A liquid-based thin-layer cell detection image recognition method based on cell distribution normalization
By generating uniform cell images through cell instance segmentation and distribution normalization, and combining them with the DINO-YOLO model, the problem of limited generalization performance of deep learning models caused by the non-uniform cell distribution in TCT images is solved, thereby improving the stability and robustness of the model and making it suitable for abnormal cell detection in cervical cancer screening.
Patent Information
- Application Number
- CN202611437926.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies ignore the non-uniformity of the spatial distribution of cell populations in TCT images, which limits the generalization performance of deep learning models in cervical cancer screening. Traditional data augmentation strategies cannot alleviate this problem at the cell instance level.
By segmenting and normalizing cell instances, a standardized liquid-based thin-layer cell image with a uniform spatial distribution of cell populations is generated, and the DINO-YOLO recognition model is used to detect abnormal cells.
It achieves consistent distribution characteristics between training and testing data, improves the learning stability and robustness of the model, solves the inherent defects of datasets that traditional data augmentation methods cannot address, and is suitable for various medical cell image analysis scenarios.
Smart Images

Figure CN123072020A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing and deep learning technology, specifically involving an image recognition method for liquid-based thin-layer cell detection (ThinPrepCytologic Test, TCT) based on cell-distribution normalization (Cell-Norm), which is suitable for the automatic detection and classification of abnormal cells in cervical cancer screening. Background Technology
[0002] Cervical cancer is the fourth leading cause of cancer death among women worldwide, posing a serious threat to women's lives and health. Liquid-based thin-layer cytology is a primary method for cervical cancer screening, enabling early detection and prevention through microscopic examination of abnormal cell morphology. However, manual cytology examination is labor-intensive, time-consuming, and prone to significant inter-observer variability—different experts may interpret the same slide differently, with reported inter-observer agreement (Kappa) values ranging from 0.16 to 0.70. False negative rates also vary considerably across different clinical settings, posing a risk of delayed diagnosis and treatment.
[0003] In recent years, deep learning technology has made significant progress in the field of medical image analysis. In cervical cytology image analysis, existing work can be divided into two main directions: First, at the model architecture level, researchers are continuously advancing detection paradigms, from early cell segmentation methods based on encoder-decoder architectures (such as U-Net and its attention-enhanced variants), to two-stage detectors (such as the Faster R-CNN family combined with the Feature Pyramid Network (FPN), and more recently, Transformer architecture-based models (such as UniCAS). These works mainly focus on designing dedicated network modules and comparative learning mechanisms to capture multi-scale cellular features and alleviate feature mismatch problems caused by cell overlap, clustering, and lesion category ambiguity. Second, at the data optimization level, existing research is dedicated to alleviating the inherent severe long-tail class imbalance problem in clinical TCT samples. Related solutions include cost-sensitive loss function design (such as Focal Loss), feature space resampling (such as SMOTE), and sample synthesis based on generative models (such as GANs), aiming to address the scarcity of high-level abnormal cell samples and the model's prediction bias towards normal cell categories.
[0004] However, existing research generally overlooks a fundamental problem—the non-uniformity of the spatial distribution of cell populations in TCT images. For example... Figure 1As shown, significant differences exist in cell density, morphology, scale, and aggregation among different TCT images. Unlike natural images, medical cell images contain only cell entities, and adjusting the spatial layout of cells does not alter their inherent semantic information. However, current state-of-the-art methods neglect the inherent imbalance in cell population distribution within TCT samples. This distributional disparity persists even after traditional pixel-level and image-level data augmentation, violating the core assumption of independent and identically distributed (i.i.d.) data between training and testing data. This fundamentally limits the generalization performance of deep learning models in TCT image analysis. Traditional data augmentation strategies operate only at the image or pixel level and cannot alleviate the cell distribution imbalance problem at the cell instance level. This explains why such methods offer very limited performance improvements in cell analysis tasks. Summary of the Invention
[0005] To address the fundamental problem of neglecting the uneven spatial distribution of cell populations within TCT images in existing technologies, and the resulting limitation in the generalization performance of deep learning models, this invention proposes a liquid-based thin-layer cell detection image recognition method based on cell distribution normalization, comprising the following steps:
[0006] Acquire raw liquid-based thin-layer cell detection images and extract raw region of interest images containing abnormal cells from them;
[0007] A cell instance segmentation model is used to segment the original region of interest image into cell instances, and the mask of each cell instance is obtained.
