Deep learning based stromal mitotic cell recognition and counting method

Abstract

The present application provides a stromal tumor mitotic cell recognition and counting method based on deep learning, comprising: acquiring and preprocessing the stained gastrointestinal stromal tumor tissue section image; the preprocessed section image is cut; the generated image block is expanded by using data enhancement technology; a target detection model based on YOLOv10 is constructed; a multi-source domain adaptation technology is applied; the recognized mitotic cells are counted, and corresponding statistical data are generated; the prediction results of the model are displayed on the functional platform, and the labeled results are verified and corrected by doctors; the model is retrained by using the data corrected by doctors. Through automatic detection and statistics, the model can generate detailed cell density distribution map and heat map, helping doctors better understand the proliferation pattern and regional activity of the tumor. These data provide a scientific basis for doctors' diagnosis and treatment plan, and promote the implementation of personalized treatment.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics, specifically to a method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning. Background Technology

[0002] In the medical field, gastrointestinal stromal tumors (GISTs) are relatively common soft tissue tumors with a high degree of malignancy. Early diagnosis and treatment are crucial for improving patient survival rates. When performing pathological analysis on GISTs, identifying and counting mitotic cells in the tumor tissue is an important basis for assessing tumor proliferative activity and determining the degree of malignancy.

[0003] Currently, the identification and counting of mitotic cells in clinical practice mainly relies on manual operation by pathologists. Doctors observe stained tissue sections under a microscope and manually label and count mitotic cells. This method demands a high level of professional experience from the doctor, and the process is tedious, time-consuming, and labor-intensive. Furthermore, the manual identification process is limited by the doctor's energy and attention; prolonged repetitive work can easily lead to fatigue, thus affecting the accuracy and efficiency of the identification.

[0004] Although deep learning models have demonstrated superior performance in image recognition with the development of computer vision technology and are gradually being introduced into medical image analysis, some studies have attempted to use deep learning models to automatically identify mitotic cells in tissue sections. However, existing automated identification systems still have some problems:

[0005] First, existing deep learning models are typically trained on specific types of pathological slides, exhibiting insufficient generalization ability when dealing with slides of different types and staining conditions. This makes the models unable to handle complex and diverse pathological images in actual clinical applications, easily leading to missed or false positives.

[0006] Secondly, because mitotic cells constitute a very small proportion of pathological sections and are often mixed with surrounding normal cells, existing models perform poorly in identifying these small targets. Convolutional neural networks (CNNs) are easily limited by convolution operations when processing small targets and cannot accurately detect all mitotic cells.

[0007] Furthermore, existing systems lack automated error correction mechanisms. Even within model predictions, manual verification and correction by doctors are still required. However, the lack of effective error detection and correction mechanisms in existing systems means doctors need to spend a significant amount of time checking each prediction individually, which is almost indistinguishable from manual operation, resulting in limited overall efficiency improvements.

[0008] These problems lead to consequences such as low identification accuracy, heavy workload for doctors, and limited clinical application. Due to insufficient generalization ability of the models and poor recognition of small targets, existing systems often cannot achieve the accuracy of manual operation by doctors in actual clinical applications. False positives or missed positives occur frequently, which may directly affect the assessment of tumor proliferative activity, leading to misdiagnosis or delayed treatment. Although existing systems can initially screen out some mitotic cells, doctors still need to perform a large amount of manual verification and correction, resulting in no significant reduction in their workload. Furthermore, repetitive operations can easily lead to fatigue, further increasing the risk of errors. Due to issues with identification accuracy and efficiency, the promotion of existing automated systems in actual clinical applications is limited. Doctors and hospitals have low trust in them, and the systems cannot fully play their due role, failing to achieve true automated diagnostic assistance and affecting the overall efficiency and diagnostic quality of pathological analysis. Summary of the Invention

[0009] To overcome the shortcomings of existing technologies, this invention proposes a deep learning-based method for identifying and counting mitotic cells in gastrointestinal stromal tumors. By combining deep learning models and various optimization techniques, it achieves efficient and accurate identification of mitotic cells in gastrointestinal stromal tumors, which not only greatly improves work efficiency and reduces the possibility of mislabeling, but also provides solid technical support for clinical diagnosis and treatment.

[0010] To achieve the above objectives, this invention provides a deep learning-based method for identifying and counting mitotic cells in stromal tumors, comprising the following steps:

[0011] Step 1: Acquire and preprocess stained gastrointestinal stromal tumor tissue slice images. Preprocessing includes color normalization, noise reduction, and contrast enhancement to ensure image quality and consistency, and to eliminate human variations during the staining process. The preprocessing step also includes converting the images to grayscale to simplify the model's input format.

[0012] Step 2: The preprocessed sliced ​​image is segmented to generate image patches suitable for deep learning model processing. The segmentation operation uses a step size of 256 pixels to divide the 512×512 pixel slice into multiple smaller sub-image patches to ensure that the target cells are completely preserved in different image patches and to reduce the impact of edge effects on the model's recognition accuracy. Image patches that do not contain target cells or contain very few cells are further filtered out using RGB pixel values.

[0013] Step 3: Expand the generated image patches using data augmentation techniques, including but not limited to color adjustments (such as hue, saturation, sharpness, and contrast), random rotation, translation, scaling, cropping, affine transformation, and copy-paste techniques to randomly copy target cells to other image locations. This data augmentation step aims to simulate various image variations that may occur in real-world scenarios, thereby improving the robustness and generalization ability of the deep learning model.

[0014] Step 4: Construct a YOLOv10-based object detection model to identify and locate mitotic cells in images. The YOLOv10 model includes several innovative modules, including a C2fUIB module, a Partial Self-Attention (PSA) module, a Spatial-Channel Decoupled Downsampling (SCDown) module, and an improved Spatial Pyramid Pooling for Fast interface (SPPF) module. The model achieves NMS-free (Non-Maximum Suppression) training through a dual allocation strategy, significantly improving the model's inference efficiency and accuracy.

[0015] Step 5: Apply multi-source domain adaptation technology to enhance the model's generalization ability by introducing data from multiple source domains. This technology involves using multiple target detection heads during source domain training, each corresponding to specific source domain data, and utilizing a gradient inversion module to generate domain adversarial features, thereby mitigating mutual interference between source domains. Subsequently, a student model is trained on the target domain, and pseudo-labels are generated using a teacher model to further improve detection accuracy in the target domain.

