A bladder cancer cell segmentation method, device and storage medium based on urine sediment microscopy

Abstract

The application discloses a bladder cancer cell segmentation method and device based on urine sediment microscopy and a storage medium; the method comprises the following steps: taking a urine sediment microscopy image as an input image, inputting the input image into a detection model, marking a high-suspected cancer cell area in the input image through the detection model, and generating a boundary box; segmenting the high-suspected cancer cell area in the boundary box through a segmentation model, obtaining a mask of each cancer cell, deleting the corresponding boundary box if the mask is empty or the area ratio is lower than a preset threshold, mapping the effective mask back to the input image, merging all the effective masks, and generating a segmentation result of the cancer cells. The detection model can effectively reduce the processing range of the segmentation model, avoid the waste of computing resources caused by full-image search, and improve the processing speed. The segmentation model can automatically filter out false detection or invalid areas, and ensure the accuracy and reliability of the segmentation result.

Description

Technical Field

[0001] This application relates to the field of cancer cell recognition technology, and in particular to a method, device and storage medium for segmenting bladder cancer cells based on microscopic examination of urine sediment. Background Technology

[0002] Bladder cancer is a common malignant tumor of the urinary system, and urine sediment microscopy, as a non-invasive early screening method, has important clinical value. Traditional manual microscopy relies on liquid-based thin-layer cytology techniques (such as SurePath and ThinPrep) to prepare slides, after which pathologists observe the morphology of cancer cells under a microscope. However, this method has problems such as high subjectivity, low efficiency, poor reproducibility, and high false positive / false negative rates. Especially when there is cell overlap and severe background interference, young physicians are prone to missing or misdiagnosing the disease, and even experienced physicians experience visual fatigue due to prolonged operation.

[0003] In recent years, deep learning has developed rapidly in medical image analysis. However, in the context of liquid-based urinary sediment slides, cancer cells exhibit highly heterogeneous morphology, severe overlap, and complex backgrounds. Existing models are mostly single or simple cascades, lacking false positive filtering of bounding boxes based on segmentation results. This leads to insufficient clinical specificity and low segmentation accuracy of cancer cells in liquid-based urinary sediment slides, making it difficult to directly apply them to auxiliary diagnosis. Summary of the Invention

[0004] The main technical problem addressed by this application is to provide a method for segmenting bladder cancer cells based on microscopic examination of urine sediment, which solves the problem of insufficient clinical specificity due to the lack of false positive filtering of the bounding box based on the segmentation results, resulting in low accuracy of cancer cell segmentation in liquid-based slides of urine sediment.

[0005] To address the aforementioned technical problems, this application provides a method for bladder cancer cell segmentation based on urinary sediment microscopy, comprising the following steps: using a urinary sediment microscopy image as an input image; inputting the input image into a detection model; marking highly suspected cancer cell regions in the input image using the detection model to generate bounding boxes; segmenting the highly suspected cancer cell regions within the bounding boxes using a segmentation model to obtain a mask for each cancer cell; deleting the corresponding bounding box if the mask is empty or its area percentage is below a preset threshold; mapping the valid masks back to the input image; merging all valid masks to generate the cancer cell segmentation result.

[0006] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method.

[0007] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described.

[0008] The beneficial effects of this application are as follows: This application utilizes a detection model to perform preliminary screening of input urinary sediment microscopic images, accurately locating highly suspected cancerous areas in the image and generating bounding boxes for labeling. This effectively narrows the processing range of the segmentation model, avoids the waste of computational resources caused by full-image search, and improves processing speed. The segmentation model performs fine segmentation on highly suspected areas within the bounding boxes, obtaining an accurate mask for each cancer cell. By setting a preset threshold for the proportion of mask area, the system can automatically filter out false detections or invalid areas, ensuring the accuracy and reliability of the segmentation results. The effective masks are mapped back to the original input image, and all effective masks are merged to generate a complete segmentation result for bladder cancer cells. This achieves efficient and accurate segmentation of bladder cancer cells, improving the diagnostic value of urinary sediment microscopic examination. Attached Figure Description

