A semi-supervised lymphoma cell segmentation system and method with pseudo-label screening

Abstract

The application relates to the technical field of medical image segmentation, and discloses a semi-supervised lymphoma cell segmentation system and method based on pseudo label screening, proposes small block patch cutting of different resolutions, realizes multi-scale instance feature and image feature collaborative representation through multi-scale deep feature extraction and fusion of an EfficientNets model, introduces a trained XGBoost model as an accurate quality filter, carries out uncertainty, fixed threshold and adaptive threshold screening on the preliminary results of an LC-YOLO model, the adaptive threshold screening of the XGBoost model can dynamically judge the authenticity of each predicted target according to the fusion features of the predicted target, effectively eliminates noise prediction inconsistent with the features of real cells, and ensures the reliability and stability of the final high-quality pseudo label.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical image segmentation, in particular to a semi-supervised lymphoma cell segmentation system and method based on pseudo-label screening. BACKGROUND

[0002] Lymphoma is a malignant tumor derived from lymphocytes, and different subtypes have their own characteristics in pathological morphology and cell composition. Lymphoma cells show high heterogeneity, mainly in cell lineage diversity, significant size difference and irregular morphology. Therefore, whole-slide imaging has become an important basis for pathologists to diagnose lymphoma, and lymphoma pathological whole-slide images WSI are usually on the order of tens of billions of pixels, and the number of cells in the image is often extremely large.

[0003] At present, the traditional lymphoma cell segmentation method relies heavily on pathologists to label cells one by one on the whole-slide image of tens of billions of pixels, which is time-consuming and labor-intensive, resulting in high labeling cost and difficulty in large-scale application. Although there are semi-supervised methods that use models to generate pseudo-labels to reduce labeling dependence, the cell size in the whole-slide image is very different from the overall image size, and the model is difficult to perceive both the global organizational structure and the local cell morphology, resulting in insufficient feature extraction and a large number of missed (small cells are submerged) and false positives (background impurities are misjudged) in the generated pseudo-labels. In addition, existing methods usually use a single fixed threshold to screen pseudo-labels, which cannot adapt to changes in cell density, morphology and staining depth in different image regions, and the quality of pseudo-labels fluctuates greatly, which directly leads to error accumulation and limited performance improvement when used for model iterative training.

[0004] In addition, in lymphoma cell segmentation, traditional models have obvious shortcomings: in terms of scale adaptation, a single resolution processing strategy cannot balance global context and local details. If a low resolution is used, small cells and morphological features are lost, and if a high resolution is used, the computational burden increases dramatically and the receptive field is limited, making it difficult to capture long-range intercellular associations. In terms of pseudo-label utilization, a simple self-training framework lacks effective quality evaluation and screening mechanisms for pseudo-labels generated by the model itself, introducing noise and errors into the training set, which severely restricts the further improvement of model performance. SUMMARY

[0005] The present application aims to improve the low correlation between lymphoma whole-slide images and cell scales, high artificial labeling cost, and unstable pseudo-label quality in the prior art, and provides a semi-supervised lymphoma cell segmentation system and method based on pseudo-label screening.

[0006] In order to achieve the above application purpose, the embodiments of the present application provide the following technical solutions:

[0007] A semi-supervised lymphoma cell segmentation method based on pseudo-label screening, comprising the following steps:

[0008] The whole slice image is divided into a plurality of resolution small block tiles;

[0009] The multi-resolution small block tiles are input into an EfficientNets model for multi-scale feature extraction, and multi-scale image features are extracted;

[0010] The single scale small block tiles are manually labeled with cell instances to obtain instance structure features;

[0011] The instance structure features and the multi-scale image features are fused to train an XGBoost model;

[0012] The manually labeled small block tiles are input into an LC-YOLO model for full supervision training;

[0013] The LC-YOLO model based on full supervision training is used to infer the small block tiles that are not manually labeled to obtain preliminary pseudo labels;

[0014] The preliminary pseudo labels are filtered through uncertainty filtering, fixed threshold filtering of non-maximum suppression, and adaptive threshold filtering of the XGBoost model to obtain high-quality pseudo labels.

[0015] In view of the low correlation between the whole slice and the cell scale, the present application solves the problem by multi-resolution tile division and multi-scale feature fusion: the whole slice image is first decomposed into tiles of different resolutions, so that the EfficientNets model can capture macroscopic tissue distribution at low resolution and finely depict cell morphology at each resolution; then the multi-scale image features are fused with the key cell instance structure features obtained by manual annotation to train the XGBoost model, so that the model can understand the tissue environment and the cell itself at the same time, effectively bridging the semantic gap between macroscopic images and microscopic cells.

[0016] In view of the high cost of manual annotation and the unstable quality of pseudo labels, the present application constructs a collaborative filtering system based on the LC-YOLO model and the XGBoost model. First, only a small amount of tiles need to be manually annotated to train the initial LC-YOLO model, which greatly reduces the annotation cost. Then, the model is used to generate preliminary pseudo labels from a large amount of unlabeled data. The key is to introduce the trained XGBoost model as a precise quality filter to filter the preliminary results of the LC-YOLO model in terms of uncertainty, fixed threshold, and adaptive threshold. The adaptive threshold filtering of the XGBoost model can dynamically determine the authenticity of each predicted target based on its fusion features (multi-scale image features + instance structure features), effectively eliminating noise predictions that do not match the characteristics of real cells, and ensuring the reliability and stability of the final high-quality pseudo labels.

[0017] Compared with the prior art, the beneficial effects of the present application are: the present application realizes multi-resolution cutting and feature fusion through the closed-loop design of multi-scale feature fusion and double-model collaborative pseudo-label screening, realizes high-precision adaptation from large full slice to tiny cells, and improves the feature perception ability of the model under different scales; through the semi-supervised process started by only a small amount of labeling, the dependence on expensive manual labeling is significantly reduced, so that the method has the potential for large-scale application; through the collaborative work of the LC-YOLO model and the XGBoost model, the pseudo-label quality screening is constructed, the XGBoost model uses its fused rich features for adaptive screening, which fundamentally guarantees the accuracy and reliability of the pseudo-label, breaks the performance bottleneck caused by pseudo-label noise in traditional semi-supervised learning, and finally realizes high-precision, high-robust lymphoma cell segmentation under low labeling cost.

[0018] Further, a pseudo-label screening semi-supervised lymphoma cell segmentation method, the plurality of resolutions include 256x256, 128x128, 64x64 pixels.

[0019] Further, a pseudo-label screening semi-supervised lymphoma cell segmentation method, the plurality of resolutions include 256x256, 128x128, 64x64 pixels.

[0020] Read the format label in the same small block patch, obtain the bounding box and polygon contour;

[0021] Calculate the area A, perimeter P, shape factor of the bounding box and polygon contour in the OpenCV library to form the instance structure feature , ;

[0022] Among them, 3 is the dimension of the instance structure feature, and the dimension of the instance structure feature includes A, P, .

[0023] In the above scheme, by constructing the instance-level structural features composed of area, perimeter and shape factor, the technical problem that the existing lymphoma cell segmentation method excessively relies on image-level deep features and ignores the key geometric morphological information of cell instances, resulting in difficulty for the model to effectively depict and distinguish lymphoma cells in size, irregularity and other aspects, is solved. The application forms a set of low-dimensional but high-discriminatory instance-level structural features by calculating three core geometric parameters of area, perimeter and shape factor. The area directly quantifies the cell scale and is the basis for distinguishing cell subclasses. The perimeter and shape factor together depict the complexity and regularity of the cell contour. The application realizes the complementation and enhancement of the cell appearance texture features and the inherent geometric morphological features by fusing the instance-level structural features with the multi-scale image features in the XGBoost model.

[0024] Further, a semi-supervised lymphoma cell segmentation method for pseudo-label screening, the instance structure features and the multi-scale image features are fused to train the XGBoost model, and the specific steps are as follows:

[0025] The multi-scale image features are fused with the instance-level structural features to obtain the fused feature vector .

[0026] The EfficientNets model generates a training feature matrix X containing positive and negative samples from all the fused feature vectors, and a label vector corresponding to the training feature matrix.

[0027] Wherein, N is the total number of fused feature vectors, D is the dimension of multi-scale image features, and 3 is the dimension of instance structure features.

[0028] The training feature matrix and the corresponding label vector are used to train the XGBoost model to obtain the confidence score, so that the XGBoost model can distinguish positive and negative samples.

[0029] Wherein, the positive sample is the cell instance segmentation region, and the negative sample is the created non-cell region.

[0030] In the above scheme, by constructing the fusion representation of multi-scale image features and instance structure features and the XGBoost model training process, the problem of single feature expression in the traditional lymphoma cell segmentation method is effectively solved: the existing method usually only relies on the image-level deep features extracted by the convolutional neural network, and ignores the geometric morphological properties of the cell instance itself, which makes it difficult to accurately distinguish lymphoma cells with different morphologies and background tissues in the pseudo-label generation link, especially in the areas with high cell density and abnormal morphology. The traditional deep features can capture texture information, but lack modeling of key discriminant factors such as cell size and shape regularity, which limits the model's representation ability for highly heterogeneous lymphoma cells. The present application fuses multi-scale image features and instance-level structure features to form enhanced feature representation, so that the model has the ability to perceive cell surface texture and geometric morphology; then the XGBoost model is used to train the fused high-dimensional features, and the complex nonlinear boundary between positive and negative samples is automatically learned through the gradient boosting decision tree architecture, and through multiple rounds of iteration optimization, the cell probability of each candidate region can be accurately evaluated, wherein the positive sample corresponds to the real cell instance segmentation area, and the negative sample systematically covers various non-cell instance segmentation area interference. The present application improves the accuracy of pseudo-label screening through the collaborative design of feature fusion and XGBoost model: multi-scale image features ensure the sensitivity of cell detection, instance structure features enhance the model's ability to recognize cell morphological specificity, and XGBoost model realizes the organic combination of the advantages of the two through ensemble learning strategy, so that the model can realize the segmentation learning of lymphoma cells while keeping the low cost of manual annotation.