[0016] Step 6: Count the identified mitotic cells and generate corresponding statistical data. This statistical data includes the total number of mitotic cells in each slide, cell density distribution, confidence ranking results, and the distribution pattern of mitotic cells in different pathological slides. The statistical results can be directly used to assist doctors in assessing the condition and formulating treatment plans.

[0017] Step 7: Display the model's prediction results on the functional platform, where doctors can verify and correct the annotations. The platform includes an intuitive user interface, allowing doctors to view the specific location and annotation results of each cell identified by the model. Doctors can correct annotations through click and drag operations. The platform also supports batch operations for doctors to improve work efficiency.

[0018] Step 8: Retrain the model using the data corrected by the doctors to continuously optimize the model's recognition and counting performance. The retraining step includes dividing the corrected dataset into multiple cross-validation sets, adjusting model hyperparameters (such as learning rate, batch size, number of training epochs, etc.), and introducing new data augmentation strategies to further improve the model's stability and accuracy in practical applications.

[0019] Furthermore, the data augmentation techniques applied in step 3 include the following operations:

[0020] Step 3.1: Color adjustment, including adjusting the hue, saturation, sharpness and contrast of the image to simulate the image changes under different staining conditions, and to ensure that the model can effectively identify target cells under different staining effects;

[0021] Step 3.2: Affine transformation, including random rotation (0°, 90°, 180°, 270°), translation, scaling, and cropping operations, to simulate possible changes in slicing and shooting angles in actual operation;

[0022] Step 3.3: The copy-paste operation randomly copies the identified target cell regions to other locations in the image, thereby increasing the diversity of target cells in the training data and enhancing the robustness of the model;

[0023] Step 3.4: Data denoising and background removal, eliminating pure white or almost cell-free areas to reduce irrelevant information from interfering with the model.

[0024] Furthermore, the innovative modules of the YOLOv10 model used in step 4 include:

[0025] Step 4.1: The C2fUIB module increases the receptive field of the convolutional kernel and reduces computational cost by employing reparameterization techniques, thereby improving the detection accuracy and inference speed of the model.

[0026] Step 4.2: Partial Self-Attention Module (PSA) is used to divide the feature map into two parts. One part is processed by the Multi-Head Self-Attention Module (MHSA), while the other part retains its original features. The two parts are then fused to enhance the model's ability to detect small targets.

[0027] Step 4.3: Spatial Channel Decoupling Downsampling Module (SCDown) reduces computational cost and the number of parameters by decoupling spatial reduction and channel addition operations, while preserving information to the maximum extent and enhancing the ability to identify small targets during downsampling.

[0028] Step 4.4: Improved Spatial Pyramid Pooling Fast Interface (SPPF) reduces the number of parameters while maintaining the ability to aggregate multi-scale features by changing the parallel structure to a serial structure, thereby improving the model's ability to detect targets at different scales.

[0029] Furthermore, the multi-source domain adaptation techniques in step 5 include:

[0030] Step 5.1: Source domain training process. Multiple source domains correspond one-to-one with multiple target detection heads. The feature encoders share the data. During backpropagation, domain adversarial features are generated through the gradient backpropagation module, thereby reducing mutual interference between source domains and optimizing the adaptive performance of the target domain.

[0031] Step 5.2: Target domain training process. The pseudo-labels generated by the teacher model are used to train the student model in the target domain. By gradually updating the weights of the student model, its detection accuracy in the target domain is enhanced.

[0032] Step 5.3: Domain adaptation effect evaluation. The t-SNE dimensionality reduction technique is used to visualize the features of the source domain and the target domain, verify the model's adaptability to the feature distribution between different domains, and evaluate the actual effect of the multi-source domain adaptation technique by comparing it with the baseline model.

[0033] Furthermore, the statistical data in step 6 further includes:

[0034] Step 6.1: Detailed location distribution map of mitotic cells in each pathological section. This distribution map forms a two-dimensional coordinate system by marking the positions of the detected mitotic cells in the original section image, displaying the precise coordinate position of each cell. This distribution map can intuitively show the spatial distribution of mitotic cells in tumor tissue, helping doctors better understand the tumor's proliferation pattern and regional activity.

[0035] Step 6.2: A heatmap is generated based on cell density. This heatmap calculates the density of mitotic cells in the slice and displays the cell density in different regions using a color gradient. The intensity of the color corresponds to the density of the cells. The heatmap helps doctors quickly identify the most actively proliferating areas in a tumor, providing important information for pathological diagnosis and treatment planning.

[0036] Step 6.3: A list of mitotic cells sorted by confidence level. This list is based on the confidence value of each cell detected by the model. The confidence value indicates the model's certainty that the cell is a mitotic cell. Doctors can prioritize verifying and correcting identification results with lower confidence levels. The list should include the confidence score, location information, and related image patches for each cell to facilitate quick location and identification by doctors.

[0037] Step 6.4: Calculate the total number and average density of mitotic cells in different pathological sections. The statistical data should include the total number of cells per section, the cell density per square millimeter, and the local density calculated for different tumor regions. These statistical results can be used to compare the tumor proliferation characteristics of different patients or different sections, providing reference data for personalized treatment.

[0038] Step 6.5: Generate a summary report containing multiple patient slides. This report integrates all test results, including the total number of cells, average density, distribution of active areas, and confidence analysis for each patient, forming a complete test report to help doctors make overall assessments and comparisons.

[0039] Furthermore, the retraining process in step 8 further includes:

[0040] Step 7.1: Collection and organization of corrected data. This data is generated by doctors after correcting the model's predictions on the functional platform. The data includes corrected cell annotations, confidence adjustments, and information on the removal of error detections. The organization process includes labeling the corrected data, performing error analysis, and merging the corrected data into a new training set.

[0041] Step 7.2: Training set partitioning and preparation. The collected corrected data is divided into multiple cross-validation sets according to a certain ratio (e.g., 70% training set, 15% validation set, 15% test set) to ensure the model's generalization ability on different datasets. Training set preparation also includes secondary data augmentation to ensure data diversity and randomness.

[0042] Step 7.3: Optimization and tuning of model hyperparameters. Before each retraining, the model's hyperparameters (such as learning rate, batch size, number of training epochs, etc.) are optimized and tuned based on the scale and characteristics of the corrected data. Hyperparameter tuning tools such as Optuna are used to automatically search within the set range, and the optimal hyperparameter combination is determined through multiple rounds of experiments.

