A method of screening for liver stem cells

Abstract

This invention relates to the field of cell image analysis and biological cell screening technology, and discloses a method for screening liver stem cells. The method includes: acquiring original images; performing effective image screening; performing background correction, noise reduction and enhancement, and contrast compensation; performing single-cell instance segmentation; performing single-cell feature extraction and trajectory association; determining target cells using a neural network model; and outputting the liver stem cell determination result. Compared to existing screening methods that rely on manual observation or single-frame static image discrimination, especially in the case of primary hepatocyte suspensions containing mature hepatocytes, bile duct epithelial cells, stromal cells, and cell debris, this method addresses the technical problem of unstable and automated identification of liver stem cells. This application improves the accuracy and automation of liver stem cell screening by constructing a continuous processing flow of image quality gating, instance segmentation, and dual-branch feature fusion judgment.

Description

Technical Field

[0001] This invention relates to the field of cell image analysis and biological cell screening technology, and in particular to a screening method for liver stem cells. Background Technology

[0002] Currently, liver stem cell screening typically relies on manual microscopic examination, immunolabeling sorting, single morphological threshold discrimination, or coarse identification based on static images. For primary hepatocyte suspensions formed after enzymatic digestion of liver tissue, the samples often contain mature hepatocytes, bile duct epithelial cells, stromal cells, dead cells, and cell debris. Different cell types overlap in size, outline, grayscale distribution, and local texture, making it difficult to accurately distinguish liver stem cells based on a single feature. Furthermore, during bright-field microscopy, primary suspensions commonly exhibit cell overlap, localized defocusing, blurred edges, uneven background illumination, and slight cell drift, further increasing the difficulty of screening.

[0003] For example, when using manual observation or fixed threshold segmentation, preliminary identification can usually only be made based on cell area, brightness, or edge intensity. When there are cell debris attached, cell clusters, or discontinuous refractive edges in the sample, it is very easy to misidentify non-target cells as liver stem cells or to mistakenly exclude real target cells. When using single-frame image recognition, because the information on the positional changes and motion continuity of cells during continuous imaging is not combined, the existing technology cannot fully meet the needs for stable identification and accurate screening of liver stem cells under complex primary hepatocyte mixed suspension conditions.

[0004] Therefore, there is an urgent need for a method that can still achieve automatic identification and accurate screening of liver stem cells even in the presence of cell overlap, blurred focus, blurred edges, lighting fluctuations and debris interference in primary hepatocyte suspensions, in order to improve the accuracy, stability and automation level of liver stem cell screening. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the present invention aims to propose a method for screening liver stem cells. This method addresses the technical problem that existing screening methods, which rely on manual observation or single-frame static image discrimination, cannot achieve stable and automated identification of liver stem cells, especially when mature hepatocytes, bile duct epithelial cells, stromal cells, and cell debris are present simultaneously in a primary hepatocyte suspension.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for screening liver stem cells.

[0007] The method for screening liver stem cells includes:

[0008] Step S10: Obtain the original image, focal plane position, exposure time and light source intensity of the primary hepatocyte suspension formed after liver tissue digestion by enzyme in the bright field microscopy imaging area. Combine the image quality gating mechanism to perform the effective image screening task and output the effective image, image quality score and image quality threshold.

[0009] Step S20: Based on the valid image, image quality score, and image quality threshold, a quality correlation preprocessing mechanism is used to perform background correction, noise reduction and enhancement, and contrast compensation tasks, and output the preprocessed image;

[0010] Step S30: Based on the preprocessed image, a candidate detection and instance segmentation coupling mechanism is used to perform single-cell instance segmentation task, and the candidate cell region, single-cell instance mask and segmentation quality score are output.

[0011] Step S40: Based on candidate cell regions, single-cell instance masks, and segmentation quality scores, a joint modeling mechanism of static morphology and texture and temporal consistency is used to perform single-cell feature extraction and trajectory association tasks, and output static feature vectors and trajectory state parameters.

[0012] Step S50: Target cells are determined by combining static feature vectors and trajectory state parameters with a dual-branch feature fusion neural network model, and the liver stem cell determination result is output.

[0013] Preferably, step S10, which involves acquiring the original image, focal plane position, exposure time, and light source intensity of the primary hepatocyte suspension formed after enzymatic digestion of liver tissue in the bright-field microscopy imaging region, and performing an effective image screening task in conjunction with an image quality gating mechanism to output effective images, image quality scores, and image quality thresholds, specifically includes:

[0014] Step S101: Obtain the time Original image of a primary hepatocyte suspension formed after enzymatic digestion of lower liver tissue in a bright-field microscopy region. Focal plane position Exposure time and light source intensity And based on the original image Calculate local variance Based on local variance Extracting high-information pixel set : ;in, Represents the x-coordinate of a pixel; Represents the ordinate of a pixel; The local variance threshold is determined based on the image grayscale distribution of the original image;

[0015] Step S102: Based on the set of high-information pixels The edge sharpness component of the current image is calculated using gradient magnitude statistics. Based on high-information pixel sets The Laplace normalized components are calculated using the Laplace response variance normalization method. Based on high-information pixel sets The background uniformity component is calculated using a low-frequency background variance suppression method. Based on high-information pixel sets Foreground drift penalty components are calculated using a time-series drift penalty method. ; and based on the Laplace normalized components Background uniformity component and foreground drift penalty component Constructing image quality scores ;

[0016] Step S103: Score the image quality Constructing image quality thresholds ;

[0017] ;

[0018] in, The length of the history window; For a moment At that time Historical image quality score sequence; MAD is an adjustment coefficient used to adjust the responsiveness of the image quality threshold to historical fluctuations. The median absolute deviation function is used to characterize the degree of dispersion of a historical image quality score sequence relative to its median value; This is a median function used to characterize the central location of a historical image quality score sequence;

[0019] Step S104: Satisfy The original image is determined as a valid image, and the valid image, image quality score, and image quality threshold are output.

