A method for cell instance segmentation of multiple channels of immunofluorescence multi-label domains for haec

Abstract

The application discloses a kind of cell instance segmentation methods for HAEC immunofluorescence multi-channel cross-marking field, it is related to the technical field of biomedical image processing and computer vision, the method is first constructed multi-channel segmentation model and executes field generalization training, subsequently extract the quality index of the measured multi-channel image and nuclear channel center seed, input model obtains initial mask chart, then based on multi-channel quality and morphological index screening low confidence mask, utilize boundary and skeleton channel effective pixel to implement discrete point set switching correction to defect edge, complete multi-source boundary evidence correction, utilize the local boundary band fitting cutting line generated by boundary channel one, reconstruct single cell mask by ray intersection point intercept mechanism, realize adhesion split, finally based on comprehensive quality score mining difficult example, realize closed loop training evolution;The application can solve the problems of missed detection, contour not fit and under-segmentation in HAEC image segmentation.

Description

Technical Field

[0001] This invention relates to the technical field of biomedical image processing and computer vision, and more specifically, to a method for cell instance segmentation using HAEC immunofluorescence multichannel cross-label domain. Background Technology

[0002] Cell instance segmentation is fundamental to biomedical image processing and batch quantitative analysis. Morphological parameters of human aortic endothelial cells (HAECs), such as area, perimeter, roundness, and aspect ratio, are crucial for vascular biology, drug screening, and pathological research. However, the dense growth, close proximity, morphological variability, and inconsistent scale among individual HAECs present significant challenges to automated cell instance segmentation.

[0003] Currently, existing cell image segmentation methods largely rely on general-purpose deep learning instance segmentation models (such as the cyto3 model in Cellpose 3.0). These general-purpose models typically preprocess the input image and directly output a cell mask, and are highly dependent on manually set single empirical scale parameters (such as fixed cell diameter) and general geometric priors. When dealing with batch analysis of HAEC immunofluorescence multichannel microscopy images, existing techniques suffer from the following significant drawbacks:

[0004] 1) Traditional models suffer from inaccurate identification and contour extraction, resulting in mask edges that do not conform to the real cell body. For example, under the F-actin (microfilament backbone) labeling channel of HAEC, the fluorescence signal mainly appears as stress fibers and cortical actin structures inside the cell, while the fluorescence signal at the actual cell edge is extremely weak, or even continuous with the background. General models cannot identify this specific biological structure and easily misjudge the coarse stress fiber texture inside the cell as the cell boundary, leading to the cell being "halved" or over-segmented. In addition, in areas with weak signals, broken boundaries, or thin protrusions, due to the lack of specific boundary evidence, the mask boundary output by the model deviates significantly from the real cell body contour, resulting in problems such as boundary shrinkage, outward expansion, truncation of thin protrusions, or over-simplification and flattening of the contour, making accurate conformation impossible.

[0005] 2) Inability to reliably resolve the adhesion problem in densely populated cell regions (i.e., undersegmentation); Endothelial cells are typically arranged in a single, tightly packed layer under physiological conditions, with cells closely packed and overlapping each other. Traditional general segmentation models lack prior constraints on biological structures (such as the lack of a rigid physical rule of "one nucleus, one instance"). When encountering discontinuous fluorescence expression at cell boundaries, diffusion caused by point spread function (PSF), or background noise interference, the fitted pixel gradient flow is easily confused. This leads to multiple adjacent cells often being incorrectly merged into a single mask, resulting in severe undersegmentation (multiple nuclei, one cell) and causing complete distortion in cell counting and instance division.

[0006] 3) Existing improved techniques do not effectively utilize multi-channel information, and even after secondary segmentation, the results remain inaccurate. To address the cell adhesion problem caused by the aforementioned deep learning models, some existing improved techniques attempt to perform secondary post-processing segmentation after adhesion is detected (e.g., when an excessively large mask is detected), but these approaches still have significant limitations:

[0007] Multichannel semantic loss and grayscale texture interference: Existing techniques often introduce multichannel merging mechanisms, but simply and crudely convert multichannel images into single-channel grayscale inputs. This leads to the complete loss of key biological channel semantics (such as the DAPI nuclear channel for locating the number of cells, the PECAM channel for locating suture boundaries, and the F-actin channel for locating the skeleton). Post-processing algorithms rely solely on the local texture or local extrema of the confused grayscale intensity for watershed segmentation, which is highly susceptible to fluctuations in local image brightness (uneven illumination) or the influence of messy fiber textures within cells, resulting in incorrectly drawn dividing lines. Blind proportional segmentation at the center: Some techniques, after detecting adhesion, blindly perform proportional mechanical segmentation or geometric cutting only at the geometric center of irregular connected regions. This method completely ignores real biological boundary evidence, disregarding the varying sizes and scale inconsistencies among individual cells within the same field of view. This results in mechanical and rigid dividing lines that cannot accurately conform to real cell boundaries. Summary of the Invention

[0008] The purpose of this invention is to provide a cell instance segmentation method and training closed-loop system for HAEC immunofluorescence multi-channel cross-label domain, in order to solve the problems of existing general segmentation models and methods in HAEC images, such as difficulty in separating adhered cells, weak / broken boundaries, semantic confusion of channels, or contour misfitting caused by mechanical geometric segmentation. This reduces the workload of manual correction, improves the segmentation accuracy of one cell per mask, and enhances the repeatability of batch quantitative analysis.

[0009] The technical solution of this invention is: to provide a method for cell instance segmentation across labeling domains in HAEC immunofluorescence multichannel, the method comprising:

[0010] Step 1: Construct a multi-channel cell instance segmentation model for human aortic endothelial cells (HAEC) and perform domain generalization enhancement training;

[0011] Step 2: Acquire the multi-channel immunofluorescence image of the HAEC to be tested, preprocess the multi-channel immunofluorescence image of the HAEC to be tested, and calculate the multi-channel quality index corresponding to the preprocessed image, including signal-to-noise ratio, saturation ratio and background variance.

[0012] Step 3: Extract the channel image corresponding to the nuclear channel from the multi-channel immunofluorescence image of the HAEC to be tested, set the center point of the cell nucleus in the foreground region of the channel image as the seed, and then input the multi-channel immunofluorescence image of the HAEC to be tested into the multi-channel cell instance segmentation model to obtain the initial cell instance mask image.

[0013] Step 4 involves fusing multi-source boundary evidence and separating adhesions from the initial cell instance mask image, specifically including:

[0014] Step 41: For each instance mask in the initial cell instance mask image, extract morphological indicators based on its edge pixel set, and calculate the local quality score of the instance mask by combining the edge pixel set with the multi-channel quality indicators corresponding to the current initial cell instance mask image. Then, calculate the boundary comprehensive confidence of the instance mask based on the morphological indicators and the local quality score.

[0015] Step 42: Modify the boundary of instance masks with a boundary comprehensive confidence score lower than a preset confidence score threshold using multi-source evidence. Generate connected components in the neighborhood of the instance mask using effective pixels from boundary channel 1 and skeleton channel. Extract the boundary pixels of the connected components as a calibration edge set. Calculate the shortest distance from the pixels in the calibration edge set to the edge of the current instance mask. Replace the point sequence segments on the edge of the instance mask whose shortest distance exceeds a preset distance threshold using the pixels in the calibration edge set. Smooth the replacement results to obtain the cell instance mask map with modified boundary.

[0016] Step 43: For the modified cell instance mask, the instance masks with more than 1 seed covering are taken as the adherent cell group. For any two adjacent cells in each adherent cell group, a local boundary band is generated using the effective pixels of boundary channel 1 and boundary channel 2 between them. The boundary lines corresponding to the two cells are fitted based on the local boundary band, and the two cells are segmented using the two boundary lines. The instance mask is reconstructed based on the segmentation results to obtain the final cell instance mask.

[0017] Step 5: Calculate the overall quality score for the final cell instance mask image. If the overall quality score is lower than the preset quality threshold, it is manually labeled and modified for closed-loop evolution training of the multi-channel cell instance segmentation model.

[0018] Furthermore, step 2 specifically includes:

[0019] The multi-channel immunofluorescence images of the HAEC to be tested include nuclear channels, boundary channel I, boundary channel II, and skeleton channels; and independent single-channel images I for k=2, 4. k k=2 represents the boundary channel 1, k=4 represents the skeleton channel, and their corresponding signal-to-noise ratios (SNR) are calculated in parallel. k , saturation ratio R sat,k and background variance .

[0020] Furthermore, step 41 specifically includes: combining the current instance mask C i Corresponding edge pixel set and multi-channel quality index SNR k R sat,k , Calculate the current instance mask C i The corresponding local quality score is expressed as:

[0021] ;

[0022] In the formula, For the current instance mask C i Local quality score in the k-th channel. , and These are preset empirical weighting coefficients. For the k-th channel in instance mask C i edge Average gradient magnitude within a 5-pixel outward neighborhood; based on morphological metrics and local quality scoring. The boundary comprehensive confidence of the instance mask is calculated and expressed as:

[0023] ;

[0024] In the formula, For the current instance mask C i The corresponding boundary comprehensive confidence level, For morphological indicators, , For the current instance mask C i Local quality scores in channels 2 and 4. It is a dynamic adjustment factor.