[0043] Step 7.4: Execution of the retraining process, which includes inputting the corrected training data into the model for retraining, while monitoring the model's accuracy and loss value through the validation set, and employing an early stopping strategy (stopping training when the validation set accuracy no longer improves) to avoid overfitting. After training, the model's final weights will be saved and used for the next detection task.

[0044] Step 7.5: Model evaluation after retraining. This evaluation includes a comprehensive analysis of the model's performance on the test set, with particular attention to improvements in performance on corrected data, including metrics such as recall, precision, and F1 score. The effectiveness of retraining is assessed by comparing the model with the initial model and other baseline models. The evaluation results are then fed back to the development and application teams to determine whether further optimization is necessary.

[0045] Step 7.6: Model Deployment and Updates. Deploy the retrained and evaluated model to the functional platform and update it in real time to the system used by doctors. Ensure the new model can automatically replace the old model and guarantee its stability and accuracy in actual use. Simultaneously, record feedback on the new model's usage for future retraining reference.

[0046] Furthermore, the model's recognition accuracy and robustness are further optimized through multi-scale feature aggregation and model augmentation training, the steps of which include:

[0047] Step 8.1: Extract multi-scale features. Input the preprocessed image patch into the deep learning model and extract feature maps of different scales through specific network layers to ensure that the model can capture multiple morphological features of cells.

[0048] Step 8.2: Perform feature fusion. The extracted features at different scales are fused together. The spatial pyramid pooling module or its improved version is used to aggregate multi-scale information to enhance the model's ability to recognize small targets.

[0049] Step 8.3: Introduce model augmentation training techniques. By combining loss functions and data augmentation strategies, enhance the model's robustness to complex samples and noise, and improve recognition accuracy.

[0050] Step 8.4: Using transfer learning techniques, the pre-trained model trained in the relevant domain is used as a basis to improve the model's adaptability to the target task by fine-tuning it in the target domain.

[0051] Furthermore, it also includes an automated error detection and correction mechanism, the steps of which include:

[0052] Step 9.1: Automatic verification of model prediction results. The system performs preliminary verification of the mitotic cells detected by the model and identifies labels with low confidence or possible errors.

[0053] Step 9.2: Generate an error detection report. The system will generate an error report for the low-confidence or suspected error labels, which will list the cell locations that may have problems and their corresponding confidence values ​​in detail.

[0054] Step 9.3: Manual correction and feedback. Doctors manually correct the annotations in the error reports, confirm or correct the model's detection results, and feed the corrected data back to the system.

[0055] Step 9.4: Dynamically adjust model weights. The system automatically adjusts the model weights and parameters based on the doctor's calibration results to improve the model's accuracy in subsequent tasks.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] 1. This invention provides a deep learning-based method for identifying and counting mitotic cells in stromal tumors. By shifting the doctor's workload from tedious manual identification to correcting the model's predictions, doctors no longer need to search for mitotic cells one by one on complete pathological slides. This significantly saves doctors' time and effort, improving overall work efficiency. During model retraining, the model is optimized using the corrected data from doctors, continuously improving its recognition accuracy. As the model's recognition capabilities increase, its stability and robustness on different pathological slides are effectively guaranteed. The model can adapt to diverse pathological slide data while maintaining high accuracy.

[0058] 2. This invention provides a deep learning-based method for identifying and counting mitotic cells in stromal tumors. Because the model can accurately identify the vast majority of mitotic cells during detection, doctors only need to correct the model's predictions. This correction process only targets parts of the model's annotation with low confidence or potential errors, significantly reducing the possibility of mislabeling. In this way, the model's performance can be continuously optimized, and the error rate can be gradually reduced.

[0059] 3. This invention provides a deep learning-based method for identifying and counting mitotic cells in stromal tumors. In actual testing, the model can predict hundreds of thousands of image patches from multiple patients in a very short time. Through automated calculation and processing, the model summarizes the detection results according to confidence levels and generates accurate statistical data on mitotic cells. Compared with traditional manual operation, the model demonstrates significant advantages in processing large-scale pathological data, greatly shortening the detection time.

[0060] 4. This invention provides a deep learning-based method for identifying and counting mitotic cells in stromal tumors. The model's detection results are comparable to those obtained manually by doctors, demonstrating its ability to effectively identify mitotic cells. The model's prediction accuracy is highly close to that of manual annotation, providing reliable technical support for pathological analysis and clinical diagnosis. Through automated detection and statistics, the model can generate detailed cell density distribution maps and heat maps, helping doctors better understand tumor proliferation patterns and regional activity. This data provides a scientific basis for doctors' diagnosis and treatment planning, promoting the implementation of personalized treatment. Attached Figure Description

[0061] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0062] Figure 1 This is a structural diagram of the YOLOv10 invention.

[0063] Figure 2 This is a schematic diagram of the SPPF structure of the present invention.

[0064] Figure 3 This invention relates to a multi-source domain adaptation method.

[0065] Figure 4-5 This is a schematic diagram of the feature domain migration of the present invention.

[0066] Figure 6 This is a schematic diagram of data enhancement for the present invention.

[0067] Figure 7 This is a schematic diagram of file upload for the present invention.

[0068] Figure 8 This is a schematic diagram of the lifecycle labeling of the present invention.

[0069] Figure 9 This is a schematic diagram of the document download for this invention.

[0070] Figure 10 This is a schematic diagram of an embodiment of the present invention. Detailed Implementation

[0071] The technical solution of the present invention will be more clearly and completely explained below with reference to the accompanying drawings and through the description of preferred embodiments of the present invention.

[0072] Terminology Explanation:

[0073] Explanation of abbreviations:

[0074] GIST (Gastrointestinal Stromal Tumor): A common soft tissue tumor.

[0075] CNN (Convolutional Neural Network): A deep learning model architecture used for image recognition and classification.

[0076] YOLOv10 (You Only Look Once, version 10): An advanced target detection model characterized by high accuracy and high speed.

[0077] C2fUIB: A module in the YOLOv10 model used to replace the Bottleneck structure in YOLOv8. It employs a complex convolutional kernel design to increase the receptive field and reduce computational cost.

[0078] PSA (Partial Self-Attention): This module enhances small object detection capabilities by segmenting the feature map and processing a portion of it using a multi-head self-attention module.

[0079] SCDown (Spatial: Channel Decoupled Downsampling): This module decouples spatial channel downsampling operations to reduce computational costs and the number of parameters while retaining information.