[0020] Preferably, step S20, which involves performing background correction, denoising enhancement, and contrast compensation tasks using a quality-correlated preprocessing mechanism based on the valid image, image quality score, and image quality threshold, and outputting the preprocessed image, specifically includes:

[0021] Step S201: Based on the valid images, cache a window of historical valid images according to a preset time frame. Online flat-field estimation is performed using low-frequency quantile statistical modeling to output the background model; dark-field images are also obtained. According to the dark field diagram Background correction was performed using a flat-field normalization correction method on the background model to obtain the background-corrected image. ;

[0022] Step S202: Based on image quality scoring Generate denoising intensity parameters And based on the denoising intensity parameter Background correction image Perform edge-preserving denoising to obtain a denoised image. ;in, , and These represent the minimum and maximum noise reduction strengths, respectively.

[0023] Step S203: Based on image quality scoring Generate contrast enhancement parameters And based on contrast enhancement parameters For denoised images Perform local contrast enhancement processing to obtain the preprocessed image. .

[0024] Preferably, in step S202, based on the denoising intensity parameter Background correction image During the edge-preserving denoising process, bilateral filtering and guided filtering are employed, along with image quality scoring. By controlling the strength of grayscale similarity constraints, noise suppression is improved when the image quality score is low, while more edge details are preserved when the image quality score is high. Among them, bilateral filtering and guided filtering are used to maintain a balance between cell contour continuity and background stability.

[0025] Preferably, step S30, which involves performing single-cell instance segmentation based on the preprocessed image using a candidate detection and instance segmentation coupling mechanism, and outputting candidate cell regions, single-cell instance masks, and segmentation quality scores, specifically includes:

[0026] Step S301: Based on the preprocessed image Constructing a difference response map with the background model : ;in, , This is a low-pass filtering operation used to extract low-frequency background components from the preprocessed image; and it is applied to the differential response map. Implementation scale is limited Candidate detection yields a set of candidate centers. Candidate cell regions are obtained by cropping around each candidate center in the candidate center set;

[0027] Step S302: Input the candidate cell region into the preset instance segmentation network, and the instance segmentation network outputs a foreground probability map;

[0028] Step S303: Obtain image quality score Based on image quality score Constructing a foreground probability threshold , , This is a truncation function used to limit the foreground probability threshold within a preset range; This is a threshold adjustment coefficient used to control the degree of influence of image quality score on the foreground probability threshold; and based on the foreground probability threshold... The foreground region is obtained by performing threshold segmentation on the foreground probability map;

[0029] Step S304: Perform distance transformation on the foreground region using Euclidean distance transformation to obtain a distance map; extract local maximum markers in the distance map according to the preset minimum radius constraint, and perform marker-controlled watershed segmentation based on the local maximum markers to obtain a single-cell instance mask;

[0030] Step S305: Construct a segmentation quality score based on the cross-union ratio between the single-cell instance mask and the preset standard instance mask template, and finally output the candidate cell region, single-cell instance mask and segmentation quality score.

[0031] Preferably, in step S40, the step of performing single-cell feature extraction and trajectory association tasks based on candidate cell regions, single-cell instance masks, and segmentation quality scores, using a joint modeling mechanism of static morphology and temporal consistency, and outputting static feature vectors and trajectory state parameters, specifically includes:

[0032] Step S401: Based on the single-cell instance mask, use Python's scikit-image library to perform region attribute measurement processing, and extract the single-cell instance area, single-cell instance perimeter, single-cell instance equivalent diameter, single-cell instance roundness, and single-cell instance eccentricity.

[0033] Step S402: Construct inner and outer ring regions based on candidate cell regions and single-cell entity masks, and calculate the refractive edge score based on the mean difference of grayscale in the ring region and the outer ring region, while extracting the cytoplasmic texture features of the candidate cell regions.

[0034] Step S403: Combine single-cell instance area, single-cell instance perimeter, single-cell instance equivalent diameter, single-cell instance roundness, single-cell instance eccentricity, refractive edge score, segmentation quality score and cytoplasmic texture features to generate a static feature vector;

[0035] Step S404: Obtain the historical trajectory state parameters and current trajectory state parameters of the center position of the candidate cell region. Based on the historical trajectory state parameters and current trajectory state parameters, perform trajectory association through Hungarian matching joint association to construct the trajectory state vector.

[0036] Preferably, step S50, which involves determining the target cell based on the static feature vector and trajectory state parameters combined with a dual-branch feature fusion neural network model and outputting the liver stem cell determination result, specifically includes:

[0037] Step S501: Construct a dual-branch feature fusion neural network model based on the multilayer perceptron branch and the gated recurrent unit branch; wherein, the multilayer perceptron branch is used to perform nonlinear mapping and high-dimensional feature extraction on the static feature vector to obtain a static representation vector representing the morphology, texture and refractive edge features of the candidate cells; the gated recurrent unit branch is used to perform temporal dependency modeling and dynamic change pattern extraction on the trajectory state parameters to obtain a temporal representation vector representing the continuity of the candidate cell motion and the trajectory evolution features;

[0038] Step S502: Obtain historical static feature vectors, historical trajectory state parameters, and cell category labels corresponding to historical samples. Use historical static feature vectors and historical trajectory state parameters as inputs to the dual-branch feature fusion neural network model, and use cell category labels corresponding to historical samples as outputs of the dual-branch feature fusion neural network model. Use the binary cross-entropy loss function to perform the pre-training process in a supervised learning manner.

[0039] Step S503: Input the static feature vector and trajectory state parameters into the pre-trained dual-branch feature fusion neural network model, and the dual-branch feature fusion neural network model outputs the liver stem cell determination result.

[0040] The present invention also provides a liver stem cell screening system comprising:

[0041] The image quality gating module is used to acquire the original image, focal plane position, exposure time and light source intensity of the primary hepatocyte suspension formed after liver tissue digestion by enzyme in the bright field microscopy imaging area. Combined with the image quality gating mechanism, it performs the effective image screening task and outputs the effective image, image quality score and image quality threshold.

[0042] The image quality gating module is used to perform background correction, noise reduction and enhancement and contrast compensation tasks based on the valid image, image quality score and image quality threshold, and output the preprocessed image using a quality correlation preprocessing mechanism.