[0025] Furthermore, step 42 specifically includes:

[0026] Instance masks with a boundary comprehensive confidence score lower than a preset confidence threshold in the initial cell instance mask image are selected as instance masks to be corrected. For the j-th instance mask to be corrected, C... jThe region extending E pixels outward from its edge is taken as its neighborhood. In the corresponding HAEC immunofluorescence multi-channel image, the first local interest region corresponding to the current instance mask to be corrected and its neighborhood is extracted and denoted as ROI. j In ROI j Internal extraction of boundary channels and grayscale matrix With skeleton channel grayscale matrix Set channel segmentation thresholds respectively and Pixels with grayscale values ​​greater than the corresponding thresholds in both channels are identified as valid biological target points, while the remaining pixels are identified as background, based on the Region of Interest (ROI). j All valid target points within the ROI j Internally construct a unified binary target feature map;

[0027] Introduce a square structural element S with a preset size of 3×3. 3×3 The target feature map is sequentially subjected to morphological dilation and erosion operations to construct a regular closed feature connected region. The set of pixels at the outer edge of this connected region is extracted and defined as the calibration edge set B. modify It contains A discrete edge points arranged according to spatial adjacency, denoted as B. modify ={P a ( )}, a=1,2,...,A,P a To calibrate the a-th edge point in the edge set, ( P is the b-th edge point. a The horizontal and vertical coordinates.

[0028] Furthermore, step 42 also includes:

[0029] Extract the current instance mask C j The original edge set, denoted as B init ={q b ( )},b=1,2,...,B,q b Let b be the b-th edge point in the original edge set. ) represents the b-th edge point q b The x and y coordinates; for the calibration edge set B modify Each edge point P in a Calculate its distance to the original edge set B init The shortest distance D(P) along the direction perpendicular to the geometric normal a B init ); The preset distance threshold is e, and the calibration edge set B is traversed. modify If a continuous subsequence of edge points satisfies D(P) a Binit If )>e, then the sequence of edge points in that segment is determined to be a valid target segment B to be replaced. sub , represented as B sub = r is a miniature of the edge points in this subsequence of edge points. To determine the starting endpoint index of the target segment to be replaced within the calibration edge set, The index of the terminating endpoint of the target segment to be replaced in the calibration edge set.

[0030] Furthermore, step 42 also includes:

[0031] Using the shortest distance mapping relationship in the original edge set B init Find the starting endpoint and termination endpoint The two closest corresponding anchor points are denoted as the original edge replacement starting points. and the original edge replacement endpoint , The index of the original edge replacement starting point. The index of the original edge replacement endpoint, based on and Determine the anchoring interval, and then use the original edge set B init Using the clockwise direction as a reference, place B init Zhong Cong arrive All original edge points in the middle are removed, and the valid target segment B to be replaced is removed. sub The edge points are inserted and covered into the anchoring interval in the same clockwise order to obtain the initial mixed edge point set B. temp For the initial set of blending edge points B temp Smooth the edges of the splice, especially at the joints. , If two points coincide, remove either one; if the two points are adjacent, do nothing; if the two points do not coincide and are not adjacent, connect the two points with a straight line, and add the points on the straight line as edge points to the initial mixed edge point set B. temp At the splicing point [ , Perform the same process to obtain the smoothed set of cell edge points B'. temp .

[0032] Furthermore, step 43 specifically includes:

[0033] Traverse each instance mask in the modified cell instance mask map, and identify instance masks with more than 1 seed as adherent cell groups. For each adherent cell group, first in the current large instance mask region C... *Within the large instance mask, adjacent cells are partitioned: Let S = { H seed coordinate point sets contained within the large instance mask.} For each seed in the seed coordinate set, establish a system with itself as the center and R as the root. s Given a coarse topological control domain with radius , if the control domains corresponding to two seeds intersect spatially, then these two seeds are considered as a pair of adjacent cells, and the intersecting region is considered as the second local region of interest (ROI). g , g is the index of the second local region of interest within the current large instance mask; for each identified pair of adjacent cells, in the ROI g Internal extraction of boundary channels and grayscale matrix grayscale matrix with boundary channel Set a uniform channel segmentation threshold for boundary channel 2. The grayscale matrices in the two channels are binarized and segmented to extract a binarized local boundary band composed of discrete biological target points.

[0034] Furthermore, step 43 also includes:

[0035] For ROI g Extract any local boundary zone within the cell and calculate the seed s of the two adjacent cells. f With s h The spatial azimuth angle of the connecting line is set with an angular tolerance range. Within the current local boundary zone, an initial centerline L is fitted based on points in the local boundary zone, passing through the boundary zone and extending within the angular tolerance range. cen .

[0036] Furthermore, step 43 also includes:

[0037] With the initial centerline L cen Centerline, in ROI g Inner extraction distance L cen For binarized effective target points within a predetermined distance range, a third-order curve equation is set. The extracted effective target point set is used to fit the third-order polynomial curve equation. Then, the curve corresponding to this equation is used as a reference, and points are moved closer to the first seed s. f The direction and proximity to the second seed h The direction is vertically shifted by a preset pixel distance, thereby explicitly generating two curves in space, defined as the first cell clipping line L. f With the second cell trimming line L h Complete the masking region C of the current large instance. * After obtaining the cell trimming lines of all adjacent cell pairs within C, for C *Calculate the relationship between any cell cut line in a single seed and the current large instance mask region C. * All intersections of the edge pixel set and the intersections of the current seed with other cell clipping lines are denoted as the first and second intersections. Then, starting from the intersection of the current cell clipping line and the line connecting the two adjacent cell seeds, the line is extended along both ends of the current cell clipping line. For any end, the first intersection with the closest Euclidean distance is taken as the true endpoint. The current cell clipping line is truncated at both ends using the true endpoints to form clipping line segments. All pixels on the clipping line segments corresponding to the current seed are assigned to its edge point set B. fin Simultaneously, taking the two first intersection points of the current seed as the starting endpoints, the large instance mask region C is... * The sequence of edge points located within the starting endpoint and surrounding the current seed in the edge pixel set is included in the edge point set B of the current seed. fin .

[0038] Furthermore, step 5 specifically includes:

[0039] For the z-th single-cell instance mask C in the final cell instance mask map z Extract the pixel set at the edge of its mask. ,based on Computational instance mask C z morphological irregularity factor RI z The preset threshold range for morphological irregularity of human aortic endothelial cells is [ ]. If instance mask C z The value of the morphological irregularity factor is in [ If the instance mask C is within the range, then its morphological index is determined to be 1. z The value of the morphological irregularity factor exceeds [ If ], then its morphological index is determined to be 0;

[0040] Based on the nucleus miss rate and cell morphology parameters within the entire cell instance mask image, the overall quality score Q of the current HAEC immunofluorescence multichannel image to be tested is calculated. image :

[0041] ;

[0042] In the formula, This represents the overall quality score of instance segmentation in the current cell instance mask image, with a value ranging from [0, 1]. This represents the number of seeds in the current cell instance mask image. The number of instance masks in the current cell instance mask map. , This is the preset weight adjustment coefficient.

[0043] The beneficial effects of this invention are:

[0044] First, the technical solution of this invention can utilize multi-channel pixel features across the labeling domain to perform multi-source evidence adaptive edge correction on cell instance masks that produce edge deviation and contour distortion through morphological reconstruction and edge point set discrete replacement. This solves the problems of weak / broken boundaries, channel semantic confusion, and contour misfitting in existing general segmentation models and methods in HAEC images. Specifically, this invention introduces multi-channel quality indicators for dynamic weighted scoring, combines cell morphological irregularity factors to screen instance masks to be corrected, and extracts multi-source biologically effective pixels from boundary channel 1 (DLL4) and skeleton channel (F-actin) within the local neighborhood of a single cell. Morphological closure operation reconstruction generates a high-confidence calibration edge set. Then, using a clockwise sequential point set discrete replacement method, the initially misfit defect point sequence segments are precisely replaced and the edges are smoothed. Finally, a final cell instance mask that highly matches the real physical features of the multi-channel is obtained. This invention effectively eliminates the defects of traditional general segmentation models when facing cross-label domain scenarios (such as F-actin microfilament skeleton channels), which misjudge the internal coarse fiber texture as the boundary due to the lack of biological structural priors, resulting in the cell being halved or over-segmented. It can automatically fill the fluorescence signal loss in weak signal, boundary breakage or thin protrusion areas, and achieve a high degree of fit between the mask edge and the macroscopic topological contour of the real cell body.