[0080] SPP (Spatial Pyramid Pooling): Spatial pyramid pooling is a technique that aggregates multi-scale feature information to enhance the model's ability to detect targets at different scales.

[0081] SPPF (Spatial Pyramid Pooling for Fast interface): An improved spatial pyramid pooling module that changes the parallel structure to a serial structure, reducing the number of parameters while maintaining the ability to aggregate features at multiple scales.

[0082] NMS (Non: Maximum Suppression): Non-maximum suppression is a post-processing step used in object detection algorithms to reduce duplicate detection boxes.

[0083] MHSA (Multi-Head Self-Attention): Multi-head self-attention is often used to enhance the contextual information processing capabilities of neural network models.

[0084] EMA (Exponential Moving Average): Used to smooth the updates of time series data or model weights.

[0085] T-SNE (t-Distributed Stochastic Neighbor Embedding): t-distributed stochastic neighbor embedding is a technique for dimensionality reduction and visualization of high-dimensional data.

[0086] TPU (Tensor Processing Unit): A hardware device developed by Google to accelerate the training and inference of machine learning models.

[0087] GPU (Graphics Processing Unit): A graphics processing unit typically used to accelerate image processing and the training and inference of deep learning models.

[0088] Optuna: A Python library for hyperparameter optimization that improves model performance by automatically searching for the optimal combination of hyperparameters.

[0089] FLOPs (Floating Point Operations per Second): The number of floating-point operations per second, used to measure the computational complexity of a model.

[0090] KL Divergence (Kullback: Leibler Divergence): KL divergence is a statistic that measures the difference between two probability distributions.

[0091] MMD (Maximum Mean Discrepancy): The maximum mean discrepancy is a method for measuring the difference between different distributions, and it is often used in domain adaptation and transfer learning.

[0092] TF: IDF (Term Frequency: Inverse Document Frequency): A statistical method used in information retrieval and text mining to assess the importance of words in a document.

[0093] JSON (JavaScript Object Notation): A lightweight data exchange format commonly used for network data transmission and storage.

[0094] Labelme: An open-source tool for image annotation, often used for creating deep learning datasets.

[0095] SVS (Scalable Vector Graphics): Scalable Vector Graphics is an XML-based vector graphics format typically used for storing and displaying high-resolution images.

[0096] Specifically, the present invention is as follows:

[0097] 1. Image preprocessing

[0098] 1.1 Slice Acquisition and Image Preprocessing

[0099] Starting with tissue sections from a patient's gastrointestinal stromal tumor, staining is first performed to ensure that mitotic cells can be clearly identified in the microscopic images. The stained tissue sections are then scanned to generate high-resolution digital images. These images are typically very large, and processing the raw images is inefficient; furthermore, noise and uneven staining in the images can interfere with the model's recognition accuracy.

[0100] 1.2 Image Segmentation

[0101] Large images are divided into 512×512 pixel patches with a step size of 256 pixels. This segmentation method helps maintain cell integrity and facilitates the model's recognition of cell morphological features. The segmented patches are then preliminarily screened to remove those with almost no cells or large edges, reducing the computational burden on the model when processing invalid data.

[0102] 1.3 Data Augmentation

[0103] Data augmentation is performed on the selected image patches, such as... Figure 6 As shown, data augmentation methods are used to improve the robustness and generalization ability of the model. These methods include:

[0104] Color adjustment: Randomly adjusts the hue, saturation, sharpness, and contrast of an image to simulate changes in the image under different coloring conditions.

[0105] Affine transformations: Randomly rotate (0°, 90°, 180°, 270°), translate, scale, and crop the image to enhance the model's ability to handle geometric transformations of the image.

[0106] Copy-paste technique: randomly copies the identified target cells to other locations in the image, increasing the diversity of target cells in the training data.

[0107] 2. Model Building and Training

[0108] 2.1 Selection of backbone model

[0109] YOLOv10 was chosen as the base detection model due to its high detection speed and accuracy. The structure of the YOLOv10 model has been optimized and includes the following key modules:

[0110] The C2fUIB module replaces the Bottleneck structure in YOLOv8 with a more complex convolutional kernel design to increase the receptive field and reduce computational cost.

[0111] Partial Self-Attention Module (PSA): The feature map is divided into two parts. One part is processed by a multi-head self-attention module (MHSA), while the other part retains the original features. The two parts are then fused to enhance the detection capability of small targets.

[0112] Spatial Channel Decoupled Downsampling Module (SCDown): By decoupling spatial reduction and channel addition operations, it reduces computational cost and the number of parameters while retaining information, thereby improving the small target detection performance during the downsampling process.

[0113] 2.2 Model Training

[0114] During training, a multi-source domain adaptation technique is employed to enhance the model's generalization ability. Specifically:

[0115] Source domain training: The model is trained on multiple source domains, with each source domain data corresponding to an independent target detection head, in order to reduce mutual interference between source domains.

[0116] Target domain training: Pseudo-labels are generated by the teacher model and the student model is trained on the target domain. The model weights are gradually adjusted to improve the detection accuracy on the target domain.

[0117] 2.3 Model Optimization

[0118] Model augmentation training techniques, including the use of hybrid loss functions and data augmentation strategies, were employed to enhance the model's robustness to complex samples and noise. Building upon this, transfer learning techniques were used to fine-tune the model's performance in the gastrointestinal stromal tumor mitotic cell identification task by fine-tuning a pre-trained model in a relevant domain.

[0119] 3. Identification and Counting of Mitotic Cells

[0120] 3.1 Image Patch Input and Recognition

[0121] The preprocessed image patch is input into the trained YOLOv10 model. The model identifies mitotic cells in the image through multi-scale feature aggregation and generates corresponding detection results, including the position coordinates and confidence score of each cell.

[0122] 3.2 Counting and Statistics

[0123] After the model identifies all mitotic cells, the system automatically counts these cells and generates detailed statistical data. The statistical data includes:

[0124] Total number of cells in each slice

[0125] Cell density distribution in different regions

[0126] Detection results ranked by confidence level

[0127] Generate a heatmap for each slice to show the density and distribution of cells.

[0128] 4. Functional Platform and Doctor Feedback

[0129] 4.1 Functional Platform Display and Doctor Correction

[0130] The model's predictions are displayed on the functional platform, allowing doctors to view the test results for each slice and manually verify and correct the annotations. The platform provides an intuitive user interface, enabling doctors to click and drag to correct annotations.