[0043] The candidate detection and instance segmentation module is used to perform single-cell instance segmentation tasks based on preprocessed images using a candidate detection and instance segmentation coupling mechanism, and outputs candidate cell regions, single-cell instance masks and segmentation quality scores.

[0044] The feature extraction and trajectory association module is used to perform single-cell feature extraction and trajectory association tasks based on candidate cell regions, single-cell instance masks and segmentation quality scores, and adopts a joint modeling mechanism of static morphological texture and temporal consistency, and outputs static feature vectors and trajectory state parameters.

[0045] The target cell determination module is used to determine the target cells based on the static feature vector and trajectory state parameters combined with a dual-branch feature fusion neural network model, and output the liver stem cell determination result.

[0046] The present invention also provides a liver stem cell screening device, comprising: a memory, a processor, and a liver stem cell screening program stored in the memory and executable on the processor, wherein the liver stem cell screening program, when executed by the processor, implements a liver stem cell screening method.

[0047] The present invention also provides a computer program product, including a liver stem cell screening program, wherein the liver stem cell screening program, when executed by a processor, implements the liver stem cell screening method.

[0048] The beneficial effects of this invention are as follows: By sequentially introducing image quality gating, quality correlation preprocessing, candidate detection and instance segmentation, and joint determination of static morphology, texture, and trajectory status into the bright-field microscopic image processing of primary hepatocyte mixed suspension, this invention can stably identify liver stem cells under conditions of cell overlap, out-of-focus, edge blurring, fragment interference, and uneven local illumination. This reduces screening errors caused by low-quality images directly participating in the determination, inaccurate cell adhesion segmentation, and misjudgment of single static features, thereby improving the accuracy and stability of liver stem cell screening.

[0049] This invention identifies target cells by inputting the static feature vectors and trajectory state parameters of candidate cells into a dual-branch feature fusion neural network model. It can not only utilize image characterization information such as the area, perimeter, roundness, eccentricity, refractive edge, and cytoplasmic texture of a single cell, but also combine the motion continuity and trajectory evolution information of the cell during continuous imaging for comprehensive discrimination. This avoids the category confusion problem caused by relying solely on single-frame image features and improves the reliability and automation level of liver stem cell screening results under complex mixed suspension conditions. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic flowchart of the first embodiment of a liver stem cell screening method according to the present invention.

[0052] Figure 2 This is a schematic diagram of the candidate cell region in a preprocessed image of a first embodiment of a liver stem cell screening method of the present invention.

[0053] Figure 3 This is a schematic diagram of the single-cell instance mask output effect of a first embodiment of a liver stem cell screening method of the present invention.

[0054] Figure 4 This is a schematic diagram of single-cell instance mask segmentation quality scoring, representing a first embodiment of a liver stem cell screening method according to the present invention.

[0055] Figure 5 This is a schematic diagram of the equipment used in the screening method for liver stem cells according to the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the liver stem cell screening method of the present invention, which presents the first embodiment of the liver stem cell screening method of the present invention.

[0058] In the first embodiment, the method for screening liver stem cells includes:

[0059] Step S10: Obtain the original image, focal plane position, exposure time and light source intensity of the primary hepatocyte suspension formed after liver tissue digestion by enzyme in the bright field microscopy imaging area. Combine the image quality gating mechanism to perform the effective image screening task and output the effective image, image quality score and image quality threshold.

[0060] It should be noted that the "original image" in this step refers to cell image data directly acquired within the bright-field microscopic imaging area, without background correction, noise reduction enhancement, or contrast compensation processing; "focal plane position" refers to the focal axis position parameter of the microscopic imaging system at the current acquisition moment; "exposure time" refers to the duration of exposure sampling performed by the image acquisition unit on the current field of view; and "light source intensity" refers to the illumination intensity parameter output by the bright-field illumination unit at the current acquisition moment. Furthermore, the "image quality gating mechanism" in this step refers to the process of determining whether the currently acquired image meets the conditions for subsequent identification processing based on the edge sharpness, local grayscale fluctuations, low-frequency background uniformity, saturated pixel ratio, and continuous changes in the focal plane in the original image. Its purpose is to first eliminate low-quality images with severe defocus, local overexposure, excessive illumination fluctuations, or significant focal plane drift, thereby outputting images that meet the subsequent analysis conditions as valid images. The image quality score is used to characterize the credibility of the current image for subsequent cell identification analysis, and the image quality threshold is used to determine whether the current image proceeds to the next step.

[0061] Understandably, this step performs a preliminary quality screening of the image before it enters subsequent cell detection and instance segmentation. This prevents subsequent steps from directly processing abnormal images with obvious out-of-focus areas, localized high-brightness saturation, uneven illumination, and focal plane shifts. This reduces false positives, false negatives, and segmentation boundary drift caused by fluctuations in image acquisition conditions, improving the stability of the entire screening process from the input end. Furthermore, since this step not only outputs valid images but also simultaneously outputs image quality scores and image quality thresholds, subsequent steps can dynamically adjust denoising intensity, contrast enhancement intensity, and candidate region thresholds based on these scores, forming a continuous processing chain rather than processing each step in isolation. For example, in a primary hepatocyte suspension sample, 10 consecutive images were acquired. Frames 3 and 7 showed focal plane shifts due to slight fluctuations in the sample liquid surface, while frame 5 showed localized overexposure due to light source reflection. If the traditional single edge sharpness threshold screening method is used, the 3rd and 5th frames may still be retained because they have high local edge response. However, in this embodiment, the 3rd frame has a larger change in focal plane position compared to the previous and next frames, resulting in an increased drift penalty component. The 5th frame has a lower image quality score due to an excessively high saturation pixel ratio. Ultimately, both of them are below the image quality threshold and are not included in subsequent processing.