[0008] The segmented cell instances are extracted and classified, distinguishing between abnormal cells and normal cells;
[0009] Perform cell distribution normalization processing, randomly resample and arrange the extracted cell instances in a non-overlapping layout on a blank canvas to generate a standardized liquid-based thin-layer cell image with a uniform spatial distribution of the cell population.
[0010] Standardized liquid-based thin-layer cell images are input into a pre-trained DINO-YOLO recognition model to obtain the detection results of abnormal cell types and locations.
[0011] Compared with the prior art, the present invention has the following beneficial effects:
[0012] 1. This invention is the first to identify the long-neglected core issue of uneven spatial distribution of cell populations in TCT images, revealing that this distribution inconsistency violates the fundamental assumption of independent and identically distributed (i.i.d.) relationships between training and test data in machine learning, and is the root cause of the limited generalization performance of deep learning models. Based on this original discovery, the proposed Cell-Norm method optimizes the spatial distribution at the cell instance level, achieving consistent distribution characteristics between training and test data, and overcoming the inherent limitation of traditional pixel-level and image-level data augmentation, which only perform surface transformations and cannot optimize the spatial layout of cells.
[0013] 2. The Cell-Norm cell distribution normalization method proposed in this invention has two unique advantages: First, it standardizes the cell population distribution in all TCT images, ensuring it strictly satisfies the core iid assumption for effective machine learning training, thereby guaranteeing the stability and robustness of model learning; second, it achieves cell-level dataset expansion by repeatedly running the random resampling pipeline with independent random seeds, generating sufficient diverse standardized training samples with limited additional computational overhead. The synergistic effect of these two advantages enables the method to solve the inherent dataset defects that traditional pixel-level and image-level enhancement pipelines cannot address. This method can serve as a general-purpose standardized preprocessing pipeline, applicable to various medical cell image analysis scenarios, providing a universal and efficient solution to the generalization bottleneck of medical vision models. Attached Figure Description
[0014] Figure 1 This is a visualization of the severely uneven spatial distribution of cell populations in a TCT image.
[0015] Figure 2 This is a schematic diagram illustrating an example of obtaining standardized liquid-based thin-layer cell images in a liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to the present invention.
[0016] Figure 3 This is a schematic diagram of the framework of the DINO-YOLO recognition model in the liquid-based thin-layer cell detection image recognition method based on cell distribution normalization of the present invention;
[0017] Figure 4 This is a schematic diagram of the cell instance resampling strategy of the present invention, wherein (a) cell instances are randomly placed; (b) cell instances are randomly rotated; and (c) a limited range of random scaling is performed. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] This invention proposes a liquid-based thin-layer cell detection image recognition method based on cell distribution normalization, comprising the following steps:
[0020] Acquire raw liquid-based thin-layer cell detection images and extract raw region of interest images containing abnormal cells from them;
[0021] A cell instance segmentation model is used to segment the original region of interest image into cell instances, and the mask of each cell instance is obtained.
[0022] The segmented cell instances are extracted and classified, distinguishing between abnormal cells and normal cells;
[0023] Perform cell distribution normalization processing, randomly resample and arrange the extracted cell instances in a non-overlapping layout on a blank canvas to generate a standardized liquid-based thin-layer cell image with a uniform spatial distribution of the cell population.
[0024] Standardized liquid-based thin-layer cell images are input into a pre-trained DINO-YOLO recognition model to obtain the detection results of abnormal cell types and locations.
[0025] This invention consists of two parts. The first part processes the original liquid-based thin-layer cell detection image to generate a standardized liquid-based thin-layer cell image with a uniform spatial distribution of cell populations. The second part performs recognition based on the standardized liquid-based thin-layer cell image. This embodiment selects the DINO-YOLO recognition model. Those skilled in the art can also choose other visual models for training and inference based on the standardized liquid-based thin-layer cell image.