[0045] Secondly, the technical solution in this invention can also utilize the physical topological constraints of one nucleus per instance and multi-channel pixel features for explicit cropping, overcoming the defect of cell-dense region adhesion (undersegmentation) in images. Specifically, this invention uses the nuclear channel (DAPI) to extract the geometric centroid of the cell nucleus center as the instance anchoring seed, establishing a hard biological physical topological constraint of one nucleus per instance. For adhered cell groups with a seed coverage greater than 1, the invention uses the spatial azimuth angle and angle tolerance range of the connection between adjacent seeds to fit an explicit cropping method within the binary local boundary band composed of boundary channel one (DLL4) and boundary channel two (endothelial junction marker PECAM / CD31). The invention employs explicit clipping lines to dynamically segment and reconstruct instance masks of adherent cell groups. It addresses the problems of existing technologies that rely on single-channel grayscale images, leading to loss of biological semantics, susceptibility to uneven local illumination, and the rigid, mechanical nature of blindly centering segmentation, deviating from real boundaries. The introduction of multi-channel boundary bands and clipping lines adaptively accommodates the scale inconsistencies between individual endothelial cells, enabling segmentation at real cell physical boundaries regardless of whether cells are located on the periphery of clusters or completely embedded within the tissue. This reliably solves the problem of adhesion (undersegmentation) of densely packed endothelial cells, ensuring absolute accuracy in cell counting and batch quantitative analysis. Attached Figure Description

[0046] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0047] Figure 1 This is a schematic flowchart of a cell instance segmentation method for HAEC immunofluorescence multichannel across labeling domains according to Embodiment 1 of the present invention;

[0048] Figure 2 This is a fluorescence image of HAEC cells containing only boundary channel one according to Embodiment 2 of the present invention;

[0049] Figure 3 This is an example mask image processed using the second segmentation method according to Embodiment 2 of the present invention;

[0050] Figure 4 This is a fluorescence image of HAEC cells containing only the skeletal channel according to Embodiment 2 of the present invention;

[0051] Figure 5 This is an example mask image processed using the segmentation method of Scheme 1 according to Embodiment 2 of the present invention;

[0052] Figure 6 This is a fluorescence image of HAEC cells containing multiple channels according to Embodiment 2 of the present invention;

[0053] Figure 7 This is an example mask image processed using the segmentation method of the present invention according to Embodiment 2 of the present invention. Detailed Implementation

[0054] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0055] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0056] Example 1:

[0057] like Figure 1 As shown, this embodiment provides a method for cell instance segmentation across labeling domains in HAEC immunofluorescence multichannel imaging. The method includes:

[0058] Step 1: Construct a multi-channel cell instance segmentation model for human aortic endothelial cells (HAEC) and perform domain generalization enhancement training.

[0059] Specifically, a multi-channel cell instance segmentation model for HAEC is constructed using the cyto3 model (an open-source cell instance segmentation network) of Cellpose 3.0. Structurally, this model includes a sequentially cascaded feature input layer, an encoder module, a decoder module, and a prediction output layer. The feature input layer consists of a multi-channel parallel convolution operator and a feature remapping unit. The multi-channel parallel convolution operator is used to perform preliminary feature extraction across multiple channels of the HAEC immunofluorescence multi-channel image. The feature remapping unit is used to concatenate and fuse the preliminary extracted features to obtain a shallow feature map. The encoder module employs a dense convolutional neural network with residual connections (preferably using a U-Net architecture, containing a downsampling pipeline layer composed of multiple cascaded residual convolutional layers and max-pooling layers) to extract multi-scale spatial geometric features of HAEC cells in different channels from bottom to top based on the shallow feature map. During this process, the spatial resolution of the feature map is gradually halved (e.g., downsampling from the original size to 1 / 2, 1 / 4, 1 / 8, 1 / 4, 1 / 5, 1 / 6 ... / 16), while the number of channels gradually doubles (e.g., to 256 or 512 channels), ultimately obtaining high-order semantic features that can characterize the global morphology and physical topological relationship of cells; the decoder module consists of cascaded bilinear interpolation upsampling units and cascaded convolutional smoothing layers, used to gradually restore the spatial resolution of the feature map through step-by-step upsampling operations. Skip connections are set in each upsampling level. The skip connections cascade the high-dimensional features restored by the current upsampling level with the corresponding symmetrical level features with the same spatial resolution in the encoder module, and feed them into the convolutional smoothing layer for dimensionality reduction and smoothing processing, ultimately obtaining a dense high-dimensional feature map with spatial resolution completely restored to the same as the original input image and fused with global semantics and micro-edge boundary features; the prediction output layer contains three parallel 1×1 convolutional prediction heads, used to perform pixel-level classification and vector regression calculations in parallel through the three prediction heads, predicting and outputting the cell center probability map, horizontal gradient field (Flow X), and vertical gradient field (Flow Y) respectively.

[0060] It should be noted that, in order to achieve pixel-level aligned calculation of the target loss function, before entering the model training iteration, the manually labeled discrete cell instance mask ground truth (Mask) must first be converted into a continuous intermediate topological feature field. The conversion employs a distance transformation and gradient inverse generation algorithm based on the multi-source heat diffusion equation: firstly, the geometric centroid of each independent cell mask is used as an energy source, and a Gaussian kernel is used to isotropically diffuse towards the outer cell boundary, generating a pixel-level continuous cell center probability ground truth map; subsequently, the first-order partial derivatives of the probability ground truth map are calculated and normalized in a two-dimensional spatial coordinate system, thereby automatically generating the true horizontal gradient field (Flow X ground truth) and the true vertical gradient field (Flow Y ground truth) with perfectly aligned spatial pixels. The above conversion method is a common technique in this field and will not be elaborated further here.

[0061] It should also be noted that after the model completes inference, the cell center probability map, horizontal gradient field and vertical gradient field output by the predicted output layer can be externally post-processed to produce a cell instance mask map using the existing Euler integral particle dynamic tracking method. This will not be described in detail here.

[0062] Multiple HAEC immunofluorescence multichannel images from different experimental batches were obtained from an existing microscopic image database. These multichannel images are complete images, meaning that the same field of view simultaneously contains specific fluorescence signals from four channels: nuclear channel (DAPI), used to locate the cell nucleus; boundary channel 1 (DLL4), membrane-associated marker, used to provide evidence of patchy boundaries; boundary channel 2 (PECAM / CD31): endothelial junction marker, used to provide evidence of jigsaw puzzle-like continuous boundaries; and scaffold channel (F-actin / phalloidin), used to provide constraints on the consistency of cell body extent and internal fibrous structure. To eliminate uneven illumination and background noise caused by differences in microscopes and exposure times among different batches, the acquired multichannel images underwent uniform domain normalization preprocessing. First, the rolling ball algorithm was used to estimate and subtract the low-frequency features of unevenly dark backgrounds in the images. Then, 1%-99th quantile intensity stretching was performed to map the gray levels of all image pixels to a uniform standard dynamic range, eliminating contrast differences and forcibly correcting images from different batches to the same feature distribution baseline. For each batch of images after normalization, multi-domain obfuscation enhancement processing is performed. That is, for the images in a given batch, non-uniform noise and blurring interference are introduced through an explicit algorithm. The processed noisy images are then paired with the corresponding original images to serve as sample images. All sample images are manually labeled to form the corresponding cell instance mask ground truth.

[0063] To prevent overfitting in a single batch and ensure the generalization representativeness of the validation set, the sample images were grouped according to experimental batches. A stratified random sampling strategy was adopted, and 0.7% of the sample images (all sampled images are image pairs containing a noisy image and its corresponding original image) were extracted from each group (i.e., each independent batch) as the training set, with the remaining 0.3% serving as the validation set. The multi-channel cell instance segmentation model was iteratively trained using the training set images. Each input consisted of an image pair containing a noisy image and its corresponding original image (the model's expanded batch size can be set to 1-4; when it is 4, the input tensor actually contains 8 images, i.e., four original images and their corresponding noisy images). During the training iterations, the model optimized the target loss function L. F Apply constraints, L F In addition to including binary cross-entropy loss and pixel prediction gradient flow mean squared error, this paper also introduces a custom prediction consistency loss formula to force the model to ignore noise and ambiguity, learn the essential form independent of batch style, and optimize the objective loss function L. F Represented as:

[0064] ;

[0065] In the formula, For binary classification, cross-entropy loss, The mean square error of the gradient flow for pixel prediction. To predict consistency loss, , , These are the weights corresponding to the binary classification cross-entropy loss, the mean square error of the pixel prediction gradient flow, and the prediction consistency loss; where... , , They are represented as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] In the formula, The binary cross-entropy loss value between the cell center probability map output by the characterization model and the cell center probability map corresponding to the manually labeled true membrane mask, where M is the total number of sample pixels involved in the calculation in the current training batch, and m is the spatial index of the pixel. Let y be the label value of the m-th pixel in the manually labeled ground truth (if the pixel belongs to the cell interior, then y is the label value of the m-th pixel in the ground truth). m =1, if it belongs to the background, then y m =0), This is the predicted probability value (range 0-1) of the m-th pixel belonging to the cell, output by the first prediction head of the prediction layer. The mean squared error loss value characterizes the difference between the predicted gradient flow field and the actual cell pixel flow topology field. and Let represent the centripetal convergence vectors of the m-th pixel, calculated from the manually labeled ground truth, along the horizontal (X-axis) and vertical (Y-axis) directions, respectively. and These represent the horizontal and vertical predicted gradient flow vector values ​​of the m-th pixel, respectively, output in parallel by the second and third prediction heads of the prediction output layer. This characterizes the consistency deviation of the intrinsic mathematical features of the network's predicted output when faced with a clear original image and a degraded noisy image for the same cell sample. K represents the total number of paired image samples in the current batch, k is the group index of the paired image sample, J is the total number of pixels in a single image, and j is the pixel spatial index in a single image. The feature input value at the j-th pixel of the original image that is not contaminated by noise in the k-th sample group. The feature input value at the j-th pixel of the noise-enhanced image in the k-th sample group. This represents the forward propagation operator of the model.