[0062] Step S20: Based on the valid image, image quality score, and image quality threshold, a quality correlation preprocessing mechanism is used to perform background correction, noise reduction and enhancement, and contrast compensation tasks, and output the preprocessed image;

[0063] It should be noted that the "quality-correlation preprocessing mechanism" in this step refers to using the image quality score and image quality threshold output from step S10 as the basis for adjusting preprocessing parameters. This allows background correction, denoising enhancement, and contrast compensation to no longer use a single fixed processing intensity, but rather to adaptively adjust based on the current image quality state. Specifically, "background correction" refers to compensating for low-frequency illumination distribution, local shadows, and field-of-view brightness bias in bright-field microscopic images to reduce the impact of uneven illumination on cell boundary and texture determination; "denoising enhancement" refers to suppressing random gray-level fluctuations caused by sensor noise, liquid refraction interference, and small debris while preserving cell contour edges; and "contrast compensation" refers to enhancing the local gray-level difference between cell regions and background regions, as well as cell edge neighborhoods, to facilitate subsequent candidate cell detection and instance segmentation. The preprocessed image refers to image data suitable for subsequent single-cell candidate detection and instance segmentation after background correction, denoising enhancement, and contrast compensation.

[0064] Understandably, this step allows the effective images selected in step S10 to be further transformed into standardized inputs more suitable for cell detection and segmentation. Because primary hepatocyte suspensions under bright-field microscopy often exhibit local illumination deviations, uneven liquid refraction, extracellular debris scattering, and insufficient edge contrast, obtaining only "usable images" is insufficient to directly support stable segmentation. This step adaptively controls the preprocessing intensity based on image quality scores. For images of lower quality but still usable, it increases denoising and local enhancement; for higher quality images, it reduces excessive smoothing and enhancement, thereby suppressing interference while preserving as much of the original information as possible about cell contours and cytoplasmic texture. This preprocessing method, linked to image quality, helps provide a more stable grayscale distribution, more continuous edges, and a more uniform background as an input foundation for subsequent candidate cell detection.

[0065] It should be understood that, compared to traditional techniques that use fixed flat-field correction, uniform filtering parameters, or uniform histogram enhancement to process all images, this step improves upon the significant quality fluctuations of primary hepatocyte mixed suspension images. Traditional fixed-parameter methods often have two problems: first, for samples with low image quality, the enhancement intensity is insufficient, failing to effectively highlight cell edges; second, for samples with high image quality, the enhancement intensity is too high, which can amplify fragmented noise or alter the original cell texture distribution. In this embodiment, background correction, denoising enhancement, and contrast compensation are all correlated with the quality evaluation results of step S10, adjusting images of different quality levels to a more uniform analysis state before entering subsequent segmentation. This reduces image differences under different acquisition conditions without causing information destruction due to fixed processing, making it more suitable for continuous screening tasks under current primary mixed suspension conditions. For example, in a certain field of view, the image quality score is 0.61, which meets the conditions for proceeding to the next step, but there is slight illumination bias and a large amount of scattered noise. If traditional fixed denoising parameters are used, a significant amount of background noise remains, making it easy to misidentify noise points as small candidate cells later. Conversely, using excessively strong uniform filtering can dull cell edges. In this embodiment, the denoising intensity and local contrast enhancement parameters are automatically increased based on a score of 0.61. This significantly suppresses fine, discrete noise in the background region while preserving the transition between light and dark areas at cell boundaries. For example, for another valid image with a score of 0.87, this step automatically reduces the denoising intensity to avoid over-smoothing the originally clear cell outlines. Through this differentiated processing, the final preprocessed image output exhibits stable boundary representation and texture readability across different viewpoints.

[0066] Step S30: Based on the preprocessed image, a candidate detection and instance segmentation coupling mechanism is used to perform single-cell instance segmentation task, and the candidate cell region, single-cell instance mask and segmentation quality score are output.

[0067] It should be noted that the "candidate detection and instance segmentation coupling mechanism" in this step refers to first locating suspected cell regions in the preprocessed image, then performing single-cell-level foreground separation and target segmentation on each suspected cell region to obtain a single-cell instance mask, and providing a quality evaluation for each segmentation result. Specifically, "candidate cell region" refers to a local image region with the potential for cell appearance, obtained through response map calculation and scale constraint detection; "single-cell instance mask" refers to the pixel-level foreground region labeling result given for each candidate cell target, used to characterize the specific contour range of a single cell in the image; "segmentation quality score" refers to the score value used to evaluate whether the obtained single-cell instance mask has continuous boundaries, reasonable morphology, and low fragmentation. This score reflects the reliability of the current segmentation result and also provides a reference for subsequent feature extraction and judgment in steps S40 and S50.

[0068] Understandably, this step allows the preprocessed image output from step S20 to be further converted into structured data at the single-cell level. On one hand, candidate detection can eliminate a large number of background, fragmented, and irrelevant regions in the full-field image, retaining only local regions more likely to contain cells, thus reducing the computational burden of subsequent instance segmentation. On the other hand, instance segmentation can separate cells that are close to each other, partially touching, or even slightly overlapping from the candidate regions as much as possible, forming independent masks corresponding to single cells. By outputting candidate cell regions, single-cell instance masks, and segmentation quality scores, not only is the conversion from image to single-cell objects achieved, but a unified data foundation is also laid for the subsequent extraction of morphological parameters, texture parameters, refractive edge features, and trajectory state parameters.

[0069] It should be understood that this step further transforms the preprocessed image output from step S20 into structured data at the single-cell level. On the one hand, candidate detection can eliminate a large number of background, fragmented, and irrelevant regions in the full-field image, retaining only local regions more likely to contain cells, thus reducing the computational burden of subsequent instance segmentation. On the other hand, instance segmentation can separate cells that are close to each other, partially touching, or even slightly overlapping from the candidate regions as much as possible, forming independent masks corresponding to single cells. By outputting candidate cell regions, single-cell instance masks, and segmentation quality scores, not only is the transformation from image to single-cell objects achieved, but a unified data foundation is also laid for the subsequent extraction of morphological parameters, texture parameters, refractive edge features, and trajectory state parameters.