[0009] Figure 1 This is a flowchart according to an embodiment of this application;

[0010] Figure 2 This is a flowchart according to an embodiment of the present application;

[0011] Figure 3 This is a flowchart of a detection model according to an embodiment of this application;

[0012] Figure 4 This is a flowchart of an attention gating mechanism according to an embodiment of this application;

[0013] Figure 5 This is a flowchart of a nuclear mass sensing module according to an embodiment of this application;

[0014] Figure 6 This is an image data diagram of a urine sediment specimen according to an embodiment of this application;

[0015] Figure 7 This is a schematic diagram of the prediction result according to an embodiment of this application. Detailed Implementation

[0016] To facilitate understanding of this application, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of this application are shown in the drawings. However, this application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of this application.

[0017] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0018] Figures 1-7 This application illustrates an embodiment of its method for segmenting bladder cancer cells based on microscopic examination of urine sediment, including:

[0019] Step S1: Use the microscopic image of urine sediment as the input image, input the input image into the detection model, and use the detection model to mark the areas of high suspicion of cancer cells in the input image to generate bounding boxes;

[0020] Step S2: Using the segmentation model, segment the regions with high suspicion of cancer cells within the bounding box to obtain the mask for each cancer cell. If the mask is empty or the area ratio is lower than a preset threshold, delete the corresponding bounding box, map the valid mask back to the input image, merge all valid masks, and generate the segmentation result of the cancer cells.

[0021] This application utilizes a detection model to initially screen input urinary sediment microscopic images, accurately locating regions with high suspicion of cancer cells and generating bounding boxes for labeling. This effectively narrows the processing scope of the segmentation model, avoiding the waste of computational resources from full-image search and improving processing speed. The segmentation model performs fine segmentation on high-susceptibility regions within the bounding boxes, obtaining precise masks for each cancer cell. By setting a preset threshold for the proportion of mask area, the system can automatically filter out false detections or invalid regions, ensuring the accuracy and reliability of the segmentation results. The valid masks are mapped back to the original input image, and all valid masks are merged to generate a complete segmentation result for bladder cancer cells. This achieves efficient and accurate segmentation of bladder cancer cells, enhancing the diagnostic value of urinary sediment microscopic examination.

[0022] A detection model is used to quickly locate regions with a high probability of cancer cells and obtain bounding boxes.

[0023] The detection model can use the YOLO (You Only Look Once) series of networks, such as YOLO9, YOLO10 and YOLO11, with YOLO11 being preferred.

[0024] The detection model was custom-trained using the official Ultralytics framework, with key parameters set to "task=detectmode=train model=yolo11n.pt epochs=200 imgsz=1024". Training employed Ultralytics' default optimization strategies, including Automatic Mixed Precision (AMP), Mosaic data augmentation, and Automatic Anchor Box Clustering. The weight file was used for subsequent prediction. During prediction, the pre-trained model weight file was loaded, with a confidence threshold of YOLO_CONF=0.3, and only cancer cell detection was performed.

[0025] After detection, a further step is to set an expansion threshold for the bounding boxes, and then expand the bounding boxes based on this threshold. The expansion threshold can be 10%-30%, expanding each bounding box outwards by 10%-30%. For example, it can be expanded by 20% (expand_ratio=0.2) to obtain a cropped region containing the complete cell and its surrounding background, ensuring sufficient background information when inputting to the segmentation model.

[0026] In some embodiments, such as Figure 3 As shown, the segmentation model includes an encoder and a decoder. The encoder comprises an initial convolutional module, a max-pooling layer, and multiple feature extraction modules connected in sequence. The initial convolutional module preserves high-resolution information in the input image, helping to recover detailed features in the urine sediment image. The max-pooling layer reduces spatial resolution while maintaining the same number of channels. The feature extraction modules extract semantic features from the image. The decoder includes multiple upsampling modules, each upsampling low-resolution features and skip-connecting them with high-resolution features in the same layer for feature fusion. The segmentation model demonstrates excellent segmentation capabilities for complex boundaries and small targets, addressing challenges such as diverse cancer cell morphologies, large scale variations, severe background interference, and cell overlap in urine sediment images.