[0070] In this embodiment, the weight is preferably set to... =1.0、 =5.0、 =0.5, used to balance the parameter update step size between probabilistic classification, gradient field regression and noise resistance consistency.

[0071] In each epoch of training, the training set images are input into the model to complete the forward propagation and the optimization target loss function L is calculated. F (Total loss), then based on L F The gradients of the parameters of each layer of the neural network are calculated using the backpropagation algorithm. A pre-defined optimizer (such as the AdamW optimizer) is then used to update the network parameters via gradient descent. After each training iteration, the accuracy is verified using validation set images, and the validation set loss value L is calculated. F When the loss fluctuation of the validation set stabilizes within the preset minimum error threshold range (indicating that the model no longer overfits or underfits), or when the number of training iterations reaches the preset maximum number of training iterations, the model is determined to have reached the global optimum and converged. Training is stopped, the current network parameters are saved, and the pre-trained multi-channel cell instance segmentation model is obtained.

[0072] In this embodiment, non-uniform noise and blurring interference are introduced through an explicit algorithm. This can be achieved by randomly applying Gaussian / Poisson noise to the original image and using a preset multi-scale Gaussian kernel to perform spatial convolution modulation on the image, in order to simulate the point spread function (PSF) degradation behavior caused by optical distortion in a microscopic imaging system. This processing method is a common image processing method and will not be described in detail here.

[0073] Step 2: Acquire the multi-channel immunofluorescence image of the HAEC to be tested, preprocess the multi-channel immunofluorescence image of the HAEC to be tested, and calculate the multi-channel quality index corresponding to the preprocessed image, including signal-to-noise ratio, saturation ratio and background variance, for subsequent adaptive fusion weight calculation.

[0074] Specifically, the HAEC immunofluorescence multichannel image to be tested is of the same type as the HAEC immunofluorescence multichannel image used as the sample in step 1. It is also a complete image with all channels, including the nuclear channel (DAPI), boundary channel one (DLL4), boundary channel two (PECAM / CD31), and backbone channel (F-actin / phalloidin). The HAEC immunofluorescence multichannel image to be tested is subjected to normalization preprocessing, that is, abnormalities are processed by the rolling ball algorithm and intensity stretching of 1%-99% quantile. This normalization preprocessing process is the same as the normalization preprocessing process in step 1, and will not be described in detail here.

[0075] To provide dynamic weighting for subsequent adaptive fusion of multi-source boundary evidence, a quantitative evaluation of the imaging quality of each independent channel in the HAEC immunofluorescence multi-channel image under test is required. In the system, the image under test is represented as a high-dimensional multi-channel tensor I containing kernel channels, boundary channel one, boundary channel two, and skeleton channels. The system decouples and segments the multi-channel tensor I along the channel dimension in the algorithm's memory into four spatially aligned two-dimensional single-channel grayscale matrices, defining each as the k-th channel image I. k k {1,2,3,4}, k=1 represents the kernel channel, k=2 represents boundary channel one, k=3 represents boundary channel two, k=4 represents the skeleton channel; for the independent single-channel image I when k=2,4 k The corresponding signal-to-noise ratio (SNR) is calculated in parallel. k , saturation ratio R sat,k and background variance Signal-to-noise ratio (SNR) k Represented as:

[0076] ;

[0077] In the formula, This represents the signal-to-noise ratio (SNR) value of the k-th channel image. To represent the average grayscale value of all pixels within the cell body foreground region in the k-th channel, this cell body foreground region is segmented by the system using a preset grayscale threshold. is the average grayscale value of all pixels within the pure background region in the k-th channel. is the standard deviation of the gray values ​​of all pixels in the pure background region in the k-th channel, used to characterize the intensity of the inherent imaging electronic noise in the background.

[0078] Saturation ratio R sat,k Represented as:

[0079] ;

[0080] In the formula, represents the proportion of fluorescent saturated pixels in the foreground region of the k-th channel image (range 0-1), and W and H are the horizontal width and vertical height pixel values ​​of the image under test, respectively. Let be the grayscale intensity value of the k-th channel image at pixel coordinates (x, y). The preset absolute threshold for fluorescence intensity saturation (e.g., for an 8-bit grayscale image). The preferred setting is 254 or 255. This is a binary mask labeling function for the foreground region. Its value is 1 if the coordinates (x, y) belong to the cell body foreground region, and 0 otherwise. Let be the total number of pixels in the foreground region of the cell body in the k-th channel. = , This represents a mathematical indicator function. The function outputs 1 when the condition within the parentheses is met, and 0 otherwise.

[0081] Background variance Represented as:

[0082] ;

[0083] In the formula, The gray-level variance value representing the pure background region of the k-th channel image is used to quantitatively characterize the power of high-frequency random noise. Let k be the set of spatial coordinates of pure background pixels in the k-th channel. Within the pure background coordinate set The total number of background pixels included.

[0084] Finally, we obtain the independent single-channel images I when k=2,4. k The corresponding channel quality indicators are summarized to form multi-channel quality indicators.

[0085] It should be noted that the rolling ball algorithm and the 1%-99th quantile intensity stretching are commonly used algorithms in image processing. The rolling ball algorithm uses a structuring element with a preset radius to roll below the three-dimensional grayscale surface of the image. It dynamically estimates and subtracts low-frequency uneven background illumination interference through nonlinear morphological opening operations. The 1%-99th quantile intensity stretching, on the other hand, statistically analyzes the grayscale histogram of all pixels in the image, truncates extreme noise at both ends, and performs linear mapping on the core grayscale range in the middle to broaden the contrast of the effective features of the cell body. The processing procedures of these two algorithms will not be described in detail here.

[0086] Step 3: Extract the channel image corresponding to the nuclear channel (DAPI) in the multi-channel immunofluorescence image of the HAEC to be tested, set the center point of the cell nucleus in the foreground region of the channel image as the seed, and then input the multi-channel immunofluorescence image of the HAEC to be tested into the trained multi-channel cell instance segmentation model to obtain the initial cell instance mask map of the image to be tested.

[0087] Specifically, a two-dimensional single-channel grayscale matrix corresponding to the nuclear channel is extracted from the multi-channel tensor corresponding to the HAEC immunofluorescence multi-channel image to be tested. The optimal segmentation threshold is calculated based on the local grayscale gradient of the grayscale matrix using a local adaptive threshold operator. The grayscale matrix is ​​then transformed into a primary binary mask image with the cell nucleus as the foreground and the non-nucleus region as the background using the optimal segmentation threshold. Then, morphological closing operation and hole filling algorithm are performed on the primary binary mask image to eliminate the empty dark spots inside the cell nucleus caused by uneven staining and to remove the small high-frequency noise patches in the background caused by antibody impurities. The two-dimensional connected component labeling algorithm is then used to identify all the mutually discrete two-dimensional pixel patches representing independent cell nuclei. Finally, each patch is assigned a unique index label representing the cell nucleus, and the geometric centroid coordinates of each independent cell nucleus patch are calculated. The geometric centroid coordinates are set as a seed corresponding to a single cell instance so that a "one nucleus, one instance" hard topological constraint can be established using the seed, which serves as the core physical criterion for determining whether the cell bodies are adhered and whether forced splitting is required.

[0088] The multi-channel HAEC immunofluorescence image to be tested is input into the trained multi-channel cell instance segmentation model. Feature extraction is performed using the multi-channel cell instance segmentation model to obtain cell center probability map, horizontal gradient field and vertical gradient field result data. The results are then post-processed (Euler integral particle dynamic tracking method) to generate an initial cell instance mask map.

[0089] It should be noted that if the image does not have a DAPI channel, i.e. the staining is ineffective or the effective cell nuclear boundary cannot be extracted, the pseudonucleus center can be calculated as the instance anchoring seed by performing a distance transformation on the internal space of the cell and retrieving the local maxima (morphological geometric center peak).

[0090] In this embodiment, a multi-channel cell instance segmentation model is first used, which can adaptively output a preliminary cell center probability map and an initial mask (geometric contour). Through the forward feature extraction capability of the deep learning network, the macroscopic topological range of the cell body can be quickly and extensively delineated.

[0091] Step 4: Perform multi-source boundary evidence correction and adhesion separation on the initial cell instance mask map.