[0070] For example, such as Figure 2As shown, using the preprocessed image as input, and preserving the main outline and local texture information of the cells, candidate detection is performed on regions in the image that meet the preset scale range and local response conditions, thereby outputting multiple candidate cell regions. It can be seen that after candidate detection, the main suspected cell targets in the preprocessed image are locally bounded, while smaller, weaker gray-scale response debris interference regions are not output as valid candidate cell regions. Therefore, this step can preferentially narrow down the target range for subsequent segmentation in the full field of view image, reduce the interference of background regions and cell debris on the single-cell instance segmentation process, and provide a more focused local input basis for subsequent instance segmentation, thus improving the target localization accuracy in complex primary hepatocyte mixed suspension scenarios. Figure 3 As shown, based on candidate detection, further instance segmentation is performed on the candidate cell regions, thereby converting the cell targets in the image into independent single-cell instance masks. As can be seen in the figure, each candidate cell is distinguished and displayed using instance masks of different colors, transforming cell regions that were originally just brightness clusters in the grayscale image into single-cell segmentation results with clear boundaries and well-defined object classifications. For cell targets that are close in location, adjacent at the edges, or even slightly touching, this step can still separate them into different instances, rather than mistakenly merging multiple cells into the same foreground region. Therefore, this step can achieve the transformation from image-level input to single-cell object-level output, providing a precise boundary basis for the extraction of area, perimeter, roundness, refractive edge scoring, and cytoplasmic texture features in subsequent steps, thus improving the reliability of subsequent liver stem cell identification. Figure 4 As shown, after outputting the candidate cell regions and single-cell instance masks, step S30 further generates corresponding segmentation quality scores for each single-cell instance mask. As can be seen from the figure, the segmentation quality scores differ for different candidate regions. Single-cell instance masks with intact boundaries, compact morphology, and low fragmentation receive higher scores, while instances with discontinuous edges, local defects, or weak segmentation stability receive relatively lower scores. This score output quantifies the reliability of the segmentation results, enabling subsequent step S40 to differentiate the reliability of different instances when extracting static feature vectors, and reducing the interference of low-reliability segmentation results on the final determination in step S50 when performing target cell determination. Therefore, this step not only completes instance segmentation but also further achieves a quantitative evaluation of the reliability of the segmentation results.

[0071] Step S40: Based on candidate cell regions, single-cell instance masks, and segmentation quality scores, a joint modeling mechanism of static morphology and texture and temporal consistency is used to perform single-cell feature extraction and trajectory association tasks, and output static feature vectors and trajectory state parameters.

[0072] It should be noted that, in this step, the "candidate cell region" refers to the local image region cropped around the suspected single-cell target output in step S30, used to carry the edge, internal grayscale, and local background information of the single-cell target; the "single-cell instance mask" refers to the pixel-level target region labeling result corresponding one-to-one with the candidate cell region, used to limit the specific contour range of the single cell in the image; and the "segmentation quality score" refers to the credibility index obtained by quantifying the boundary closure degree, shape integrity degree, local fragmentation degree, and instance independence degree of the single-cell instance mask. Furthermore, the "static morphology and texture joint modeling mechanism with temporal consistency" in this step refers to: on the one hand, extracting static parameters that characterize the current image morphology, edge brightness difference, and cytoplasmic texture distribution of the cell based on the single-cell instance mask and candidate cell region; on the other hand, associating and describing the state of the same single cell in the temporal dimension based on the spatial positional change relationship of the candidate cell region in consecutive image frames, the target matching relationship between neighboring frames, and the continuous motion change relationship. Among them, the "static feature vector" refers to the feature set formed by the combination of single-cell instance area, single-cell instance perimeter, single-cell instance equivalent diameter, single-cell instance roundness, single-cell instance eccentricity, refractive edge score, segmentation quality score, and cytoplasmic texture features, which is used to characterize the appearance state of a single cell from the perspective of the current frame; the "trajectory state parameters" refers to the parameter set composed of information such as the center position of the candidate cell region, velocity components, inter-frame position offset, and continuous association markers, which is used to characterize the motion state and temporal continuity of a single cell in consecutive image frames.

[0073] It should also be noted that the "refractive edge scoring" in this step refers to the quantification of the difference between the average gray level of the inner ring region and the average gray level of the outer ring region after constructing the inner and outer ring regions based on the single-cell instance mask. It is used to reflect the refractive performance at the cell edge. "Cytoplasmic texture features" refers to the features extracted from the internal region of a single cell, such as local gray-level distribution patterns, granularity distribution patterns, directional change patterns, or co-occurrence distribution patterns. These features include at least one of local binary pattern texture features and gray-level co-occurrence matrix texture features. "Trajectory association" refers to the process of matching the observation results of the same cell in different frames between the candidate cell region in the current frame and the candidate cell region in the historical frame based on positional proximity, morphological similarity, and motion continuity.

[0074] Understandably, this step transforms the single-cell instance mask output in step S30 from a "segmentation result" into a "structured representation result that can be directly used by the decision-making model." In other words, step S30 addresses the problem of "separating cells," while this step addresses the problem of "clearly describing the separated cells." Specifically, the static feature vector can refine the description of candidate cells from three levels: geometric morphology, edge refraction, and internal texture. This allows the model to not only perceive whether a single cell is "large or small," "round or long," and "clear or unclear its edges," but also to further perceive whether the gray-scale distribution within its cytoplasm is uniform, whether the edge brightness difference matches the appearance characteristics of the target cell, and whether the current segmentation result is reliable. Simultaneously, the trajectory state parameters supplement each candidate cell with continuous information in the temporal dimension, enabling the model to identify whether a candidate cell's movement is smooth, its positional changes are continuous, and whether it maintains a similar appearance and motion pattern in adjacent frames. Because different cell types in primary hepatocyte suspensions often exhibit some degree of apparent overlap in a single frame image—for example, some mature hepatocytes may resemble liver stem cells in size, and some fragment aggregates may show similar brightness differences to target cells at local edges—relying solely on morphology or grayscale at a single moment can easily lead to misjudgment. However, through the joint modeling process in this step, candidate cells are endowed with both static morphological and textural descriptions and temporal continuity descriptions, thus allowing subsequent target cell determination to be based on a more complete feature set.