[0031] Further, a semi-supervised lymphoma cell segmentation method for pseudo-label screening, the LC-YOLO model comprises a backbone network, a neck network, and a star segmentation head;

[0032] The backbone network extracts multi-scale backbone features from the input lymphoma cell image;

[0033] The neck network fuses the multi-scale features through the top-down and bottom-up feature pyramid structure to obtain the fused features;

[0034] The star segmentation head completes the bounding box regression, foreground classification and star contour decoding in parallel on three scales, and outputs a GEOJSON format file including a polygon.

[0035] In the above scheme, by constructing the LC-YOLO model architecture comprising a backbone network, a neck network and a star-shaped segmentation head, the key problems existing in the current lymphoma cell instance segmentation method are effectively solved: the segmentation method based on rectangular bounding box and fixed template of the traditional instance segmentation model (such as Mask R-CNN) cannot accurately fit the actual contour of the cell when dealing with the characteristics of highly irregular shape and significant size difference of lymphoma cells, resulting in inaccurate cell boundary positioning and serious detail loss; at the same time, the existing single-stage segmentation model is prone to problems such as overlapping of predicted boxes and degradation of mask quality when processing cell dense areas, and cannot meet the demand of pathological diagnosis for accurate depiction of cell morphology. The backbone network of the present application extracts the multi-scale backbone features of the image, ensuring that the model can perceive complete information from local cell details to global tissue structure; the neck network adopts a top-down and bottom-up feature pyramid structure, realizing effective fusion of multi-scale features and enhancing the detection ability of the model for cells of different sizes; the star-shaped segmentation head innovatively completes the bounding box regression, foreground classification and star-shaped contour decoding in parallel at three scales, realizes accurate fitting of the irregular cell contour through the ray length regression method starting from the cell center point, and finally outputs a GEOJSON format file containing a complete polygon contour. Through the progressive architecture design of "multi-scale feature extraction-feature pyramid fusion-star-shaped contour decoding", the present application constructs an end-to-end lymphoma cell instance segmentation solution: the backbone network ensures the perception ability of the cell basic features, the neck network solves the detection problem of multi-scale cells, and the star-shaped segmentation head realizes high-precision restoration of the real shape of the cell through its unique contour representation method, which not only significantly improves the accuracy of cell boundary segmentation, but also provides a directly usable data basis for subsequent morphological analysis through the output of the structured GEOJSON format, so that the model realizes accurate segmentation of lymphoma cell instances while maintaining high-speed inference.

[0036] Further, a pseudo-label screening semi-supervised lymphoma cell segmentation method, the star-shaped segmentation head divides the segmentation head input features into three branches, left branch, middle branch and right branch;

[0037] The left branch is used for calculating the bounding box regression of the features, comprising CBS-b1 layer, CBS-b2 layer, Conv2d-b1 layer and bounding box loss module connected in sequence;

[0038] The middle branch is used for calculating the foreground classification of the features, comprising DWConv-f1 layer, CBS-f1 layer, DWConv-f2 layer, CBS-f2 layer, Conv2d-f3 layer and classification loss module connected in sequence;

[0039] The right branch is used for decoding the star-shaped contour of the feature, and includes a CBS-x1 layer, a CBS-x2 layer, a Conv2d-x1 layer, a mask coefficient module, a mask module, a star polygon module and a high-pass screening module connected in sequence.

[0040] In the above scheme, by constructing a three-branch star-shaped segmentation head architecture, the technical limitations of existing instance segmentation methods in processing lymphoma cells are effectively solved: the traditional segmentation head simultaneously processes the positioning, classification and segmentation tasks, resulting in mutual interference of feature representation, and it is difficult to balance the boundary box accuracy, classification confidence and contour detail integrity; especially in the cell dense area, it is easy to cause mask overlap, boundary blur and other problems. The segmentation head is designed as three independent branches: the left branch focuses on the accurate regression of the bounding box; the middle branch reduces the computational complexity while ensuring the feature expression ability, and realizes accurate foreground classification; the right branch decodes the features into accurate polygon contour representation through the cooperation of the mask coefficient module and the star polygon module. Through the parallel processing architecture of the three branches, the present application not only avoids the feature interference between tasks, but also ensures the quality of the output contour through the high-pass screening module, and provides accurate geometric structure information for subsequent pseudo-label generation.

[0041] Further, a semi-supervised lymphoma cell segmentation method for pseudo-label screening, the processing process of the right branch is:

[0042] The segmentation head input feature input is extracted by the CBS-x1 layer and the CBS-x2 layer to the Conv2d-x1 layer;

[0043] The Conv2d-x1 layer generates a low-pass feature output of the mask to the mask coefficient module;

[0044] The mask coefficient module performs linear mapping and outputs to the mask module, and the mask module obtains an instance mask and outputs to the star polygon;

[0045] The star polygon module extracts the boundary through the instance mask, projects the boundary into a star vertex sequence, and splices the confidence, area and perimeter of the instance into an instance feature vector;

[0046] A plurality of instance feature vectors are input into the high-pass screening module, the high-pass screening module screens out low-quality instances according to a threshold and a morphological constraint, and converts the screened instance feature vectors into a GEOJSON format file for export.

[0047] In the above scheme, through the star-shaped contour decoding flow of the right branch, the technical problems that the traditional method adopts fixed rectangular boundary box or simple mask prediction, which is difficult to accurately fit the highly irregular morphological characteristics of lymphoma cells, and boundary adhesion is easy to occur in cell dense area are effectively solved. The present application strengthens the boundary details, optimizes the feature representation through linear mapping, generates accurate instance masks, and the star-shaped polygon module innovatively projects the continuous boundary into a discrete star-shaped vertex sequence, realizing the conversion from pixel-level mask to vector polygon; finally, through the high-pass screening module, the instance is quality filtered based on the morphological constraints such as confidence, area and perimeter, and the standardized GEOJSON format file is output. Through the progressive processing flow of "feature extraction-mask generation-vector conversion-quality screening", the present application constructs an efficient cell contour analysis scheme: the deep separable convolution reduces the computational complexity while maintaining the boundary details, the representation mode of the star-shaped vertex sequence not only retains the integrity of the cell morphology but also is convenient for subsequent analysis, and the high-pass screening mechanism ensures the reliability of the output instance, which significantly improves the standardization and practicality of the contour representation while ensuring the segmentation accuracy.

[0048] Further, a semi-supervised lymphoma cell segmentation method for pseudo label screening, the preliminary pseudo label is screened by uncertainty, and the specific steps are:

[0049] The LC-YOLO model under full supervision training is used for k times of repeated reasoning on the small block patches Patch without manual annotation, and the segmentation mask set generated each time is collected The confidence score is compared with the confidence score ;

[0050] The standard deviation of each confidence score and the average IoU between multiple mask predictions are calculated, and the uncertainty index is defined:

[0051] ;

[0052] Wherein, is the uncertainty index, is the standard deviation of the confidence score, is the average value of the mask consistency;

[0053] The threshold value is set, the uncertainty screening is carried out, the low-noise instances with the uncertainty index < in the preliminary label are retained, and the preliminary labels with large confidence fluctuation or inconsistent masks are removed.

[0054] In the above scheme, by constructing an uncertainty screening mechanism based on repeated inference, the quality fluctuation problem existing in traditional pseudo-label generation methods is effectively solved. Existing methods usually rely only on the results of a single model inference, and cannot evaluate the reliability of the prediction results. This leads to low-quality predictions with large confidence fluctuations or inconsistent masks being mistakenly classified as pseudo-labels, thereby introducing noise and causing error accumulation in semi-supervised training. This invention obtains the segmentation mask set and confidence score through k repeated inferences, and constructs an uncertainty index by combining the standard deviation of confidence and the average IoU of the mask. This index quantifies both the stability and consistency of the model prediction. By setting a threshold for screening, high uncertainty instances are systematically eliminated, while reliable pseudo-labels are retained.

[0055] Furthermore, a semi-supervised lymphoma cell segmentation method using pseudo-label screening, wherein the fixed threshold screening with non-maximum inhibition comprises the following steps:

[0056] Segmentation mask set A soft mask is synthesized through K inferences. The representative mask is then obtained by binarizing the soft mask, as shown in the formula:

[0057] ;

[0058] ;

[0059] in, This is a soft mask, with pixel values ​​ranging from 0 to 1. It represents the average probability that the i-th instance is in the foreground at pixel x, where x is a pixel. Let i be the set of mask indices matched in K repeated inferences for the same instance i. For set The number of elements, Let i be the set of segmentation masks. As a representative mask, is the soft mask binarization threshold, and 1 is an indicator function that takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise.

[0060] In the embodiments, The value is 0.5;

[0061] Confidence score The ranking score of candidate instances is synthesized through K inferences, and the formula is:

[0062] ;

[0063] Where k is the index of the k-th repeated reasoning (k=1,…,K), and K is the total number of repeated reasoning operations performed on the same tile. For candidate instance i in The average confidence score obtained from the inference process. a final ranking score of the candidate instance i, an uncertainty indicator of the candidate instance i;

[0064] obtain a candidate set through the ranking score and the soft mask of the candidate instance , in descending order of the final ranking score, put the candidate instance with the highest final ranking score into the result set, and calculate the mask for other final ranking scores according to the formula:

[0065] ;

[0066] wherein, the mask of the other candidate instance q, the representative mask of the other instance q, the candidate instance p with the highest final ranking score;

[0067] if , the other candidate instance q is suppressed to ensure the sparsity and diversity of the candidate set.

