Search

CN-122023345-A - MMC chromosome fracture detection method and system based on deep learning

CN122023345ACN 122023345 ACN122023345 ACN 122023345ACN-122023345-A

Abstract

The application provides a method and a system for detecting MMC chromosome breakage based on deep learning. The method comprises the steps of firstly, carrying out Gaussian filtering denoising on a single cell visual field image, inputting a preprocessed image into an improved 24-class chromosome instance segmentation model, outputting instance masks and 1-22-X, Y class probabilities, judging complete/fragment/pseudo structures of an instance ROI through a deep learning three-class model, removing noise to obtain a fragment set, skeletonizing the fragment masks to extract two end fracture endpoints, inputting the fragment ROI into a twin metric learning network to obtain embedded characterization, calculating connection scores and fusing the connection scores with class consistency to obtain edge scores, carrying out unique pairing on the endpoints, allowing global matching under empty matching constraint to obtain optimal association, generating fracture events and outputting a score value. The method improves the accuracy of small fragment detection and association, reduces false positive of misspelled and the result can be rechecked and is convenient for quality control.

Inventors

  • TUO FUJUN
  • GUO QIANQIAN
  • NIE FANGXING

Assignees

  • 中科伊和智能医疗科技(广西)有限公司

Dates

Publication Date
20260512
Application Date
20260130

Claims (10)

