Cancerous cell identification method in TCT detection mode

Abstract

The invention discloses a cancerous cell identification method in a TCT detection mode. The cancerous cell identification method comprises the following steps: step 1, establishing a data set and finishing arrangement; step 2, constructing a detection model; step 3, using CSPDarknet53 in the YOLO to replace a feature extraction trunk in the Faster R-CNN detection network, and optimizing the performance; and step 4, training and deploying the detection model, the stage is divided into two processes of detection model training and detection model deploying, after preliminary verification, large-scale training is performed on the detection model to obtain parameters of the final detection model, and the trained detection model is put into practical application to realize cell canceration detection. The invention belongs to the technical field of medical image analysis and artificial deep learning, and solves the problems of low cancer cell recognition efficiency and insufficient precision caused by deep feature loss and data imbalance due to complex cell boundaries in cancer cell recognition in the prior art.

Description

Technical Field

[0001] The present invention belongs to the field of medical image analysis and artificial deep learning technology, and relates to a method for identifying cancerous cells under TCT detection mode. Background Art

[0002] Cancer is one of the major diseases that threaten human health worldwide. With the advancement of medical technology, early cancer detection and diagnosis are crucial for improving cure rates and patient outcomes. Currently, TCT (thin-layer cytology test), a leading method for cancer cell screening, has been widely used in medical pathology analysis due to its painless, non-invasive, and efficient nature. However, existing technologies still face two major challenges in cancer cell identification: loss of deep feature location information and data imbalance.

[0003] Among them, the loss of deep feature position information refers to the gradual loss of deep feature position information during the multiple pooling processes using traditional neural networks. The morphological changes of cancerous cells in the early stages are usually subtle, which places higher demands on the extraction of deep features. This makes it difficult for existing detection network models to accurately identify cancerous cells. In addition, data imbalance is reflected in the sample data sets for cancer cell detection. Cancerous cells usually only account for a small proportion, while healthy cells account for the vast majority, making the distribution of abnormal and normal cells extremely unbalanced. As a result, existing network models are easily biased towards normal cells during training, reducing the detection rate of cancerous cells. Therefore, there is an urgent need to develop a new method to overcome the two technical difficulties mentioned above. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying cancerous cells under the TCT detection mode, which solves the problem of low efficiency and insufficient accuracy in the recognition of cancer cells in the existing technology due to the complex cell boundaries, loss of deep features and data imbalance.

[0005] The technical solution adopted by the present invention is a method for identifying cancerous cells under TCT detection mode, which is implemented according to the following steps: Step 1: Create a data set and complete the organization; Step 2: Build a detection model and design a deep learning network architecture for cell detection; Step 3: Use CSPDarknet53 in YOLO to replace the feature extraction backbone in the Faster R-CNN detection network and optimize the performance; Step 4: Train and deploy the detection model. This stage is divided into two processes: detection model training and detection model deployment. After obtaining the complete detection model structure and preliminary verification, the detection model is trained on a large scale to obtain the parameters of the final detection model, and the trained detection model is put into practical application to realize the automation of cell canceration detection.

[0006] The beneficial effect of this invention lies in its application in the detection and labeling of abnormal cells during early cervical cancer screening, based on TCT (thin layer cytology testing) technology. This method, combined with deep learning network model optimization, data enhancement, and feature visualization techniques, enables rapid, efficient, and automated cancer cell identification. It has broad applications in medical pathology analysis, early cancer screening and diagnosis, and the development of auxiliary diagnostic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 is a flow chart of constructing a data set in the method of the present invention; Figure 2 It is a flow chart of constructing a detection model in the method of the present invention; Figure 3 It is a block diagram of the backbone network designed in the method of the present invention; Figure 4 This is an example diagram of the cancer cell region obtained by the method of the present invention; Figure 5 It is a deployment flow chart of the detection model adopted by the method of the present invention. DETAILED DESCRIPTION

[0008] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0009] The overall concept of the method of the present invention is to improve the existing Faster R-CNN detection network as the main framework of the detection model of the present invention. In combination with the FPN (Feature Pyramid Network) module and the DCN (Deformable Convolution) module, it can adaptively extract the features of complex cell boundaries and efficiently fuse multi-scale features, thereby improving the detection accuracy of cancerous cells. To improve detection efficiency and detection accuracy, CSPDarknet53 from YOLO is also introduced as the feature extraction backbone network to achieve rapid target detection while ensuring detection accuracy.

