Deep learning based junction relationship network training and cell occlusion determination method

Abstract

The application provides a deep learning-based junction relationship network training and cell occlusion determination method, performs instance segmentation on a cell microscopic image to obtain a cell instance mask, extracts a cell boundary based on the cell instance mask, performs junction point detection on the cell boundary to obtain a cell occlusion junction point, and classifies the cell occlusion junction point into a cell T-shaped occlusion point, a cell Y-shaped converging point or a cell X-shaped crossing point; a local image patch is cropped with the cell occlusion junction point as the center, and a multi-channel sample containing the local image patch is constructed; the multi-channel sample is labeled with a unified relationship label, the unified relationship label including an occlusion relationship, a same-layer relationship and an undeterminable; a deep learning network is trained with the labeled multi-channel sample, the deep learning network outputting a probability distribution of a junction point relationship category; and the multi-channel sample of a to-be-tested cell microscopic image is input into the trained deep learning network to output a junction point relationship category and a confidence thereof.

Description

Technical Field

[0001] This invention relates to the field of biomedical image processing technology, specifically to a method for training a deep learning-based boundary point relationship network and determining cell occlusion. Background Technology

[0002] Cell microscopic images are a crucial data source for biomedical research and clinical diagnosis. In applications such as cell culture monitoring, pathological section analysis, and drug screening, researchers need to accurately count, measure morphology, and analyze behavior of cells. However, cells inevitably clump together, stack, and partially overlap during growth, posing a significant challenge to automated cell analysis. When multiple cells spatially overlap in an image, determining their relative vertical or occlusion relationships becomes a critical prerequisite for subsequent analysis.

[0003] Currently, the technical solutions for determining cell occlusion relationships mainly fall into three categories. The first category focuses on cell instance segmentation, using deep learning networks to obtain cell masks and boundaries. While this type of method can effectively separate overlapping cells, it typically only addresses the problem of separating cells and struggles to stably output directional occlusion relationships, failing to provide clear spatial order information for subsequent hierarchical analysis.

[0004] The second type of approach determines occlusion based on geometric cues at intersection points. These methods typically detect T-shaped intersection structures in images and combine manually designed image features such as edge strength, grayscale contrast, and texture continuity to distinguish between long and short sides, thereby inferring the occlusion direction. However, in practical applications of cell microscopy, this type of method faces numerous challenges: cell boundaries are often blurred, adjacent cells have highly similar textures, and the contrast is not significant when transparent cells are superimposed. Furthermore, structures such as X-shaped intersections and Y-shaped convergences are easily confused with T-shaped occlusion points, leading to insufficient stability of threshold-based determination methods and poor adaptability to different microscopic equipment, cell types, and imaging qualities.

[0005] The third type of approach attempts to use deep learning for end-to-end occlusion detection, or relies on experts to delineate regions of interest and annotate them point by point to train the model. While these methods can achieve good results on specific datasets, they have two prominent problems: First, the model is prone to learning pseudo-features related to specific device parameters or staining conditions, limiting its generalization ability; second, the cost of producing data for expert frame-by-frame interpretation and fine annotation is extremely high, making it difficult to support the high-throughput processing needs of large-scale image data generated by long-term sequential imaging in incubators.

[0006] A comprehensive analysis of existing technologies reveals the following main problems in the current field of cell occlusion relationship determination: First, there is a lack of standardized data organization methods and labeling systems for the key local structure of cell occlusion boundaries, making it difficult for models to effectively learn around the local evidence that truly determines occlusion relationships; second, traditional long / short side discrimination based on handcrafted image features is sensitive to threshold parameters and lacks robustness when faced with changes in imaging conditions; finally, existing methods often cannot simultaneously and uniformly handle three common boundary structures: cell T-shaped occlusion points, cell X-shaped intersection points, and cell Y-shaped convergence points, especially lacking a systematic and learnable solution for handling Y-shaped structures where there is no distinction between upper and lower layers and X-shaped intersection disambiguation. Summary of the Invention

[0007] This invention proposes a deep learning-based method for training boundary point relationship networks and determining cell occlusion, to address the following technical problems: how to automatically and stably detect occlusion boundary points between cell boundaries in a large number of cell images and perform reproducible localization and deduplication; how to uniformly handle three common structures—cell T-shaped occlusion points, cell X-shaped intersection points, and cell Y-shaped convergence points—under the same training framework; and how to design a low-cost annotation strategy to enable annotators to efficiently annotate boundary point relationship categories.