[0092] To address the inherent mask boundary misalignment (rough and distorted boundary drawing) and undersegmentation defects caused by multi-cell adhesion in general deep learning models in cross-labeled domain scenarios, this embodiment designs an adaptive boundary correction and topological constraint hard splitting algorithm driven by local image features, which specifically executes the following sub-steps.

[0093] Step 41: For each instance mask in the initial cell instance mask image, extract morphological indicators based on its edge pixel set, and calculate the local quality score of the instance mask by combining the edge pixel set with the multi-channel quality indicators corresponding to the current initial cell instance mask image. Then, calculate the boundary comprehensive confidence of the instance mask based on the morphological indicators and the local quality score.

[0094] Specifically, let C be the mask of the i-th instance in the initial cell instance mask map. i i {1,2,...,N}, where N is the number of instance masks in the current initial cell instance mask map, and for instance mask C... i Extract the pixel set at the edge of its mask. (That is, directly extracting the pixels corresponding to the mask edges from the image), based on Computational instance mask C i morphological irregularity factor RI i To quantify the regularity of the spatial morphology of this instance mask, it is expressed as:

[0095] ;

[0096] In the formula, Representative example mask C i area, Represents the perimeter of the edge pixel set; sets the morphological irregularity threshold range for human aortic endothelial cells (HAEC). If instance mask Ci Morphological irregularity factor The value is within the threshold range [ If the instance mask C is within the range, then its morphological index is determined to be 1. i Morphological irregularity factor The value exceeds the threshold range. If ], then its morphological index is determined to be 0.

[0097] In this embodiment, a predetermined number of standard contour samples of known human aortic endothelial cells (HAECs) can be obtained in advance from a cell sample database, and the morphological irregularity factor of each standard contour sample can be calculated. Then, based on the statistical distribution characteristics of the predetermined number of morphological irregularity factors, the morphological irregularity threshold range is determined. Morphological indices are used to evaluate the anatomical similarity between the geometric topology of the instance mask and the morphology of real human aortic endothelial cells. If the morphological index is 0, it indicates that the morphology of the current instance mask has deviated or has segmentation defects and does not conform to the inherent morphological characteristics of real human aortic endothelial cells. In this case, subsequent boundary evolution or mask correction procedures will be triggered to dynamically adjust and optimize the boundary of the current instance mask.

[0098] The multi-channel quality metrics corresponding to the current initial cell instance mask image include signal-to-noise ratio (SNR). k , saturation ratio R sat,k and background variance Where k=2,4, k=2 represents boundary channel one, and k=4 represents skeleton channel; combined with the current instance mask C i Corresponding edge pixel set and multi-channel quality index SNR k R sat,k , Calculate the current instance mask C i The corresponding local quality score is expressed as:

[0099] ;

[0100] In the formula, For the current instance mask C i Local quality score in the k-th channel. , and The preset empirical weighting coefficients are preferably set to [value]. =0.4、 =0.2、 =0.4, For the k-th channel in instance mask C i edge The average gradient magnitude within a 5-pixel outward neighborhood is used to characterize the actual boundary steepness provided by the channel in the current cell locality.

[0101] In this embodiment, the average gradient magnitude is obtained by extracting the image gradient operator and subtracting the local morphological mask (this method is commonly used, namely: first, the gradient components of the channel image in the orthogonal direction are calculated and synthesized into a global gradient magnitude map; then, the instance mask is morphologically dilated using a circular structuring element, and the difference between the dilated mask and the original instance mask is calculated to accurately peel out the annular local neighborhood mask that only covers the area 5 pixels outside the mask edge; finally, the total number of pixels in the annular neighborhood is counted, and the average gradient magnitude of all gradient magnitudes in the neighborhood is calculated). Existing technologies usually only focus on and stretch the global image quality, ignoring the spatial heterogeneity between individual cells and imaging channels. In this embodiment, the local boundary steepness (average gradient magnitude) is used as a dynamic feedback mechanism and organically introduced into the single-cell level comprehensive weight coefficient that integrates signal-to-noise ratio, saturation, and background variance. This achieves adaptive graded quality scoring for different imaging channels and different individual cells with strong heterogeneity. This graded and channel-specific local weighted mapping method can effectively solve the problem that it is difficult to balance the single-cell segmentation quality in complex imaging environments.

[0102] Based on morphological indicators and local quality scores The boundary comprehensive confidence of the instance mask is calculated and expressed as:

[0103] ;

[0104] In the formula, For the current instance mask C i The corresponding boundary comprehensive confidence level, For morphological indicators, , For the current instance mask C i Local quality scores in channels 2 and 4. As a dynamic adjustment factor, the preferred one is... Take 0.5.

[0105] In this embodiment, a multi-channel quality index is introduced to construct a multi-source channel comprehensive weight coefficient at the single-cell level, which adaptively weights and corrects morphological irregularity factors. This coefficient can be used to determine whether the initial mask conforms to the geometric features of real cells under complex factors such as cell overlap, uneven illumination, strong background noise, and artifacts. If the comprehensive weight coefficient is lower than a threshold (indicating local image quality degradation, high noise, and low neural network recognition accuracy), its boundary comprehensive confidence is dynamically lowered, and it is identified as a difficult case to be corrected to trigger subsequent mask correction procedures; if the comprehensive weight coefficient is higher than or equal to the threshold (indicating clear local image features and high recognition accuracy), its boundary comprehensive confidence is maintained or increased. In this mechanism, the weight coefficient has the ability to dynamically adjust locally, avoiding the local underfitting or overfitting problems caused by global stretching in traditional algorithms.

[0106] Step 42: For instance masks with a boundary comprehensive confidence score lower than a preset confidence threshold, perform multi-source evidence boundary modification. In the neighborhood of the instance mask, generate connected components using effective pixels from boundary channel 1 and skeleton channel. Extract the boundary pixels of the connected components as a calibration edge set. Calculate the shortest distance from the pixels in the calibration edge set to the current instance mask edge. Use the pixels in the calibration edge set to replace point sequence segments on the instance mask edge whose shortest distance exceeds a preset distance threshold. Smooth the replacement results to obtain the cell instance mask map with modified boundaries.

[0107] Specifically, instance masks with a boundary comprehensive confidence score lower than a preset confidence threshold (which can be set to 0.75) in the initial cell instance mask image are used as instance masks to be corrected. For the j-th instance mask to be corrected, C... j The region extending outwards from its edge by a distance of E pixels is taken as its neighborhood. E can be selected as an integer between 18 and 30 to ensure that the constructed neighborhood range is larger than the outer range of normal cells. The first local interest region corresponding to the mask of the instance to be corrected and its neighborhood is extracted from the corresponding HAEC immunofluorescence multi-channel image and denoted as ROI. j In ROI j Internal extraction of boundary channels and grayscale matrix With skeleton channel grayscale matrix Set channel segmentation thresholds respectively and Pixels with grayscale values ​​greater than the corresponding thresholds in both channels are identified as valid biological target points, while the remaining pixels are identified as background. To fuse boundary evidence from multiple channels and eliminate signal breakage defects across label domains, a ROI-based approach is used. j All valid target points within the ROI j Internally construct a unified binary target feature map F target (x,y) is represented as:

[0108] ;

[0109] In the formula, For ROI j The pixel space coordinates within.

[0110] Introduce a square structural element S with a preset size of 3×3. 3×3 For the target feature map F target (x,y) undergoes morphological dilation and erosion operations sequentially to construct a regular closed feature connected region. The set of pixels representing the outer edges of this connected region is extracted and defined as the calibration edge set B. modify It contains A discrete edge points arranged according to spatial adjacency (which is a set of closed edge pixels), denoted as B. modify ={P a ( )}, a=1,2,...,A,P a To calibrate the a-th edge point in the edge set, ( P is the b-th edge point. a The horizontal and vertical coordinates.

[0111] It should be noted that the closing operation, consisting of morphological dilation and erosion, is a commonly used operation in image processing, and its process will not be described in detail here. Modifying the instance mask using the calibration edge set corresponding to the feature connected components can eliminate local noise caused by the fiber structure and fill in minor breaks in the membrane channels; in this embodiment, the instance mask refers to the single mask corresponding to a cell.

[0112] Extract the current instance mask C j The original set of edges (which is the set of closed edge pixels) is denoted as B. init ={q b ( )},b=1,2,...,B,q b Let b be the b-th edge point in the original edge set. ) represents the b-th edge point q b The x and y coordinates; for the calibration edge set B modify Each edge point P in a Calculate its distance to the original edge set B init The shortest distance along the direction perpendicular to the geometric normal (i.e., the Euclidean geometric distance) D(P) a B init ), represented as:

[0113] ;

[0114] The preset distance threshold is e (preferably set to 3 to 5 pixels), and the calibration edge set B is traversed. modify If a continuous subsequence of edge points satisfies D(P) a B init If the value of the edge point sequence is greater than or equal to 'e', ​​then the model is considered to have suffered severe undersegmentation or misalignment in the corresponding local region (not matching the actual cell membrane edge), and this edge point sequence is determined as the valid target segment B to be replaced. sub , represented as B sub = r is a miniature of the edge points in this subsequence of edge points. The starting endpoint index of the effective target segment to be replaced in the calibration edge set (corresponding to the first point whose shortest distance exceeds the threshold in the subsequence of edge points of the segment). The index of the termination endpoint of the target segment to be replaced in the calibration edge set (corresponding to the last point whose shortest distance exceeds the threshold in the subsequence of edge points of the segment).