[0068] In the above scheme, in view of the problems of mask overlap and inaccurate confidence evaluation of the traditional pseudo-label screening method in the dense cell instance scene, the prior art usually uses a single threshold or a simple NMS strategy, which is difficult to effectively distinguish highly overlapping cell instances and cannot balance the detection recall rate and the precision rate. The present application realizes accurate screening of dense cell instances through soft mask synthesis and improved non-maximum suppression strategy: first, the soft mask is synthesized through K times reasoning, and a more representative binary mask is obtained based on probability average; second, the final ranking score is calculated based on the confidence score and the uncertainty indicator to ensure that high-quality instances are selected first; finally, through the improved NMS mechanism, redundant detection is dynamically suppressed based on the mask overlap rate, while retaining real instances with different shapes. Through the three-stage processing flow of "soft mask synthesis-comprehensive scoring-dynamic suppression", the present application constructs an adaptive pseudo-label screening mechanism, the soft mask synthesis improves the accuracy of boundary positioning, the comprehensive scoring mechanism ensures the scientificity of the screening standard, and the dynamic suppression based on the mask overlap rate effectively balances the sensitivity and specificity of the detection, while maintaining a high recall rate and significantly reducing false positives, providing a high-quality pseudo-label basis for subsequent model training.

[0069] A semi-supervised lymphoma cell segmentation system for pseudo-label screening includes a split image module, a feature extraction module, an instance labeling module, a feature fusion module, an LC-YOLO model training module, an LC-YOLO model inference module, and a preliminary pseudo-label screening module.

[0070] The split image module splits the whole slice image into a plurality of small block tiles with different resolutions.

[0071] The feature extraction module inputs the multi-resolution small block patches into the EfficientNets model for multi-scale feature extraction to extract multi-scale image features;

[0072] The instance labeling module manually labels the cell instances of the single-scale small block patches to obtain instance structure features;

[0073] The feature fusion module fuses the instance structure features and the multi-scale image features for training the XGBoost model;

[0074] The LC-YOLO model training module inputs the manually labeled small block patches into the LC-YOLO model for full-supervised training;

[0075] The LC-YOLO model inference module performs inference on the small block patches that are not manually labeled based on the LC-YOLO model trained by full supervision to obtain preliminary pseudo labels;

[0076] The preliminary pseudo label screening module obtains high-quality pseudo labels after the preliminary pseudo labels are screened by uncertainty screening, fixed threshold screening of non-maximum suppression, and adaptive threshold screening of the XGBoost model.

[0077] In the above scheme, by constructing a modular lymphoma cell segmentation system, the multiple challenges faced by traditional semi-supervised learning systems in WSI cell segmentation are effectively solved: existing systems usually use a single model for pseudo label generation and optimization, resulting in insufficient feature extraction, unstable pseudo label quality, and poor performance in lymphoma images with highly heterogeneous cell morphology. The present application realizes multi-scale processing of the whole slide image through the image segmentation module, the feature extraction module and the instance labeling module respectively obtain the deep image features and the cell morphology features, the feature fusion module effectively integrates the multi-source features, the LC-YOLO model training and inference module realizes preliminary detection, and finally the three screening mechanisms ensure the quality of the pseudo labels. Through the systematic modular design, the present application realizes the optimization from data preprocessing to high-quality pseudo label generation, the multi-resolution segmentation ensures the integrity of the features of different scales, the multi-source feature fusion enhances the model's representation ability for heterogeneous cells, and the three screening mechanisms guarantee the reliability of the pseudo labels from different dimensions, which reduces the cost of manual labeling, and significantly improves the accuracy and stability of cell segmentation through continuous iterative optimization. BRIEF DESCRIPTION OF DRAWINGS

[0078] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. It should be understood that the following drawings only show some of the embodiments of the present application, and therefore should not be regarded as a limitation on the scope, and for those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0079] Figure 1 Flow chart of semi-supervised lymphoma cell segmentation method for pseudo label screening.

[0080] Figure 2 Flow chart of multi-scale feature extraction and fusion.

[0081] Figure 3 Structure diagram of LC-YOLO model.

[0082] Figure 4 Principle diagram of adaptive threshold screening of XGBoost model.

[0083] Figure 5 Technical principle diagram of traditional pseudo label simplification.

[0084] Figure 6 Technical principle diagram of pseudo label simplification of the present application.

[0085] Figure 7 Structure diagram of semi-supervised lymphoma cell segmentation system for pseudo label screening.

[0086] Figure 8 Effect diagram of semi-supervised effect verification.

[0087] Figure 9 Effect diagram of pseudo label ablation experiment. DETAILED DESCRIPTION

[0088] The technical solutions of the embodiments of the present application will be described clearly and completely in the embodiments of the present application combined with the drawings. Obviously, the described embodiments are only some of the embodiments of the present application, not all. The components of the embodiments of the present application described and shown in the drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of the present application.

[0089] It should be noted that similar reference numerals and letters refer to like items in the accompanying drawings, and once an item is defined in one drawing, it need not be further defined and explained in subsequent drawings. Also, in the description of the present application, the terms "first", "second", and the like are used only to distinguish different descriptions, and cannot be understood as indicating or implying relative importance or implying any such actual relationship or order between these entities or operations. In addition, the terms "connected", "connected" and the like can be direct connection between elements, or indirect connection via other elements.

[0090] Embodiment 1: Semi-supervised lymphoma cell segmentation method with pseudo label screening.

[0091] The present application is realized by the following technical solutions, as shown in a semi-supervised lymphoma cell segmentation method with pseudo label screening, comprising the following steps: Figure 1

[0092] S1: The whole slice image WSI is divided into a plurality of small block patches Patch of different scales.

[0093] Specifically, the whole slice image WSI is divided into small block patches Patch of 256x256, 128x128, and 64x64 pixels.

[0094] S2: Input the small block patches of different scales into the EfficientNets model to extract multi-scale image features.

[0095] Specifically, as shown in Figure 2 S2 includes the following steps:

[0096] S21: The small block patches Patch of 256x256, 128x128, and 64x64 pixels are sequentially input into the EfficientNets model after fixed ratio scaling, normalization and tensorization processing;

[0097] S22: The EfficientNets model respectively performs complex scaling on each small block patch Patch, and performs forward inference on three scales to obtain multi-scale feature maps Patch;

[0098] S23: The EfficientNets model extracts output vectors of three scales, concatenates the output vectors, and obtains high-dimensional multi-scale image features .

[0099] S3: Artificially label cell instances of a single scale small block patch Patch to obtain instance structure features.

[0100] Specifically, S3 includes the following steps: ​

[0101] S31: read the format label in the same small block tile Patch, obtain the bounding box and polygon contour;

[0102] S32: calculate the area A, perimeter P, shape factor of the bounding box and polygon contour in the OpenCV library to form the instance structure feature , ;

[0103] wherein 3 is the dimension of the instance structure feature, and the dimension of the instance structure feature includes A, P, .

[0104] It should be noted that the small block tile Patch is configured with the format label COCO, and the format label COCO includes the bounding box, polygon contour and area information.

[0105] S4: fuse the instance structure feature and the multi-scale image feature for training the XGBoost model.

[0106] Specifically, S4 includes the following steps:

[0107] S41: splice the multi-scale image feature and the instance-level structure feature to obtain a fused feature vector ;

[0108] S42: the EfficientNets model collects all the fused feature vectors to generate a training feature matrix X containing positive and negative samples, and a label vector corresponding to the training feature matrix;

[0109] wherein N is the total number of fused feature vectors, D is the dimension of the multi-scale image feature, and 3 is the dimension of the instance structure feature;

[0110] S43: use the training feature matrix and the corresponding label vector to train the XGBoost model to obtain a confidence score, so that the XGBoost model can distinguish between positive and negative samples;

[0111] wherein the positive sample is a cell instance segmentation region, and the negative sample is a created non-cell region.

[0112] The prior art scheme is generally difficult to realize precise characterization of the high heterogeneity of lymphoma cells while ensuring the generalization ability of the model, and there is a lot of noise in the pseudo-label screening process, which affects the effect of subsequent semi-supervised training. The existing method often uses a single model to extract image features, ignoring the geometric morphological properties of the cell instances themselves, resulting in incomplete feature expression; in the pseudo-label screening, it often relies on a fixed threshold, and cannot dynamically adjust the screening standard according to the cell morphology and context information. The present application adopts a method of multi-scale feature extraction based on EfficientNets and fusion of instance structure features, first captures the local details and global distribution of cells through parallel processing of three scales (256x256, 128x128, 64x64); then calculates the area, perimeter and shape factor and other instance-level geometric features to quantify the morphological properties of the cells; finally, the multi-scale image features and instance structure features are spliced to form a high-discriminability fusion feature vector for training the XGBoost model. The present application not only comprehensively covers the apparent information of the cells at different resolutions by means of multi-scale feature extraction, but also explicitly describes the morphological heterogeneity of the cells by introducing instance structure features, effectively making up for the shortcomings of single image features; at the same time, the XGBoost model is used to adaptively screen the fusion features, realizing precise control of the quality of pseudo-labels and providing a high-quality, low-noise training basis for subsequent semi-supervised model iteration.

[0113] S5: inputting the artificially labeled small patch into the LC-YOLO model for full-supervised training.