[0010] The cancerous cell identification method based on TCT technology of the present invention is implemented according to the following steps: Step 1: Create a data set and complete the organization, refer to Figure 1 The specific process is: 1.1) Collect data, Obtain a large number of TCT cell images from a hospital or professional pathology testing institution to ensure that the sample contains normal cells and cancerous cells of various grades; 1.2) Label samples, Experienced pathologists manually annotate the collected TCT cell images, including boxing out the area where the cancerous cells are located and noting the cell type; The annotation process follows unified standards to ensure the accuracy and consistency of the annotation results.

[0011] 1.3) Divide the dataset, The labeled dataset is divided into training set, validation set and test set. For example, the training set accounts for 80% of the total data and is used for model training and parameter optimization; the validation set accounts for 10% and is used to monitor model performance and prevent overfitting during training; the test set accounts for 10% and is used to finally evaluate the recognition accuracy and generalization ability of the model.

[0012] Step 2: Build a detection model, such as Figure 2 As shown in the figure, the deep learning network architecture for cell detection is designed. The specific process is: 2.1) Improve the existing Faster R-CNN detection network as the main framework, Faster R-CNN (existing technology) is a region-based convolutional neural network, which includes three main parts: convolutional feature extraction, candidate region proposal (RPN) and region detection (RoI layer).

[0013] First, the input cervical cell image undergoes a series of convolution and downsampling operations to extract a multi-channel feature map. Then, an RPN is used to slide on the feature map to generate a large number of anchor boxes. A lightweight binary classification network is used to determine whether each anchor box contains the target cell. The position and size of the anchor box are regressed to propose candidate detection regions. Finally, for the selected candidate regions, the corresponding region features are intercepted from the feature map at the RoI layer and classified and fine-tuned position regression is performed to output a preliminary detection result, which includes the predicted box position and category.

[0014] Faster R-CNN consists of three modules: feature extraction → proposal box generation (RPN) → RoI region classification and regression; The expression for generating the candidate box is: (1) in, is the predicted value of the anchor's foreground probability, is the true label of the anchor; is the regression prediction box of the anchor, Predict the target for the anchor real box; is the classification loss, is the total loss of the Region Proposal Network (RPN); is the loss balance coefficient, , Normalization term for the number of positive samples for classification / regression; The expression for RoI region classification is: (2) in, is the classification loss, K is the total number of categories, For the k The true label of the class, For the model k The predicted probability of the class; The expression of RoI regression is: (3) in, Total regression loss, is the offset predicted by the model, is the actual offset; 2.2) Introducing FPN into the main framework, To address the multi-scale cell detection problem, this step integrates the FPN architecture (Feature Pyramid Network) into the feature extraction portion of the Faster R-CNN framework. The specific implementation process involves building on the multi-scale features extracted by the original convolutional backbone and establishing a top-down path to progressively upsample high-level features. These are then horizontally connected and fused with the underlying features at the corresponding scale.

[0015] The advantage of FPN is that large cell clusters can benefit from high-level semantic features, while small early-stage lesion cells are highlighted by fused low-level detail features. Therefore, after integrating FPN, the detection model has multi-scale detection capabilities, improving the detection rate of small lesion areas, especially when early cancer cells manifest as small-scale changes, which improves the detection accuracy of the detection model. The expression of FPN is: (4) in, It is the first The output feature map of the layer, Represents the bottom-up backbone network in the The feature map extracted by the layer, Indicates that the feature map of the previous layer (higher semantic layer) is upsampled to the current layer size. Conv (⋅) is a convolution operation, preferably a 3×3 convolution, which is used to remove the aliasing effect caused by upsampling; 2.3) Integrate DCN (Deformable Convolution) into the main framework, To enhance the detection model's ability to model complex cell edges and morphologies, the present invention incorporates a Deformable Convolution (DCN) module. This module dynamically adjusts the position of the convolution kernel based on cell morphology and boundaries, overcoming the limitation of traditional convolution that only samples on a regular grid. This allows the model to adaptively align with the edge contours of cancerous cells to extract features, thereby improving the accuracy of cancer cell identification.

[0016] The expression of the DCN module is: (5) in is the output feature map, is the input feature map, is the convolution kernel weight, is the receptive field of the standard convolution kernel, It is the relative position within the convolution window, and P0 is the spatial coordinate of the current output value to be calculated.

