Search

CN-121999007-A - Cell segmentation method based on multichannel fluorescence feature fusion

CN121999007ACN 121999007 ACN121999007 ACN 121999007ACN-121999007-A

Abstract

The invention discloses a cell segmentation method based on multichannel fluorescence feature fusion, which relates to the technical field of cell segmentation and comprises the following steps of data preprocessing and channel synchronization, network architecture design, cell nucleus positioning, mask generation and the like. Through deep learning technology, combined with multichannel image data such as DAPI and SpGreen, the method realizes the automatic segmentation of accurate cell nuclei and cytoplasmic areas, effectively solves the challenges of cell adhesion and overlapping by introducing a attention mechanism and multi-scale feature fusion, simultaneously optimizes cell boundaries by distance transformation and conditional random field CRF, improves segmentation precision, and enables a user to accurately evaluate the distribution of cells and the coexpression condition of different fluorescent channels by a subsequent cell counting and coexpression analysis function.

Inventors

  • WANG QIAN
  • Qiu Zhipei
  • LIU ZHIYAO

Assignees

  • 北京工业大学

Dates

Publication Date
20260508
Application Date
20260128

Claims (10)

  1. 1. The cell segmentation method based on multichannel fluorescence feature fusion is characterized by comprising the following steps of: s1, carrying out standardization, noise removal and space alignment treatment on multi-channel image data; S2, designing a double-flow structure and attention mechanism model based on a deep convolutional neural network; S3, generating an accurate cell nucleus mask by using the DAPI signal and performing thresholding; S4, extracting a cytoplasmic region by using SpGreen channels and predicting a probability map; S5, combining the DAPI mask and the cytoplasmic area map, and executing accurate mask mapping; S6, optimizing cell boundaries and removing noise, and repairing cell morphology and segmentation accuracy; s7, evaluating the segmentation precision and optimizing the model based on IoU, the Dice coefficient and other indexes; s8, performing cell counting and coexpression analysis, and evaluating interaction among different channels; s9, generating a visual report of the segmentation result and data export, and supporting subsequent analysis and sharing.
  2. 2. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in the step S1, the normalization, noise removal and spatial alignment of the multi-channel image data: s1.1, loading all four fluorescence channel data from imaging equipment, ensuring that the image data of each channel is consistent in format, denoising and standardization processing are carried out, and unifying the image size and resolution; S1.2, normalizing the image of each channel to eliminate brightness difference among different channels, wherein a common normalization method is Z-score normalization: ; Wherein, the For the gray value of each pixel, Is the mean value of the pixels and, Is the standard deviation of the pixel; S1.3, cutting the image into a concerned area for reducing the complexity of processing, and performing space alignment to ensure that all image channels are consistent in space position; S1.4, carrying out local contrast enhancement on the image of each channel, and increasing details of the image by using a CLAHE (clear-cut-off) namely a contrast-limited self-adaptive histogram equalization algorithm; s1.5, splicing the information of all channels to form a multi-channel composite image, and preparing an input network.
  3. 3. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in the step S2, the model design of the dual-flow structure and the attention mechanism based on the deep convolutional neural network is characterized in that: S2.1, the network adopts a double-flow structure, wherein one flow processes DAPI channel data, namely cell nucleus information, and the other flow processes SpGreen channel data, namely cell cytoplasm information; S2.2, extracting high-level characteristics of respective channels by using a convolutional neural network CNN for each stream, wherein ResNet or VGG structures are generally adopted as a basic network; S2.3, introducing an attention mechanism in the network to strengthen the attention of the network to important areas, and using a channel attention mechanism to improve the attention to DAPI high-signal areas: ; Wherein, the For the weight matrix to be learned, In order to input the characteristics of the feature, As a result of the bias term, Activating a function for Sigmoid; S2.4, combining the multi-scale features, and carrying out context information fusion through a pyramid network so that the network can capture cell information of different scales; s2.5, outputting a cell nucleus mask as the network, and representing the spatial position of the cell nucleus.