[0075] For example, in a primary hepatocyte suspension sample, step S30 outputs five candidate cell regions and their corresponding single-cell instance masks. Candidate cells 1 and 2 are similar in area and roundness; based solely on traditional static morphological parameters, both could be identified as target cells. Further execution of this step reveals that candidate cell 1 has a higher refractive edge score, a relatively uniform and continuous grayscale distribution in its cytoplasm texture, and a stable center position across six consecutive frames with minimal trajectory velocity fluctuations. Candidate cell 2, while similar in area and roundness, exhibits a smaller average grayscale difference between its outer and inner rings, a discrete granular cytoplasm texture, and significant center position fluctuations across consecutive frames, resulting in unstable velocity changes. Based on the joint modeling results, candidate cell 1 is closer to liver stem cells in both static feature vectors and trajectory state parameters, while candidate cell 2 is more likely to be a non-target cell or an abnormal object formed by debris interference. For another example, in another set of continuously sampled images, a candidate cell's single-cell instance mask is slightly irregular due to slight blurring of its local edges. Traditional methods might exclude candidates with low roundness; however, this step considers segmentation quality score, refractive edge score, and trajectory stability across consecutive frames. If a candidate cell exhibits smooth positional changes and continuous velocity across five consecutive frames, and its cytoplasmic texture and edge brightness / darkness differences still match the target cell characteristics, then its static feature vector and trajectory state parameters, when combined, can maintain a high target relevance.

[0076] Step S50: Target cells are determined by combining static feature vectors and trajectory state parameters with a dual-branch feature fusion neural network model, and the liver stem cell determination result is output.

[0077] It should be noted that the "static feature vector" in this step refers to the feature set output by step S40, which is used to characterize the appearance state of the candidate cell in the current frame image. It includes at least one or more of the following: single cell instance area, single cell instance perimeter, single cell instance equivalent diameter, single cell instance roundness, single cell instance eccentricity, refractive edge score, segmentation quality score, and cytoplasmic texture features. The "trajectory state parameters" refer to the parameter set output by step S40, which is used to characterize the position change, velocity change, and state continuity of the candidate cell in consecutive image frames. It includes at least one or more of the following: candidate cell region center position, velocity component, inter-frame position offset, and continuous association marker.

[0078] Furthermore, the "dual-branch feature fusion neural network model" in this step refers to a neural network model that processes the static feature vector and trajectory state parameters in parallel and then performs a fusion determination. One branch receives the static feature vector to extract high-order representation information related to the current appearance of the candidate cell; the other branch receives the trajectory state parameters to extract temporal representation information related to the continuous motion state of the candidate cell. The outputs of the two branches are concatenated, mapped, or weighted and fused to output the category determination result corresponding to the current candidate cell. The "target cell determination" in this step refers to the process of classifying whether the current candidate cell belongs to liver stem cells; the "liver stem cell determination result" refers to the final classification output corresponding to the candidate cell, which can be expressed as a category label of liver stem cells / non-liver stem cells, or as a probability value, confidence value, or score value of the candidate cell belonging to liver stem cells. The dual-branch feature fusion neural network model preferably consists of a static branch, a temporal branch, and a fusion output layer. The static branch can adopt a multilayer perceptron structure to perform nonlinear mapping on the static feature vector to obtain the static representation vector; the temporal branch can adopt a gated recurrent unit structure to perform temporal dependency modeling on the trajectory state parameter sequence to obtain the temporal representation vector; the fusion output layer is used to map the static representation vector and the temporal representation vector together to the target cell category output.

[0079] Understandably, this step allows for joint analysis of the static feature vector and trajectory state parameters output from step S40 within the same decision framework, thus preventing subsequent target cell classification from relying solely on a single type of feature. For primary hepatocyte suspensions, candidate cells often exhibit similar morphology, locally similar textures, and overlapping edge brightness differences. If judgment is based solely on static appearance features, misclassification may occur for some mature hepatocytes, bile duct epithelial cells, or fragment aggregates. Similarly, judging solely on trajectory state parameters may overlook key morphological differences in the current frame, leading to insufficient classification criteria. This step, through a dual-branch feature fusion approach, integrates the appearance of candidate cells in the current image with their continuous motion across consecutive image frames, thereby establishing target cell classification based on a more comprehensive set of criteria. Specifically, the static branch enhances the model's ability to recognize candidate cell contour morphology, edge brightness differences, and cytoplasmic texture differences; the temporal branch enhances the model's ability to recognize the inter-frame motion stability, trajectory evolution continuity, and positional change patterns of candidate cells; and the fusion output layer can convert both types of information into the final judgment result. Since this step directly outputs the liver stem cell judgment result, it effectively performs the final comprehensive utilization function of a series of preceding results, including image quality gating, preprocessing, instance segmentation, static feature construction, and trajectory state association. For example, in a batch of primary hepatocyte mixed suspension samples, after processing through steps S10 to S40, a total of 120 candidate cell static feature vectors and trajectory state parameters were obtained, of which 32 were manually verified as genuine liver stem cells. If a traditional fixed threshold rule based on area, roundness, and average grayscale is used for judgment, 40 target candidate cells are ultimately screened out, but this includes many non-target cells, resulting in significant misjudgments. If only a single-branch neural network based on static feature vectors is used, although it reduces some misjudgments compared to the fixed threshold method, it is still difficult to reliably exclude some candidate objects that are similar in appearance to the target cells in a single frame but have obviously abnormal motion states. In this embodiment, after inputting the static feature vector and trajectory state parameters into the dual-branch feature fusion neural network model, the following can be observed: For candidate cell number 12, its area, roundness, and refractive edge scores are all close to those of the target cell, but its center position changes significantly over 6 consecutive frames, and its speed fluctuation is significantly higher than that of other real target cells in the same batch. After combining the trajectory state parameters, the dual-branch feature fusion neural network model classifies this candidate cell as a non-liver stem cell, thereby avoiding misjudgments caused by relying solely on static appearance.Conversely, for candidate cell number 27, although its current frame edge is slightly blurred, resulting in insufficient prominence of single-frame static features, its position changes smoothly and its velocity changes continuously in consecutive frames. Furthermore, its cytoplasmic texture features and refractive edge scores are generally consistent with the target cell. The dual-branch feature fusion neural network model can still identify it as a liver stem cell, thereby avoiding the mis-exclusion of the real target cell by the traditional single-frame identification method.