[0114] Specifically, as shown in the figure, Figure 3 The LC-YOLO model includes a backbone network, a neck network and a star segmentation head.

[0115] The backbone network extracts multi-scale backbone features (P3, P4, P5) from the input lymphoma cell image;

[0116] The neck network fuses the multi-scale features through the feature pyramid structure from top to bottom and from bottom to top to obtain fusion features (F3, F4, F5);

[0117] The star segmentation head completes the bounding box regression, foreground classification and star contour decoding in parallel at three scales, and outputs a GEOJSON format file including a polygon.

[0118] Specifically, the backbone network includes CBS-E1 layer, CBS-E2 layer, C3k2-E1 layer, CBS-E3 layer, C3k2-E2 layer, CBS-E4 layer, C3k2-E3 layer, CBS-E5 layer, C3k2-E4 layer, SPPF layer and C2PSA layer connected in series.

[0119] In the embodiment, asFigure 3 input lymphoma cell image to the input end of the CBS-E1 layer, the output end of the CBS-E1 layer outputs a feature map to the input end of the CBS-E2 layer, the output end of the CBS-E2 layer outputs a feature map to the input end of the C3k2-E1 layer, the output end of the C3k2-E1 layer outputs a feature map to the input end of the CBS-E3 layer, the output end of the CBS-E3 layer outputs to the input end of the C3k2-E2 layer, the first output end of the C3k2-E2 layer outputs a high-resolution stem feature P3 , the second output end of the C3k2-E2 layer outputs a high-resolution stem feature P3 to the input end of the CBS-E4 layer, the output end of the CBS-E4 layer outputs a feature map to the input end of the C3k2-E3 layer, the first output end of the C3k2-E3 layer outputs a medium-resolution stem feature P4 , the second output end of the C3k2-E3 layer outputs a scale stem feature medium-resolution stem feature P4 to the input end of the CBS-E5 layer, the output end of the CBS-E5 layer outputs a feature map to the input end of the C3k2-E4 layer, the output end of the C3k2-E4 layer outputs a feature map to the SPPF layer, the output end of the SPPF layer outputs a feature map to the input end of the C2PSA layer, the first output end and the second output end of the C2PSA layer output a low-resolution stem feature P5 .

[0120] It should be noted that C is the number of channels of the lymphoma cell image, and the subscript and superscript represent whether the number of channels changes in this level (for example )

[0121] Specifically, the neck network comprises an up-sampling 1 module, an up-sampling 2 module, a C3k2-N1 layer, a C3k2-N2 layer, a C3k2-N3 layer, a C3k2-N4 layer, a CBS-N1 layer, and a CBS-N2 layer.

[0122] In the embodiment, as Figure 3 indicated, the first output end of the C2PSA layer outputs a low-resolution stem feature P5 to the input end of the up-sampling 1 module, the output end of the up-sampling 1 module outputs a feature map and the first output end of the C3k2-E3 layer outputs a medium-resolution stem feature P4 are spliced, and the spliced feature input end of the C3k2-N1 layer module, and the first output end of the C3k2-N1 layer module outputs a feature map input end of the upsample2 module, and the output end of the upsample2 module outputs a feature map and the first output end of the C3k2-E2 layer outputs a high-resolution backbone feature P3 perform splicing, and splice the features input end of the C3k2-N2 layer, and the first output end of the C3k2-N2 layer outputs a high-resolution fusion feature F3 , and the second output end of the C3k2-N2 layer outputs the high-resolution fusion feature F3 input end of the CBS-N1 layer, and the output end of the CBS-N1 layer outputs a feature map and the second output end of the C3k2-N1 layer module outputs a feature map perform splicing, and splice the features input end of the C3k2-N3 layer, and the first output end of the C3k2-N3 layer outputs a medium-resolution backhaul fusion feature F4 , and the second output end of the C3k2-N3 layer outputs the medium-resolution backhaul fusion feature F4 input end of the CBS-N2 layer, and the output end of the CBS-N2 layer outputs a feature map and the second output end of the C2PSA layer outputs a low-resolution backbone feature P5 perform splicing, and splice the features input end of the C3k2-N4 layer, and the output end of the C3k2-N4 layer outputs a low-resolution fusion feature F5 .

[0123] Specifically, the star-shaped segmentation head includes a star-shaped segmentation head X1, a star-shaped segmentation head X2, and a star-shaped segmentation head X3.

[0124] The first output end of the C3k2-N2 layer outputs a high-resolution fusion feature F3 input end of the star-shaped segmentation head X1, and the first output end of the C3k2-N3 layer outputs a medium-resolution backhaul fusion feature F4 input end of the star-shaped segmentation head X2, and the output end of the C3k2-N4 layer outputs a low-resolution fusion feature F5 input end of the star-shaped segmentation head X3.

[0125] It should be noted that the formula for the segmentation head input feature is:

[0126] ;

[0127] wherein, is a feature map entering the s-th scale segmentation head, height, width, channel number of the input feature of the segmentation head respectively, corresponding to high resolution, medium resolution and low resolution scale respectively;

[0128] In the embodiment, stride of the is usually ;

[0129] The processing formula of the Conv layer is:

[0130] ;

[0131] wherein, is the input feature tensor , is the convolution kernel weight , is the channel-wise bias, , is a two-dimensional convolution operation, H is the height of the input feature tensor, and W is the width of the input feature tensor, is an input function, is an output function, is a kernel;

[0132] The processing formula of the DSConv layer is:

[0133] ;

[0134] wherein, is a 3x3 kernel for deep convolution , is linear mixing of the channels after deep convolution to ;

[0135] The CBS layer includes a Conv layer, a BN layer and a SiLU activation function.

[0136] The processing formula of the CBS layer is:

[0137] ;

[0138] wherein, op is to specify that the layer uses a Conv layer or a DSConv layer, BN is a BN layer, SiLU is a SiLU activation function, is a 3x3 depth separable convolution, is a 3x3 ordinary convolution.

[0139] Since the processing processes of the star-shaped segmentation head X1, the star-shaped segmentation head X2 and the star-shaped segmentation head X3 are the same, only the input feature sizes are different, the processing processes are uniformly described.

[0140] Specifically, as shown in Figure 3 The processing procedure of the star-shaped segmentation head is as follows:

[0141] The star-shaped segmentation head divides the input feature of the segmentation head into three branches, i.e., a left branch, a middle branch and a right branch.

[0142] The left branch is used for calculating the bounding box regression of the feature, and includes CBS-b1 layer, CBS-b2 layer, Conv2d-b1 layer and bounding box loss module connected in sequence.

[0143] The middle branch is used for calculating the foreground classification of the feature, and includes DWConv-f1 layer, CBS-f1 layer, DWConv-f2 layer, CBS-f2 layer, Conv2d-f3 layer and classification loss module connected in sequence.

[0144] The right branch is used for decoding the star-shaped contour of the feature, and includes CBS-x1 layer, CBS-x2 layer, Conv2d-x1 layer, mask coefficient module, mask module, star-shaped polygon module and high-pass screening module connected in sequence.

[0145] Specifically, the processing procedure of the left branch is as follows:

[0146] The input feature of the segmentation head is input into the CBS-b1 layer to extract the positioning feature, and the output is input into the CBS-b2 layer. The CBS-b2 layer strengthens the regression representation, and the output is input into the Conv2d-b1 layer. The Conv2d-b1 layer generates the bounding box regression amount, and the output is input into the bounding box loss module to calculate the regression loss (such as IoU / GIoU / CIoU) with the labeled box.

[0147] The processing procedure is that the input feature of the segmentation head is input into the CBS-b1 layer. The CBS-b1 layer extracts the positioning feature from the input feature of the segmentation head, and the formula is as follows:

[0148] ;

[0149] wherein, is the output feature of the CBS-b1 layer;

[0150] The output feature of the CBS-b1 layer is input into the CBS-b2 layer. The CBS-b2 layer strengthens the regression representation of the output feature of the CBS-b1 layer, and the formula is as follows:

[0151] ;

[0152] wherein, is the output feature of the CBS-b2 layer;

[0153] The output features of the CBS-b2 layer are input into the Conv2d-b1 layer. The Conv2d-b1 layer generates bounding box regressions from the output features of the CBS-b2 layer, using the following formula:

[0154] ;

[0155] in, For bounding box regression values, , can be or , This represents the distance from the current grid point / anchor point to the left, top, right, and bottom boundaries of the target bounding box. The center coordinates of the target box Width w and height h;

[0156] The Conv2d-b1 layer outputs the bounding box regression values ​​to the bounding box loss module, which calculates the bounding box loss using the following formula:

[0157] ;

[0158] in, For bounding box loss, , All are scalar weighted hyperparameters. The Cross-Union Ratio (CIoU) loss function can be replaced with the GIoU / CIoU loss function. For the reason The predicted set obtained from decoding, For a set of truth boxes, For the corresponding true regression parameters, It is the L1 norm, which is the sum of the absolute values ​​of the four regression components.

[0159] Specifically, the processing procedure for the intermediate branches is as follows:

[0160] The segmentation head input features are sequentially passed through the DWConv-f1 layer, CBS-f1 layer, and DWConv-f2 layer for parameter compression, local response enhancement, and semantic information integration before being output to the CBS-f2 layer. The CBS-f2 layer performs classification adaptation and outputs to the Conv2d-f3 layer. The Conv2d-f3 layer outputs a classification / center heatmap to the classification loss module. The classification loss module aligns with the labeled classification / center heatmap and calculates the classification loss.