[0026] Example 1: Cell distribution normalization treatment.
[0027] This embodiment details the cell distribution normalization process, which includes the following steps:
[0028] Step 1: Region of Interest (ROI) extraction to obtain the original ROI image. Specifically, from the original TCT image, multi-scale ROI local patches are cropped uniformly outward from each labeled aberrant cell. These local image patches are used as the original ROI image. In this embodiment, three local patch sizes are used: 640×640, 1024×1024, and 1536×1536. To ensure the integrity of the annotations, the bounding box coordinates of the aberrant cells are linearly transformed to the new ROI coordinate system.
[0029] Step 2: Cell Instance Segmentation. This embodiment uses SAM (Segment Anything Model) released by Meta AI to perform cell instance segmentation on the cropped ROI local patches. Specifically, local patches of 640×640, 1024×1024, and 1536×1536 sizes are adjusted to a uniform size. For example, the resize function in Python can be used to adjust image patches. In this embodiment, the three image patches are uniformly adjusted to 640×640 ROI local patches. A 16×16 cue point grid is evenly distributed on each 640×640 ROI local patch. The total number of cue points can be adaptively adjusted by those skilled in the art based on the image size. The SAM model outputs an independent segmentation mask for each cue point and executes a three-stage filtering pipeline:
[0030] (1) IoU threshold filtering: Set the IoU threshold to 0.88 to ensure strong alignment between the predicted mask and the real cell contour;
[0031] (2) Stability filtering: Set the mask stability score threshold to 0.95 to ensure the reliability of the segmentation results;
[0032] (3) Area filtering: Limit the minimum mask area to 500 pixels to eliminate fragmented noise and invalid small areas;
[0033] All filtered masks are merged into a single binary mask by performing a pixel-wise maximum value operation. This mask is then overlaid onto the original ROI local patch, with the background area set to pure black and the cell area retaining its original pixel values.
[0034] Step 3: Cell Instance Extraction and Classification. Eight-connected domain analysis is used to locate independent cell regions within the foreground mask, enabling spatially separated cell instance identification. This also achieves preliminary separation of overlapping and adjacent cell bodies. Specifically, this includes the following steps:
[0035] The octet analysis method is used to locate independent cell regions in the foreground mask, and to complete the identification of spatially separated cell instances and the preliminary separation of overlapping and adjacent cell bodies;
[0036] Morphological filtering is performed on connected components, and regions with a connected component area of less than 100 pixels are regarded as noise and discarded. For valid connected components, single-cell patches are cropped by their minimum bounding rectangle, retaining only the pixels within the target component and filling the remaining areas with a black background. At the same time, key data of each patch can be recorded, including the horizontal and vertical offsets, bounding box width and height, and the actual pixel area of the corresponding cell in the original ROI local patch of the patch.
[0037] Cell coverage is used as a matching index to distinguish abnormal cells from normal cells. Cell coverage is defined as the ratio of the overlap area between the segmented cell bounding box and the labeled abnormal box to the total area of the segmented cell patch. When the coverage is not less than 0.5, it is marked as an abnormal cell; otherwise, it is considered a normal cell.
[0038] When an anomaly bounding box covers multiple segmented cell instances, the discrete cells are merged into a unified tile, creating a blank canvas that covers the minimum bounding rectangle of all target cells, and mapping them while maintaining their original relative positions within the ROI local tile.
[0039] After completing the matching, filtering and merging process, two standardized datasets, one for abnormal cells and one for normal cells, are obtained. All entries are sorted in descending order of tile area.
[0040] Step 4: Random resampling of cell instances, the specific process includes:
[0041] For each original ROI local patch, a blank layout canvas with the same resolution as the original ROI local patch is created, and the background color is a solid color. In this embodiment, the background color is set to the statistical mean of the background hue of the original TCT microscopic image (light pink, RGB: 253, 194, 255). It can be flexibly replaced with other solid color backgrounds according to actual needs. In addition, if more than half of the area of the minimum bounding rectangle of the current cell instance falls within the minimum bounding rectangle of another cell instance, the two cell instances are merged to obtain the merged minimum bounding rectangle. If one of the cell instances in the merged cell instance is an abnormal cell, the merged cell instance is also an abnormal cell; otherwise, the merged cell instance is a normal cell.
