Cervical cell characterization classification method based on structure constraint network

Abstract

The invention provides a cervical cell characterization classification method based on a structure constraint network, and the method comprises the following steps: A, carrying out the data enhancement of a cell image, and generating different enhanced views of the same cell; b, performing feature extraction on the enhanced view by adopting a convolutional neural network to obtain cell space structure features; c, introducing a bulldozer distance EMD (Earth Mover's Distance) to apply spatial structure consistency constraint on the enhanced view, and taking the spatial structure consistency constraint as a measurement standard for structure distribution alignment; and D, inputting the features subjected to structure alignment into a classifier to complete cell category identification. According to the method, the EMD measurement is combined to apply the structural consistency constraint in the self-supervised learning framework, so that the network can extract stable and discriminative features under an unsupervised condition, and the problems of cellular morphology diversity, dyeing difference, insufficient labeling and the like are effectively solved. According to the method, excellent accuracy is obtained in an open-source cervical cell data set experiment, good robustness is shown, and an efficient and reliable technical scheme is provided for large-scale automatic analysis of clinical cervical cells.

Description

A Cervical Cell Characterization and Classification Method Based on Structural Constraint Networks Technical Field

[0001] This invention relates to a cervical cell characterization and classification method based on structural constraint networks, specifically relating to the fields of medical image processing and deep learning. Background Technology

[0002] Cervical cancer is one of the most common malignant tumors in women, primarily caused by human papillomavirus (HPV) infection. Therefore, efficient and accurate early screening and diagnosis of cervical cancer are of great significance.

[0003] In recent years, with the continuous development of artificial intelligence and deep learning, researchers have proposed various automated computer-aided cell diagnostic methods, which have improved the efficiency and accuracy of diagnostic classification to some extent and reduced reliance on manual intervention. However, existing methods perform poorly in handling cases of uneven staining, blurred boundaries, or overlapping cells, resulting in limited classification accuracy and still exhibiting the following problems:

[0004] 1. High dependence on annotation: Most methods require a large amount of high-quality annotation data, while the annotation of medical images usually needs to be done by professionals, which is time-consuming, costly and easily affected by human error.

[0005] 2. Sensitive to interference: Cervical cell images are easily affected by factors such as staining method and microscope imaging conditions during the acquisition process. Existing models are not robust enough in the face of color changes, rotation, noise interference, etc., which can easily lead to unstable classification results.

[0006] 3. Insufficient utilization of structural feature information: Existing methods often focus on overall or local texture features, making it difficult to effectively capture the spatial structural relationship between cell nucleus and cytoplasm, and thus have limitations in distinguishing cell types with similar morphology. Summary of the Invention

[0007] The purpose of this invention is to provide a cervical cell characterization and classification method based on structural constraint networks.

[0008] This invention includes the following steps:

[0009] S1: Preprocessing of the acquired cervical cell image data includes operations such as size standardization and color correction. Subsequently, various data enhancements are applied to the cell images to simulate the diversity that may occur in clinical practice. Enhancement methods include rotation, flipping, scaling, color perturbation, elastic deformation, and grayscale conversion, thereby generating multiple cell images with different views. This step effectively improves the model's adaptability to staining differences and changes in imaging conditions.

[0010] S2: The different enhanced views from step S1 are input into a ResNet-18 convolutional neural network with shared parameters to extract multi-level feature maps. Unlike traditional methods that directly use global pooling, this invention preserves two-dimensional spatial features, enabling the feature maps to reflect morphological information such as the cell nucleus, cytoplasm, and their spatial distribution. Furthermore, the feature maps are normalized and divided into multiple local regions (patches), each corresponding to a portion of the cell's structure, such as nuclear boundaries and cytoplasmic texture, to capture fine-grained morphological differences.

[0011] S3: For different enhanced views of the same cell, this invention constructs a structural consistency constraint mechanism. By calculating the weights of each local region feature in step S2 to reflect its relative importance in the overall cell morphology, and constructing a similarity matrix between the two views based on cosine similarity, a matching cost between local features is formed. Then, the Earth Mover's Distance (EMD) method is used to solve for the optimal transport plan, achieving alignment of the structural distribution between the two views. By minimizing structural differences, the model can focus on diagnostically significant spatial structural features within the cell, weakening interference from non-critical factors such as imaging conditions and color changes.

[0012] S4: During training, the structural constraint loss obtained in step S3 is combined with the classification loss for joint optimization, enabling the network to distinguish between different types of cells while maintaining structural consistency across different views of the same cell. After training convergence, inputting an image of cervical cells to be detected will output its category label, achieving automated classification of normal cells, abnormal cells, etc.