[0027] In some embodiments, such as Figure 2 As shown, the specific structure and steps of the encoder are as follows:

[0028] Initial Convolutional Module: The input image is first passed through a 7×7 convolutional layer (stride=2), followed by batch normalization and ReLU activation, reducing the spatial resolution of the input image from 1024×1024 to 512×512, while increasing the number of channels from 3 to 64. The output feature of this module is denoted as F0, serving as the first skip connection feature. This preserves high-resolution information and helps recover detailed features in urine sediment images.

[0029] Max pooling layer: Perform 3×3 max pooling (stride=2) on F0 to further reduce the spatial resolution of the input image to 256×256, while keeping the number of channels unchanged at 64.

[0030] Feature extraction module 1: Consists of 3 Bottleneck residual blocks with a planes parameter of 64. Since the expansion factor parameter of the Bottleneck structure is 4, the number of output channels in this stage is 64 × 4 = 256. No spatial downsampling is performed in this stage (stride = 1), and the output feature result is denoted as F1, with a spatial size of 256 × 256 × 256, serving as the second skip connection feature.

[0031] Feature extraction module 2 consists of four Bottleneck residual blocks with a planes parameter of 128. The first residual block is spatially downsampled with a stride of 2. The output feature is denoted as F2, with a spatial size of 128×128×512, and serves as the third skip connection feature.

[0032] Feature extraction module 3 consists of 6 Bottleneck residual blocks with a planes parameter of 256. The first residual block is spatially downsampled with a stride of 2. The output feature is denoted as F3, with a spatial size of 64×64×1024, and serves as the fourth skip connection feature.

[0033] Feature extraction module 4 consists of three Bottleneck residual blocks with a planes parameter of 512. The first residual block is spatially downsampled with a stride of 2. The output feature is denoted as F4, with a spatial size of 32×32×2048. It is the deepest feature data of the coding layer, containing the most semantic feature information, but with the lowest spatial resolution. F4 is not used as a skip connection feature, but rather as the input to the deepest upsampling module of the decoder.

[0034] Each Bottleneck residual block uses a compression-expansion approach. First, the number of image channels is compressed using a 1×1 convolution, then features are extracted using a 3×3 convolution, and finally the number of channels is expanded using a 1×1 convolution.

[0035] In some embodiments, such as Figure 3 As shown, the decoder includes multiple upsampling modules. Each upsampling module upsamples low-resolution features and performs skip connections with high-resolution features in the same layer to achieve feature fusion.

[0036] Upsampling module 4: First, F4 is transposed and upsampled to restore it to the same size as F3. Then, the weighted result of F3 is calculated through the attention gating mechanism and concatenated with the upsampled F4. Finally, the features are fused through two 3×3 convolutional layers, and the output result has a spatial size of 64×64×512, which is then passed to upsampling module 3.

[0037] Upsampling module 3: Upsamples the output of upsampling module 4, calculates the weighted result of F2 through attention gating mechanism, and splices and fuses it with the upsampled features. The output result has a spatial size of 128×128×256 and is passed to upsampling module 2.

[0038] Upsampling module 2: Upsamples the output of upsampling module 3, calculates the weighted result of F1 through attention gating mechanism, and splices and fuses it with the upsampled features. The output result has a spatial size of 256×256×128 and is transmitted to upsampling module 1.

[0039] Upsampling Module 1: Upsamples the output of Upsampling Module 2, calculates the weighted result of F0 through the attention gating mechanism, and splices and fuses it with the upsampled features. The output result has a spatial size of 512×512×64 and is passed to Upsampling Module 0.

[0040] Upsampling Module 0: Upsamples the output of Upsampling Module 1 to 1024×1024 using bilinear interpolation, then passes it through two 3×3 convolutional layers. Each convolution is followed by batch normalization and a ReLU activation function to further optimize gradient calculation. Finally, a 1×1 convolutional layer generates the final segmentation result, with the number of output channels equal to the number of classes.