[0080] Example 2: Furthermore, the present invention provides a liver stem cell screening system that employs a liver stem cell screening method described in the above embodiments, thereby solving a technical problem in liver stem cell screening. The beneficial effects of the liver stem cell screening system provided by the present invention are the same as those of the liver stem cell screening method described in the above embodiments, and other technical features of the liver stem cell screening system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0081] Example 3: This invention provides a liver stem cell screening device, please refer to... Figure 5A liver stem cell screening device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a liver stem cell screening method as described in Embodiment 1 above. The liver stem cell screening device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. This liver stem cell screening device is merely an example and should not be construed as limiting the functionality or scope of the embodiments of this invention. A liver stem cell screening device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. Random access memory 1004 also stores various programs and data required for the operation of a liver stem cell screening device. Processing device 1001, read-only memory 1002, and random access memory 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows a liver stem cell screening device to communicate wirelessly or wiredly with other devices to exchange data. Although a liver stem cell screening device with various systems is shown in the figures, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0082] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the liver stem cell screening method described above. The computer program product provided by this invention can solve a technical problem in liver stem cell screening. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the liver stem cell screening method provided in the above embodiments, and will not be repeated here.

[0083] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0084] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for screening liver stem cells, characterized in that, The methods include: Step S10: Obtain the original image, focal plane position, exposure time and light source intensity of the primary hepatocyte suspension formed after liver tissue digestion by enzyme in the bright field microscopy imaging area. Combine the image quality gating mechanism to perform the effective image screening task and output the effective image, image quality score and image quality threshold. Step S20: Based on the valid image, image quality score, and image quality threshold, a quality correlation preprocessing mechanism is used to perform background correction, noise reduction and enhancement, and contrast compensation tasks, and output the preprocessed image; Step S30: Based on the preprocessed image, a candidate detection and instance segmentation coupling mechanism is used to perform single-cell instance segmentation task, and the candidate cell region, single-cell instance mask and segmentation quality score are output. Step S40: Based on candidate cell regions, single-cell instance masks, and segmentation quality scores, a joint modeling mechanism of static morphology and texture and temporal consistency is used to perform single-cell feature extraction and trajectory association tasks, and output static feature vectors and trajectory state parameters. Step S50: Target cells are determined by combining static feature vectors and trajectory state parameters with a dual-branch feature fusion neural network model, and the liver stem cell determination result is output.

2. The method for screening liver stem cells as described in claim 1, characterized in that, Step S10 involves acquiring the original image, focal plane position, exposure time, and light source intensity of the primary hepatocyte suspension formed after enzymatic digestion of liver tissue in the bright-field microscopy imaging region. This is combined with an image quality gating mechanism to perform an effective image screening task, outputting effective images, image quality scores, and image quality thresholds. Specifically, this includes: Step S101: Obtain the time Original image of a primary hepatocyte suspension formed after enzymatic digestion of lower liver tissue in a bright-field microscopy region. Focal plane position Exposure time and light source intensity And based on the original image Calculate local variance Based on local variance Extracting high-information pixel set : ;in, Represents the x-coordinate of a pixel; Represents the ordinate of a pixel; The local variance threshold is determined based on the image grayscale distribution of the original image; Step S102: Based on the set of high-information pixels The edge sharpness component of the current image is calculated using gradient magnitude statistics. Based on high-information pixel sets The Laplace normalized components are calculated using the Laplace response variance normalization method. Based on high-information pixel sets The background uniformity component is calculated using a low-frequency background variance suppression method. Based on high-information pixel sets Foreground drift penalty components are calculated using a time-series drift penalty method. ; and based on the Laplace normalized components Background uniformity component and foreground drift penalty component Constructing image quality scores ; Step S103: Score the image quality Constructing image quality thresholds ; ； in, The length of the history window; For a moment At that time Historical image quality score sequence; MAD is an adjustment coefficient used to adjust the responsiveness of the image quality threshold to historical fluctuations. The median absolute deviation function is used to characterize the degree of dispersion of a historical image quality score sequence relative to its median value; This is a median function used to characterize the central location of a historical image quality score sequence; Step S104: Satisfy The original image is determined as a valid image, and the valid image, image quality score, and image quality threshold are output.

3. The method for screening liver stem cells as described in claim 2, characterized in that, Step S20, based on the valid image, image quality score, and image quality threshold, employs a quality-correlated preprocessing mechanism to perform background correction, noise reduction, and contrast compensation tasks, and outputs the preprocessed image. Specifically, this includes: Step S201: Based on the valid images, cache a window of historical valid images according to a preset time frame. Online flat-field estimation is performed using low-frequency quantile statistical modeling to output the background model; dark-field images are also obtained. According to the dark field diagram Background correction was performed using a flat-field normalization correction method on the background model to obtain the background-corrected image. ; Step S202: Based on image quality scoring Generate denoising intensity parameters And based on the denoising intensity parameter Background correction image Perform edge-preserving denoising to obtain a denoised image. ;in, , and These represent the minimum and maximum noise reduction strengths, respectively. Step S203: Based on image quality scoring Generate contrast enhancement parameters And based on contrast enhancement parameters For denoised images Perform local contrast enhancement processing to obtain the preprocessed image. .

4. The method for screening liver stem cells as described in claim 3, characterized in that, In step S202, based on the denoising intensity parameter Background correction image During the edge-preserving denoising process, bilateral filtering and guided filtering are employed, along with image quality scoring. By controlling the strength of grayscale similarity constraints, noise suppression is improved when the image quality score is low, while more edge details are preserved when the image quality score is high. Among them, bilateral filtering and guided filtering are used to maintain a balance between cell contour continuity and background stability.