[0131] 4.2 Feedback Data Collection and Retraining

[0132] After the doctor completes the annotation correction, the system automatically collects and aggregates the corrected data to generate a new training dataset. The system then automatically initiates a retraining process, adjusting the model's weights and parameters based on the doctor's feedback to improve the model's accuracy and stability.

[0133] 4.3 Model Update and Deployment

[0134] After retraining, the system generates a new model and automatically pushes it to the doctor's end, replacing the old model. The system also backs up the old model so that it can be quickly restored to the previous version if the new model performs poorly.

[0135] 5. Automated error detection and correction

[0136] 5.1 Automatic Validation and Error Report Generation

[0137] The system automatically verifies each model prediction result, identifies labels with low confidence or potential errors, and generates a detailed error detection report, listing the suspected erroneous cell locations and their corresponding confidence values.

[0138] 5.2 Manual Correction and Dynamic Adjustment

[0139] Doctors manually correct the annotations based on error reports and feed the corrected data back to the system. The system automatically adjusts the model's weights and parameters based on the correction results, further improving the model's accuracy.

[0140] This invention can effectively realize the automated identification and counting of mitotic cells in gastrointestinal stromal tumors, significantly improving the efficiency and accuracy of detection, and providing important technical support for pathological diagnosis and treatment.

[0141] As one specific implementation method,

[0142] 1. Pre-processing of the image

[0143] The patient's stained biological tissue sections have the following two problems:

[0144] Compared to the input size of a typical depth model, the slice size is too large.

[0145] To make the images suitable for deep learning model analysis, they are segmented. To reduce the potential loss of target cells (severance) during segmentation, the images are divided into 512×512 patches with a stride of 256. Larger images facilitate doctors' observation and annotation of target cells. Labelme is used for annotation, with target cells outlined using rectangular boxes. Furthermore, to improve the recall rate of target recognition, the 512x512 patches are further divided into 256x256 sub-patterns (still with a stride), and only the sub-patterns containing target annotations are retained.

[0146] Patches containing very few or no cells are discarded because these patches, representing the edges of the tissue slice image, often appear pure white in the final SVS file. A patch with very few cells is converted to grayscale, and the pixel value is calculated by dividing by the number of pixels to obtain a threshold. Other patches are processed in the same way. Therefore, thresholding the average RGB pixel values ​​of the slice removes portions with no or very few cells. After examination by a doctor, it was found that the cells separated at the cut boundaries were very few of the desired mitotic cells, so this processing can be considered not to affect the final result.

[0147] 2) Different degrees of dyeing result in different background colors.

[0148] To avoid potential interference from different backgrounds on cell recognition, image preprocessing operations other than segmentation were performed, including copy-paste (randomly copying the target to other parts of the image to generate more data), color adjustments (hue, saturation, sharpness, and contrast), and minor affine transformations (random cropping, rotation, etc., with rotation mainly including clockwise 90°, 180°, and 270°). The effects are shown above. Figure 2 As shown, the goal is to minimize background discrepancies.

[0149] The overall style (size, color) of cell section images is relatively uniform, but differences in imaging equipment and staining procedures can lead to slight perturbations (such as changes in cell size scale, staining degree, and imaging clarity). Therefore, the data augmentation methods described above are good data augmentations for these changes, while not disrupting the morphological characteristics of cells, and thus do not disrupt the meaning of the images.

[0150] Detection of dividing cells

[0151] Dataset: The stromal tumor mitotic cell dataset (hereinafter referred to as the stromal tumor dataset) obtained after the above processing consists of tissue sections from a total of 14 patients, which are divided into 4252 sections containing the target regions, with a total of 5409 target regions. The training, validation and test sets are divided in a 3:1:1 ratio.

[0152] Cell detection:

[0153] Backbone Model: The model employs YOLOv10, the most mature industrial-grade detection network currently available, which has achieved a new level of performance in both accuracy and speed. In this work, the authors first propose a continuous dual assignment method for NMS-free training of YOLOs, which delivers competitive performance and low inference latency. Furthermore, they introduce a comprehensive efficiency-accuracy driven model design strategy for YOLOS. All components of YOLOS are comprehensively optimized from both efficiency and accuracy perspectives, significantly reducing computational overhead and enhancing performance. The specific model is as follows: Figure 1 As shown:

[0154] C2fUIB simply replaces the Bottleneck structure of C2f in YOLOv8 with the CIB structure. The Bottleneck structure is a common convolutional network structure since ResNet and will not be discussed here. The CIB structure similarly adopts a 5-layer network of 3×3, 1×1, 3×3 (or 7×7 if using the RepVGGDW convolutional structure, which uses reparameterization to increase the convolutional kernel to increase the receptive field of the convolutional network while reducing computational cost and optimization issues), 1×1, and 3×3.

[0155] Partial self-attention (PSA): After a 1×1 convolution, the features are uniformly divided into two parts. Only one part is fed into an NPSA block consisting of a multi-head self-attention module (MHSA) and a feedforward network (FFN). The two parts are then connected and fused via a 1×1 convolution. Furthermore, the dimensions of the query and key are assigned to half that of the values, and LayerNorm is replaced with BatchNorm for fast inference.

[0156] Spatial-channel decoupled downsampling (SCDown): Conventional downsampling uses a standard 3×3 convolution with a stride of 2, simultaneously performing spatial downsampling (from H×W to H / 2×W / 2) and channel transformation (from C to 2C). This introduces a non-negligible computational cost of O(9HWC^2) and a parameter count of O(18C^2). V10 decouples the spatial reduction and channel augmentation operations for more efficient downsampling. Specifically, it first uses point-to-point convolution to adjust the channel dimensions, and then uses depthwise separable convolution for spatial downsampling. This reduces the computational cost to O(2HWC^2 + 9HWC) and the parameter count to O(2C^2 + 18C). Simultaneously, it maximizes the preservation of information during the downsampling process, thus maintaining competitive performance while reducing latency.