[0017] After the combined processing in step 2, the present invention constructs an improved detection model for cancerous cells, and the detection model will continue to optimize its core feature extraction unit in subsequent steps.

[0018] Step 3: Use CSPDarknet53 in YOLO to replace the feature extraction backbone in the Faster R-CNN detection network and optimize the performance, such as Figure 3 As shown, the specific process is: 3.1) Replace the feature extraction backbone in the original detection network, CSPDarknet53 is an efficient backbone (existing technology) proposed in the YOLO series. It combines a dense residual structure and a cross-stage partial connection strategy. It can reduce the amount of computation and memory usage while maintaining the expressiveness of the model, while improving the accuracy of the model and using the extracted features for classification and positioning tasks.

[0019] Use CSPDarknet53 in the YOLO network as the feature extraction backbone of Faster R-CNN. The main feature extraction of CSPDarknet53 is: First, the input features are divided, and the expression is: (6) in, represents the field of real numbers, Indicates the number of channels, represents the height of the feature map, Indicates the width of the feature map; Secondly, the input features Divide into two parts on the input dimension , among which Some are passed directly through short connections, The high-level features are extracted through several layers of Bottleneck residual modules, and the expression is: (7) in, Represents the composite mapping of the Bottleneck residual module; Finally, feature fusion is performed, and the expression is: (8) Among them, Conv(⋅) represents the fused convolution operation; By replacing CSPDarknet53, the detection model of the present invention achieves faster computing speed and stronger feature representation capability, so that the detection model has both the efficiency of YOLO and the high precision of the Faster R-CNN structure.

[0020] 3.2) After replacing the feature extraction backbone in the original detection network, we introduced a composite loss function, multi-task training, and improved feature fusion methods at the training level to further improve the model's ability to identify cancerous cells. The process is as follows: 3.2.1) Calculate and apply the composite loss function, In object detection tasks, particularly cervical cell classification and localization, the use of a composite loss function effectively combines the optimization of multiple target tasks, improving model performance in classification, localization, and other areas. A composite loss function combines the losses of multiple different tasks into a single overall loss function, optimizing these tasks simultaneously during training. Common tasks for object detection include classification loss, bounding box regression loss, and optional other tasks (such as keypoint regression and occlusion detection).

[0021] During model training, this step uses a composite loss function that combines classification loss, positioning loss, and deformable convolution loss. The weight coefficients balance the impact of different tasks, ultimately achieving efficient detection and precise positioning of cervical cells.

[0022] The expression of the composite loss function is: (9) in, Represents the total loss function represents the classification loss function, represents the confidence loss function, Bounding box regression loss function,

[0023] 3.2.2) Introduce the PANet module and implement feature fusion. The Path Aggregation Network (PANet), a state-of-the-art technology, maintains the FPN structure while further integrating PANet modules to enhance the utilization of features at different levels. In implementation, the PANet module progressively transfers features from the lowest layers to the highest layers of the feature maps generated by the FPN at each scale in step 2.2. This further enhances the contextual information within the feature maps and improves the detection of small and complex objects. After PANet processing, the feature maps at each scale simultaneously contain local information at that scale, detailed information from all lower layers, and contextual information from all higher layers, truly achieving global and local fusion. Effective recognition of small and complex cervical lesions is particularly crucial, as these small and complex objects require sufficient contextual information. The core of this approach is to enhance the contextual information of feature maps through multi-path information aggregation, performing feature aggregation and fusion at each scale. Furthermore, the PANet module facilitates gradient propagation between different layers, enabling better training of parameters in previous layers.

[0024] The expression of feature fusion is: (10) in, is the fused feature map, N is the number of fused features, , It is i Input feature maps

[0025] Applying the PANet module together with the aforementioned FPN and DCN to the feature extraction backbone of the original detection network replaced by CSPDarknet53 can be seen as a comprehensive upgrade of the detection model. This not only solves the multi-scale problem, but also considers spatial deformation and uses rich context to improve decision-making, which provides a guarantee for the network model's ultimate high-precision detection.

[0026] From then on, an optimized detection model was obtained.

[0027] 3.3) Use the optimized detection model to identify cancer cell areas, Select TCT cell images containing cancer cells, use the optimized detection model to identify the cancer cell area, and perform feature extraction and target detection on the cancer cell area of ​​the input TCT cell image.

[0028] The detection model performs the same forward reasoning on the input TCT cell image as in the training phase, proposes candidate regions through RPN on feature maps of each scale, and then the RoI layer gives the final detection result.