  4. 4. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in the step S3, the accurate nuclear mask is generated by using the DAPI signal and thresholding is performed: S3.1, learning the spatial position of the cell nucleus through a deep network, generating a cell nucleus mask by taking a DAPI channel as a main input, and separating a cell nucleus region by using an adaptive thresholding method; s3.2, extracting core areas from the DAPI channel image, wherein the areas correspond to the centers of cell nuclei, and further refining the boundary of the core areas through a watershed algorithm; S3.3, optimizing threshold setting according to the local brightness information of the image so as to dynamically adjust the extraction range of the cell nucleus; S3.4, removing small noise in the cell nucleus mask through morphological operation, and further enhancing the accuracy of the cell nucleus region; s3.5, transferring the processed cell nucleus mask into a subsequent step for extracting the cytoplasmic boundary.
  5. 5. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in the step S4, the SpGreen channels are used for extracting cytoplasmic regions and carrying out probability map prediction: S4.1, extracting a probability map of a cytoplasmic area by using a convolutional neural network based on SpGreen channels; s4.2, performing depth feature extraction on SpGreen channels by using a convolution layer, so as to obtain multi-layer feature representation of cytoplasm; s4.3, obtaining a predicted probability map of cytoplasm through a Sigmoid activation function: ; Wherein, the Is a convolution kernel for cytoplasmic region prediction, Is a SpGreen feature of the input and, Is a Sigmoid activation function; S4.4, further optimizing the boundary of a cytoplasmic area through a conditional random field CRF, and correcting a fuzzy area between cells; s4.5, taking the cytoplasmic probability map after CRF optimization as the input of the subsequent step.
  6. 6. The method of claim 1, wherein the step S5, combining DAPI mask and cytoplasmic map, performs an accurate mask mapping: S5.1, mapping by using a nuclear mask of a DAPI channel and a cytoplasmic probability map of a SpGreen channel; s5.2, performing distance transformation based on a cell nucleus mask to generate a probability map of a cytoplasm boundary: ; Wherein, the Is a point on the nuclear mask, Is the result of the distance transformation, Is each pixel point in the space, Representing euclidean distance; s5.3, ensuring that the cytoplasm boundary does not exceed a SpGreen fluorescence area through constraint conditions; s5.4, optimizing a mask according to the distance transformation result, and mapping the cell nucleus mask to a cytoplasmic area to form a more accurate cell boundary; s5.5, the final cell boundary indicates the exact cell location.
  7. 7. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in the step S6, cell boundary optimization and noise removal, cell morphology and segmentation accuracy are repaired: s6.1, denoising the segmentation result, removing an area with an unclear boundary, and further repairing the cell edge; s6.2, correcting the shape of the cell boundary through morphological repair technology, namely expansion and corrosion operation; s6.3, applying a multi-level image restoration method, including removing artifacts and enhancing the accuracy of a real boundary; S6.4, ensuring that the cell structure accords with biological characteristics by adjusting the morphological information of each cell; and S6.5, outputting the corrected cell boundary image as a segmentation result.
  8. 8. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in step S7, segmentation accuracy evaluation and model optimization based on IoU and Dice coefficients and other indexes: S7.1, selecting an index IoU and a Dice coefficient to evaluate the precision of the segmentation result; S7.2, optimizing the super parameters of the network model to improve the segmentation accuracy; s7.3, adjusting a network structure according to the evaluation result, and adding or reducing certain layers to improve the performance; s7.4, testing the adjusted model through a verification set, and ensuring the performance on unseen data; And S7.5, outputting a final segmentation accuracy evaluation result.
  9. 9. The method of claim 1, wherein in step S8, cell counting and coexpression analysis are performed to evaluate interactions between different channels: s8.1, counting cells based on a cell boundary segmentation result, and counting the number of single cells; s8.2, calculating the co-expression rate between different cells by analyzing the overlapping areas of different fluorescent channels; s8.3, evaluating intersection and difference between different channels by using a multi-component analysis method; s8.4, performing cell coexpression analysis by using a statistical method; s8.5, generating a final cell count and coexpression analysis report.
  10. 10. The method for cell segmentation based on multi-channel fluorescence feature fusion according to claim 1, wherein in step S9, a visual report and data derivation of segmentation results are generated, and the following analysis and sharing are supported: S9.1, visualizing the segmented image, and displaying each cell and the area thereof; s9.2, outputting counting and coexpression data of each cell, so that subsequent analysis is facilitated; S9.3, archiving the segmentation result and the statistical data, and preparing for sharing; s9.4, generating an interactive visual report, so that the segmentation result can be understood conveniently; S9.5, fine tuning and improvement are carried out on the method according to application requirements.