[0013] The method proposed in this invention mainly includes four stages: image processing, feature extraction, structural constraints, and classification decision. These stages form a progressive relationship and work together to ensure the accuracy and robustness of classification.

[0014] In the input stage of step S1, the cervical cell images are standardized and diversified to better reflect the actual clinical situation. This step not only eliminates differences in acquisition conditions but also expands the sample space by introducing various enhancement methods, thereby reducing the model's dependence on a limited sample size.

[0015] Furthermore, in step S2, the feature extraction stage, this invention does not only extract overall texture information but also consciously preserves the spatial structure of cells. By dividing the feature map, the morphological features of different local regions can be highlighted, enabling the model to differentiate regions with diagnostic significance, such as cell nucleus morphology and cytoplasmic distribution. This design improves the fineness and discriminative power of the features.

[0016] Furthermore, in step S3, the structural consistency constraint stage, this invention introduces a structural constraint module to measure and maintain the internal structural consistency of the same cell under different imaging conditions. This mechanism can automatically reduce the interference of non-critical factors (such as illumination and staining differences) on classification, allowing the model to focus more on the spatial structural relationships of the cell itself, thereby obtaining a stable feature representation.

[0017] Furthermore, in step S4, the training phase combines structural consistency constraints with classification objectives to form a dual-driven optimization strategy. This strategy ensures effective differentiation between different cell categories while maintaining consistent representation of the same cell under different views, enabling automatic classification of cervical cells in various complex environments.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. Reduced reliance on labeled data: This invention employs a self-supervised structural constraint learning approach, enabling network training without relying on large amounts of high-quality manually labeled data. In the field of medical imaging, manual annotation typically requires pathologists, which is time-consuming and prone to introducing human biases. This invention establishes structural consistency constraints between different enhanced views of the same cell image, enabling the network to learn robust feature representations from unlabeled or minimally labeled data. This significantly reduces reliance on manual annotation, thereby reducing manpower and time costs.

[0020] 2. Improved Classification Robustness: By introducing multi-view data augmentation and structural consistency constraints, this invention enables the model to maintain stable performance under various complex conditions. It effectively resists the influence of non-critical factors such as different staining methods, differences in microscope imaging conditions, image rotation, lighting changes, or noise interference, thereby ensuring the stability and reliability of classification results. This characteristic makes this invention more applicable in actual clinical settings.

[0021] 3. Fully Utilizing Cellular Structural Information: Unlike traditional methods that rely solely on texture or global features, this invention preserves the two-dimensional spatial information of cell images during the feature extraction stage. It achieves fine-grained feature modeling by dividing local regions and utilizes the bulldozer distance method to optimally match and align the local feature distributions of different enhanced views. This enables the model to accurately capture cell nucleus morphology, cytoplasmic distribution, and the spatial relationship between the nucleus and cytoplasm. This deep utilization of structural features helps distinguish morphologically similar cell types with different diagnostic significance.

[0022] 4. Significantly Improved Classification Accuracy: This invention has been validated on different publicly available cervical cell datasets (SIPaKMeD and Herlev datasets). Experimental results show that the method of this invention significantly outperforms existing mainstream models in terms of classification accuracy, recall, and F1 score. In particular, when distinguishing morphologically similar cell types (such as parabasal cells and metaplastic cells), this invention can still maintain high classification accuracy, effectively reducing false positives and false negatives, and improving diagnostic reliability.

[0023] 5. Practical Clinical Application Value: This invention demonstrates significant performance advantages on multiple publicly available cervical cell datasets and maintains high accuracy in distinguishing morphologically similar cell categories. This indicates that the method can adapt to the complex and varied image quality conditions in real clinical environments, providing efficient and reliable technical support for early screening and large-scale automated detection of cervical cancer. Attached Figure Description

[0024] Figure 1 is a schematic diagram of the process of the present invention.

[0025] Figure 2 is a schematic diagram of the specific structure of the structural constraint module of the present invention. Detailed Implementation

[0026] The specific embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0027] As shown in Figure 1, the cervical cell characterization and classification method based on structural constraint networks of the present invention includes the following steps:

[0028] S1: Preprocessing of the acquired cervical cell image data includes operations such as size standardization and color correction. Subsequently, various data enhancements are applied to the cell images to simulate the diversity that may occur in clinical practice. Enhancement methods include rotation, flipping, scaling, color perturbation, elastic deformation, and grayscale conversion, thereby generating multiple cell images with different views. This step effectively improves the model's adaptability to staining differences and changes in imaging conditions.