[0161] The specific processing involves inputting the segmentation head input features into the DWConv-f1 layer. The DWConv-f1 layer then performs parameter compression and local response enhancement operations on the segmentation head input features, as shown in the following formula:

[0162] ;

[0163] wherein, is the output feature of the DWConv-f1 layer;

[0164] The output feature of the DWConv-f1 layer is input into the CBS-f1 layer, and the CBS-f1 layer integrates semantic information for the output feature of the DWConv-f1 layer, and the formula is:

[0165] ;

[0166] wherein, is the output feature of the CBS-f1 layer;

[0167] The output feature of the CBS-f1 layer is input into the DWConv-f2 layer, and the DWConv-f2 layer performs pressure parameter and enhances local response operation for the output feature of the CBS-f1 layer, and the formula is:

[0168] ;

[0169] wherein, is the output feature of the DWConv-f2 layer;

[0170] The output feature of the DWConv-f2 layer is input into the CBS-f2 layer, and the CBS-f2 layer performs classification adaptation, and the formula is:

[0171] ;

[0172] wherein, is the final adaptive feature of the classification branch;

[0173] The final adaptive feature of the classification branch is input into the Conv2d-f3 layer, and the Conv2d-f3 layer performs 1x1 convolution on the classification feature to obtain a classification / center heat map, and the formula is:

[0174] ;

[0175] wherein, is the classification / center heat map, , when , represents a foreground / background binary classification center heat map (single-channel foreground probability map), K is the number of categories, and corresponds to K output channels, is a Sigmoid activation function, which normalizes the convolution output to ;

[0176] The classification / center heat map (predicted heat map) is aligned with the true value classification / center heat map (labeled true heat value map), and a classification loss is calculated at each pixel position, and the average is calculated in space, and the formula is:

[0177] ;

[0178] wherein, is the classification loss of the scale, x is the pixel / grid position index on the feature map, may be Focal or BCE, Focal is a Focal loss function, and BCE is a BCE loss function, is the true value classification / center heat map.

[0179] Specifically, the processing procedure of the right branch is as follows:

[0180] The segmentation head input feature is extracted by the CBS-x1 layer and the CBS-x2 layer to the Conv2d-x1 layer, the Conv2d-x1 layer generates a low channel feature output of the mask to the mask coefficient module, the mask coefficient module performs linear mapping to output to the mask module, the mask module obtains an instance mask output to the star polygon, the star polygon module extracts the boundary through the instance mask, projects the boundary as a star vertex sequence, splices the confidence, area and perimeter of the instance into an instance feature vector, and inputs a plurality of instance feature vectors into the high-pass screening module. The high-pass screening module screens out low-quality instances according to the threshold and morphological constraints, converts the screened instance feature vector into a GEOJSON format file and exports.

[0181] The specific processing procedure is that the segmentation head input feature is input into the CBS-x1 layer, the CBS-x1 layer extracts features from the segmentation head input feature, and the formula is as follows:

[0182] ;

[0183] wherein, is the output feature of the CBS-x1 layer;

[0184] The output feature of the CBS-x1 layer is input into the input end of the CBS-x2 layer, and the CBS-x2 layer performs enhanced representation, and the formula is as follows:

[0185] ;

[0186] The output feature of the CBS-x2 layer is input into the Conv2d-x1 layer, and the Conv2d-x1 layer performs 1×1 two-dimensional convolution, and the formula is as follows:

[0187] ;

[0188] wherein, is the mask base, , is the number of channels of the mask base (prototype mask);

[0189] The mask base is input to the mask coefficient module, and the branch or candidate point selection strategy generates a set of candidate positions on the feature map At each candidate position Local aggregation is performed on the features, and a mask coefficient vector of the instance is obtained through 1x1 convolution linear mapping, and the formula is:

[0190] ;

[0191] Among them, is the weight coefficient of instance j on the prototype channel (mask coefficient vector of instance j), is the candidate position The mask base and the segmentation head input feature are sampled or locally aggregated to obtain a vector describing the semantic of the candidate instance;

[0192] The mask coefficient is used to weight and sum the prototype features, and the soft mask of instance j is obtained through the Sigmoid activation function, and the formula is:

[0193] ;

[0194] Among them, is the mask of instance j at pixel position x, , is the value of the cth channel of the mask base at pixel position x, is the component of the mask coefficient vector of instance j on the cth channel, is the Sigmoid activation function;

[0195] The instance mask is input to the star polygon module, and the star polygon module extracts the boundary through the instance mask, and projects the boundary into a star vertex sequence, and the formula is:

[0196] ;

[0197] Among them, M is the total number of rays (the number of vertices of the star polygon), is the maximum radius of instance j in the mth ray direction from the instance center to the mask boundary, r is the radius search variable along the ray direction, is the instance center, obtained by the branch or candidate point, is the boundary threshold, is the star vertex sequence of instance j, is the ray angle;

[0198] ​For each instance j, calculate the confidence, area, perimeter, and splice into instance-level feature vector, formula is:

[0199] ;

[0200] Wherein, is the instance j feature vector, is the polygon formed by closed, is the confidence of instance j (for example, heat map peak / classification branch score), is the area of the polygon, is the perimeter of the polygon;

[0201] Multiple instance feature vectors are input into the high-pass screening module, and the high-pass screening module screens out low-quality instances according to threshold and morphological constraints:

[0202] ;

[0203] Wherein, is 0, the low-quality instance j is screened out, , is the confidence threshold, is the minimum value of the allowed area of the polygon, is the maximum value of the allowed area of the polygon, is the lower limit of the perimeter of the polygon;

[0204] For all instances j that pass the high-pass screening, write the star vertex coordinates and attributes to the GEOJSON file, formula is:

[0205] ;

[0206] ;

[0207] is the polygon vertex coordinate sequence of instance j, is the first vertex coordinate of the polygon, is the Mth vertex coordinate (closed at the beginning and end), is the score and geometric attribute of instance j, score is the score, is the confidence of instance j, area is the polygon area, perimeter is the polygon perimeter, and cale is the scale mark, is the step size.

[0208] The existing instance segmentation method has a fundamental limitation in the technical architecture when facing biological samples such as lymphoma cells with highly heterogeneous morphology. The mainstream method is usually based on two technical paths: one is based on rectangular bounding box detection after segmentation, which first locates and then refines; the other is directly performing pixel-level semantic segmentation and then obtaining instances through connected component analysis. These two paths may be effective in conventional cell segmentation, but when facing specific objects such as lymphoma cells with a large disparity in size, different shapes, and irregular boundaries, they expose many shortcomings.

[0209] Specifically, the existing method has three levels of defects in technical implementation. In the feature extraction layer, the traditional single branch network structure is difficult to optimize the three tasks of positioning, classification and segmentation with different feature requirements, leading to network learning goal conflict and model convergence difficulty; in the contour representation layer, whether based on fixed vertex number polygon regression or pixel-level prediction based on binary mask, it is difficult to achieve a good balance between representation accuracy and computational efficiency. Fixed vertex number polygon is difficult to adapt to different complexity of cell morphology, while pixel-level prediction produces boundary sawtooth and huge computational overhead; in the quality evaluation layer, the existing method only relies on classification confidence for instance screening, completely ignoring the key indicator of geometric integrity of instances, resulting in a large number of poor instances with incomplete boundaries and abnormal shapes being retained, which seriously affects the accuracy of subsequent analysis.

[0210] The star-shaped segmentation head of the application fundamentally reconstructs the technical path of instance segmentation through an innovative three-branch architecture. The left branch focuses on bounding box regression, gradually refines the positioning features, and ensures the accurate determination of cell position; the middle branch is responsible for foreground classification, using an alternating structure of depth separable convolution and conventional convolution to significantly reduce computational complexity while ensuring feature expression ability; the most critical right branch builds a complete contour representation and quality evaluation system: from feature extraction to mask base feature generation, to linear mapping of mask coefficient module, and finally to precise conversion from feature to geometry through star polygon module. The technical innovation of the star polygon module lies in its unique contour representation method. Based on the biological characteristics of cells, most cells have approximate radial symmetry of star shape. By emitting M rays from the center of the cell, the maximum boundary distance in each direction is calculated to form an adaptive star vertex sequence, which can accurately fit various irregular shapes and naturally maintain the continuity and smoothness of the boundary. By calculating the multi-dimensional features such as confidence, area, and perimeter of each instance, a comprehensive quality evaluation vector is formed. The high-pass filtering module sets intelligent thresholds based on these geometric features, systematically filtering out poor instances with abnormal area, small perimeter or discontinuous boundary. Compared with the traditional method of relying solely on classification confidence, the screening mechanism based on geometric integrity can more accurately identify and retain high-quality cell instances.

[0211] In the embodiment, the mask refinement is performed using the shrinking star convex polygon method, considering that there is usually a small amount of H&E dye penetration around the cells in the pathological section under H&E staining, and the profile of the star convex polygon is shrunk by 2-4 pixel values to the centroid, so that the cell instance segmentation result is more consistent with the reality.

[0212] Finally, the three key links of instance segmentation, detection, segmentation and screening are organically integrated in a unified framework, while maintaining the independence of each link, the three-branch architecture ensures the specialized processing of different tasks and avoids feature interference; the star polygon representation provides flexible and accurate boundary description; and the multi-dimensional quality evaluation ensures the reliability of the output result. This systematic design approach enables the application to maintain high segmentation accuracy while ensuring the biological reasonableness of the output instances when processing highly heterogeneous lymphoma cells, laying a solid technical foundation for subsequent pathological analysis.

[0213] It should be noted that the data of the LC-YOLO model is stored in a GEOJSON format file, which stores geometric and associated attribute information.

[0214] S6: Based on the LC-YOLO model trained under full supervision, the small block patches Patch without manual annotation are inferred to obtain preliminary pseudo-labels.