Description

Cell segmentation method based on multichannel fluorescence feature fusion Technical Field The invention relates to the technical field of cell segmentation, in particular to a cell segmentation method based on multichannel fluorescence feature fusion. Background Cell segmentation is a key fundamental technology in cell biology, pathology analysis, drug screening and precision medical research, and the main purpose of the cell segmentation is to accurately identify and separate single cell areas from microscopic imaging data so as to realize quantitative research of cell counting, morphological analysis and multichannel fluorescence signals. With the development of multichannel fluorescence microscopy imaging technology, researchers can obtain spatial distribution information of cell nuclei, cytoplasm and various biomarkers at the same time, and fluorescence channels such as DAPI, spGreen, spGold, CY and the like are widely applied to cell structure and function research. In the prior art, the cell segmentation method mainly comprises a threshold value-based method, a traditional image processing-based method and a segmentation method based on deep learning, and the methods realize automatic segmentation of cell nuclei or cell areas to a certain extent and are applied to the fields of high-throughput imaging analysis and biomedical image processing. However, in the practical application process, the conventional cell segmentation method still depends on a single fluorescent channel or a simple characteristic extraction strategy, complementary information contained in a multi-channel fluorescent image is difficult to fully utilize, when cells are dense, the growth state is complex or adhesion and overlapping occur among cells, the problems of incomplete segmentation, fuzzy cell boundary and even wrong segmentation are easy to occur, the conventional partial deep learning segmentation method lacks an effective constraint mechanism when the relationship between the cell nucleus and the cytoplasm is processed, an accurate mapping relationship between the cell nucleus and the cytoplasm is not established, the cell boundary expansion is inaccurate, the reliability of the subsequent cell counting and the coexpression analysis is influenced, brightness difference, noise interference and spatial offset exist among different fluorescent channels, and the prior art often lacks systematic standardization, alignment and boundary optimization procedures, so that the stability and generalization capability of a segmentation result are insufficient, and the accuracy and the practicability of the cell image analysis are reduced. We therefore provide a cell segmentation method based on fusion of multi-channel fluorescent features. Disclosure of Invention In view of the above-mentioned drawbacks of the prior art, a first object of the present invention is to provide a cell segmentation method based on multi-channel fluorescence feature fusion, which solves the above-mentioned problems in the prior art. In order to achieve the above purpose, the present invention provides the following technical solutions: a cell segmentation method based on multichannel fluorescence feature fusion comprises the following steps: s1, carrying out standardization, noise removal and space alignment treatment on multi-channel image data; S2, designing a double-flow structure and attention mechanism model based on a deep convolutional neural network; S3, generating an accurate cell nucleus mask by using the DAPI signal and performing thresholding; S4, extracting a cytoplasmic region by using SpGreen channels and predicting a probability map; S5, combining the DAPI mask and the cytoplasmic area map, and executing accurate mask mapping; S6, optimizing cell boundaries and removing noise, and repairing cell morphology and segmentation accuracy; s7, evaluating the segmentation precision and optimizing the model based on IoU, the Dice coefficient and other indexes; s8, performing cell counting and coexpression analysis, and evaluating interaction among different channels; s9, generating a visual report of the segmentation result and data export, and supporting subsequent analysis and sharing. The invention is further arranged that in the step S1, the normalization, noise removal and spatial alignment processing of the multi-channel image data: S1.1, loading all four fluorescence channel data (DAPI, spGreen, spGold, CY) from imaging equipment, ensuring that the image data of each channel is consistent in format, denoising and standardizing, and unifying the image size and resolution; s1.2, normalizing the image of each channel to eliminate brightness difference among different channels. A common normalization method is Z-score normalization: Wherein, the For the gray value of each pixel,Is the mean value of the pixels and,Is the standard deviation of the pixel; S1.3, cutting the image into a concerned area for reducing the complexity of processing, and p