[0042] Each cell instance undergoes two random geometric transformations before layout: first, it is scaled using bilinear interpolation with a random scaling factor in the range of 0.9 to 1.1 to maintain pixel fidelity and prevent distortion of cell morphology and structure; then, it is rotated with a random rotation angle in the range of 0° to 360°, during which the canvas boundary is adaptively expanded to prevent cell truncation; all transformation parameters (scaling factor, rotation angle, and affine matrix) are fully recorded to ensure experimental traceability.
[0043] A non-overlapping layout mechanism is adopted. Candidate coordinates are randomly sampled for each cell instance. These coordinates serve as the center of the minimum bounding rectangle of the cell instance. Axis-aligned bounding box collision detection is used to determine whether the minimum bounding rectangle of the current cell instance overlaps with the minimum bounding rectangle of the already laid-out cells (i.e., whether the projections of the detection boxes of the two cell instances on the x-axis and y-axis overlap simultaneously. If the projections on both axes overlap, it is considered an overlap; otherwise, it is considered a non-overlap). If an overlap is detected or the minimum bounding rectangle of the current cell instance exceeds the canvas boundary, it is necessary to iteratively resample the new coordinates. Abnormal cell instances are prioritized for layout and given double the number of placement attempts relative to normal cells. Normal cells fill the remaining canvas space in descending order of tile area, taking into account both space utilization and the rationality of sample distribution. If the cell instance is not successfully placed within the set number of placement attempts, the placement of the cell instance is abandoned. The number of placement attempts can be determined by those skilled in the art based on experience or set to a fixed number. In this embodiment, it is set to 60 times. Figure 4 A schematic diagram of the cell instance resampling strategy is given, where the first image in rows (a), (b), and (c) is the original cell image, the other images in row (a) are four instances after the original cell is randomly placed, the other images in row (b) are four instances after the original cell is randomly rotated, and the other images in row (c) are four instances after the original cell is randomly scaled within a priority range.
[0044] Each original TCT image undergoes cell instance segmentation and extraction only once. The random resampling pipeline runs repeatedly with an independent random seed, achieving efficient data expansion with limited computational overhead. The cell entities in each standardized image maintain a one-to-one correspondence with the original TCT image, and the detection and recognition results can be accurately mapped back to the original image.
[0045] Figure 2 A schematic diagram illustrating an example of obtaining standardized liquid-based thin-layer cell images is provided. Figure 2 The first row in the middle is the image on the right obtained by segmenting the original liquid-based thin-layer cell detection image on the left. That is, the image on the right includes every cell instance in the original liquid-based thin-layer cell detection image. Figure 2 The second row extracts and classifies the segmented cell instances, distinguishing them into abnormal cells and normal cells. Figure 2 The third row in the middle is a standardized liquid-based thin-layer cell image with a uniform spatial distribution of cell populations, generated by randomly resampling and arranging the extracted cell instances in a non-overlapping layout on a blank canvas.
[0046] Example 2: Construction of the DINO-YOLO recognition model.
[0047] The DINO-YOLO recognition model includes the YOLOv12 detection framework and the DINOv3 visual model. Generally, the DINOv3 visual model is used to enhance the feature map of the input DINO-YOLO recognition model and the feature map at an intermediate scale, or only the feature map of the input DINO-YOLO recognition model. However, the first type of model has a large number of parameters and slow inference speed. In dense target scenes such as pathological cells, the enhancement of the second type of structure will introduce background noise.
[0048] To reduce computational overhead while improving cell recognition accuracy, this invention utilizes only the DINOv3 visual model to enhance feature maps at intermediate scales. Specifically, in this embodiment, the DINO-YOLO recognition model uses YOLOv12-Large as its backbone network framework, with a maximum channel count limited to 512 to achieve a balance between detection accuracy and computational efficiency. A pre-trained DINOv3-ViT-B / 16 module of specification DINOv3-ViT-B / 16 is embedded after the backbone network as a feature enhancement unit. The YOLOv12-Large backbone network continuously downsamples the input standardized liquid-based thin-layer cell image to obtain feature maps at three different scales. The pre-trained DINOv3-ViT-B / 16 module is used to enhance the feature maps at intermediate scales. The data processing specifically includes:
[0049] YOLOv12-Large extracts convolutional feature maps from standardized liquid-based thin-layer cell images and uses these convolutional feature maps as the original Z1 feature maps.