  1. 1. The MMC chromosome breakage detection method based on deep learning is characterized by comprising the following steps: S1, carrying out Gaussian filtering denoising on a single-cell image to obtain a preprocessed image; S2, inputting the preprocessed image into a trained improved 24-class chromosome instance segmentation model to obtain a chromosome instance set, wherein each instance at least comprises an instance mask and a corresponding chromosome class probability vector, and the chromosome class comprises No. 1-22 chromosomes and X, Y chromosomes; S3, constructing an instance ROI for each instance in the chromosome instance set, inputting the instance ROI into a trained deep learning three-classification model to output a classification result of a complete, fragment or pseudo structure, removing the pseudo structure instance, and reserving the fragment instance to form a fragment set; S4, skeletonizing each fragment instance mask in the fragment set to extract two fracture endpoint coordinates; S5, constructing a segment ROI for each segment instance, and inputting the segment ROI into a trained twin depth neural network metric learning model to obtain a segment embedded representation for representing the similarity of the segments; S6, generating a candidate associated edge set from the fragment set based on a fracture endpoint spatial proximity relation of the fragment instance and the class consistency determined according to the chromosome class probability vector, calculating a connection score of the candidate associated edge based on the fragment embedding characterization, and fusing the connection score and the class consistency to obtain a fused edge score; s7, under the constraint that each breaking endpoint is paired with at most one breaking endpoint and the breaking endpoints are allowed to be unpaired, determining the matching cost by fusion edge scores and performing global matching to obtain an optimal endpoint association set; And S8, generating a breaking event set according to the optimal endpoint association set, and obtaining a breaking score value according to the breaking event set, wherein the breaking score value is a score value obtained by mapping the number of the breaking events.
  2. 2. The method for detecting the break of the MMC chromosome based on deep learning of claim 1, wherein the trained improved 24-class chromosome instance segmentation model comprises a convolution downsampling Stem, a trunk feature extraction network, a feature pyramid fusion network and an instance prediction head, wherein the trunk feature extraction network is a window self-attention transducer trunk and outputs at least 4-scale feature graphs, the feature pyramid fusion network is of a bidirectional fusion structure and comprises a top-down path and a bottom-up path, so that the instance segmentation capability of chromosomes and fragments of different scales is improved.
  3. 3. The method for detecting MMC chromosome breakage based on deep learning of claim 2, wherein the trained improved 24-class chromosome instance segmentation model further comprises small-patch enhancement branches, which hold and strengthen high-resolution shallow features and inject fine grain texture and boundary information of small target patches into an instance pre-probe through cross-scale fusion with a feature pyramid fusion network, thereby improving recall rate of small patch instances and reducing missed detection.
  4. 4. The method for detecting the break of the MMC chromosome based on the deep learning of claim 3, wherein the instance prediction head at least comprises a candidate instance generation branch, a mask branch and 24 class classification branches, wherein the mask branch adopts a dynamic convolution kernel generation mechanism or an instance query driven mask prediction mechanism, and a mask boundary is secondarily refined through a point-level boundary refinement sub-network after the mask is output so as to reduce mask adhesion errors caused by chromosome mutual attachment or fracture blurring.
  5. 5. The method for detecting the break of the MMC chromosome based on the deep learning of claim 4, wherein the training loss of the trained improved 24-class chromosome instance segmentation model at least comprises (a) 24-class classification loss, adopting class imbalance self-adaptive weighting to improve the recognition of low-frequency classes and small fragments, (b) mask loss, adopting a combination of Dice class loss and binary cross entropy class loss, (c) boundary consistency loss, and based on mask boundary area or boundary point supervision, highlighting the segmentation precision improvement of the improved model at the break boundary.
  6. 6. The method for detecting the break of the MMC chromosome based on the deep learning of claim 1, wherein the trained deep learning three-classification model is a mask-guided dual-flow network structure and comprises an image flow encoder and a mask flow encoder, wherein the image flow encoder performs feature extraction on an example ROI image, the mask flow encoder performs feature extraction on an example mask or a distance transformation diagram thereof, and the structure priori of the mask flow is guided into the image flow features through a cross-attention fusion module so as to distinguish complete, fragmented and pseudo structures.
  7. 7. The method for detecting the break of the MMC chromosome based on the deep learning of claim 1, wherein the trunk of the trained deep learning three-classification model is a lightweight network and comprises a convolution Stem, a plurality of residual blocks or lightweight transducer coding blocks and a classification head, wherein a channel-space joint attention module is arranged in the trunk and is used for strengthening the difference characteristics of a pseudo structure and real fragments in texture continuity, edge morphology and density distribution so as to reduce the probability of false structure misentering the fragment set, and the strengthening pseudo structure comprises dyeing particles, scratches and background fragments.
  8. 8. The method for detecting the break of the MMC chromosome based on the deep learning of claim 1, wherein the trained deep learning three-classification model adopts multi-task training, outputs at least one auxiliary result besides the three-classification result, wherein the auxiliary result is a boundary probability map, a noise particle density score or a dyeing non-uniformity score, and optimizes the loss corresponding to the auxiliary result and the three-classification loss in a combined way so as to improve the separability of a pseudo structure and enhance the generalization capability of the model.
  9. 9. The method for detecting the fracture of the MMC chromosome based on the deep learning, as set forth in claim 1, wherein the trained twin depth neural network metric learning model comprises a shared encoder and a projection head, wherein the shared encoder is a convolutional neural network or a lightweight transform network, multi-scale cavity convolution or deformable convolution is introduced into at least one layer of the encoder to enhance the robustness to the fracture morphology change, the projection head is a multi-layer perceptron and normalizes output embedded vectors to obtain the segment embedded representation, the input of the twin depth neural network metric learning model comprises at least one structure guide channel besides a segment ROI image, the structure guide channel is an example mask channel, a skeleton channel or an endpoint thermodynamic diagram channel, and chromosome class probability vectors or low-dimensional embedded thereof are injected into the shared encoder or the projection head as condition information to realize the joint measurement of texture similarity, structural compatibility and class priori.
  10. 10. An MMC chromosome fracture detection system based on deep learning, comprising: the preprocessing module is used for carrying out Gaussian filtering denoising on the single-cell image to obtain a preprocessed image; the example segmentation module inputs the preprocessed image into a trained improved 24-class chromosome example segmentation model to obtain a chromosome example set, wherein each example at least comprises an example mask and a corresponding chromosome class probability vector, and the chromosome class comprises chromosome numbers 1-22 and X, Y; The segment set generation module constructs an instance ROI for each instance in the chromosome instance set, inputs the instance ROI into a trained deep learning three-classification model to output a classification result of a complete, fragment or pseudo structure, eliminates the pseudo structure instance, and reserves the fragment instance to form a segment set; The fracture endpoint determination module is used for skeletonizing each fragment instance mask in the fragment set to extract two fracture endpoint coordinates; The segment embedding representation acquisition module constructs a segment ROI for each segment instance, and inputs the segment ROI into a trained twin depth neural network metric learning model to obtain a segment embedding representation for representing the similarity of the segments; The fusion edge score calculation module is used for generating a candidate association edge set from the fragment set based on the spatial proximity relation of the breaking end points of the fragment instance and the class consistency determined according to the chromosome class probability vector, calculating the connection score of the candidate association edge based on the fragment embedding characterization, and fusing the connection score and the class consistency to obtain a fusion edge score; The optimal endpoint association set module is used for determining the matching cost by fusion edge scores and performing global matching under the constraint that each breaking endpoint is matched with at most one breaking endpoint and the breaking endpoints are allowed to be unpaired, so as to obtain an optimal endpoint association set; And the score calculating module is used for generating a breaking event set according to the optimal endpoint association set, obtaining a breaking score according to the breaking event set, and the breaking score is a score obtained by mapping the number of the breaking events.