[0029] S2: The different enhanced views from step S1 are input into a ResNet-18 convolutional neural network with shared parameters to extract multi-level feature maps. Unlike traditional methods that directly use global pooling, this invention preserves two-dimensional spatial features, enabling the feature maps to reflect morphological information such as the cell nucleus, cytoplasm, and their spatial distribution. Furthermore, the feature maps are normalized and divided into multiple local regions (patches), each corresponding to a portion of the cell's structure, such as nuclear boundaries and cytoplasmic texture, to capture fine-grained morphological differences.

[0030] S3: For different enhanced views of the same cell, this invention constructs a structural consistency constraint mechanism. By calculating the weights of each local region feature in step S2 to reflect its relative importance in the overall cell morphology, and constructing a similarity matrix between the two views based on cosine similarity, a matching cost between local features is formed. Then, the Earth Mover's Distance (EMD) method is used to solve for the optimal transport plan, achieving alignment of the structural distribution between the two views. By minimizing structural differences, the model can focus on diagnostically significant spatial structural features within the cell, weakening interference from non-critical factors such as imaging conditions and color changes.

[0031] S4: During training, the structural constraint loss obtained in step S3 is combined with the classification loss for joint optimization, enabling the network to distinguish between different types of cells while maintaining structural consistency across different views of the same cell. After training convergence, inputting an image of cervical cells to be detected will output its category label, achieving automated classification of normal cells, abnormal cells, etc.

[0032] The cervical cell characterization and classification method based on structural constraint networks is obtained through network training, specifically including:

[0033] In step S1, cervical cell images are acquired and input into the preprocessing module. For input image I, enhancement operations yield two different views, I1 and I2. This process ensures that key cell structures are preserved under different conditions. The data enhancement method in the preprocessing module of this invention includes:

[0034] Random cropping and scaling: The cropped area is 0.8 to 1.0 of the original image area and scaled to 224×224 to increase spatial robustness and simulate different shooting scales and positions in real-world scenes.

[0035] Randomized grayscale transformation: Converts the image to grayscale with an 80% probability to force the model to learn more essential texture or structural features without relying on color information, thereby enhancing robustness.

[0036] Random color perturbation: The brightness, contrast, saturation, and hue of the image are randomly changed with a perturbation strength of 0.4 to enhance the model's robustness to color changes.

[0037] Random horizontal flip: Flips the image horizontally with a 50% probability to increase data diversity.

[0038] In a preferred experimental protocol, the cervical cell images are single-cell images manually cropped from clustered cell images extracted from Pap smears under a microscope. This ensures high-resolution capture of cell morphological features, and the image size is uniformly set to 224×224. The cervical cell image classification labels are roughly divided into five categories, including dyskeratosis, vacuolated cells, metaplasia, parabasal cells, and superficial intermediate types.

[0039] In step S2, different enhanced views of the preprocessed cervical image are input into a ResNet-18 residual network with the global pooling layer removed to extract the convolutional features in the global two-dimensional space of the image, denoted as:

[0040] F (k) ∈R H×W×C k = 1, 2

[0041] Where H and W are the feature space dimensions, and C is the number of channels. Global biases in image features are eliminated by normalizing the features along the channel axis, emphasizing the relative spatial structure of the image features.

[0042]

[0043] The normalized features are then planarized in space, divided into N = H × W local regions, forming a set of local features for fine-grained structural representation.

[0044] In step S3, as shown in Figure 2, a structural constraint module for better identification of cervical cell characteristics is described in the present invention.

[0045] The set of feature vectors f obtained in step S2 for flattening different views (k) The structural constraint module calculates the correspondence between different views using cosine similarity to evaluate how each morphological region of the cervical cell aligns in the two enhanced views.

[0046]

[0047] Among them, s i，j This indicates the degree of matching between a specific structure in View 1 (such as a dense kernel region) and its corresponding structure in View 2. Furthermore, a cost matrix is ​​constructed for each view based on similarity; the lower the cost, the higher the structural correspondence, meaning the cost of representing node pairs with similar representations should be lower than the cost of other pairs.

[0048] d i，j =1-S i，j

[0049] Furthermore, the goal of structural constraints is to find the most efficient method to transform the structural distribution of one enhanced view into the structural distribution of another, thereby aligning two slightly different perspectives of the same cell. Therefore, based on this, a weighting mechanism is used to assign weights to local features, forming the optimal transport problem:

[0050]

[0051] Among them, the constraint condition guarantees the conservation of mass, g i，j This represents the optimal transfer rate, which is the extent to which the focus is shifted from each salient structure in one view to the corresponding structure in another view. This represents the normalized weight of the view.