[0008] To address the aforementioned technical problems, this invention provides a method for training a deep learning-based boundary point relationship network and determining cell occlusion, comprising the following steps:

[0009] Step S1: Perform instance segmentation on the cell micrograph to obtain a cell instance mask, and extract cell boundaries based on the cell instance mask;

[0010] Step S2: Detect the intersection points of the cell boundaries to obtain cell occlusion intersection points, and classify the cell occlusion intersection points as cell T-shaped occlusion points, cell Y-shaped convergence points, or cell X-shaped intersection points;

[0011] Step S3: Crop a local image patch centered on the cell occlusion boundary point, and construct a multi-channel sample containing the local image patch;

[0012] Step S4: Label the multi-channel samples using unified relation labels, which include occlusion relationship, same layer relationship, and undeterminable relationship;

[0013] Step S5: Train a deep learning network using the labeled multi-channel samples, and the deep learning network outputs the probability distribution of the relationship categories of the boundary points;

[0014] Step S6: Input the multi-channel samples of the microscopic image of the cell to be tested into the trained deep learning network, and output the boundary point relationship category and its confidence level.

[0015] Preferably, in step S1, the cell microscopic image is preprocessed before instance segmentation is performed. The preprocessing includes brightness normalization, contrast normalization, and noise suppression. After performing hole filling, spot removal, and morphological smoothing on the cell instance mask, the cell boundary is extracted and the cell boundary is refined into boundary line segments with a single pixel width.

[0016] Preferably, in step S2, the boundary point detection includes: screening candidates for spatially close cell pairs or those that come into contact after mask expansion; discretizing the boundaries of the candidate cell pairs into a set of line segments and performing intersection detection to obtain boundary candidate points; and clustering and merging the spatially close boundary candidate points to obtain the cell occlusion boundary points.

[0017] Preferably, in step S2, the boundary line segments in the neighborhood of the cell occlusion boundary point are classified according to their direction, number, and geometric distribution: when there is a pair of boundary line segments extending approximately 180 degrees in opposite directions and a boundary line segment terminating at the boundary point, it is classified as a cell T-shaped occlusion point; when there are three boundary line segments distributed approximately symmetrically, it is classified as a cell Y-shaped convergence point; when there are two pairs of boundary line segments extending approximately 180 degrees in opposite directions and intersecting at a cross, it is classified as a cell X-shaped intersection point.

[0018] Preferably, in step S3, the multi-channel sample includes an original image patch, an instance mask patch, a boundary map patch, and a direction encoding patch; the direction encoding patch encodes the boundary direction within the neighborhood of the cell occlusion boundary point in the form of discrete angles, unit vector fields, or direction heatmaps.

[0019] Preferably, in step S4, the multi-channel samples classified as Y-shaped cell convergence points are labeled as being in the same layer by default; when manual verification confirms that there is occlusion, the same layer relationship is modified to an occlusion relationship.

[0020] Preferably, step S4 further includes: when the unified relationship label is an occlusion relationship, labeling the long side and short side of the multi-channel sample as auxiliary supervision information, wherein the long side is a through boundary extending approximately 180 degrees in opposite directions within the neighborhood of the intersection point, and the short side is a boundary terminating at the intersection point.

[0021] Preferably, in step S5, the deep learning network adopts multi-task learning, with the main task being boundary point relationship classification and the auxiliary task being long-side and short-side attribution prediction; when the unified relationship label is an occlusion relationship, the loss of the main task and the loss of the auxiliary task are calculated simultaneously; when the unified relationship label is a same-layer relationship or cannot be determined, the loss of the auxiliary task is masked.

[0022] Preferably, before step S5, the method further includes: dividing the labeled multi-channel samples into training set, validation set and test set according to the image source; using resampling or class weights to balance the class imbalance; and performing data augmentation operations such as rotation, translation, scaling, brightness perturbation and noise injection on the multi-channel samples.