[0157] Spatial Pyramid Pooling for Fast Interface (SPPF): An improved version of Spatial Pyramid Pooling (SPP), it changes the parallel structure to a serial structure. Here, the receptive field of the second 5x5 maxpool in the serial structure is equivalent to a parallel 9x9 maxpool, reducing the number of parameters while maintaining SPP's ability to aggregate multi-scale features. Figure 2 As shown;

[0158] Its innovative aspects compared to previous work:

[0159] 1. NMS-free training

[0160] A consistent dual-assignment strategy is proposed to achieve NMS-free training while maintaining competitive performance and low inference latency. This step essentially saves time by filtering out duplicate predicted boxes during inference. Specifically, during training, YOLO typically uses TAL to assign multiple positive samples to each instance. This one-to-many assignment generates rich supervision signals, facilitating optimization and enabling the model to achieve superior performance. However, this requires YOLO to rely on NMS post-processing, which leads to suboptimal inference efficiency at deployment time. Unlike one-to-many assignment, one-to-one matching assigns only one prediction to each ground truth, avoiding NMS post-processing. However, this results in weak supervision, leading to suboptimal accuracy and convergence speed.

[0161] The dual assignment strategy combines the advantages of the two strategies mentioned above. This research introduces another one-to-one head for YOLO. It retains the same structure and optimization objective as the original one-to-many branch, but utilizes one-to-one matching to obtain label assignments. During training, the two heads jointly optimize to provide rich supervision; during inference, YOLOv10 discards the one-to-many head and uses the one-to-one head to make predictions. This enables YOLO to be deployed end-to-end without incurring any additional inference cost.

[0162] As one specific implementation method,

[0163] 2. Model Design Strategy

[0164] A comprehensive efficiency-accuracy driven model design strategy was introduced, comprehensively optimizing various components of YOLO. This significantly reduced computational overhead and enhanced model capabilities from both efficiency and accuracy perspectives. Each design component, including the lightweight classification head, SCDown, and sorting-guided module design, contributes to reducing the number of parameters, FLOPs, and latency. Importantly, these improvements were achieved while maintaining excellent performance. Figure 3 As shown:

[0165] As one specific implementation method,

[0166] 3. Multi-source domain adaptation:

[0167] Multi-source domain adaptation is a machine learning method that leverages data from multiple source domains to improve the performance of a task (such as classification or regression) in a target domain. This approach is particularly suitable when there are distributional differences between the source and target domains, but these source domains share the same objective. The key to multi-source domain adaptation lies in mining and utilizing the shared information between different source and target domains while handling the differences between the domains. A major challenge of multi-source domain adaptation is how to effectively integrate information from multiple source domains while taking into account the heterogeneity between them. A significant advantage of this method is its ability to utilize multiple related source domains to improve the performance of the target domain, often more effectively than single-source domain adaptation or traditional machine learning methods. For this reason, the MIDOG2022 dataset was chosen as the source domain, as it also records mitotic cells, but for different diseases, representing a typical example of similar types with different distributions.

[0168] The MIDOG2022 dataset is provided by the MITosis DOmain Generalization Challenge 2022. It consists of images from six different tumor types, five of which are labeled. The dataset contains a total of 405 cases, with 9501 mitotic figure annotations included in the training set. The task is to detect mitotic figures from histopathological images. Model performance is evaluated on 10 different tumor types, with the F1 score as the primary metric.

[0169] Method: First, a teacher model is trained on the source domain. Then, the teacher model is used to generate pseudo-labels on the target domain. Finally, the pseudo-labels are used to train the final student model on the target domain.

[0170] 1. Source Domain Training

[0171] Multiple source domains correspond one-to-one with multiple object detection heads, sharing the same feature encoder. During backpropagation, sample features that do not correspond to the source domain of an object detection head pass through a gradient inversion module before being input to the object detection head to generate domain adversarial features. In other words, an object detection head for a specific source domain encodes the features of that source domain and simultaneously uses gradient inversion to confuse other object detection heads. Each object detection head is separated from other source domains, thus encoding knowledge specific to that source domain. This approach reduces mutual interference between source domains and helps minimize knowledge degradation.

[0172] Figure 4 There are three different methods of multi-source domain adaptation: (a) Simple combination: Due to domain shifts between sources, simple combination of sources may cause mutual interference. (b) Alignment: Multi-source alignment reduces domain offset, but reduces target-related knowledge when performing alignment without target domain guidance. (c) Decoupling: Decoupling multiple source domains and preventing their mutual interference preserves target-related knowledge specific to a particular domain. Naively, target detection heads from different source domains can use an averaging method to obtain the final result. However, the correlation between different source domains and the target domain is not the same, and source domains with higher correlation should have higher weights. Specifically, the similarity between two domains can be considered at two granularities: "slice level" and "target level." Mathematical formulas for measuring the data distance between source and target domains include KL Divergence, MMD, H-divergence, and Wasserstein distance.

[0173] 2. Target Domain Training

[0174] To mitigate the overfitting of pseudo-labels in the teacher model to the target domain, EMA is used to average the weights of the student model in the target domain to the teacher model in the source domain.

[0175] Training settings:

[0176] First, a common range was set for hyperparameters such as learning rate and batch size. Then, the Optuna tool (a Python library for hyperparameter optimization) was used to search for hyperparameters within the range. The obtained hyperparameters were then fine-tuned. Finally, the Adam optimizer was used for optimization, with an initial learning rate of 0.01 and a final learning rate cosine reduced to 0.001. The batch size was 128, and a total of 200 epochs were trained.

[0177] Evaluation method:

[0178] The effectiveness of multi-source domain adaptation can be evaluated from the following perspectives:

[0179] 1. The degree of alignment between features from different domains for domain-adaptive target detectors. Can multiple source domains better adapt to target domain features?

[0180] 2. Final object detection performance. Does it represent a significant performance improvement compared to baseline methods (such as methods without domain adaptation)?

[0181] Experimental results:

[0182] Feature domain transfer, using t-SNE to reduce feature dimensionality and visualize, such as... Figure 4-5 As shown:

[0183] GIST is a stromal tumor, and the other 5 categories are the 5 categories of MIDOG. Compared with the baseline method without domain adaptation, the feature encoder of the domain-adapted target detector enables the source domain features to be better adapted to the target domain.

[0184] 2) Target detection results

[0185] Baseline 0.811 0.239 0.369 MDA 0.683 0.489 0.570

[0186] It can be observed that, compared to baseline methods without domain adaptation, domain-adapted object detectors exhibit significantly superior object detection performance, particularly in recall, which can significantly reduce false negatives in clinical diagnosis.