5. The method for screening liver stem cells as described in claim 3, characterized in that, In step S30, the single-cell instance segmentation task is performed based on the preprocessed image using a candidate detection and instance segmentation coupling mechanism, and the candidate cell region, single-cell instance mask, and segmentation quality score are output. Specifically, this includes: Step S301: Based on the preprocessed image Constructing a difference response map with the background model : ;in, , This is a low-pass filtering operation used to extract low-frequency background components from the preprocessed image; and it is applied to the differential response map. Implementation scale is limited Candidate detection yields a set of candidate centers. Candidate cell regions are obtained by cropping around each candidate center in the candidate center set; Step S302: Input the candidate cell region into the preset instance segmentation network, and the instance segmentation network outputs a foreground probability map; Step S303: Obtain image quality score Based on image quality score Constructing a foreground probability threshold , , This is a truncation function used to limit the foreground probability threshold within a preset range; This is a threshold adjustment coefficient used to control the degree of influence of image quality score on the foreground probability threshold; and based on the foreground probability threshold... The foreground region is obtained by performing threshold segmentation on the foreground probability map; Step S304: Perform distance transformation on the foreground region using Euclidean distance transformation to obtain a distance map; extract local maximum markers in the distance map according to the preset minimum radius constraint, and perform marker-controlled watershed segmentation based on the local maximum markers to obtain a single-cell instance mask; Step S305: Construct a segmentation quality score based on the cross-union ratio between the single-cell instance mask and the preset standard instance mask template, and finally output the candidate cell region, single-cell instance mask and segmentation quality score.

6. The method for screening liver stem cells as described in claim 1, characterized in that, In step S40, based on candidate cell regions, single-cell instance masks, and segmentation quality scores, a joint modeling mechanism of static morphological texture and temporal consistency is used to perform single-cell feature extraction and trajectory association tasks, outputting static feature vectors and trajectory state parameters. Specifically, this includes: Step S401: Based on the single-cell instance mask, use Python's scikit-image library to perform region attribute measurement processing, and extract the single-cell instance area, single-cell instance perimeter, single-cell instance equivalent diameter, single-cell instance roundness, and single-cell instance eccentricity. Step S402: Construct inner and outer ring regions based on candidate cell regions and single-cell entity masks, and calculate the refractive edge score based on the mean difference of grayscale in the ring region and the outer ring region, while extracting the cytoplasmic texture features of the candidate cell regions. Step S403: Combine single-cell instance area, single-cell instance perimeter, single-cell instance equivalent diameter, single-cell instance roundness, single-cell instance eccentricity, refractive edge score, segmentation quality score and cytoplasmic texture features to generate a static feature vector; Step S404: Obtain the historical trajectory state parameters and current trajectory state parameters of the center position of the candidate cell region. Based on the historical trajectory state parameters and current trajectory state parameters, perform trajectory association through Hungarian matching joint association to construct the trajectory state vector.

7. The method for screening liver stem cells as described in claim 1, characterized in that, Step S50, which involves determining the target cell based on the static feature vector and trajectory state parameters combined with a dual-branch feature fusion neural network model and outputting the liver stem cell determination result, specifically includes: Step S501: Construct a dual-branch feature fusion neural network model based on the multilayer perceptron branch and the gated recurrent unit branch; wherein, the multilayer perceptron branch is used to perform nonlinear mapping and high-dimensional feature extraction on the static feature vector to obtain a static representation vector representing the morphology, texture and refractive edge features of the candidate cells; the gated recurrent unit branch is used to perform temporal dependency modeling and dynamic change pattern extraction on the trajectory state parameters to obtain a temporal representation vector representing the continuity of the candidate cell motion and the trajectory evolution features; Step S502: Obtain historical static feature vectors, historical trajectory state parameters, and cell category labels corresponding to historical samples. Use historical static feature vectors and historical trajectory state parameters as inputs to the dual-branch feature fusion neural network model, and use cell category labels corresponding to historical samples as outputs of the dual-branch feature fusion neural network model. Use the binary cross-entropy loss function to perform the pre-training process in a supervised learning manner. Step S503: Input the static feature vector and trajectory state parameters into the pre-trained dual-branch feature fusion neural network model, and the dual-branch feature fusion neural network model outputs the liver stem cell determination result.

8. A liver stem cell screening system, applied to the liver stem cell screening method according to any one of claims 1 to 7, characterized in that, The liver stem cell screening system includes: The image quality gating module is used to acquire the original image, focal plane position, exposure time and light source intensity of the primary hepatocyte suspension formed after liver tissue digestion by enzyme in the bright field microscopy imaging area. Combined with the image quality gating mechanism, it performs the effective image screening task and outputs the effective image, image quality score and image quality threshold. The image quality gating module is used to perform background correction, noise reduction and enhancement and contrast compensation tasks based on the valid image, image quality score and image quality threshold, and output the preprocessed image using a quality correlation preprocessing mechanism. The candidate detection and instance segmentation module is used to perform single-cell instance segmentation tasks based on preprocessed images using a candidate detection and instance segmentation coupling mechanism, and outputs candidate cell regions, single-cell instance masks and segmentation quality scores. The feature extraction and trajectory association module is used to perform single-cell feature extraction and trajectory association tasks based on candidate cell regions, single-cell instance masks and segmentation quality scores, and adopts a joint modeling mechanism of static morphological texture and temporal consistency, and outputs static feature vectors and trajectory state parameters. The target cell determination module is used to determine the target cells based on the static feature vector and trajectory state parameters combined with a dual-branch feature fusion neural network model, and output the liver stem cell determination result.

9. A screening device for liver stem cells, characterized in that, The liver stem cell screening device includes: a memory, a processor, and a liver stem cell screening program stored in the memory and executable on the processor. When the liver stem cell screening program is executed by the processor, it implements a liver stem cell screening method according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a liver stem cell screening program, which, when executed by a processor, implements a liver stem cell screening method according to any one of claims 1 to 7.

Images