[0052] Furthermore, this module assigns different weights to each local feature of the view to quantify the relative morphological importance of the feature across the entire cell. After calculation, all weights are normalized so that the sum of the weights is 1. These normalized weights serve as the quality distribution of spatial locations in the feature map, ensuring that the weights are assigned based on relevance, not the size of the local feature.

[0053]

[0054] in, The global feature average vector for another view, ∈ is set to 10. -3 To maintain numerical stability and prevent gradient vanishing, the max function ensures that the weights are always non-negative. Similarly, the weights of the other view can also be calculated using this formula.

[0055] The minimum transmission cost obtainable through the optimization process of the structural constraint module in step S3 is the structural consistency loss, which is used to calculate the overall structural consistency score.

[0056] Step S4 performs global pooling and full connection operations on the view features obtained in step S2, and jointly optimizes them with the overall structural consistency score obtained in step 3, classifying them by outputting category names.

[0057] To verify the practical feasibility of the method described in this invention, the following experiment was designed and completed:

[0058] The experiment used the SIPaKMeD single-cell cervical cell dataset, which consists of 4049 manually separated and annotated cell images, labeled into 5 different categories. In the experimental evaluation, SIPaKMeD was used to train, validate and test the dataset using an 8:1:1 splitting method.

[0059] All experiments in this study were conducted on a device equipped with an NVIDIA GeForce RTX 4060 graphics card and using PyTorch 2.1.2 as the primary deep learning framework. The network was trained using the Adam optimizer with an initial learning rate of 0.01, a batch size of 64, and a weight decay of 5e-4. All experiments were repeated three times using different random seeds, and the average performance was reported to ensure the stability of the results.

[0060] To comprehensively evaluate the performance of this invention in cervical cell image classification, this experiment used four widely used evaluation metrics: accuracy, precision, recall, and F1-score.

[0061] This experiment compares the results with a large number of mainstream baseline models on the SIPaKMeD dataset. As shown in Table 1, the present invention achieves more significant performance and provides a major improvement compared with traditional convolutional neural networks and Transformer-based models.

[0062] Table 1 Comparison results of different baselines

[0063]

[0064] To verify the universality of this invention, the Herlev cervical cell dataset was also introduced for additional experiments. Following the same protocol and hyperparameter settings as the SIPaKMeD experiment, the dataset was re-divided into normal and abnormal classes, with the training and test sets split in an 8:2 ratio. The experimental results are shown in Table 2.

[0065] Table 2 Performance of the present invention on the Herlev dataset

[0066]

[0067] The results show that the present invention maintains strong performance and robustness in experiments on other datasets, further demonstrating the effectiveness of the present invention for automatic cervical cell classification in different clinical scenarios.

Claims

1. A cervical cell characterization and classification method based on structural constraint networks, characterized in that, Includes the following steps: The acquired cervical cell images are sized and color-corrected, and multiple different views are generated through rotation, flipping, scaling, color perturbation, grayscale conversion, and elastic deformation to enhance data diversity. The multiple views are input into a convolutional neural network with shared parameters for feature extraction, while preserving two-dimensional spatial distribution information. The feature map is further divided into multiple local regions to highlight fine-grained features such as cell nuclear boundaries and cytoplasmic texture. A similarity matrix is ​​constructed based on the similarity measure between local features, and the bulldozer distance (EMD) method is used to calculate the optimal matching relationship between different views, thereby maintaining the structural consistency of the same cell under different views. The structural constraint loss and classification loss are combined for training and optimization, so that the model can automatically classify cervical cell categories while maintaining the consistency of multiple views.

2. The method according to claim 1, characterized in that, The image enhancement methods include at least one of random rotation, random flipping, random scaling, random grayscale transformation, random color perturbation, and elastic deformation.

3. The method according to claim 1, characterized in that, The convolutional neural network is a residual neural network ResNet-18 with the global pooling layer removed.

4. The method according to claim 1, characterized in that, The local region corresponds to a portion of the cell's structure, including the nuclear boundary, cytoplasmic region, or cytoplasmic texture.

5. The method according to claim 1, characterized in that, The structural constraints assign weights to each local feature through a weighting mechanism to reflect its relative importance in the overall cell morphology and maintain mass conservation in optimal transport calculations.

6. The method according to claim 1, characterized in that, The classification process is implemented through global pooling and full connection operations, and outputs the category labels of cervical cells.

7. The method according to claim 1, characterized in that, The method described is applicable to distinguishing between normal and abnormal cervical cells and has high classification accuracy and robustness on publicly available cervical cell datasets.

8. The method according to any one of claims 1-7, characterized in that, The method is applied to early cervical cancer screening and large-scale automated cell detection.

9. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements a cervical cell characterization and classification method based on a structural constraint network, as described in any one of claims 1-7.

Images