[0041] In some embodiments, such as Figure 4 As shown, in the upsampling module, the weighted results are calculated using an attention gating mechanism. The attention gating mechanism utilizes the detailed information provided by the decoder features to generate an attention weight map α, which is used to weight the encoder features. The specific calculation process is as follows:

[0042] The decoder and encoder feature channels have different dimensions. First, a 1×1 convolution is used to map them to an intermediate dimension. This process can be represented as:

[0043]

[0044]

[0045] Where g represents the decoder feature and x represents the encoder feature. and These represent the decoder features respectively. The encoder features x are subjected to a 1×1 convolution operation and batch normalization (BN) to transform them, and the resulting output is the convolutional decoder features. and encoder features The two channels have the same dimension, which prepares for the subsequent fusion process. After unifying the channel dimensions, the decoder features need to be... Upsampled to encoder features For the same size, upsampling is performed using bilinear interpolation, and new pixels are filled by linear weighting of neighboring pixels. The size alignment process can be represented as:

[0046]

[0047] in, These represent the height and width of the predicted bounding box, respectively. Features aligned to spatial dimensions. and An attention weight map α is generated through feature interaction to quantify the feature importance at each location. First, the attention weight map α is generated... and Pixel-by-pixel summation is performed to fuse high semantic information and high-resolution detail information. Then, a non-linear relationship is introduced through the ReLU activation function, outputting a feature map φ. Finally, the feature map φ is weighted and normalized, then passed through a 1×1 convolutional layer. The channels are compressed to one dimension, and then the output values ​​are mapped to the [0,1] interval using the Sigmoid activation function, ultimately generating an attention weight map α, representing the importance of the corresponding spatial location. The closer the value is to 1, the higher the probability that the corresponding spatial location is the target region; the closer it is to 0, the more likely the region is background noise. After this, the generated attention weight map α is compared with the original encoder features. Pixel-by-pixel multiplication is performed to achieve adaptive weighting of the encoder features, resulting in weighted features. It preserves the spatial details of the encoder while suppressing noise data in the background area, thus enhancing the feedback in the target area.

[0048]

[0049]

[0050]

[0051] in, Indicates the feature map results. This represents the weighted feature map result.

[0052] The upsampling module calculates weighted results through an attention gating mechanism, which can fully preserve the spatial detail information of the encoder, while suppressing noise data in the background area, enhancing the feedback of the target area, and improving the detection accuracy of the target.

[0053] In microscopic images of urine sediment, a key morphological feature of cancer cells is a significantly enlarged nucleus with a nucleus-to-cytoplasm ratio significantly higher than that of normal cells. This characteristic manifests in the image as a high-contrast, dense image in the nucleus region, while the cytoplasm region exhibits a low-contrast, sparse image. To fully utilize the morphological features of cancer cells, this application proposes a nucleocytoplasmic sensing module. This module extracts features from the nucleus and cytoplasm separately by designing different convolutional kernel sizes, and performs adaptive feature fusion based on nucleocytoplasmic probability estimation, thereby enhancing the ability to identify the nuclear region of cancer cells. The nucleocytoplasmic sensing module acts as a pre-step for skip connections, applying high-resolution features to the coding layer.

[0054] In some embodiments, such as Figure 5 As shown, the nucleoplasmic sensing module adopts a dual-branch architecture, extracting feature representations of the cell nucleus and cytoplasm respectively.

[0055] The nuclear branch is used to extract nuclear features. Since the nuclear region appears as a high-contrast, dense structure in images, to capture the overall pattern of this structure, the nuclear branch uses a large receptive field convolutional kernel (5×5) for feature extraction. The formula for the nuclear branch is as follows:

[0056]

[0057]

[0058] in, Indicates the input feature map, The output characteristics representing the kernel branch. This represents the intermediate features of the first stage of nuclear branching.