[0023] Preferably, step S6 further includes: filtering valid boundary point relationship categories based on the confidence level, converting occlusion relationships into directed edges of the cell relationship graph, and converting same-layer relationships into undirected edges of the cell relationship graph; performing loop detection on the directed edges and deleting the edges with the lowest confidence level to eliminate loops; and displaying the cell instance mask by color overlay based on the cell relationship graph.

[0024] The beneficial effects of the present invention include at least the following:

[0025] First, this invention integrates cell T-shaped occlusion points, cell X-shaped intersection points, and cell Y-shaped convergence points into the same training framework, and directly outputs the relationship category of the intersection points using a unified relationship labeling system. Among them, Y-shaped convergence points participate in training as same-layer relationships, which not only avoids the waste of information caused by simply discarding such samples, but also reduces the risk of misjudging same-layer contact as occlusion, making the judgment results more consistent with the actual geometric relationship of cells.

[0026] Second, compared to existing methods that rely on handcrafted image features such as edge strength, grayscale contrast, and texture consistency to distinguish between long and short sides, this invention creates a multi-channel training dataset using local images of cell occlusion boundaries and their structural information. A deep learning network then automatically learns the occlusion judgment pattern. This data-driven approach maintains more stable judgment results under common cell imaging conditions such as blurred boundaries, low contrast, uneven staining, and high noise, effectively reducing false positives and false negatives caused by rule-based threshold sensitivity.

[0027] Third, this invention automatically extracts local patches around the boundary points and simultaneously organizes multi-channel inputs such as the original image, instance mask, boundary and orientation encoding to form a structured and reusable data format. Furthermore, this invention supports a semi-automatic annotation strategy that combines pseudo-label generation with manual correction, enabling rapid accumulation of large-scale training samples with low labor costs, continuously improving model performance, and possessing good reproducibility and engineering capabilities. Attached Figure Description

[0028] Figure 1 This is a flowchart of the overall process for training a deep learning-based boundary point relationship network and determining cell occlusion.

[0029] Figure 2 A schematic diagram of the intersection of cell outlines;

[0030] Figure 3This is a schematic diagram of the training of the boundary point relationship network;

[0031] Figure 4 This is a schematic diagram of the boundary point relationship network and the determination method test. Detailed Implementation

[0032] 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 protection scope of the present invention.

[0033] This invention addresses the difficulty in stably and automatically determining the vertical occlusion relationships between cells in ordinary two-dimensional cell micrographs. It proposes a deep learning-based method for training an intersection point relationship network and determining cell occlusion, such as… Figure 1 As shown. Unlike manually designing methods that rely on image features to distinguish long / short sides, this invention uses cell occlusion boundary points as the core sample units, constructs a training dataset around the boundary point locations, trains a deep learning network through supervised learning, and automatically outputs the long / short side at the boundary point and the corresponding occlusion direction, thereby providing more robust local evidence when constructing a directed occlusion relationship graph in the future. It includes the following steps.

[0034] Step S1: Perform instance segmentation on the cell micrograph to obtain a cell instance mask, and extract cell boundaries based on the cell instance mask.

[0035] This invention first acquires cell microscopic image data sources. Data sources may include cellular regions from periodic incubator imaging sequences, microscopic images, or digital pathological images. For each input image, a standardized preprocessing workflow is performed. Preprocessing operations include brightness normalization, contrast normalization, and noise suppression, aimed at reducing the impact of different devices and experimental batches on the distribution of training data. Depending on the specific application requirements, operations such as resolution standardization and cropping strategy standardization may also be performed to ensure consistency in subsequent processing.

[0036] For each preprocessed cell image, this invention employs a cell instance segmentation method to obtain a cell instance mask. Instance segmentation can be achieved using existing deep learning segmentation networks or traditional image segmentation procedures, and its output must include at least the instance number and corresponding pixel region for each cell. After obtaining the initial segmentation results, post-processing operations are performed on the instance mask to improve boundary quality. Specifically, small hole filling, small spot removal, and morphological smoothing operations are performed to obtain instance regions with continuous boundaries and topological stability. Based on the processed instance mask, cell boundary curves are extracted and refined into boundary line segments of single pixel width, forming a set of boundary line segments and an ordered point representation that can be used for orientation analysis and intersection point detection.