Description

MMC chromosome fracture detection method and system based on deep learning Technical Field The invention relates to the technical field of chromosome fracture detection, in particular to an MMC chromosome fracture detection method and system based on deep learning. Background MMC (mitomycin C ) induced chromosome breakage detection and scoring are common means in cytogenetics, genotoxicology and DNA damage assessment, and are widely applied to relevant clinical screening, drug and chemical safety evaluation, environmental mutational factor research, cell repair pathway functional analysis and other scenes. Typically, after MMC processing, a chromosome microscopic image of the metaphase (metaphase) is obtained, and the break, fragment, connection to break, etc. are identified and counted by a professional to form an index such as a cell-level break score value (e.g., breaks/cell). The indexes are closely related to experimental batches, sample sources, dyeing schemes, microscopic equipment parameters and operator experience, so that high requirements on detection accuracy, repeatability and traceability are provided in practice. In the prior art, the most common approach is still manual sheet scoring. The manual method can comprehensively judge the conditions of complex background, overlapped adhesion, fracture morphology change and the like by combining experience, but has the obvious defects that firstly, the labor intensity of manual scoring is high, the time consumption is long, the efficiency requirement is difficult to meet particularly in high-throughput sample or large-scale research, secondly, the subjective difference among different operators is large, the stability of the same operator at different times is difficult to ensure, the consistency of results is insufficient, thirdly, the manual process lacks structural intermediate evidence, the formation basis of each fracture event is difficult to review and track, and the quality control cost is high. To improve efficiency, some research and engineering systems have attempted to introduce automated image processing. The traditional method generally extracts candidate chromosome regions from images through threshold segmentation, edge detection, morphological processing, skeleton extraction and other means, and classifies or counts the candidate chromosome regions based on manual features such as area, length, perimeter, aspect ratio, gray distribution and the like, or correlates fragments through distance rules and nearest neighbor strategies. However, in the MMC induced sample, chromosome fracture morphology is complex, fragment size span is large, noise interference such as dyeing particles, scratches, background fragments, uneven dyeing, defocus blur and the like is often accompanied, the traditional threshold value or edge method is highly sensitive to imaging conditions, and problems such as fragment omission, false structure false detection, difficulty in separating attached chromosomes and the like are easy to occur. Particularly, under the condition that fragments are dense or multiple chromosomes are overlapped with each other, fragments are paired only by local geometric adjacency or heuristic rules, fragments from different sources are easily spliced into the same fracture event in an error mode, and therefore false positives are increased, scoring is higher or unstable. In recent years, deep learning advances in microscopic image segmentation and object recognition, but the direct use of segmentation results for fracture scoring still faces key difficulties in that on one hand, a plurality of chromosome instances need to be accurately distinguished in a single cell view and reliable category information is output at the same time to assist in association, on the other hand, fracture events essentially relate to pairing and connection relations among fragments, belong to structural association problems, global consistency of connection relations is difficult to ensure only by segmentation and classification, and if unique connection constraint and empty matching mechanisms at endpoint levels are absent, simultaneous connection of a plurality of endpoints by one endpoint is easy to occur or unreasonable connection links are generated under noise interference, so that the fracture event generation lacks credibility. Meanwhile, the existing automatic method often lacks interpretable intermediate output (such as endpoint positions, candidate connection and confidence level thereof, matching constraint meeting conditions and the like), so that the result is difficult to quickly review, and the application of the method in clinical quality control or high-requirement scientific research analysis is limited. Therefore, a technical scheme for detecting and scoring MMC chromosome breaks, which can stably operate under complex noise and morphological change conditions, introduce global consistency constraints at the level of breaking association,