[0029] Thanks to multiple improvements to previous related steps, the detection model of the present invention can identify cancerous cells in images with high confidence and locate their precise position.

[0030] For example, in Figure 4 In the illustrated example, the detection model of the present invention successfully selected abnormal cancerous cells within a complex background TCT cell image and labeled them as cancerous (while surrounding normal cells were classified as negative), achieving high agreement with the manually labeled ground truth. This preliminary validation demonstrates that the optimized detection model in Step 3 is capable of detecting cancerous cells in practical applications, laying the foundation for further large-scale training and deployment.

[0031] Step 4: train and deploy the detection model of the present invention. like Figure 5 As shown in the figure, after obtaining a complete detection model structure and preliminary verification, the detection model is trained on a large scale to obtain the parameters of the final detection model. The trained detection model is then put into practical application to achieve automated cell canceration detection. This stage is divided into two processes: detection model training and detection model deployment. 4.1) Training the detection model, The detection model optimized in step 3 is fully trained using the training set obtained in step 1. During the training process, each iteration includes the following operations: a batch of labeled images is input into the detection model, the detection model outputs the predicted detection box and classification results for each image in the current batch, and then the difference between the predicted results and the actual annotations is calculated to obtain the loss value. Finally, the parameters of the detection model are adjusted according to the loss value. The specific training process is: First, the parameters of the detection model (especially the backbone network) are initialized using the CSPDarknet53 weights pre-trained on ImageNet data. Correspondingly, the parameters of the FPN module, PANet module, RPN module, and RoI layer are randomly initialized. Then, iterative training begins. Each time, several images from the training set are input into the detection model. The detection model uses convolutional forward propagation to sequentially complete feature extraction, candidate box proposal, and RoI detection, obtaining the predicted cancerous cell box position, size, category, and confidence score. Based on these predictions and the corresponding manual annotations, the classification error, confidence error, and localization error are substituted into a predefined composite loss function (see step 3.2.1 for details) to calculate the total loss for the current batch. Thirdly, backpropagation and optimization algorithms (existing technologies such as Adam or SGD) are used to update the model weights of the detection model and gradually reduce the prediction error. Early stopping mechanisms and validation sets are used during training to prevent overfitting and to evaluate the performance of the detection model during training. 4.2) Deploy the trained detection model, Deploy the trained detection model into the actual cancer cell detection system to process TCT cell images in real time; After preprocessing, the TCT cell image is input into the trained detection model for processing, and the detection frame, classification label (such as normal or cancerous) and confidence level of the cancerous cell are output; Non-maximum suppression (NMS, an existing technology) is used to optimize the detection results, remove duplicate frames and filter low-confidence detection frames, and finally mark the specific location of cancerous cells on the original TCT cell image.

[0032] Example 1 Basic detection model verification experiment.

[0033] Experimental subjects or conditions: 100 images of cervical TCT cell images were provided by a tertiary hospital in Shaanxi Province. All images were scanned at a resolution of 2048×2048, and the cell regions and cancer grade were manually annotated.

[0034] Parameter setting: The original Faster R-CNN structure is used as the main framework, and the backbone network uses ResNet50. FPN and DCN are not added, and it is only used as a baseline performance comparison model.

[0035] Training time: The model was trained for 120 rounds, each round took about 6 minutes, and the total training time was about 12 hours.

[0036] Usage effect: On the test set, the mAP is 72.4% and the recall rate is 65.3%, but there are cases where small-sized cancerous cells are missed.

[0037] Conclusion: The implementation scheme of Example 1 serves as a control group for the method of the present invention, which shows that without the use of multi-scale fusion and deformable convolution, the ability to detect complex cancer cells is limited, especially the performance of subtle abnormal morphologies.

[0038] Example 2 An improved detection model with FPN structure is introduced.

[0039] Experimental subjects or conditions: The data set from the same source as Example 1 was used to ensure experimental comparability.

[0040] Parameter setting: The FPN module is integrated on the basis of the Faster R-CNN structure to fuse multi-scale features, and the backbone network still uses ResNet50.

[0041] Training time: The model was trained for 120 rounds, with each round taking about 7 minutes, and the total training time was about 14 hours.

[0042] Usage effect: mAP increased to 78.1%, recall rate increased to 72.8%, and the detection rate of small-scale cancerous cells was significantly improved.