[0037] Step S2: Detect cell boundary points to obtain cell occlusion points, and classify cell occlusion points as cell T-shaped occlusion points, cell Y-shaped convergence points, or cell X-shaped intersection points.

[0038] This invention performs candidate screening of cell pairs at the instance level to reduce computational complexity and improve detection efficiency. The screening strategy only performs boundary detection on spatially close cell pairs or those that come into contact after mask expansion, excluding obviously non-adjacent cell combinations. For each candidate cell pair, the boundary between the two sides is discretized into a set of line segments, and line segment intersection detection and proximity detection are performed to obtain boundary candidate points. Since the boundary discretization and detection process may generate multiple close candidate points, this invention clusters and merges spatially close boundary candidate points to obtain stable and unique cell occlusion boundary points. The clustering and merging operation effectively avoids boundary point drift, missed detection, or duplicate counting problems caused by segmentation noise, boundary spikes, and dense cell scenes.

[0039] For each detected cell occlusion boundary point, this invention statistically analyzes the number and geometric distribution of the principal directions of the boundary line segments within the neighborhood of the boundary point, and classifies the boundary points into three types accordingly. Figure 2As shown. The first type is the T-shaped occlusion point: When there is a pair of boundary segments extending approximately 180 degrees in opposite directions in the neighborhood of the intersection point, and a boundary segment terminating at the intersection point, it is determined to be a T-shaped occlusion point structure. In this structure, the opposing boundary segments form a continuous long side, and the boundary segment terminating at the intersection point is the short side. The cell belonging to the long side is usually located in the upper layer and occludes the cell belonging to the short side. The second type is the Y-shaped convergence point: When there are three boundary segments approximately symmetrically distributed in the neighborhood of the intersection point, it usually corresponds to the convergence or bifurcation structure of three cells on the same plane. At this time, there is no clear upper and lower occlusion relationship. The third type is the X-shaped intersection point: When there are two pairs of boundary segments extending approximately 180 degrees in opposite directions in the neighborhood of the intersection point in a cross shape, further disambiguation is needed to determine which pair is the continuous long side and which pair is interrupted at the intersection point. The aforementioned geometric type recognition is used to organize samples and auxiliary annotations. All three types of boundary point samples are included in the training data pool, and a unified relation label system is used for supervised learning.

[0040] Step S3: Crop local image patches centered on the cell occlusion boundary and construct multi-channel samples containing the local image patches.

[0041] Around each cell occlusion boundary point, this invention crops a fixed-size local image patch from the original microscopic image, centered on the boundary point, as the basic sample input. The size of the local image patch is set to N×N pixels, where N is determined according to the cell scale and microscopic resolution, ensuring that the patch range can cover sufficient boundary structure information within the neighborhood of the boundary point. The cropping process employs a uniform pixel scale and edge filling strategy to ensure consistency across different sample inputs.

[0042] In addition to the original image patch, this invention simultaneously generates and trims structural auxiliary channels to explicitly provide association information on instances, boundaries, and orientations. Multi-channel input enables the network to directly learn the structural differences between long and short sides and the geometric features of contacts within the same layer, without relying on manually designed thresholds or handcrafted features. The specific composition of the multi-channel sample includes four parts: the first part is the original grayscale or color image patch, providing the original appearance information of the neighborhood of the intersection point; the second part is the instance mask patch, identifying the affiliation of different cell instances, enabling the network to distinguish which cell the boundary belongs to; the third part is the boundary map patch, which can be in the form of a boundary binary map or a boundary distance transformation map, emphasizing the boundary position and structure; the fourth part is the orientation encoding patch, representing the orientation information of the boundary line segments within the neighborhood of the intersection point in the form of discrete angle encoding, unit vector field encoding, or orientation heatmap, providing the network with explicit geometric orientation priors. Depending on actual needs, auxiliary channels such as gradient magnitude maps and edge response maps can also be selectively added to enhance the network's adaptability to different imaging conditions. All the channels mentioned above are strictly spatially aligned within the same clipping window and bound to the coordinates of the intersection point, forming a standardized multi-channel sample that can be directly used for training and inference.