[0215] S7: The preliminary pseudo-labels are subjected to uncertainty screening, fixed threshold screening of non-maximum suppression, and adaptive threshold screening of the XGBoost model to obtain high-quality pseudo-labels.

[0216] The preliminary pseudo-labels are subjected to uncertainty screening, and the specific steps are as follows:

[0217] S71: The LC-YOLO model under full supervision is used to perform k times of repeated inference on the small block patches Patch without manual annotation, and the segmentation mask set generated each time is collected and the confidence score ;

[0218] S72: The standard deviation of each confidence score and the average IoU between multiple mask predictions are calculated, and the uncertainty index is defined as:

[0219] ;

[0220] wherein, is the uncertainty index, is the standard deviation of the confidence score, is the average value of the mask consistency;

[0221] S73: Set the threshold , and the uncertainty index in the preliminary label is reserved < The low-noise instance of the preliminary label is removed, and the preliminary label with large confidence fluctuation or inconsistent mask is removed.

[0222] In the embodiment, the threshold value is set to 0.1.

[0223] The fixed threshold screening of non-maximum suppression, the specific steps are as follows:

[0224] S74: Divide the mask set Synthesize the soft mask through K times of reasoning, and obtain the representative mask by binarizing the soft mask, and the formula is:

[0225] ;

[0226] ;

[0227] Wherein, the soft mask is a pixel value of 0 to 1, which is the average probability of the i-th instance being foreground at pixel x, and x is the pixel, is the mask index set matched by the same instance i in K times of repeated reasoning, is the element number of the set is the segmentation mask set of the instance i, is the representative mask, is the soft mask binarization threshold value, 1 is an indicator function, which takes a value of 1 when the condition in the parentheses is met, otherwise it takes a value of 0; In the embodiment, the value of is 0.5.

[0228] In the embodiment, the value of is 0.5.

[0229] S75: Confidence score Synthesize the ranking score of the candidate instance through K times of reasoning, and the formula is:

[0230] ;

[0231] Wherein, k is the index of the k-th repeated reasoning (k=1,…, K), and K is the total number of repeated reasoning for the same tile, is the average confidence score of the candidate instance i obtained in times of reasoning, is the final ranking score of the candidate instance i, is the uncertainty index of the candidate instance i;

[0232] S76: Obtain the candidate set through the ranking score of the candidate instance and the soft mask, and sort the candidate set according to ​​​Sort from large to small, put the candidate instance with the highest final ranking score into the result set, and calculate the mask for other final ranking scores, the formula is:

[0233] ;

[0234] wherein, is the mask of other candidate instance q, is the representative mask of other instance q, is the candidate instance p with the highest final ranking score;

[0235] If , the other candidate instance q is suppressed to ensure the sparsity and diversity of the candidate set.

[0236] The adaptive threshold screening of the XGBoost model includes the following steps:

[0237] S77: input the preliminary label screened by the non-maximum suppression and the fixed threshold into the trained XGBoost model to obtain a confidence score;

[0238] S78: according to the confidence score of the XGBoost model after training, manually set an adaptive threshold, and retain instances as high-quality pseudo labels;

[0239] wherein is the confidence score, is the adaptive threshold.

[0240] In the embodiment, the confidence score is higher than 0.95, and the adaptive threshold is set to 0.9 to ensure the high quality of the pseudo labels.

[0241] Specifically, as shown in Figure 4 , the adaptive threshold screening process of the XGBoost model includes the following steps:

[0242] 1) initialization and the 0th round of training: copy the training set as training set 0, train a CART tree with the current learning objective (usually first-order / second-order gradient, and can also be regarded as residual error when regression), and obtain model 0, which is the first tree in the boosting process;

[0243] 2) generate model 1 bias degree (residual error / gradient) from model 0: use model 0 to make predictions on the training set to obtain the current fitting degree, and calculate the target to be learned in the next round according to the fitting degree, which is written as “model 1 bias degree” in the figure (the essence is residual error or negative gradient, if using second-order approximation, it will also be combined with Hessian), take the target as a new label, and form training set 1;

[0244] 3) The first round of training: retrain a CART tree with training set 1 to get model 1, which is used to correct the part that has not been fitted in the previous round, i.e. the bias degree of model 1;

[0245] 4) Iteration to the second round and beyond: repeat the above steps, predict and calculate the bias degree of model 2 from model 1 on the training set, get training set 2, train the CART tree with training set 2 to get model 2, and so on, until the bias degree of model n-1 is predicted and calculated from model n-1 on the training set, get training set n, train the CART tree with training set n to get model n, each round continues to learn on the residual gradient of the previous round, making the overall fitting gradually better (learning rate will reduce the contribution of each tree).

[0246] 5) Ensemble summation (on the test set): calculate the output of model 0, model 1, model 2, …, model n on the test set one by one, and make a sum with the learning rate as the weight, accumulate the contribution of all CART trees into the final prediction function, the formula is:

[0247] ;

[0248] Where, is the prediction function, is the number of models, , is the output of the th model, is the learning rate;

[0249] 6) Get the final result: compare the prediction function after summation with the test set label, output the classification probability / score or regression value, and calculate the evaluation index (such as AUC, Accuracy, MSE, etc.).

[0250] It is noted that the XGBoost model generates a GEOJSON pseudo-label file, all instances that pass the XGBoost model screening are marked pseudo=1, and the original and screened instance quantity logs of each image are recorded, and the LC-YOLO model can be semi-supervised trained to obtain a lymphoma cell instance segmentation model, and other models can also be trained.

[0251] As shown in Figure 5 , the technical principle of the conventional pseudo-label simplification is:

[0252] 1. Train the LC-YOLO model with labeled data set data for full supervision;

[0253] 2. The LC-YOLO model trained with full supervision obtains pseudo-labels by reasoning other data;

[0254] 3. Set a fixed threshold through the labeled dataset data, and filter out the pseudo labels less than the threshold;

[0255] 4. After screening, the pseudo labels are used for semi-supervised training of the LC-YOLO model again to obtain the final LC-YOLO model.

[0256] As shown in Figure 6 The technical principle of the pseudo label simplification of the application is:

[0257] 1. The XGBoost model is fused and trained through the labeled dataset data;

[0258] 2. The LC-YOLO model is fully supervised trained through the labeled dataset data;

[0259] 3. The LC-YOLO model of the full supervision training obtains pseudo labels by reasoning other data;

[0260] 4. The XGBoost model is trained to obtain the confidence score of the labeled dataset, an adaptive threshold is set through the confidence score, pseudo labels less than the adaptive threshold are filtered out, and high-quality pseudo labels are obtained;

[0261] 5. The LC-YOLO model is semi-supervised trained again through the high-quality pseudo labels to obtain the final LC-YOLO model.

[0262] It should be noted that after obtaining high-quality pseudo labels, not only the LC-YOLO model can be semi-supervised trained, but also other models can be semi-supervised trained. The fixed threshold screening of the non-maximum suppression is not described in detail in the simplified technical principle, and the fusion training of the XGBoost model is also not described in detail. The main purpose is to compare the difference between the traditional pseudo label technical principle and the pseudo label technical principle of the application.

[0263] In summary, the core advantage of the application compared with the traditional method is:

[0264] The pseudo label screening is upgraded from a simple threshold filtering operation to an intelligent discrimination process based on a machine learning model, which significantly improves the accuracy and robustness of the screening;

[0265] Through multi-dimensional feature fusion, the judgment basis is expanded from single classification confidence to a comprehensive evaluation system containing appearance features and geometric morphological features, so that the screening decision is more comprehensive and reasonable;

[0266] Since the XGBoost model can output higher-quality pseudo labels, this enables the subsequent semi-supervised training of the LC-YOLO model to be performed on a cleaner dataset, effectively curbing the accumulation of errors, guiding the positive iteration and continuous improvement of the model performance, and forming a benign learning closed loop.

[0267] Therefore, the present application fundamentally solves the pain points of single dimension and poor adaptability of the traditional method by introducing the XGBoost model to realize adaptive threshold screening, and lays a solid foundation for realizing high-performance and high-reliability semi-supervised lymphoma cell segmentation.

[0268] Embodiment 2: A semi-supervised lymphoma cell segmentation system with pseudo-label screening.

[0269] As shown in Figure 7 A semi-supervised lymphoma cell segmentation system with pseudo-label screening includes an image segmentation module, a feature extraction module, an instance annotation module, a feature fusion module, an LC-YOLO model training module, an LC-YOLO model inference module, and a preliminary pseudo-label screening module.

[0270] The image segmentation module divides the whole slice image into multiple small block images with different resolutions.

[0271] The feature extraction module inputs the small block images with different resolutions into the EfficientNets model for multi-scale feature extraction, and extracts multi-scale image features.

[0272] The instance annotation module manually annotates the cell instances of the small block images with a single scale to obtain instance structure features.

[0273] The feature fusion module fuses the instance structure features and the multi-scale image features to train the XGBoost model.

[0274] The LC-YOLO model training module inputs the manually annotated small block images into the LC-YOLO model for full supervision training.

[0275] The LC-YOLO model inference module infers the small block images that are not manually annotated based on the LC-YOLO model trained by full supervision to obtain preliminary pseudo-labels.

[0276] The preliminary pseudo-label screening module obtains high-quality pseudo-labels after the preliminary pseudo-labels are screened by uncertainty screening, fixed threshold screening of non-maximum suppression, and adaptive threshold screening of the XGBoost model.

[0277] Embodiment 3: Experimental verification of the semi-supervised lymphoma cell segmentation method with pseudo-label screening.