[0187] 4. Model re-validation

[0188] To improve model performance, continuous optimization is performed, and a communication platform for doctors is established. The basic process is as follows: After the pre-trained model makes predictions on the test set, the prediction results are displayed on a dedicated website. Doctors log in to the platform and correct the prediction results (i.e., annotations) (keeping correct predictions and deleting incorrect ones). The developers then download the corrected data from the website for model training, using a 50 / 50 cross-validation method (the data is randomly divided into 5 parts; one part is used as the test set and the others as the training set). During each training iteration, the hyperparameters are adjusted linearly based on the size of the training data. Training stops early when the accuracy on the test set no longer increases.

[0189] The main functions of the functional platform are as follows:

[0190] 3.1 File Upload

[0191] like Figure 7 As shown, this step receives image, patient name, image type, and storage location information via an HTTP interface. The system performs multiple processing steps: verifying the integrity and compliance of uploaded files; verifying patient information to ensure accurate data association; automatically identifying and processing file formats; optimizing storage strategies to avoid duplication and recording detailed file metadata; parsing tag information in JSON to enhance data value; and implementing end-to-end encryption to strengthen data security and privacy protection. This solution integrates upload, verification, processing, storage, and security management functions, improving the intelligence and security of medical image data processing.

[0192] 3.2 Labeling Lifecycle

[0193] like Figure 8 As shown, this step demonstrates a medical image labeling update technology that receives label update instructions via API, optimizing the review process. Its main functions include: receiving API requests, verifying the existence of image files, implementing a duplicate review protection mechanism to ensure data consistency, updating the image review status, and managing the review and truth information for each label.

[0194] 3.3 File Download

[0195] like Figure 9 As shown, medical image download can provide on-demand JSON download services for image files and accompanying tags. The process includes: receiving download requests and extracting images, encoding the image data; collecting and filtering correct tag information to form a list of LabelShape objects; constructing an ImageJson object containing image and tag metadata; and finally, converting this object into a downloadable JSON file attachment and sending it via HTTP response.

[0196] As one specific embodiment, such as Figure 10 As shown,

[0197] 1. Detection model:

[0198] The existing data is used to make predictions. After the developers obtain the prediction results Pred_data, they send them to the functional platform for inclusion.

[0199] Once the corrected dataset Recised_data is returned, a new round of training will automatically begin, updating the system's own weight parameters.

[0200] 2. Functional Platform:

[0201] The model's prediction results are obtained, displayed, and logged-in doctors complete the inspection and correction of the annotations.

[0202] The system receives a new batch of patient data (new_data) uploaded by hospital doctors and sends it back to the model for prediction.

[0203] The predicted results are analyzed, and the information is delivered to the doctors who need it to adjust the treatment plan.

[0204] 3. The hospital:

[0205] Doctors log into the platform to verify the model's predictions.

[0206] Upload new data for model prediction and receive the results analysis from the platform.

[0207] The above-described specific embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, and improvements made by those skilled in the art to the technical solutions of the present invention based on the provided textual description and drawings, without departing from the design concept and spirit of the present invention, should all fall within the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

Claims

1. A deep learning-based method for identifying and counting mitotic cells in mesenchymal tumors, characterized in that, Includes the following steps: Step 1: Acquire and preprocess stained gastrointestinal stromal tumor tissue slice images. Preprocessing includes color normalization, noise reduction, and contrast enhancement of the slice images, as well as eliminating human differences during the staining process. The preprocessing step also includes converting the images to grayscale to simplify the input format of the model. Step 2: Segment the preprocessed sliced ​​image to generate image blocks suitable for deep learning model processing. The segmentation operation is performed with a step size of 256 pixels, dividing the 512×512 pixel slice into multiple sub-image blocks, and reducing the impact of edge effects on the model's recognition accuracy. Image blocks that do not contain target cells or contain only a few cells are removed by thresholding the RGB pixel values. Step 3: Expand the generated image patch using data augmentation techniques, including but not limited to color adjustment, random rotation, translation, scaling, cropping, affine transformation, and copy-paste technology to randomly copy the target cell to other image locations; Step 4: Construct a target detection model based on YOLOv10 to identify and locate mitotic cells in images; the YOLOv10 model includes innovative modules, including a C2fUIB module, a partial self-attention module, a spatial channel decoupling downsampling module, and an improved spatial pyramid pooling fast interface module. The innovative module includes: Step 4.1: The C2fUIB module increases the receptive field of the convolutional kernel and reduces computational cost by employing reparameterization techniques, thereby improving the detection accuracy and inference speed of the model. Step 4.2: Partial self-attention module. By dividing the feature map into two parts, one part is processed by the multi-head self-attention module, and the other part retains the original features. Then the two parts of features are fused to enhance the model's ability to detect small targets. Step 4.3: Spatial channel decoupling downsampling module. By decoupling spatial reduction and channel addition operations, the computational cost and number of parameters are reduced, while information is preserved to the maximum extent, enhancing the ability to identify small targets during downsampling. Step 4.4: Improved spatial pyramid pooling fast interface, by changing the parallel structure to a serial structure, reduces the number of parameters while maintaining the ability to aggregate multi-scale features, and improves the model's ability to detect targets at different scales; Step 5: Apply multi-source domain adaptation technology to enhance the model's generalization ability by introducing data from multiple source domains. The multi-source domain adaptation technology includes using multiple target detection heads during source domain training, with each detection head corresponding to specific source domain data, and using a gradient inversion module to generate domain adversarial features, thereby reducing mutual interference between source domains. Subsequently, train the student model on the target domain and generate pseudo-labels through the teacher model to further improve the detection accuracy in the target domain. Step 6: Count the identified mitotic cells and generate corresponding statistical data; the statistical data includes the total number of mitotic cells in each slide, cell density distribution, confidence ranking results, and the distribution pattern of mitotic cells in different pathological slides; the statistical results are used to assist doctors in assessing the condition and formulating treatment plans; Step 7: Display the model's prediction results on the functional platform, where doctors can verify and correct the annotation results. The functional platform includes an intuitive user interface that allows doctors to view the specific location and annotation results of each cell identified by the model. Doctors can make annotation corrections by clicking and dragging. Step 8: Retrain the model using the data corrected by the doctors to continuously optimize the model's recognition and counting performance; the retraining step includes dividing the corrected dataset into multiple cross-validation sets, adjusting the model hyperparameters and introducing new data augmentation strategies to further improve the model's stability and accuracy in practical applications.