[0059] The cytoplasmic branch is used to extract cytoplasmic features. Since the cytoplasm region appears as a low-contrast, sparse structure in the image, it employs a small receptive field convolutional kernel (3×3) for feature extraction to preserve detail and avoid excessive smoothing. The formula for the cytoplasmic branch is as follows:

[0060]

[0061]

[0062] in, Output characteristics representing cytoplasmic branching. This represents the intermediate characteristics of the first stage of cytoplasmic branching.

[0063] To adaptively fuse nuclear and cytoplasmic features, the module introduces a nucleoplasmic probability estimation network to generate a nucleoplasmic weight map for each pixel location. This network takes the concatenation of nuclear and cytoplasmic features as input and outputs a single-channel weight map, where the value of each pixel represents how much the location resembles the cell nucleus: 0 indicates a stronger resemblance to cytoplasm, and 1 indicates a stronger resemblance to the cell nucleus. The formula for nucleoplasmic probability estimation is as follows:

[0064]

[0065]

[0066]

[0067]

[0068] in, This represents the intermediate feature map after the first round of convolutional feature transformation. This represents the intermediate feature map after the second round of convolutional feature transformation. This represents the splicing of nuclear and cytoplasmic features along the channel dimension. This represents a nucleocytoplasmic probability weighting plot. This represents the Sigmoid activation function, ensuring that the weight values ​​are between [0,1].

[0069] Based on the nucleoplasmic probability weight map, the module performs weighted fusion of nuclear and cytoplasmic features. In the nuclear region, a high weight value indicates that nuclear features dominate; in the cytoplasmic region, a low weight value indicates that cytoplasmic features dominate. This adaptive fusion strategy can fully utilize the morphological features of cancer cells and enhance the ability to identify nuclear regions. The formula for adaptive feature fusion is as follows:

[0070]

[0071]

[0072]

[0073]

[0074] in, This represents the weighted nuclear features. This represents the weighted cytoplasmic characteristics. This indicates the features after splicing and merging. This indicates pixel-by-pixel multiplication. This indicates the final output characteristics of the module.

[0075] In upsampling module 0, the output of upsampling module 1 is upsampled to 1024×1024 using bilinear interpolation. Bilinear interpolation is a fixed and non-learnable processing method. Although bilinear interpolation is computationally efficient, as a smooth interpolation method, it may lose important details, especially in boundary regions.

[0076] In this application, transposed convolution is used to replace bilinear interpolation for upsampling. The parameters of the transposed convolution can be iteratively optimized through backpropagation, capturing important details while sacrificing some time cost. During the upsampling process, the number of feature channels is gradually reduced. In the transposed convolution, the number of output channels is set to half the number of input channels, which maintains the richness of features while avoiding the computational pressure caused by too many channels.

[0077] The segmentation model incorporates a nucleoplasm perception module, an attention gating module, and a transposed convolutional upsampling module. The nucleoplasm perception module extracts features from the cell nucleus and cytoplasm separately by designing different convolutional kernel sizes, and performs adaptive feature fusion based on nucleoplasm probability estimation, thereby enhancing the ability to identify the nuclear region of cancer cells. The attention gating module performs weighted fusion of decoded features and encoded skip features to adaptively enhance key regions of cancer cells. The transposed convolutional upsampling module uses learnable weights to replace bilinear interpolation, improving the accuracy of boundary recovery.

[0078] Before use, the detection and segmentation models can be pre-trained. The input size for training both models is uniformly 1024×1024, with a weight decay parameter of 1×e⁻⁴, using the Adam optimizer with a momentum parameter of 0.9, an initial learning rate of 1×10⁻⁶, and a cosine annealing scheduling strategy. To alleviate the extreme imbalance between foreground and background pixels, a combination of Focal Loss and Dice Loss is used as the loss function. Training is conducted for 200 epochs, with the validation set Pixel Accuracy and FWIoU serving as the model selection criteria. The optimal weights are used for subsequent predictions.