[0278] 1.1 Experimental preparation:

[0279] Experiments are based on the Ubuntu distribution of the Linux system, the kernel version is 5.15.0-93-generic, running on 64-bit x86 architecture hardware, and the graphics card is NVIDIA GeForce RTX 4090;

[0280] The data of the experiment all come from the pathology department, covering 30 lymphoma subtypes, including classic Hodgkin's lymphoma, follicular lymphoma, angioimmunoblastic lymph node follicle helper T cell lymphoma, diffuse large B cell lymphoma NOS, etc. Subtypes, a total of 89 whole slide images WSI (training set: 22; validation set 21; test set: 30; pseudo label: 16) and 45060 small block patches (training set: 8054; validation set 2006; test set: 3000; pseudo label: 32000) from 30 lymphoma subtypes after segmentation.

[0281] The pathology toolbox tiatoolbox and openslide library in Python read the whole slide image WSI and segment it into small block patches of specified pixel size. All cell instance segmentation annotation datasets, except for pseudo labels, are based on manual annotation in the case image open source tool Qupath software.

[0282] 1.2 Model evaluation index:

[0283] In order to comprehensively evaluate the performance of the model, multiple evaluation indexes are adopted, including Precision, Recall, mAP50, mAP50-95, Bbox and Mask PR and F1 curves, which reflect the performance of the model in cell instance segmentation from different angles.

[0284] Among them, Precision is the precision, Recall is the recall, mAP50 is the average precision when IoU is 0.5, mAP50-95 is the average precision when IoU is between 0.5 and 0.95, Bbox is the bounding box, and Mask is the mask.

[0285] 1.3 Model comparison experiment:

[0286] In order to comprehensively evaluate the performance in the task of cell instance segmentation, multiple mainstream models are compared, including YOLO series YOLOv8, YOLOv9, YOLO11, YOLOv12, as well as U-Net, Mask R-CNN, Cellpose and StarDist. The specific hyperparameter settings are shown in Table 1-2.

[0287] Table 1: YOLO cell instance segmentation model hyperparameter table

[0288]

[0289] Table 2: Other cell instance segmentation model hyperparameter table

[0290]

[0291] Where Batchsize is the batch size of training, Epochs is the number of training rounds, Learning_rate is the learning rate; Warmup_epochs is the preheating round, Optimizer is the optimizer;

[0292] Where Patience is the patience value of early stopping, RPN_ANCHOR_SCALES is the RPN anchor box scale, TRAIN_ROIS_PER_IMAGE is the number of ROI per image for training, BACKBONE is the backbone network, DETECTION_MIN_CONFIDENCE is the detection minimum confidence threshold, model_type is the model type, pretrained_model is the pre-trained model, min_train_masks is the minimum number of training masks, chan is the input channel 1, chan2 is the input channel 2, weight_decay is the weight decay, train_learning_rate is the learning rate (training), n_rays is the number of rays, grid is the grid size, and train_steps_per_epoch is the number of training steps per epoch.

[0293] According to the hyperparameter settings, the corresponding model is the optimal result in the single model of repeated experiment optimization, as shown in Table 3.

[0294] Table 3: Performance index table of lymphoma cell instance segmentation in model

[0295]

[0296] In terms of bounding box detection, YOLO11 has the best overall performance: Bbox_Precision reaches 0.8526, Bbox_Recall is 0.8138, Bbox_mAP50 and Bbox_mAP50-95 are 0.8734 and 0.5990, respectively. The detection accuracy and recall rate of YOLOv8 and YOLOv9 are slightly lower, and the results of YOLOv12 and YOLO11 are very close, but their Bbox_mAP50 and Bbox_mAP50-95 are still slightly inferior, which are 0.21% and 0.51% lower, respectively. U-Net, Mask R-CNN, Cellpose and StarDist still have a significant gap compared with the YOLO series in this task.

[0297] In the mask segmentation task, YOLO11 still performs outstandingly, with Mask_Precision and Mask_Recall reaching 0.7311 and 0.6860, Mask_mAP50 being 0.6960, and Mask_mAP50-95 reaching 0.2445, leading other models overall.

[0298] 1.4 Semi-supervised effect verification:

[0299] To verify the effectiveness of semi-supervised learning in the cell instance segmentation task, three different datasets were selected for comparison experiments: A: only using 8000 manually labeled data; B: only using 8000 pseudo-label data; C: 8000 manually labeled data + 32000 pseudo-label data, as shown in Table 4.

[0300] Table 4: Performance indicators of lymphoma cell instance segmentation in semi-supervised learning verification experiments

[0301]

[0302] The results obtained through the comparison experiments clearly show the characteristics of semi-supervised learning. After introducing pseudo-label data, the relevant indicators of bounding box detection have decreased to some extent, for example, Bbox_Precision from 0.8526 to 0.8236, Bbox_Recall from 0.8138 to 0.7928, and Bbox_mAP50 from 0.5990 to 0.5357. In the cell instance segmentation task, mask-related indicators are more reflective of model performance than bounding box-related indicators, so the slight decline in Bbox indicators has limited impact on this task. In the mask segmentation task, Mask_Precision increased from 0.7311 to 0.7428, Mask_Recall increased from 0.6860 to 0.7045, Mask_mAP50 increased from 0.6960 to 0.7296, and Mask_mAP50-95 increased from 0.2445 to 0.2762. These data strongly prove that the model's mask segmentation accuracy has been improved in all aspects within different IoU threshold intervals, and the model still achieves excellent detection and segmentation capabilities in experiments using only 8000 pseudo-label data for training.

[0303] As Figure 8 shown, from top to bottom: full supervision effect, pseudo-label effect, semi-supervised effect, and manual annotation effect. The LC-YOLO model, under the innovative point of multi-scale feature fusion XGBoost model screening pseudo-labels, has shown significant performance advantages in the lymphoma cell instance segmentation task, achieving the highest segmentation accuracy.

[0304] 1.5 Pseudo-label screening ablation experiment:

[0305] To evaluate the impact of the pseudo-label screening semi-supervised method based on the XGBoos model on the model segmentation performance, an ablation experiment was designed. Both groups of experiments were based on the same 8,000 manually annotated images, with an additional 32,000 pseudo-labels introduced. However, in Group C, the pseudo-labels were screened using an adaptive threshold quality screening based on multi-scale instance and image feature fusion using the XGBoost model, as shown in Table 5.

[0306] Table 5: Performance Indicators of Lymphoma Cell Segmentation in Pseudolabeled Screening Ablation Experiments

[0307]

[0308] Although the unfiltered pseudo-label data combination achieved higher Bbox_Precision and Bbox_Recall in the bounding box detection task, with corresponding Bbox_mAP50 and Bbox_mAP50-95 scores of 0.8717 and 0.5983 respectively, both higher than the corresponding data of the filtered group, the XGBoost model's filtered pseudo-label combination achieved comprehensive superiority in mask segmentation metrics in Mask_Precision (0.7428 vs. 0.7348), Mask_Recall (0.7045 vs. 0.6852), Mask_mAP50 (0.7296 vs. 0.6993), and Mask_mAP50-95 (0.2762 vs. 0.2421). This demonstrates that the filtering strategy significantly promotes the accurate reconstruction of cell outlines and the characterization of morphological details.

[0309] like Figure 9 As shown, from top to bottom, the results are: fully supervised effect, pseudo-label effect, semi-supervised effect, and manual annotation effect. It can be clearly seen that the selection of pseudo-labels by the multi-scale feature fusion XGBoost model is important for the quality of pseudo-labels and the performance of the model.

[0310] 1.6 Robustness verification using semi-supervised methods:

[0311] To further explore the robustness of the pseudo-label screening semi-supervised learning strategy in the cell instance segmentation task and verify its potential to reduce manual annotation data while improving segmentation accuracy. In the experiment, multiple groups of different proportions of manual annotation data combined with pseudo-label data were set: the amount of manual annotation data was gradually reduced from 8000 to 4000 (reduced by 50%), 2000 (reduced by 75%), 800 (reduced by 90%), and 400 (reduced by 95%); the amount of pseudo-label data was increased from 32000 to 36000, 38000, 39200, and 39600, respectively. This design aims to simulate the performance changes of semi-supervised methods as the difficulty or cost of annotation increases in actual applications. C: 8000 manually annotated samples + 32000 pseudo-label data; E: 4000 manually annotated samples + 36000 pseudo-label data; F: 2000 manually annotated samples + 38000 pseudo-label data; G: 800 manually annotated samples + 39,200 pseudo-label data; H: 400 manually annotated samples + 39,600 pseudo-label data, as shown in Table 6.

[0312] Table 6: Performance indicators of lymphoma cell instance segmentation in semi-supervised protocol robustness verification

[0313]

[0314] In the bounding box detection task, when the amount of manual annotation data is reduced from 8000 to 4000, Bbox_Precision increases from 0.8236 to 0.8399, Bbox_Recall decreases from 0.7928 to 0.7907, Bbox_mAP50 increases from 0.8401 to 0.8422, and Bbox_mAP50-95 remains stable at around 0.5353. This indicates that when the amount of manual annotation data is halved, the performance of the model in the bounding box detection task fluctuates little, and the detection accuracy and recall rate remain relatively stable. However, when the amount of manual annotation data is further reduced, the model performance decreases significantly. For example, when the amount of manual annotation data is reduced to 2000, Bbox_Precision decreases to 0.7525, Bbox_Recall decreases to 0.5321, Bbox_mAP50 decreases to 0.6240, and Bbox_mAP50-95 decreases to 0.1683. This indicates that when the amount of manual annotation data is too small, the robustness of the model in the bounding box detection task is challenged, and the positioning and detection accuracy of the target are greatly affected. However, it is worth noting that when the amount of manual annotation data is 800 and 400, Bbox_Precision is 0.6330 and 0.6924, respectively, and Bbox_Recall is 0.6059 and 0.6434, respectively. Although there is some fluctuation, it still maintains a certain level, indicating that the model still has good adaptability in the extreme case of data scarcity.