[0050] The spatial dimension of the convolutional feature map is flattened to the sequence length, and the number of channels is linearly projected from 512 to 768 to align with the hidden layer dimension of DINOv3-ViT-B / 16, generating the Patch feature sequence.
[0051] A learnable CLS Token and four learnable Register Tokens are concatenated at the front end of the Patch feature sequence. The CLS Token is used to capture global feature information, and the Register Tokens are used to stabilize the Transformer training process and reduce noise interference to feature learning.
[0052] The concatenated feature sequence is input into DINOv3-ViT-B / 16 for processing. The CLS Token and Register Token in the processed feature sequence are discarded. Then, the processed feature sequence is restored to a 2D feature map through the Unflatten operation.
[0053] The number of channels is linearly projected from 768 to 512 through a projection operation and then residually fused with the original Z1 feature map to obtain a fused feature map.
[0054] The fused feature map is continuously downsampled to generate multiple feature maps at different scales. The feature maps at all scales are then input into the multi-scale detection head to obtain the final detection result.
[0055] Specifically, this embodiment uses YOLOv12-Large as the backbone network framework, and performs customized channel adaptation and structural refinement, such as... Figure 3 The specific image processing steps include:
[0056] First, the standardized liquid-based thin-layer cell image is initially downsampled through two standard convolutional layers with a stride of 2. The channel dimension is gradually expanded from 3 to 64 and then to 128. After processing, the feature map size is reduced to 1 / 4 of the size of the standardized liquid-based thin-layer cell image.
[0057] The number of channels is then expanded to 256 by a C3k2 module, and then spatially downsampled by a convolutional layer with a stride of 2 (keeping the number of channels unchanged at 256). The resulting feature map size is reduced to 1 / 8 of the size of the normalized liquid-based thin-layer cell image.
[0058] Then, a C3k2 module expands the channels to 512. The expanded channel feature map, namely YOLOv12-Large, is used to extract the convolutional feature map from the standardized liquid-based thin-layer cell image.
[0059] The convolutional feature map is fed into the pre-trained DINOv3 visual model. After the pre-trained DINOv3 visual model enhances the convolutional feature map, it is fused with the convolutional feature map using residuals to obtain the original Z1 feature map.
[0060] The original Z1 feature map is preserved with 512 channels through a convolution operation with a stride of 2, and then high-level features are encoded by stacking four A2C2f modules to obtain the Z2 scale feature map;
[0061] Spatial downsampling is performed again using a convolution operation with a stride of 2 to maintain 512 channels, and then four A2C2f modules are stacked to generate a Z3-scale feature map.
[0062] In this embodiment, the pre-trained DINOv3 visual model serves as a feature enhancement module, employing a weight freezing strategy and functioning solely as a plug-and-play attention enhancement unit. Its data processing includes:
[0063] The spatial dimensions of the convolutional feature maps are flattened into a sequence, and the number of channels is linearly projected from 512 to 768 to align with the hidden layer dimensions of DINOv3-ViT-B / 16.
[0064] A learnable CLS Token and four Register Tokens are appended to the front of the sequence: the CLS Token is used to capture global feature information, and the Register Tokens are used to stabilize the Transformer training process. After the complete input sequence is fed into the DINOv3-ViT-B / 16 backbone network for processing, only the Patch Token is retained and the CLS Token and Register Token are discarded.
[0065] The output sequence is unflattened to restore a 2D feature map, and then the number of channels is mapped back from 768 to 512 through the output projection layer.