[0043] Step S4: Use unified relation labels to label multi-channel samples. Unified relation labels include occlusion relationship, same layer relationship and undeterminable relationship.

[0044] Specifically, such as Figure 3 As shown, this invention uniformly defines all boundary point samples as boundary point relationship determination tasks, and the network output directly corresponds to the local relationship category between cells. For any boundary point sample, the two cell instances with the largest contribution within the boundary point's neighborhood are taken as the principal instance pair, denoted as cell A and cell B, and labeled with a unified relationship tag. The unified relationship tag includes three categories: the first is occlusion relationship, indicating that cell A is above and occludes cell B, denoted as A→B, or cell B is above and occludes cell A, denoted as B→A; the second is same-layer relationship, indicating that the two cells are in contact or converge on the same plane without vertical occlusion, denoted as A∥B; the third is undeterminable, which is used when there is transparent superposition, insufficient evidence, or excessive noise that makes stable determination impossible.

[0045] Under this unified relationship labeling system, the Y-shaped convergence point of cells is not excluded from the training data, but rather serves as a typical sample source for same-layer relationships. Samples geometrically identified as Y-shaped convergence points of cells are labeled as same-layer relationships A∥B by default. If, during manual verification, it is found that the intersection point is actually caused by occlusion, such as a pseudo-Y structure due to segmentation errors or a clearly visible local occlusion interruption, the label can be modified to an occlusion relationship, i.e., A→B or B→A. If it still cannot be confirmed, it is labeled as undetermined. Samples of T-shaped occlusion points and X-shaped intersection points of cells are labeled as occlusion relationships, same-layer relationships, or undetermined based on manual judgment, to cover the actual situations of cell scenarios such as true occlusion, same-layer contact, and transparent overlay.

[0046] To ensure that network learning aligns with traditional terminology in its structural interpretation, this invention introduces long-edge and short-edge attribution as auxiliary supervision information, but this is only enabled when the unified relation label is an occlusion relation. Specifically, the annotator identifies the attribution of the continuous long edge and the terminating short edge within the neighborhood of the boundary point: boundary line segments extending approximately 180 degrees in opposite directions within the neighborhood of the boundary point are identified as continuous long edges, and boundary structures terminating at the boundary point corresponding to the occlusion semantics are identified as short edges. The cell instances belonging to the long edge and the short edge are recorded respectively. For samples A∥B in the same layer, long-edge and short-edge attribution are not mandatory. During training, the loss of this auxiliary task is masked to avoid supervision noise caused by forcibly defining long and short edges for samples in the same layer.

[0047] To reduce annotation costs, this invention supports a semi-automatic annotation strategy. First, geometric rules provide candidate instance pairs and boundary direction hints. The annotation tool, by default, labels Y-shaped samples as being in the same layer and provides initial occlusion directions and pseudo-labels for T-shaped samples with typical shapes, assigning them to their long and short sides. Annotators can complete the annotation by confirming or correcting the relationship categories and long / short side assignments through an interactive interface, significantly reducing manual workload. For difficult examples, annotators can directly select undecidable examples and mark them as difficult, allowing for weight reduction or rejection strategies during the training phase.

[0048] Step S5: Train a deep learning network using labeled multi-channel samples. The deep learning network outputs the probability distribution of the relationship categories of the boundary points.

[0049] This invention organizes the labeled boundary point samples into training, validation, and test sets. Data partitioning employs a strategy of grouping by image source or field of view number to ensure that samples from the same field of view or adjacent frames are not distributed across sets, thus avoiding data leakage that could lead to overestimation of performance. During partitioning, care is taken to maintain balanced coverage across different experimental batches, different microscopic devices, different cell densities, and different imaging qualities, enabling the trained model to learn cross-conditionally stable discriminative features.

[0050] To address the class imbalance problem, this invention employs resampling or class weighting for balancing. When the proportion of samples in the same layer is too high, the proportion of undecidable samples is too high, or the occlusion direction is unevenly distributed, the sampling frequency or loss weight of each class is adjusted to prevent the model from favoring one class output while ignoring other classes.