[0115] Using the shortest distance mapping relationship in the original edge set B init Find the starting endpoint and termination endpoint The two closest corresponding anchor points are denoted as the original edge replacement starting points. and the original edge replacement endpoint , The index of the original edge replacement starting point. The index of the original edge replacement endpoint, based on and Determine the anchoring interval, and then use the original edge set B init Using the clockwise direction (clockwise is used uniformly in this embodiment) as a reference, place B... init Zhong Cong arrive All original edge points in the middle are removed, and the valid target segment B to be replaced is removed. sub The edge points in the array are inserted and covered into the anchored interval (i.e., the break interval where the original edge points are removed) in the same clockwise order, resulting in the initial mixed edge point set B. temp This allows for the replacement of erroneous edges on the initial mask with precise biological morphological edges.

[0116] Since forced replacement of point sets may cause breaks or overlaps at the splicing endpoints due to coordinate jumps, in order to conform to the continuous characteristics of real endothelial cells, the initial mixed edge point set B is modified. temp Smooth the edges of the splice, especially at the joints. , If two points coincide, remove either one to avoid duplicate pixels. If the two points are adjacent, do nothing. If the two points do not coincide and are not adjacent, connect the two points with a straight line, and add the points on the line as edge points to the initial blended edge point set B. temp At the splicing point [ , Perform the same process to obtain the smoothed set of cell edge points B'. temp (It is a set of closed edge pixels). Then, a polygon region filling algorithm is used to fill the final set of cell edge points B'. temp All pixels inside are rendered in a paint bucket style foreground (set to 1), thereby generating a modified final cell instance mask in the algorithm's memory that closely matches the real physical features of the multichannel array.

[0117] Step 43: For the modified cell instance mask, the instance masks with more than 1 seed covering are taken as the adherent cell group. For any two adjacent cells in each adherent cell group, a local boundary band is generated using the effective pixels of boundary channel 1 and boundary channel 2 between them. The boundary lines corresponding to the two cells are fitted based on the local boundary band, and the two cells are segmented using the two boundary lines. The instance mask is reconstructed based on the segmentation results to obtain the final cell instance mask.

[0118] Specifically, each instance mask in the modified cell instance mask map is traversed, and instance masks with more than 1 seed are identified as adherent cell groups. To resolve the adhesion problem between adjacent cells within each group, for each adherent cell group, the current large instance mask region C is first... * Within the large instance mask, adjacent cells are partitioned: Let S = { H seed coordinate point sets contained within the large instance mask.} For each seed in the seed coordinate set, establish a system with itself as the center and R as the root. s Given a coarse topological control domain (i.e., a neighborhood region in space) with radius , if the control domains corresponding to two seeds intersect in space, then these two seeds are considered as a pair of adjacent cells, and the area where they intersect is considered as the second local region of interest, denoted as ROI. g , g is the index of the second local region of interest within the current large instance mask; for each identified pair of adjacent cells, in order to perform shearing at its actual biological interface, in the ROI g Internal extraction of boundary channels and grayscale matrix grayscale matrix with boundary channel Set a uniform channel segmentation threshold for boundary channel 2. The grayscale matrices in the two channels are binarized and segmented, with grayscale values ​​greater than 100%. The pixels identified as valid biological target points are determined, while the remaining pixels are considered background, within the channel segmentation threshold. Under the constraints, ROI g A binary local boundary band composed of discrete biological target points is extracted from the interior.

[0119] Since the pixels of boundary channel 2 (PECAM / CD31) are continuous boundary markers in a jigsaw puzzle shape, they will present a bright band with a certain width at the junction of two cells, and they are attached together with the pixels of boundary channel 1 (DLL4) of the two adjacent cells, which can be used together as a pixel band for cell segmentation. When one channel pixel is not obvious, the other channel pixel can be used as a substitute.

[0120] For ROI g Extract any local boundary zone within the cell and calculate the seed s of the two adjacent cells. f With s h The spatial azimuth angle of the connecting line (calculated based on the coordinates of two points), and an angle tolerance range are set. (Preferably, the angle between the line connecting the two nuclei and the line connecting the nuclei is between 60° and 120°, i.e., approximately perpendicular to the line connecting the nuclei). Within the current local boundary zone, an initial centerline L is fitted based on points in the local boundary zone, passing through the boundary zone and extending within the angle tolerance range. cen y=k cen x+b cen k cen b cen Represented as:

[0121] ;

[0122] ;

[0123] ;

[0124] In the formula, , Let x and y be the geometric centroid coordinates of the set of all valid target points within the current local boundary zone. This represents the slope of the free least squares fit without angular or directional constraints. For the initial centerline L cen The fitted slope value, For the initial centerline L cen The fitting intercept value, , The lower and upper limits of the preset angle tolerance range are neither 90° nor 270°; where, when When the denominator is 0, i.e., a degenerate vertical state occurs (where the denominator is 0), check if it is within the angular tolerance range. If it is within the range, then directly set the initial centerline L. cen For y=b cen , If not, then use [ Truncate the values ​​and calculate. or corresponding .

[0125] To completely separate the adhered pixels, use the initial center line L cen Using the center line, in the ROI g Inner extraction distance L cen Binarized valid target points within a predetermined distance range (i.e., a set of valid points with a pixel value of 1 extracted from their surrounding neighborhood) are identified. A third-order curve equation is then established, and the extracted valid target point set is fitted to the third-order polynomial curve equation (specifically, a least-squares fitting algorithm can be used, which can be achieved by constructing a least-squares optimization objective function to minimize the sum of squared vertical distances from the target point set to the curve). The form of the third-order polynomial curve equation is: y(x) = a³x 3 +a2x 2 +a1x+a0, where a3, a2, a1, and a0 are the coefficients of the third-order curve fitting polynomial; using the curve corresponding to the equation as the baseline, respectively move towards the first seed s f The direction and proximity to the second seed h The direction is vertically shifted by a preset pixel distance (preferably 1-2 pixels) to explicitly generate two curves of the same shape in space, which are defined as the first cell clipping line L. f With the second cell trimming line L h Complete the masking region C of the current large instance. * After obtaining the cell trimming lines of all adjacent cell pairs within C, for C * For a single seed (corresponding to one cell) within a given instance, treat all its cell clipping lines as infinitely bidirectional curves. For any given cell clipping line, calculate its relationship with the current large instance mask region C. * All intersections of the edge pixel set and the intersections of other cell clipping lines corresponding to the current seed are denoted as the first intersection and the second intersection. Then, starting from the intersection of the current cell clipping line and the line connecting the two adjacent cell seeds, the line is extended along both ends of the current cell clipping line. For any end, the first intersection with the closest Euclidean distance is taken as the true endpoint (i.e., the first first intersection or the first second intersection). The current cell clipping line is truncated at both ends using the true endpoints to form clipping line segments. All pixels on the clipping line segments corresponding to the current seed are assigned to its edge point set B. finSimultaneously, taking the two first intersection points of the current seed as the starting endpoints, the large instance mask region C is... * The sequence of edge points located within the starting endpoint and surrounding the current seed from the edge pixel set (involving multiple paths, the sequence with the smallest sum of distances from each point in the sequence to the seed can be used for judgment) is added to the edge point set B of the current seed. fin (If there is no first intersection point, skip this step).

[0126] After traversing each seed in the above manner, the polygon region filling algorithm is then used to fill the final set of edge points B corresponding to a single seed. fin All pixels within the cell are re-rendered using a paint bucket approach (set to 1), resulting in a final single-cell instance mask in the algorithm's memory that exhibits precise physical boundary isolation, extremely low local deformation rate, and high consistency with real multi-channel microscopic evidence. Meanwhile, the original large instance mask region C... * All remaining unre-rendered points within the edge pixel set are cleared (set to 0) to serve as the background, completing the segmentation and obtaining the final cell instance mask map.

[0127] It should be noted that for a single seed, when it is located at the outer edge of the original large instance mask region, because part of its boundary is exposed on the outermost side of the cell cluster, two first intersections will inevitably occur (such as when two or three cells are chain-like or cluster-like). However, when four or more endothelial cells are densely arranged, and the seed being tested is completely enveloped by the outermost cell mask and is geometrically within the original large instance mask region, the clipping line corresponding to the seed will be intercepted by the clipping lines of adjacent cells as it extends outward, thus generating four internal inflection points formed by second intersections. In this case, there are no first intersections.