[0066] The original Z1-scale feature maps, Z2-scale feature maps, and Z3-scale feature maps are used together as input to the multi-scale detection head. This embodiment employs the native YOLOv12 PAN-FPN structure as the feature fusion network. Efficient multi-scale feature fusion is achieved through bidirectional feature transfer paths from top to bottom and bottom to top. The efficiently fused features at each scale are input into the classification head to obtain prediction results, resulting in a multi-scale prediction result. Redundant bounding boxes are then eliminated using non-maximum suppression (NMS) to obtain the final detection result. Figure 3 The specific processing steps include:
[0067] After upsampling the Z3-scale feature map to the same size as the Z2-scale feature map, it is concatenated with the Z2-scale feature map and then input into the A2C2f module for fusion to obtain the first fused feature map. After upsampling the first fused feature map to the same size as the original Z1-scale feature map, it is concatenated with the original Z1-scale feature map and then input into the A2C2f module for fusion to obtain the university fused feature map at the original Z1 scale. The classification head is then used to perform classification prediction based on the university fused feature map at the original Z1 scale.
[0068] The original Z1-scale university fusion feature map is downsampled to the size of the first fusion feature map after convolution operation, and then spliced with the first fusion feature map. Then it is input into the A2C2f module for fusion to obtain the second fusion feature map, that is, the Z2-scale university fusion feature map. The classification head is used to perform classification prediction based on the second fusion feature map.
[0069] The second fused feature map is downsampled to the same size as the Z3 scale feature map after convolution operation. It is then concatenated with the Z3 scale feature map and input into the A2C2f module for fusion to obtain the university fused feature map at the Z3 scale. The classification head is then used to perform classification prediction based on the university fused feature map at the Z3 scale.
[0070] In this embodiment, the DINOv3 visual model employs a weight freezing strategy and serves only as a plug-and-play attention enhancement unit. Its learnable parameters are limited to the YOLOv12 backbone network, neck network, detection head, and the cross-modal fusion module between the DINOv3 visual model and the YOLOv12 detection framework pipeline. The total loss function used during training of the DINO-YOLO recognition model is expressed as:
[0071]
[0072] in, The total loss function for the DINO-YOLO recognition model; For bounding box regression loss; For classification loss; For distribution focus loss; , , These are the weight coefficients for bounding box regression loss, classification loss, and distribution focus loss, respectively.
[0073] The method of this invention can serve as a general standardized preprocessing pipeline, applicable to various medical cell image analysis scenarios. By optimizing the data distribution quality at the cell level, it fundamentally satisfies the basic assumption of independent and identically distributed training and testing data, thus solving the problem of generalization bottleneck of deep learning models in different cell analysis tasks.
[0074] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A liquid-based thin-layer cell detection image recognition method based on cell distribution normalization, characterized in that, Includes the following steps: Acquire raw liquid-based thin-layer cell detection images and extract raw region of interest images containing abnormal cells from them; A cell instance segmentation model is used to segment the original region of interest image into cell instances, and the mask of each cell instance is obtained. The segmented cell instances are extracted and classified, distinguishing between abnormal cells and normal cells; Perform cell distribution normalization processing, randomly resample and arrange the extracted cell instances in a non-overlapping layout on a blank canvas to generate a standardized liquid-based thin-layer cell image with a uniform spatial distribution of the cell population. Standardized liquid-based thin-layer cell images are input into a pre-trained DINO-YOLO recognition model to obtain the detection results of abnormal cell types and locations.
2. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 1, characterized in that, The cell instance segmentation model uses the SAM model. By uniformly placing cue points on the original region of interest image, the SAM model outputs an independent segmentation mask for each cue point. A three-stage filtering pipeline then filters for high-quality masks. The three-stage filtering pipeline specifically includes the following steps: Set the IoU threshold to be no less than 0.88 to ensure the alignment of the predicted mask with the real cell contours; Set the mask stability score to no less than 0.95 to ensure segmentation reliability; To eliminate fragmentation noise, the minimum mask area is limited to 500 pixels. All masks that meet the conditions are merged into a single binary mask through pixel-wise maximum value operation, and then overlaid on the original region of interest image to obtain the segmentation result.
3. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 1, characterized in that, The process of extracting and classifying the segmented cell instances includes: The octet analysis method is used to locate independent cell regions in the foreground mask, and to complete the identification of spatially separated cell instances and the preliminary separation of overlapping and adjacent cell bodies; Morphological filtering is applied to connected components, and regions with a connected component area less than 100 pixels are treated as noise and discarded; other regions are cropped into single-cell patches using their minimum bounding rectangle, retaining only the pixels within the target component, and filling the remaining regions with a black background. Cell coverage is used as a matching index to distinguish abnormal cells from normal cells. Cell coverage is defined as the ratio of the overlap area between the segmented cell bounding box and the labeled abnormal box to the total area of the segmented cell patch. When the coverage is not less than 0.5, it is marked as an abnormal cell; otherwise, it is considered a normal cell. When an anomaly bounding box covers multiple segmented cell instances, the discrete cells are merged into a unified tile, creating a blank canvas that covers the minimum bounding rectangle of all target cells while maintaining their original relative positions on the region of interest image.
4. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 1, characterized in that, The process of performing cell distribution normalization includes: Create a blank layout canvas with the same resolution as the original region of interest image, and use a solid color background for the blank layout canvas; Each cell instance undergoes two random geometric transformations for enhancement before layout and arrangement, and the transformation parameters are recorded to ensure traceability of the transformation process; A non-overlapping layout mechanism is adopted. Candidate coordinates are randomly sampled for each cell instance. Axis-aligned bounding box collision detection is used to determine spatial overlap with already laid-out cells. When an overlap is detected, new coordinates are iteratively resampled. Abnormal cell instances are prioritized for layout and given double the number of placement attempts. Normal cells fill the remaining canvas space in descending order of tile area.
5. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 4, characterized in that, The background color of the blank layout canvas is the statistical mean of the background hue of the original liquid-based thin-layer cell detection image.
6. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 4, characterized in that, The enhancement process of cell instances before layout arrangement involves two random geometric transformations: The first random geometric transformation is scaled using bilinear interpolation with a random scaling factor in the range of 0.9 to 1.1; The second random geometric transformation rotates the object by a random rotation angle within the range of 0° to 360°.
7. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 1, characterized in that, The DINO-YOLO recognition model includes the YOLOv12 detection framework and the DINOv3 vision model. The YOLOv12 detection framework uses YOLOv12-Large as the backbone network framework, with a maximum channel count limit of 512. The DINOv3 vision model adopts the DINOv3-ViT-B / 16 specification and is only used to enhance features at intermediate scales.
8. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 7, characterized in that, The DINO-YOLO recognition model processes the input standardized liquid-based thin-layer cell images as follows: Convolutional feature maps were extracted from standardized liquid-based thin-layer cell images using YOLOv12-Large. The spatial dimension of the convolutional feature map is flattened to the sequence length, and the number of channels is linearly projected from 512 to 768 to align with the hidden layer dimension of DINOv3-ViT-B / 16, generating the Patch feature sequence. A learnable CLS Token and four learnable Register Tokens are concatenated at the front end of the Patch feature sequence. The CLS Token is used to capture global feature information, and the Register Tokens are used to stabilize the Transformer training process and reduce noise interference to feature learning. The concatenated feature sequence is input into DINOv3-ViT-B / 16 for processing. The CLSToken and Register Token in the processed feature sequence are discarded. Then, the processed feature sequence is restored to a 2D feature map through the Unflatten operation. The number of channels is linearly projected from 768 to 512 through a projection operation and then residually fused with the convolutional feature map to obtain a fused feature map. The fused feature map is continuously downsampled to generate multiple feature maps at different scales. The feature maps at all scales are then input into the multi-scale detection head to obtain the final detection result.
9. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 7, characterized in that, During the training of the DINO-YOLO recognition model, the parameters of the DINOv3 vision model are frozen.
10. The liquid-based thin-layer cell detection image recognition method based on cell distribution normalization according to claim 9, characterized in that, The total loss function used during training of the DINO-YOLO recognition model is expressed as: in, The total loss function for the DINO-YOLO recognition model; For bounding box regression loss; For classification loss; For distribution focus loss; , , These are the weight coefficients for bounding box regression loss, classification loss, and distribution focus loss, respectively.