[0051] Data augmentation operations are performed without altering the topological semantics of the boundaries. Specific augmentation techniques include rotation, translation, slight scaling, brightness perturbation, contrast perturbation, and noise injection. All augmentation operations are applied simultaneously to the original image patch, instance mask patch, boundary map patch, and orientation encoding patch to ensure that the input channels and labels remain strictly aligned after augmentation, thereby improving the model's robustness to different imaging conditions and cell morphology variations.

[0052] This invention trains a deep learning network for local patching of boundary points. Taking multi-channel boundary point samples as input, it outputs the probability distribution of boundary point relationship categories. The network output includes four or three probability values: occlusion relationship (A→B or B→A), same-layer relationship A∥B, and undecidable. Cell A and cell B are determined by the dominant instance pair in the samples. The network structure can employ a lightweight convolutional neural network or a local classification network with attention mechanisms. The network design ensures that it can effectively learn the texture features and geometric patterns of cell boundary interactions within a small window.

[0053] To enhance the network's ability to learn the semantics of long and short edges, this invention sets up an auxiliary output head to predict the attribution information of connecting long edges and terminating short edges, and employs a multi-task learning strategy for joint training. The main task is boundary point relationship classification, optimized using cross-entropy loss or focus loss; the auxiliary task is long / short edge attribution prediction. During training, when the unified relationship label of a sample is an occlusion relationship, both the main task loss and the auxiliary task loss are calculated simultaneously; when the unified relationship label of a sample is a same-layer relationship or undecidable, only the main task loss is calculated and the auxiliary task loss is masked, thus ensuring that the supervision of long and short edges focuses on serving the learning of occlusion relationships without interfering with the learning process of same-layer samples.

[0054] For undecidable categories, this invention employs reasonable weight reduction or a threshold rejection strategy to prevent undecidable samples from dominating the training process. During training, threshold selection, early stopping judgment, and hyperparameter adjustment are performed based on the validation set, ultimately resulting in a model that can stably output results for occlusion relationships, same-layer relationships, and undecidable relationships under different cell densities and imaging conditions.

[0055] Step S6: Input the multi-channel samples of the microscopic image of the cell to be tested into the trained deep learning network, and output the boundary point relationship category and its confidence level.

[0056] like Figure 4As shown, after training is complete, the present invention performs the same preprocessing, instance segmentation, boundary extraction, intersection point detection, and multi-channel sample construction steps on the microscopic images of the cells to be tested during the inference phase as during the training phase. Each intersection point sample to be tested is input into the trained deep learning network to obtain the intersection point relationship category and its confidence level. For intersection points with a confidence level higher than a preset threshold, the corresponding cell occlusion or same-layer relationship determination result is output; for intersection points that cannot be determined or have insufficient confidence, no definitive conclusion is output or only a weak output is made.

[0057] After obtaining the relationship determination results for each boundary point, this invention transforms occlusion relationships into directed edges in the cell relationship graph and same-layer relationships into undirected edges. Directed edges indicate a vertical occlusion relationship between two cells, with the edge direction determined by the occlusion direction; undirected edges indicate that two cells are in contact on the same plane and there is no hierarchical difference. Since local determinations may be inconsistent, this invention performs cycle detection on directed edges. When an occlusion relationship forms a cycle, the edge with the lowest confidence in the cycle is deleted to eliminate logical contradictions, thereby obtaining a stable and consistent directed graph of cell occlusion relationships.

[0058] Finally, this invention uses a cell relationship diagram to color-overlay and display cell instance masks. The coloring strategy assigns different display colors or transparency based on the relative hierarchical relationship of the cells, allowing observers to intuitively identify which cells are located in the upper layer and which are located in the lower layer. The system output includes a color-visualized image as well as boundary point-level and cell pair-level relationship data, which can be directly used by downstream applications such as cell counting, morphological statistics, culture monitoring, and pathological analysis.