[0079] The training dataset can be obtained as follows: Samples are collected from the second morning urine or fresh urine (volume > 30 mL) to ensure sufficient formed elements. The slide preparation process is as follows: First, the sample is vigorously shaken for 10 min to ensure uniform distribution of formed elements; then, the sample is placed in a 50 mL graduated centrifuge tube and centrifuged at 2000 r / min for 5 min, discarding the supernatant and retaining the sediment; automated slide preparation is performed using an automated liquid-based thin-layer cell preparation and staining machine and its consumables, following standard operating procedures, including cell enrichment, Papanicolaou staining, dehydration with anhydrous ethanol, xylene clearing, and mounting with neutral resin. The prepared liquid-based thin sections are used for optical microscopy imaging to obtain a high-resolution urine sediment image dataset.

[0080] This slide preparation method is based on liquid-based cytology, which can effectively reduce cell overlap and background interference, improve the clarity of cancer cell morphology, and is suitable for subsequent deep learning analysis.

[0081] However, the large image size of urine sediment microscopic examination significantly increases model training time and cell identification difficulty. Therefore, this application adopts a sliding window slicing training and inference method to solve the problem caused by the large image data size. Figure 6 As shown, the input data will be divided into multiple 1024×1024 pixel slices, which will then be input into the object detection model for training. In the inference stage, the image will also be divided into multiple slices of the same size for prediction, and then visualized on the whole image using a sliding window.

[0082] All image annotations were performed by physicians to ensure accuracy. The YOLOv11 detection model training data was annotated with bounding boxes using the X-AnyLabeling tool, which supports interactive annotation and direct export of YOLO annotation format. The annotation results are horizontal bounding boxes for cancer cells, with hundreds to thousands of cancer cell instances annotated per specimen image. The NCA-UNet segmentation model training data was annotated with polygons using the Labelme tool. Physicians outlined the cancer cells, generating JSON-formatted annotation files.

[0083] For a single input image, prediction is performed through the following steps:

[0084] (1) The detection model outputs bounding boxes of highly suspected cancer cells;

[0085] (2) After expanding each bounding box, trim the patch and perform normalization preprocessing consistent with the training process;

[0086] (3) Input the segmentation model to obtain the prediction result, and then calculate it through the Softmax function. Binarize it with a threshold of 0.6 to obtain the initial mask.

[0087] (4) The segmentation result of the segmentation model is used to determine the detection result of the detection model. If the number of predicted cancer cell pixels in the bounding box of the detection model is less than 0.01 of the total number of pixels in the predicted box, the bounding box is deleted. If it is greater than 0.01, it can be used as a valid mask.

[0088] (5) Based on the expansion and stretching ratio, restore the effective mask to the original image and perform accurate coordinate mapping.

[0089] (6) Combine all valid masks and overlay them on the input image.

[0090] This application is compared with the prior art as shown in Tables 1-4 and Figure 7 As shown.

[0091] Table 1 Performance comparison of the segmentation model with other models

[0092] method Mean IoU (%) Dice (%) FW IoU (%) Average inference time (ms / figure) MC-Unet 83.81 87.3 98.24 390.45 U-Net (ResNet50) 81.50 84.47 98.86 301.57 Mask R-CNN 78.9 81.7 96.59 562 Segmentation Model 84.17 88.08 98.23 331.91

[0093] Table 2 Performance comparison between this application and YUSEG

[0094] method mAP@0.5:0.95 Precision Mean IoU (%) Recall Dice (%) Average inference time (ms / figure) YUSEG 0.918 0.8920 82.4 0.8263 84.49 329 This application 0.941 0.9200 82.9 0.7500 84.35 392

[0095] Table 3 Performance comparison with using YOLOv11 alone

[0096] method Precision Recall F1-Score TP FP FN YOLOv11 0.8500 0.7800 0.835 156 28 44 This application 0.9200 0.7500 0.8261 150 13 50

[0097] Table 4. Results of ablation experiments using the segmentation model

[0098] Model Configuration Attention gating Nucleus-Cytoplasmic Sensing Module Mean IoU (%) Dice (%) FW IoU(%) Average inference time (ms / figure) U-Net × × 81.50 84.47 98.86 301.57 + Attention Gating √ × 81.47 86.83 98.74 311.06 + Nucleocytoplasmic Sensing Module × √ 82.19 85.36 98.69 314.47 This application √ √ 84.17 88.08 98.73 331.91

