual review.
9. The chromosome karyotype abnormality detection system based on capability gating and consistency adjudication according to claim 8, characterized in that, The capability-gated fusion module uses hard gating to shield anomaly claims output by anomaly detection methods that fail to meet minimum sensitivity, maximum false alarm rate, or minimum sample support. Abnormal claims that pass hard gating and point to the same chromosome instance, the same abnormality type, and the same compatibility breakpoint region are merged, and candidate scores are calculated based on capability weights and detection confidence. When at least one anomaly detection method in the effective capability domain reports an anomaly declaration and the candidate score reaches the candidate threshold, the corresponding anomaly declaration is added to the anomaly candidate set.
10. The chromosome karyotype abnormality detection system based on capability gating and consistency adjudication according to claim 6, characterized in that, The consistency scoring module generates corresponding verifiable inferences based on chromosome number abnormalities, deletion or duplication abnormalities, reciprocal translocation abnormalities, or isochromatic abnormalities. The degree of satisfaction of each inference is determined by using multiple chromosome instances within the same karyotype; A consistency score is calculated based on satisfaction and reliability weights. The hierarchical output module calculates a comprehensive score based on the gating fusion result and the consistency score, and outputs high-confidence anomalies, anomalies to be reviewed, or anomalies not detected based on the high-confidence threshold and low-confidence threshold calibrated by the validated dataset. The average number of review candidates per unit sample is limited to within the pre-set manual review burden budget.
Citation Information
Patent Citations
Chromosome anomaly detection method, device and equipment and computer readable storage medium
CN112037174A
Automatic chromosome karyotype analysis and anomaly detection method
CN112288706A
Automatic chromosome segmentation and classification method based on deep learning
CN113658150A