[0059] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described; only preferred embodiments of the present invention are illustrated. The descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. As long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0060] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for training a deep learning-based boundary point relationship network and determining cell occlusion, characterized in that, Includes the following steps: Step S1: Perform instance segmentation on the cell micrograph to obtain a cell instance mask, and extract cell boundaries based on the cell instance mask; Step S2: Detect the intersection points of the cell boundaries to obtain cell occlusion intersection points. Classify the cell occlusion intersection points according to the direction, number and geometric distribution of the boundary line segments in the neighborhood of the cell occlusion intersection points. Classify the cell occlusion intersection points as cell T-shaped occlusion points, cell Y-shaped convergence points or cell X-shaped intersection points. Step S3: Crop a local image patch centered on the cell occlusion boundary point and construct a multi-channel sample containing the local image patch; the multi-channel sample includes the original image patch, instance mask patch, boundary map patch and orientation encoding patch; the orientation encoding patch encodes the boundary orientation in the neighborhood of the cell occlusion boundary point in the form of discrete angle, unit vector field or orientation heatmap; Step S4: Label the multi-channel samples using unified relation labels, which include occlusion relationship, same layer relationship, and undeterminable relationship; The multi-channel samples classified as Y-shaped cell convergence points are labeled as being in the same layer by default; Step S5: Train a deep learning network using the labeled multi-channel samples, and the deep learning network outputs the probability distribution of the relationship categories of the boundary points; Step S6: Input the multi-channel samples of the microscopic image of the cell to be tested into the trained deep learning network, output the boundary point relationship category and its confidence, filter the valid boundary point relationship category according to the confidence, convert the occlusion relationship into the directed edge of the cell relationship graph, and convert the same layer relationship into the undirected edge of the cell relationship graph; Cycle detection is performed on the directed edges, and the edges with the lowest confidence are deleted to eliminate loops; the cell instance mask is colored and overlaid based on the cell relationship graph.

2. The method according to claim 1, characterized in that, In step S1, the cell microscopic image is preprocessed before instance segmentation. The preprocessing includes brightness normalization, contrast normalization, and noise suppression. After performing hole filling, spot removal, and morphological smoothing on the cell instance mask, the cell boundary is extracted and refined into boundary line segments with a width of one pixel.

3. The method according to claim 1, characterized in that, In step S2, the boundary point detection includes: screening candidates for spatially close cell pairs or those that come into contact after mask expansion; discretizing the boundaries of candidate cell pairs into a set of line segments and performing intersection detection to obtain boundary candidate points; and clustering and merging the spatially close boundary candidate points to obtain the cell occlusion boundary points.

4. The method according to claim 1, characterized in that, In step S2, the method for classifying the boundary line segments in the neighborhood of the cell occlusion boundary point according to their direction, number, and geometric distribution includes: when there is a pair of boundary line segments extending approximately 180 degrees in opposite directions and a boundary line segment terminating at the boundary point, it is classified as a cell T-shaped occlusion point; when there are three boundary line segments distributed approximately symmetrically, it is classified as a cell Y-shaped convergence point; when there are two pairs of boundary line segments extending approximately 180 degrees in opposite directions and intersecting at a cross, it is classified as a cell X-shaped intersection point.

5. The method according to claim 1, characterized in that, In step S4, when manual verification confirms the existence of occlusion, the same-layer relationship is modified to an occlusion relationship.

6. The method according to claim 1, characterized in that, Step S4 further includes: when the unified relationship label is an occlusion relationship, the long side and short side are labeled for the multi-channel samples as auxiliary supervision information, wherein the long side is a through boundary extending approximately 180 degrees in opposite directions within the neighborhood of the intersection point, and the short side is a boundary that terminates at the intersection point.

7. The method according to claim 6, characterized in that, In step S5, the deep learning network adopts multi-task learning, with the main task being boundary point relationship classification and the auxiliary task being long-side and short-side attribution prediction. When the unified relationship label is an occlusion relationship, the loss of the main task and the loss of the auxiliary task are calculated simultaneously. When the unified relationship label is a same-layer relationship or cannot be determined, the loss of the auxiliary task is masked.

8. The method according to claim 1, characterized in that, Before step S5, the method further includes: dividing the labeled multi-channel samples into training set, validation set and test set according to the image source; using resampling or class weights to balance the class imbalance; and performing data augmentation operations such as rotation, translation, scaling, brightness perturbation and noise injection on the multi-channel samples.

Images