2. The method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning according to claim 1, characterized in that, The data augmentation techniques applied in step 3 include the following operations: Step 3.1: Color adjustment, including adjusting the hue, saturation, sharpness and contrast of the image to simulate the image changes under different staining conditions, and to ensure that the model can effectively identify target cells under different staining effects; Step 3.2: Affine transformation, including random rotation, translation, scaling, and cropping operations, to simulate possible changes in slicing and shooting angles in actual operation; Step 3.3: The copy-paste operation randomly copies the identified target cell regions to other locations in the image, thereby increasing the diversity of target cells in the training data and enhancing the robustness of the model; Step 3.4: Data denoising and background removal, eliminating pure white or almost cell-free areas to reduce irrelevant information from interfering with the model.

3. The method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning according to claim 1, characterized in that, The multi-source domain adaptation techniques in step 5 include: Step 5.1: Source domain training process. Multiple source domains correspond one-to-one with multiple target detection heads. The feature encoders share the data. During backpropagation, domain adversarial features are generated through the gradient backpropagation module, thereby reducing mutual interference between source domains and optimizing the adaptive performance of the target domain. Step 5.2: Target domain training process. The pseudo-labels generated by the teacher model are used to train the student model in the target domain. By gradually updating the weights of the student model, its detection accuracy in the target domain is enhanced. Step 5.3: Domain adaptation effect evaluation. The t-SNE dimensionality reduction technique is used to visualize the features of the source domain and the target domain, verify the model's adaptability to the feature distribution between different domains, and evaluate the actual effect of the multi-source domain adaptation technique by comparing it with the baseline model.

4. The method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning according to claim 1, characterized in that, The statistical data in step 6 further includes: Step 6.1: Detailed location distribution map of mitotic cells in each pathological section. The distribution map forms a two-dimensional coordinate system by marking the positions of the detected mitotic cells in the original section image, showing the precise coordinate position of each cell. This distribution map can intuitively show the spatial distribution of mitotic cells in tumor tissue, helping doctors to better understand the proliferation pattern and regional activity of the tumor. Step 6.2: A heat map is generated based on cell density. The heat map calculates the density of mitotic cells in the slice and displays the cell density in different regions in the form of color gradient. The depth of the color corresponds to the density of cells. The heat map can help doctors quickly identify the most active proliferating areas in the tumor and provide an important basis for pathological diagnosis and treatment planning. Step 6.3: A list of mitotic cells sorted by confidence level, which is sorted based on the confidence value of each cell detected by the model; the confidence value represents the degree of certainty that the model is that the cell is a mitotic cell, and doctors prioritize verifying and correcting identification results with low confidence levels; Step 6.4: Statistically analyze the total number and average density of mitotic cells in different pathological sections. The statistical data includes the total number of cells in each section, the cell density per square millimeter, and the local density calculated according to different tumor regions. Step 6.5: Generate a summary report containing multiple patient slides. The summary report integrates all test results, including the total number of cells, average density, distribution of active areas, and confidence analysis for each patient, forming a complete test report to help doctors make overall assessments and comparisons.

5. The method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning according to claim 1, characterized in that, The retraining process in step 8 further includes: Step 7.1: Collection and organization of corrected data. The corrected data is generated by doctors after correcting the model prediction results on the functional platform. The data includes corrected cell annotations, confidence adjustments, and information on the removal of error detections. The organization process includes labeling the corrected data, error analysis, and merging the corrected data into a new training set. Step 7.2: Training set division and preparation. The collected corrected data is divided into multiple cross-validation sets according to the proportion to ensure the model's generalization ability on different datasets. The preparation of the training set also includes secondary augmentation of the data to ensure the diversity and randomness of the data. Step 7.3: Optimization and adjustment of model hyperparameters. Before each retraining, the hyperparameters of the model are optimized and adjusted according to the scale and characteristics of the corrected data. Hyperparameter tuning tools are used to automatically search within a set range, and the best combination of hyperparameters is determined through multiple rounds of experiments. Step 7.4: The retraining process includes inputting the corrected training data into the model for retraining, while monitoring the model's accuracy and loss value through the validation set, and adopting an early stopping strategy, i.e., stopping training when the accuracy of the validation set no longer improves, in order to avoid overfitting; after training is completed, the final weights of the model will be saved and used for the next detection task. Step 7.5: Evaluation of the retrained model, including a comprehensive analysis of the model's performance on the test set and the improvement in the model's performance on the corrected data; evaluate the effect of retraining by comparing it with the initial model and other baseline models, and feed the evaluation results back to the development and application teams to decide whether to further optimize; Step 7.6: Model Deployment and Update. Deploy the retrained and evaluated model to the functional platform and update it in real time to the system used by doctors. Ensure that the new model can automatically replace the old model. At the same time, record the feedback on the use of the new model for reference in the next retraining.

6. The method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning according to claim 1, characterized in that, The model's recognition accuracy and robustness are further optimized through multi-scale feature aggregation and model augmentation training, the steps of which include: Step 8.1: Extract multi-scale features. Input the preprocessed image patch into the deep learning model and extract feature maps of different scales through specific network layers to ensure that the model can capture multiple morphological features of cells. Step 8.2: Perform feature fusion. The extracted features at different scales are fused together. The spatial pyramid pooling module or its improved version are used to aggregate multi-scale information to enhance the model's ability to recognize small targets. Step 8.3: Introduce model augmentation training techniques. By combining loss functions and data augmentation strategies, enhance the model's robustness to complex samples and noise, and improve recognition accuracy. Step 8.4: Using transfer learning techniques, the pre-trained model trained in the relevant domain is used as a basis to improve the model's adaptability to the target task by fine-tuning it in the target domain.

7. The method for identifying and counting mitotic cells in mesenchymal tumors based on deep learning according to claim 1, characterized in that, It also includes an automated error detection and correction mechanism, the steps of which include: Step 9.1: Automatic verification of model prediction results. The system performs preliminary verification of the mitotic cells detected by the model and identifies labels with low confidence or possible errors. Step 9.2: Generate an error detection report. The system will generate an error report for the low-confidence or suspected error labels, which will list the cell locations that may have problems and their corresponding confidence values ​​in detail. Step 9.3: Manual correction and feedback. Doctors manually correct the annotations in the error reports, confirm or correct the model's detection results, and feed the corrected data back to the system. Step 9.4: Dynamically adjust model weights. The system automatically adjusts the model weights and parameters based on the doctor's calibration results to improve the model's accuracy in subsequent tasks.

Images