[0043] Conclusion: The implementation scheme of Example 2 serves as a control group for the method of the present invention. It proves that after integrating the FPN module on the basis of the Faster R-CNN structure, it has a significant performance improvement effect when processing targets of different sizes (such as small cells in TCT cell images), and is suitable for early cancer cell identification.

[0044] Example 3 DCN is introduced to enhance cell boundary feature extraction.

[0045] Experimental subjects or conditions: 300 high-complexity images containing overlapping cells, fuzzy boundaries, and irregularly shaped cells.

[0046] Parameter setting: Based on Example 2, a DCN module is further added and applied to the high-level feature map output by FPN.

[0047] Training time: The training time is extended to 15 hours (due to the increased computational complexity of DCN), and 120 rounds of training are maintained.

[0048] Usage effect: The model's mAP reached 82.7%, the recall rate increased to 76.5%, and the detection ability of cells with blurred boundaries was significantly improved.

[0049] Conclusion: The implementation scheme of Example 3 serves as a control group of the method of the present invention. Based on Example 2, DCN in Example 3 enhances the modeling capability of complex cell edges and improves the robustness of detection of irregular cancerous cells.

[0050] Example 4 Replace the ResNet backbone with CSPDarknet53.

[0051] Experimental subjects: Run on a public set of cervical cancer cell images (such as the Herlev Dataset, a total of 917 samples), containing various levels of cancer.

[0052] Parameter setting: Replace the original detection network backbone with CSPDarknet53, and retain the FPN and DCN structures.

[0053] Training time: 120 rounds of training, 5 minutes per round, a total of about 9 hours, which is faster than the previous structure.

[0054] Usage effect: mAP reaches 84.9%, recall rate is 79.8%, and detection time is shortened by about 28% compared with the ResNet model, making it more suitable for large-scale screening tasks.

[0055] Conclusion: The implementation scheme of Example 4 serves as a control group for the method of the present invention. The CSPDarknet53 structure adopted in Example 4 significantly improves the processing speed while ensuring detection accuracy, and is a more efficient backbone choice.

[0056] Example 5 PANet is introduced to improve context feature fusion.

[0057] Experimental subjects or usage conditions: 150 smear images were collected, of which 50 were mild lesions, 50 were high-grade lesions, and 50 were normal controls.

[0058] Parameter setting: The CSPDarknet53+FPN+DCN combination is adopted, and PANet is integrated on this basis to enhance context information and obtain a complete detection model.

[0059] Usage time: 120 rounds of training, which took about 9.5 hours.

[0060] Effect: mAP increased to 87.3%, recall rate reached 82.1%, and classification accuracy was higher among various types of lesions.

[0061] Conclusion: In Example 5, the CSPDarknet53+FPN+DCN+PANet approach was adopted. The complete detection model of the method of the present invention effectively improved the recognition accuracy under complex backgrounds, enhanced the modeling of the relationship between cells, and improved the credibility of abnormal cell identification.

[0062] Example 6 This is the complete detection model deployment and clinical simulation use of the present invention.

[0063] Experimental subjects or usage conditions: The complete detection model of the present invention was integrated into the TCT film reading auxiliary system of a tertiary hospital to perform batch detection on actual clinical smear images.

[0064] Parameter setting: The deployed version is the final integrated complete detection model: CSPDarknet53+FPN+DCN+PANet, integrating non-maximum suppression and interface display modules.

[0065] Processing time: It takes an average of 0.8 seconds to process each image, and about 13 minutes to process 1,000 images.

[0066] Effect of use: The average accuracy rate reached 90.4%, the time doctors spent reading films was reduced by nearly 60%, the misdiagnosis rate was significantly reduced, and efficiency was improved.

[0067] Conclusion: In this Example 6, the complete detection model of the present invention performs well in actual usage scenarios, can significantly improve the efficiency of cervical cancer screening, and has practical promotion value.

Claims

1. A method for identifying cancerous cells in TCT detection mode, characterized in that: Follow these steps to implement: Step 1: Create a data set and complete the organization; Step 2: Build a detection model and design a deep learning network architecture for cell detection; Step 3: Use CSPDarknet53 in YOLO to replace the feature extraction backbone in the Faster R-CNN detection network and optimize the performance; Step 4: Train and deploy the detection model. This stage is divided into two processes: detection model training and detection model deployment. After obtaining the complete detection model structure and preliminary verification, the detection model is trained on a large scale to obtain the parameters of the final detection model, and the trained detection model is put into practical application to realize the automation of cell canceration detection.