[0315] In terms of mask segmentation tasks, the model shows stronger robustness. When the amount of artificial labeled data decreases from 8000 to 4000, Mask_Precision decreases slightly from 0.7428 to 0.7305, Mask_Recall decreases slightly from 0.7045 to 0.6971, Mask_mAP50 decreases from 0.7296 to 0.7040, Mask_mAP50-95 decreases from 0.2762 to 0.2538, and the changes of various indicators are within a small range, indicating that the model performance is basically stable. When the amount of artificial labeled data decreases to 2000, Mask_Precision increases significantly to 0.8391, although Mask_Recall decreases to 0.5933, Mask_mAP50 remains at 0.6980, and Mask_mAP50-95 increases to 0.2893. When the amount of artificial labeled data is 800 and 400, Mask_Precision is 0.7429 and 0.7588 respectively, Mask_Recall is 0.7357 and 0.7051 respectively, Mask_mAP50 is 0.7233 and 0.7021 respectively, and Mask_mAP50-95 is 0.2991 and 0.2707 respectively, showing a trend of stable increase, which fully shows that in the mask segmentation task, even if the amount of artificial labeled data is greatly reduced, the model can still effectively use the pseudo-label data to ensure high segmentation precision and even improve the segmentation precision, showing strong robustness.

[0316] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A semi-supervised lymphoma cell segmentation method using pseudo-label screening, characterized in that, Includes the following steps: The full slice image is divided into multiple smaller image patches of different resolutions; Multi-resolution small image patches are input into the EfficientNets model for multi-scale feature extraction, thereby extracting multi-scale image features. Manually annotate cell instances in small, single-scale image patches to obtain instance structural features; Instance structural features and multi-scale image features are fused together to train the XGBoost model; Manually labeled small image patches are input into the LC-YOLO model for fully supervised training. The LC-YOLO model, trained under full supervision, performs inference on small, unlabeled image patches to obtain preliminary pseudo-labels. High-quality pseudo-labels are obtained by filtering the initial pseudo-labels through uncertainty screening, fixed threshold screening for non-maximum suppression, and adaptive threshold screening for the XGBoost model. The adaptive threshold selection for the XGBoost model involves the following steps: The initial labels, filtered by a fixed threshold after non-maximum suppression, are input into the trained XGBoost model to obtain confidence scores. Based on the confidence scores after XGBoost model training, an adaptive threshold is manually set to retain... Instances of these are set as high-quality pseudo-tags; in It is a confidence score. It is an adaptive threshold.

2. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 1, characterized in that, The multiple resolutions include 256×256, 128×128, and 64×64 pixels.

3. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 2, characterized in that, The specific steps for manually labeling cell instances in small-scale image patches of a single scale are as follows: Read the format annotations in the same small patch to obtain the bounding box and polygon outline; Calculate area A, perimeter P, and shape factor using bounding boxes and polygon outlines in the OpenCV library. Forming instance structural features , ; Where 3 represents the dimension of instance structure features, and the dimensions of instance structure features include A, P, ... .

4. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 1, characterized in that, The specific steps for fusing instance structural features and multi-scale image features to train the XGBoost model are as follows: Multiscale image features With instance-level structural features Perform feature concatenation to obtain the fused feature vector. ; The EfficientNets model takes all the fused feature vectors and generates a training feature matrix X containing both positive and negative samples. and the label vector corresponding to the training feature matrix ; Where N is the total number of fused feature vectors, D is the multi-scale image feature dimension, and 3 is the instance structure feature dimension; The training feature matrix and the corresponding label vector are used to train the XGBoost model to obtain a confidence score, enabling the XGBoost model to distinguish between positive and negative samples. Positive samples are segmented regions of cell instances, while negative samples are created non-cell regions.

5. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 1, characterized in that, The LC-YOLO model includes a backbone network, a neck network, and a star-shaped segmentation head; The backbone network extracts multi-scale backbone features from the input lymphoma cell image; The neck network fuses multi-scale features using top-down and bottom-up feature pyramid structures to obtain fused features; The star-shaped segmentation head performs bounding box regression, foreground classification, and star-shaped contour decoding in parallel at three scales, and outputs a GEOJSON format file including polygons.

6. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 5, characterized in that, The star-shaped segmentation head divides the segmentation head input features into three branches: the left branch, the middle branch, and the right branch. The left branch is used for bounding box regression of features, which includes CBS-b1 layer, CBS-b2 layer, Conv2d-b1 layer and bounding box loss module connected in sequence; The middle branch is used for foreground classification of computational features, including the DWConv-f1 layer, CBS-f1 layer, DWConv-f2 layer, CBS-f2 layer, Conv2d-f3 layer, and classification loss module, which are connected in sequence. The right branch is used to decode the star-shaped contour of the feature, which includes the CBS-x1 layer, CBS-x2 layer, Conv2d-x1 layer, mask coefficient module, mask module, star polygon module, and high-pass filter module connected in sequence. The processing procedure for the right branch is as follows: The segmentation head input features are extracted through CBS-x1 and CBS-x2 layers and then fed into Conv2d-x1 layer; The low-channel features of the mask generated by the Conv2d-x1 layer are output to the mask coefficient module; The mask coefficient module performs a linear mapping and outputs the result to the mask module. The mask module then obtains the instance mask and outputs it to the star polygon module. The star polygon module extracts the boundary through instance mask, projects the boundary into a star vertex sequence, and concatenates the confidence, area, and perimeter of the instance into an instance feature vector; Multiple instance feature vectors are input into the high-pass filtering module. The high-pass filtering module filters out low-quality instances based on thresholds and morphological constraints, and then converts the filtered instance feature vectors into GEOJSON format files for export.

7. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 1, characterized in that, The initial pseudo-labels undergo uncertainty screening, specifically through the following steps: The LC-YOLO model, trained under full supervision, is used to perform k repeated inferences on unannotated patch images, and the generated segmentation mask set is collected each time. With confidence score ; Calculate the standard deviation of each confidence score and the average IoU between multiple mask predictions, and define the uncertainty index: ； in, It is an indicator of uncertainty. It is the standard deviation of the confidence score. It is the average value of mask consistency; Set threshold Uncertainty screening is performed, retaining the uncertainty indicators from the initial labels. < Low-noise instances are used to eliminate initial labels with large confidence fluctuations or inconsistent masks.

8. The semi-supervised lymphoma cell segmentation method using pseudo-label screening according to claim 1, characterized in that, The fixed threshold screening for non-maximum suppression involves the following steps: Segmentation mask set A soft mask is synthesized through K inferences. The representative mask is then obtained by binarizing the soft mask, as shown in the formula: ； ； in, This is a soft mask, with pixel values ​​ranging from 0 to 1. It represents the average probability that the i-th instance is in the foreground at pixel x, where x is a pixel. Let i be the set of mask indices that are matched in K repeated inferences for the same instance i. For set The number of elements, Let i be the set of segmentation masks. As a representative mask, is the soft mask binarization threshold, and 1 is an indicator function that takes the value of 1 when the condition in parentheses is met, and takes the value of 0 otherwise. Confidence score The ranking score of candidate instances is synthesized through K inferences, and the formula is: ； Where k is the index of the k-th repeated reasoning, k=1,…,K, and K is the total number of repeated reasoning operations performed on the same tile. For candidate instance i in The average confidence score obtained from the inference process. The final ranking score for candidate instance i. The uncertainty index for candidate instance i; The candidate set is obtained by using the ranking score and soft mask of the candidate instances. ,according to Sort the data from largest to smallest, add the candidate instance with the highest final sort score to the result set, and calculate the mask for the other final sort scores using the following formula: ； in, For the mask of other candidate instances q, This serves as a representative mask for other instances q. The candidate instance p with the highest final ranking score; like This suppresses other candidate instances q, ensuring the sparsity and diversity of the candidate set.

9. A semi-supervised lymphoma cell segmentation system for pseudo-label screening, characterized in that, It includes an image segmentation module, a feature extraction module, an instance annotation module, a feature fusion module, an LC-YOLO model training module, an LC-YOLO model inference module, and a preliminary pseudo-label selection module; The image segmentation module segments the full slice image into multiple small image blocks of different resolutions. The feature extraction module inputs multi-resolution small image patches into the EfficientNets model for multi-scale feature extraction, thereby extracting multi-scale image features. The instance annotation module manually annotates cell instances in small-scale image patches at a single scale to obtain instance structural features; The feature fusion module fuses instance structural features and multi-scale image features for training the XGBoost model; The LC-YOLO model training module inputs manually labeled small image patches into the LC-YOLO model for fully supervised training. The LC-YOLO model inference module infers small, unlabeled image patches based on a fully supervised LC-YOLO model to obtain preliminary pseudo-labels; The preliminary pseudo-label screening module obtains high-quality pseudo-labels after passing the preliminary pseudo-labels through uncertainty screening, fixed threshold screening of non-maximum suppression, and adaptive threshold screening of the XGBoost model. The adaptive threshold selection for the XGBoost model involves the following steps: The initial labels, filtered by a fixed threshold after non-maximum suppression, are input into the trained XGBoost model to obtain confidence scores. Based on the confidence scores after XGBoost model training, an adaptive threshold is manually set to retain... Instances of these are set as high-quality pseudo-tags; in It is a confidence score. It is an adaptive threshold.

Images