[0128] The curve intersection truncation mechanism constructed in this embodiment adaptively assigns the minimum distance neighborhood between the first and second intersection points to the truncation endpoint. Therefore, regardless of whether the cell is located on the periphery of a cluster or completely embedded within the tissue, it can seamlessly and completely reconstruct a single-cell mask by the internal self-overlapping closure of the truncated line segment (i.e., relying on the second intersection point to form a closed polygon). This topological description method based on dynamic truncation at dual intersection points completely eliminates the industry application pain points of traditional geometric segmentation, such as "miscutting the effective area inside the cell" or "leaving local small holes due to loss of direction" when extending lines, giving the entire segmentation system extremely high industrial generalization robustness and mathematical completeness.

[0129] In this embodiment, to address the undersegmentation problem caused by multi-cell adhesion due to dense cell arrangement, the traditional global image segmentation operator is no longer relied upon. Instead, the cell nucleus center seed extracted in step 3 is used for spatial alignment verification to accurately locate adjacent cell pairs. Furthermore, for cells determined to be adhered, local boundary band extraction and explicit geometric shearing reconstruction are performed using boundary channel one (DLL4) and boundary channel two (endothelial junction marker PECAM / CD31). This ensures that the mask edge of the reconstructed single-cell instance after segmentation closely matches the actual physical topological boundary between cells, significantly reducing morphological local distortion at the interface. This provides a deterministic subpixel-level mask basis for subsequent calculation of cell morphology factors and high-precision quantitative characterization of perimeter and area.

[0130] Step 5: Calculate the overall quality score for the final cell instance mask image. If the overall quality score is lower than the preset quality threshold, manually label and modify it, and then put it into the hard example pool so that it can be fed back into the multi-channel cell instance segmentation model for closed-loop evolution training.

[0131] Specifically, for the z-th single-cell instance mask C in the final cell instance mask diagram... z Extract the pixel set at the edge of its mask. ,based on Computational instance mask C z morphological irregularity factor RI z The preset threshold range for morphological irregularity of human aortic endothelial cells (HAEC) is [[ If instance mask C z The value of the morphological irregularity factor is in [ If the instance mask C is within the range, then its morphological index is determined to be 1. z The value of the morphological irregularity factor exceeds [ If ], then its morphological index is determined to be 0.

[0132] Based on the nucleus miss rate and cell morphology parameters within the entire cell instance mask image, the overall quality score Q of the current HAEC immunofluorescence multichannel image to be tested is calculated. image :

[0133] ;

[0134] In the formula, This represents the overall quality score of instance segmentation for the current cell instance mask image, with a value ranging from [0, 1]. The closer the value is to 1, the higher the segmentation quality. This represents the number of seeds in the current cell instance mask image. This represents the number of instance masks in the current cell instance mask image (after segmentation and reconstruction). Used to characterize the cell false negative or false positive rate of the system within the map area, where z is the traversal index of a single cell instance. , This is a preset weighting adjustment coefficient used to balance the contributions of quantity topological alignment and morphological regularity to quality assessment. In this embodiment, it is preferably set to... = 0.7、 = 0.3.

[0135] Set a global quality confidence threshold T. Q (Preferred setting: 0.65) For the current cell instance mask image, if T Q If the segmentation mask quality of the image meets the predetermined accuracy standard of scientific statistics, it is directly connected to the downstream analysis pipeline and the final cell mask result image is output. <T Q (This indicates that there are large areas of undetected nuclei, abnormal adhesion caused by fluorescence quenching, or severe boundary distortion within the image). The system determines that the segmentation mask quality of the image does not meet the predetermined accuracy standard of scientific statistics. The system triggers the hard case mining mechanism, which imports the cell instance mask image into the preset hard case storage pool for manual fine-tuning and annotation updates. This corrects the misdetection boundaries and fills in the undetected topology, thereby forming a high-confidence hard case incremental ground truth mask pair. This pair is then used together with a new batch of sample images as a training set for closed-loop evolutionary training of the multi-channel cell instance segmentation model.

[0136] It should also be noted that the final single-cell instance mask output in this embodiment completely eliminates the adhesion and under-segmentation caused by dense arrangement in terms of topological structure, and undergoes adaptive smoothing stitching based on dual-channel microscopic physical evidence in terms of edge geometry. Therefore, it can approximate and restore the macroscopic topological contour of HAEC cells in real physiological state to the greatest extent. This instance mask can be used for subsequent cell quantitative analysis and can also bring powerful closed-loop evolution effect to subsequent deep learning model training.

[0137] Example 2:

[0138] To verify the practical application effect of the method of the present invention, this embodiment extracted a portion of HAEC cell images for comparative experiments, and processed them using three different schemes: Scheme 1 is a traditional model recognition method based on a single channel (only processing fluorescence images containing boundary channel DLL4, the signal is distributed in patches / discontinuously, and the boundaries of densely celled areas are weak); Scheme 2 is a traditional model recognition method based on a single channel plus a secondary segmentation method (only processing fluorescence images containing cytoskeleton channels, showing cytoskeleton fluorescence signals and weak boundary features); Scheme 3 is the method of the present invention (simultaneously processing multi-channel fluorescence images containing nuclear channels, boundary channel 1, and cytoskeleton channels).

[0139] like Figure 2 The image shown contains only the fluorescence of boundary channel 1 (DLL4). Figure 3 The output of this scheme is shown in the figure. As can be seen from the figure, the recognition effect is poor, with a large number of cells being missed and edges not fitting properly. This is because the traditional model only relies on the preset fixed cell size for mechanical extraction. The input single-channel features are too simple and one-sided. The model cannot obtain the global biological semantics and cannot achieve accurate fitting at all.

[0140] like Figure 4 The image shown is a fluorescence image containing only the skeleton channel. Figure 5 The output image shows the results of this scheme; as can be seen, its recognition results are also unsatisfactory. Because the model cannot acquire global biological semantics, it easily misjudges the coarse stress fiber texture inside cells as cell boundaries, leading to cells being "halved" or over-segmented. Furthermore, at the edges, there are issues such as boundary shrinkage, outward expansion, truncation of slender protrusions, or over-simplification and flattening of contours, resulting in numerous missed detections and boundary deviations in the image. More importantly, due to the lack of cross-validation from multi-channel physical evidence, even subsequent secondary segmentation is merely a blind mechanical cutting at the geometric center, and the results still fail to accurately match the true cell boundary contours.

[0141] like Figure 6 The image shown is a multi-channel fluorescence image containing multiple channels. Figure 7 The output diagram of the method of the present invention is shown. By comparison, it can be seen that the recognition results of the method of the present invention are significantly better than the two traditional methods. Under the deep fusion of multi-channel features and collaborative constraints, the system can accurately detect each individual cell, greatly reducing the number of missed detections. After the adaptive multi-source boundary correction and physical topology splitting in step 4, it accurately achieves "one nucleus, one instance" in the anatomical sense. The instance mask of each cell can fit its real cell body contour perfectly, perfectly restoring the real physical boundary when endothelial cells are densely arranged.

[0142] The steps in this invention can be adjusted, combined, or deleted according to actual needs.

[0143] The units in the device of the present invention can be merged, divided, or reduced according to actual needs.

[0144] In this invention, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "linking" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of these terms in this invention according to the specific circumstances.

[0145] The shapes of the components in the accompanying drawings are schematic and may differ from their actual shapes. The drawings are only used to illustrate the principles of the present invention and are not intended to limit the present invention.

[0146] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The scope of protection of the invention is defined by the appended claims and may include various modifications, alterations, and equivalents made to the invention without departing from the scope and spirit of the invention.

Claims

1. A method for cell instance segmentation across labeling domains in HAEC immunofluorescence multichannel imaging, characterized in that, The method includes: Step 1: Construct a multi-channel cell instance segmentation model for human aortic endothelial cells (HAEC) and perform domain generalization enhancement training; Step 2: Acquire the multi-channel immunofluorescence image of the HAEC to be tested, preprocess the multi-channel immunofluorescence image of the HAEC to be tested, and calculate the multi-channel quality index corresponding to the preprocessed image, including signal-to-noise ratio, saturation ratio and background variance. Step 3: Extract the channel image corresponding to the nuclear channel from the multi-channel immunofluorescence image of the HAEC to be tested, set the center point of the cell nucleus in the foreground region of the channel image as the seed, and then input the multi-channel immunofluorescence image of the HAEC to be tested into the multi-channel cell instance segmentation model to obtain the initial cell instance mask image. Step 4 involves fusing multi-source boundary evidence and separating adhesions from the initial cell instance mask image, specifically including: Step 41: For each instance mask in the initial cell instance mask image, extract morphological indicators based on its edge pixel set, and calculate the local quality score of the instance mask by combining the edge pixel set with the multi-channel quality indicators corresponding to the current initial cell instance mask image. Then, calculate the boundary comprehensive confidence of the instance mask based on the morphological indicators and the local quality score. Step 42: Modify the boundary of instance masks with a boundary comprehensive confidence score lower than a preset confidence score threshold using multi-source evidence. Generate connected components in the neighborhood of the instance mask using effective pixels from boundary channel 1 and skeleton channel. Extract the boundary pixels of the connected components as a calibration edge set. Calculate the shortest distance from the pixels in the calibration edge set to the edge of the current instance mask. Replace the point sequence segments on the edge of the instance mask whose shortest distance exceeds a preset distance threshold using the pixels in the calibration edge set. Smooth the replacement results to obtain the cell instance mask map with modified boundary. Step 43: For the modified cell instance mask, the instance masks with more than 1 seed covering are taken as the adherent cell group. For any two adjacent cells in each adherent cell group, a local boundary band is generated using the effective pixels of boundary channel 1 and boundary channel 2 between them. The boundary lines corresponding to the two cells are fitted based on the local boundary band, and the two cells are segmented using the two boundary lines. The instance mask is reconstructed based on the segmentation results to obtain the final cell instance mask. Step 5: Calculate the overall quality score for the final cell instance mask image. If the overall quality score is lower than the preset quality threshold, it is manually labeled and modified for closed-loop evolution training of the multi-channel cell instance segmentation model.

2. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 1, characterized in that, Step 2 specifically includes: The multi-channel immunofluorescence images of the HAEC to be tested include nuclear channels, boundary channel I, boundary channel II, and skeleton channels; and independent single-channel images I for k=2, 4. k k=2 represents the boundary channel 1, k=4 represents the skeleton channel, and their corresponding signal-to-noise ratios (SNR) are calculated in parallel. k , saturation ratio R sat,k and background variance .

3. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 2, characterized in that, Step 41 specifically includes: combining the current instance mask C i Corresponding edge pixel set and multi-channel quality index SNR k R sat,k , Calculate the current instance mask C i The corresponding local quality score is expressed as: ； In the formula, For the current instance mask C i Local quality score in the k-th channel. , and These are preset empirical weighting coefficients. For the k-th channel in instance mask C i edge Average gradient magnitude within a 5-pixel outward neighborhood; based on morphological metrics and local quality scoring. The boundary comprehensive confidence of the instance mask is calculated and expressed as: ； In the formula, For the current instance mask C i The corresponding boundary comprehensive confidence level, For morphological indicators, , For the current instance mask C i Local quality scores in channels 2 and 4. It is a dynamic adjustment factor.

4. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 3, characterized in that, Step 42 specifically includes: Instance masks with a boundary comprehensive confidence score lower than a preset confidence threshold in the initial cell instance mask image are selected as instance masks to be corrected. For the j-th instance mask to be corrected, C... j The region extending E pixels outward from its edge is taken as its neighborhood. In the corresponding HAEC immunofluorescence multi-channel image, the first local interest region corresponding to the current instance mask to be corrected and its neighborhood is extracted and denoted as ROI. j In ROI j Internal extraction of boundary channels and grayscale matrix With skeleton channel grayscale matrix Set channel segmentation thresholds respectively and Pixels with grayscale values ​​greater than the corresponding thresholds in both channels are identified as valid biological target points, while the remaining pixels are identified as background, based on the Region of Interest (ROI). j All valid target points within the ROI j Internally construct a unified binary target feature map; Introduce a square structural element S with a preset size of 3×3. 3×3 The target feature map is sequentially subjected to morphological dilation and erosion operations to construct a regular closed feature connected region. The set of pixels at the outer edge of this connected region is extracted and defined as the calibration edge set B. modify It contains A discrete edge points arranged according to spatial adjacency, denoted as B. modify ={P a ( )}, a=1,2,...,A,P a To calibrate the a-th edge point in the edge set, ( P is the b-th edge point. a The horizontal and vertical coordinates.

5. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 4, characterized in that, Step 42 further includes: Extract the current instance mask C j The original edge set, denoted as B init ={q b ( )},b=1,2,...,B,q b Let b be the b-th edge point in the original edge set. ) represents the b-th edge point q b The x and y coordinates; for the calibration edge set B modify Each edge point P in a Calculate its distance to the original edge set B init The shortest distance D(P) along the direction perpendicular to the geometric normal a B init ); The preset distance threshold is e, and the calibration edge set B is traversed. modify If a continuous subsequence of edge points satisfies D(P) a B init If )>e, then the sequence of edge points in that segment is determined to be a valid target segment B to be replaced. sub , represented as B sub = r is a miniature of the edge points in this subsequence of edge points. To determine the starting endpoint index of the target segment to be replaced within the calibration edge set, The index of the terminating endpoint of the target segment to be replaced in the calibration edge set.

6. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 5, characterized in that, Step 42 further includes: Using the shortest distance mapping relationship in the original edge set B init Find the starting endpoint and termination endpoint The two closest corresponding anchor points are denoted as the original edge replacement starting points. and the original edge replacement endpoint , The index of the original edge replacement starting point. The index of the original edge replacement endpoint, based on and Determine the anchoring interval, and then use the original edge set B init Using the clockwise direction as a reference, place B init Zhong Cong arrive All original edge points in the middle are removed, and the valid target segment B to be replaced is removed. sub The edge points are inserted and covered into the anchoring interval in the same clockwise order to obtain the initial mixed edge point set B. temp For the initial set of blending edge points B temp Smooth the edges of the splice, especially at the joints. , If two points coincide, remove either one; if the two points are adjacent, do nothing; if the two points do not coincide and are not adjacent, connect the two points with a straight line, and add the points on the straight line as edge points to the initial mixed edge point set B. temp At the splicing point [ , Perform the same process to obtain the smoothed set of cell edge points B'. temp .

7. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 3, characterized in that, Step 43 specifically includes: Traverse each instance mask in the modified cell instance mask map, and identify instance masks with more than 1 seed as adherent cell groups. For each adherent cell group, first in the current large instance mask region C... * Within the large instance mask, adjacent cells are partitioned: Let S = { H seed coordinate point sets contained within the large instance mask.} For each seed in the seed coordinate set, establish a system with itself as the center and R as the root. s Given a coarse topological control domain with radius , if the control domains corresponding to two seeds intersect spatially, then these two seeds are considered as a pair of adjacent cells, and the intersecting region is considered as the second local region of interest (ROI). g , g is the index of the second local region of interest within the current large instance mask; for each identified pair of adjacent cells, in the ROI g Internal extraction of boundary channels and grayscale matrix grayscale matrix with boundary channel Set a uniform channel segmentation threshold for boundary channel 2. The grayscale matrices in the two channels are binarized and segmented to extract a binarized local boundary band composed of discrete biological target points.

8. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 7, characterized in that, Step 43 further includes: For ROI g Extract any local boundary zone within the cell and calculate the seed s of the two adjacent cells. f With s h The spatial azimuth angle of the connecting line is set with an angular tolerance range. Within the current local boundary zone, an initial centerline L is fitted based on points in the local boundary zone, passing through the boundary zone and extending within the angular tolerance range. cen .

9. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 8, characterized in that, Step 43 further includes: With the initial centerline L cen Centerline, in ROI g Inner extraction distance L cen For binarized effective target points within a predetermined distance range, a third-order curve equation is set. The extracted effective target point set is used to fit the third-order polynomial curve equation. Then, the curve corresponding to this equation is used as a reference, and points are moved closer to the first seed s. f The direction and proximity to the second seed h The direction is vertically shifted by a preset pixel distance, thereby explicitly generating two curves in space, defined as the first cell clipping line L. f With the second cell trimming line L h Complete the masking region C of the current large instance. * After obtaining the cell trimming lines of all adjacent cell pairs within C, for C * Calculate the relationship between any cell cut line in a single seed and the current large instance mask region C. * All intersections of the edge pixel set and the intersections of the current seed with other cell clipping lines are denoted as the first and second intersections. Then, starting from the intersection of the current cell clipping line and the line connecting the two adjacent cell seeds, the line is extended along both ends of the current cell clipping line. For any end, the first intersection with the closest Euclidean distance is taken as the true endpoint. The current cell clipping line is truncated at both ends using the true endpoints to form clipping line segments. All pixels on the clipping line segments corresponding to the current seed are assigned to its edge point set B. fin Simultaneously, taking the two first intersection points of the current seed as the starting endpoints, the large instance mask region C is... * The sequence of edge points located within the starting endpoint and surrounding the current seed in the edge pixel set is included in the edge point set B of the current seed. fin .

10. The cell instance segmentation method for HAEC immunofluorescence multichannel cross-label domain as described in claim 1, characterized in that, Step 5 specifically includes: For the z-th single-cell instance mask C in the final cell instance mask map z Extract the pixel set at the edge of its mask. ,based on Computational instance mask C z morphological irregularity factor RI z The preset threshold range for morphological irregularity of human aortic endothelial cells is [ ]. If instance mask C z The value of the morphological irregularity factor is in [ If the instance mask C is within the range, then its morphological index is determined to be 1. z The value of the morphological irregularity factor exceeds [ If ], then its morphological index is determined to be 0; Based on the nucleus miss rate and cell morphology parameters within the entire cell instance mask image, the overall quality score Q of the current HAEC immunofluorescence multichannel image to be tested is calculated. image : ； In the formula, This represents the overall quality score of instance segmentation in the current cell instance mask image, with a value ranging from [0, 1]. This represents the number of seeds in the current cell instance mask image. The number of instance masks in the current cell instance mask map. , This is the preset weight adjustment coefficient.

Images