2. The method for identifying cancerous cells under TCT detection mode according to claim 1, characterized in that: In step 1, the specific process is: 1.1) Data acquisition: Acquire a large number of TCT cell images to ensure that the sample contains normal cells and various grades of cancerous cells; 1.2) Sample labeling: Manually label the collected TCT cell images, including selecting the area where the cancerous cells are located and indicating the cell type; 1.3) Divide the dataset: Divide the labeled dataset into training set, validation set and test set.

3. The method for identifying cancerous cells under TCT detection mode according to claim 2, characterized in that: In step 1.3), the data set is divided into the following steps: the training set accounts for 80% of the total data and is used for model training and parameter optimization; the validation set accounts for 10% and is used to monitor model performance and prevent overfitting during training; and the test set accounts for 10% and is used to finally evaluate the recognition accuracy and generalization ability of the model.

4. The method for identifying cancerous cells under TCT detection mode according to claim 1, characterized in that: In step 2, the specific process is: 2.1) Improve the existing Faster R-CNN detection model as the main framework, First, the input cervical cell image undergoes a series of convolution and downsampling operations to extract a multi-channel feature map. Then, an RPN is used to slide on the feature map to generate a large number of anchor boxes. A lightweight binary classification network is used to determine whether each anchor box contains the target cell. The position and size of the anchor box are regressed to propose candidate detection regions. Finally, for the selected candidate regions, the corresponding regional features are intercepted from the feature map at the RoI layer and classified and fine-tuned position regression is performed to output a preliminary detection result, which includes the predicted box position and category. 2.2) Introducing FPN into the main framework, Based on the multi-scale features extracted by the original convolutional backbone, a top-down path is established to upsample high-level features step by step, and then horizontally connect and fuse them with the underlying features of the corresponding scale; 2.3) Integrate the DCN module into the main frame.

5. The method for identifying cancerous cells under TCT detection mode according to claim 1, characterized in that: In step 3, the specific process is: 3.1) Replace the feature extraction backbone in the original detection network; 3.2) Introducing composite loss functions, multi-task training, and improved feature fusion methods at the training level; 3.3) Using the optimized detection model to identify cancer cell areas, Select TCT cell images containing cancer cells, use the optimized detection model to identify the cancer cell area, and perform feature extraction and target detection on the cancer cell area of ​​the input TCT cell image.

6. The method for identifying cancerous cells under TCT detection mode according to claim 5, characterized in that: In step 3.2), the specific process is: 3.2.1) Calculate and apply the composite loss function, During model training, a composite loss function is used, which combines classification loss, positioning loss and deformable convolution loss. The expression of the composite loss function is: (9) in, Represents the total loss function represents the classification loss function, represents the confidence loss function, Bounding box regression loss function, 3.2.2) Introduce the PANet module and implement feature fusion. The expression of feature fusion is: (10) in, is the fused feature map, N is the number of fused features, , It is i Input feature maps.

7. The method for identifying cancerous cells in TCT detection mode according to claim 1, characterized in that: In step 4, the detection model is trained. The specific process is: First, the parameters of the detection model are initialized using the CSPDarknet53 weights pre-trained on ImageNet data. The parameters of the FPN module, PANet module, RPN module, and RoI layer are randomly initialized. Then, iterative training begins. Each time, several images from the training set are input into the detection model. The detection model uses convolutional forward propagation to sequentially complete feature extraction, candidate box proposal, and RoI detection, obtaining the predicted cancerous cell box position, size, category, and confidence level. Based on these predictions and the corresponding manual annotations, the classification error, confidence error, and localization error are substituted into a predefined composite loss function to calculate the total loss for the current batch. Thirdly, back-propagation and optimization algorithms are used to update the model weights of the detection model and gradually reduce the prediction error; Use early stopping and validation sets during training to prevent overfitting and evaluate the performance of the model during training.

8. The method for identifying cancerous cells under TCT detection mode according to claim 1, characterized in that: In step 4, deploy the trained detection model. The specific process is: After preprocessing, the TCT cell images are input into the trained detection model for inference, which outputs the detection frame, classification label and confidence of the cancerous cells. Non-maximum suppression is used to optimize the detection results, remove duplicate frames and filter low-confidence detection frames, and finally mark the specific location of the cancerous cells on the original TCT cell images.

Images