[0099] Therefore, this application achieves high-precision instance segmentation. Experiments show that the mean IoU of the segmentation model reaches 84.17% and the Dice coefficient reaches 88.08%, significantly outperforming other models. It also boasts high inference efficiency (approximately 392 ms / image), making it suitable for real-time clinical auxiliary diagnosis and possessing strong practicality and promotional value. The innovative post-processing filtering strategy (based on mask quality to remove bounding boxes) significantly reduces false positives (precision improved by 7%), enhances clinical specificity, and reduces unnecessary anxiety and examinations for patients. The nucleocytoplasmic perception module extracts features from the cell nucleus and cytoplasm separately by designing different convolutional kernel sizes, and performs adaptive feature fusion based on nucleocytoplasmic probability estimation, thereby enhancing the ability to identify cancer cell nuclei. The synergistic improvement of attention gating and the nucleocytoplasmic perception module improves the segmentation ability for small cancer cell targets, overlapping regions, and ambiguous boundaries.

[0100] The above are merely embodiments of this application and do not limit the scope of this patent application. Any equivalent structural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of this application.

Claims

1. A method for segmenting bladder cancer cells based on microscopic examination of urine sediment, characterized in that, Including the following steps: The image of urine sediment under microscopy is used as the input image. The input image is input into the detection model. The detection model marks the regions of high suspicion of cancer cells in the input image and generates bounding boxes. The segmentation model segments the regions with high suspicion of cancer cells within the bounding box, obtaining a mask for each cancer cell. If the mask is empty or the area ratio is lower than a preset threshold, the corresponding bounding box is deleted, the valid masks are mapped back to the input image, and all valid masks are merged to generate the segmentation result of the cancer cells.

2. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 1, characterized in that, Set an expansion threshold for the bounding box, and expand the bounding box based on the expansion threshold.

3. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 1, characterized in that, The segmentation model includes an encoder and a decoder. The encoder includes an initial convolutional module, a max pooling layer, and multiple feature extraction modules connected in sequence. The initial convolutional module is used to retain high-resolution information in the input image, and the max pooling layer is used to reduce spatial resolution while keeping the number of channels constant. The feature extraction modules are used to extract semantic features from the input image. The decoder includes multiple upsampling modules. Each upsampling module upsamples low-resolution features and skips connections with high-resolution features in the same layer to perform feature fusion.

4. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 3, characterized in that, The upsampling module calculates the weighted result through an attention gating mechanism. The attention gating mechanism uses the detailed information provided by the decoder features to generate an attention weight map. The generated attention weight map is then multiplied pixel by pixel with the encoder features to adaptively weight the encoder features.

5. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 3, characterized in that, Before the skip connection, a nucleoplasmic sensing module is set up, which includes a nuclear branch and a cytoplasmic branch. The nuclear branch and the cytoplasmic branch are set with different convolution kernel sizes to extract features of the nucleus and cytoplasm respectively.

6. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 5, characterized in that, The nuclear branch was used for feature extraction using a 5×5 convolution kernel; the cytoplasmic branch was used for feature extraction using a 3×3 convolution kernel.

7. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 6, characterized in that, The nucleocytoplasmic sensing module introduces a nucleocytoplasmic probability estimation network to generate a nucleocytoplasmic weight map for each pixel location. The nucleocytoplasmic probability estimation network takes the concatenation of nuclear and cytoplasmic features as input and outputs a single-channel nucleocytoplasmic probability weight map. Based on the nucleocytoplasmic probability weight map, the nucleocytoplasmic sensing module performs weighted fusion of nuclear and cytoplasmic features.

8. The method for segmenting bladder cancer cells based on microscopic examination of urine sediment according to claim 1, characterized in that, In the last upsampling module of the segmentation model, the output of the last upsampling module is set as a transposed convolution.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 8.

Images