A lightweight cervical cell classification method based on multi-scale entropy guided feature adaptation
By introducing multi-scale entropy as an explicit texture prior in cervical cell image classification and combining it with a lightweight network, the problems of high computational resource consumption and neglect of texture structure are solved, achieving efficient and accurate cervical cell recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF SCI & TECH
- Filing Date
- 2026-05-09
- Publication Date
- 2026-07-03
AI Technical Summary
Existing deep learning methods consume high computational resources in cervical cell image classification, making it difficult to run efficiently on lightweight devices. Furthermore, they neglect important textural structural changes in abnormal cervical cells, resulting in insufficient recognition accuracy and interpretability.
By introducing multi-scale local entropy as an explicit texture statistics prior at the data input level, a lightweight entropy feature adaptive module is constructed, which fuses the multi-scale entropy map with the RGB image and uses it in the lightweight network MobileNetV2 to enhance the ability to identify abnormal cervical cells.
Without increasing computational burden, the recognition accuracy and interpretability of the lightweight model are significantly improved, achieving a classification accuracy of 97.78%, and the model interpretability is enhanced, focusing on key diagnostic areas.
Smart Images
Figure CN122336746A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing and deep learning technology, and in particular to a lightweight cervical cell classification method and system based on multi-scale entropy-guided feature adaptation. Background Technology
[0002] Cervical cancer is one of the most serious malignant tumors threatening women's health worldwide. Image screening based on cervical cytology smears is an important means of early detection of precancerous lesions and reducing mortality. Traditional interpretation of cytology images relies on the experience of pathologists, which is not only time-consuming and labor-intensive but also highly subjective. In recent years, computer-aided diagnostic methods based on deep learning, especially convolutional neural networks, have shown great potential in the automatic classification of cervical cell images, effectively improving screening efficiency and accuracy.
[0003] However, existing deep learning methods still face several challenges in practical clinical applications, especially in deployments in primary healthcare institutions with limited computing resources. First, mainstream deep networks such as ResNet and DenseNet have large parameter sizes and high computational costs, making them difficult to run efficiently on lightweight devices. While lightweight networks like MobileNetV2 have extremely high computational efficiency, their inherent receptive field limits their ability to capture multi-scale features. Second, most existing methods rely solely on the RGB color information of images for feature learning, neglecting subtle textural changes crucial for identifying abnormal cervical cells, such as uneven chromatin distribution and irregular nuclear membranes. This textural information is not readily apparent in the RGB color space, limiting model recognition accuracy and resulting in insufficient generalization ability and interpretability with limited data.
[0004] Therefore, how to explicitly introduce texture prior information with clear diagnostic significance to enhance the network's ability to perceive abnormal cells and improve the interpretability of classification decisions, while maintaining the efficiency of lightweight networks, is a key technical problem that urgently needs to be solved in the field of cervical cell image classification. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation. Unlike existing schemes that perform multi-scale feature fusion in deep feature space, this invention introduces multi-scale local entropy as an explicit texture statistical prior at the data input level, and integrates it into the pre-trained network through a very lightweight adaptive module. This significantly improves the recognition accuracy and interpretability of the lightweight model for abnormal cervical cells without significantly increasing the computational burden.
[0006] The specific technical solution for achieving the objective of this invention is as follows: 1. A lightweight cervical cell classification method based on multi-scale entropy guidance, characterized by comprising the following steps: Step 1: Obtain the cervical cell image dataset, scale and augment the images, and divide them into training, validation and test sets according to the proportions; Step 2: Construct a multi-scale entropy extraction module, use three neighborhood windows of different sizes to calculate the local entropy of the image, generate entropy maps of the corresponding scales, and fuse the multi-scale entropy maps according to preset weights to obtain a fused entropy map; Step 3: Construct a four-channel input by concatenating the R, G, and B channels of the original image with the fused entropy map along the channel dimension to form an input tensor of size H×W×4; Step 4: Construct a lightweight entropy feature adaptive module, which consists of two cascaded 1×1 convolutional layers, used to compress and map the four-channel input tensor into a three-channel feature map to adapt to the input layer of the subsequent pre-trained MobileNetV2 network. Step 5: Input the three-channel feature map output from Step 4 into the MobileNetV2 backbone network for forward propagation. Obtain the probability prediction of cell class through fully connected layers and the Softmax function, and perform backpropagation according to the loss function to optimize and fine-tune the network parameters. Step 6: In the inference phase, for a new cervical cell image, after performing steps 2 to 4, input the trained model and take the category corresponding to the highest probability value as the final classification result.
[0007] 2. The lightweight cervical cell classification method based on multi-scale entropy guidance according to claim 1, characterized in that, in step 2, the multi-scale local entropy features of the image are extracted, specifically as follows: Step 2-1: For the input image, define three different neighborhood window sizes: 5×5, 7×7, and 9×9. The 5×5 window is used to capture fine details of the nuclear chromatin; the 7×7 window is used to match typical nuclear sizes; and the 9×9 window is used to perceive cytoplasmic context information.
[0008] Step 2-2: Within each window, calculate the local entropy value H according to the following formula: in Represents the grayscale value within a local window The probability of occurrence, where n is the gray level; Steps 2-3: Traverse the entire image and generate entropy maps corresponding to the three scales respectively. , and .
[0009] Steps 2-4: Weight and fuse the generated entropy maps at the three scales according to the following formula to obtain the final fused entropy map E: This ensures that the intermediate scale window, which best matches the cell nuclear scale, receives the highest weight.
[0010] 3. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that the construction of the four-channel input tensor in step 3 is specifically as follows: The three single-channel components R, G, and B of the original RGB image are concatenated with the fused entropy map E obtained in step 2 along the channel dimension to generate a four-channel input tensor. : Among them, the value of the fusion entropy map E is normalized to the same value range as the RGB channels.
[0011] 4. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that the entropy feature adaptation module in step 4 has the following specific structure: This module consists of two consecutive 1×1 convolutional layers. Let the input four-channel tensor be... The weight tensor of the first 1×1 convolution is If the bias is b1, then the output of the first layer is: in This represents the number of intermediate feature channels. The weights of the second layer's 1×1 convolution are... With a bias of b2, the final output is: This output This is a three-channel adaptive feature map adapted to the input of the MobileNetV2 network. The total number of additional parameters introduced by this module is approximately 0.38M, effectively controlling computational overhead.
[0012] 5. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, wherein the image preprocessing in step 1 further includes data augmentation operations, wherein the data augmentation includes one or more of random rotation, horizontal flipping, width offset, and height offset.
[0013] 6. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, wherein the categories of the cervical cell images include: incompletely keratinized cells, keratotic cells, metaplastic cells, parabasal cells, and superficial-middle cells.
[0014] 7. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that the fused entropy map is an explicit texture statistical prior, which is different from the abstract multi-scale semantic features extracted by deep learning within the network in the prior art. The present invention completes the extraction and fusion of multi-scale texture information at the data input stage, and has stronger interpretability and plug-and-play versatility.
[0015] The beneficial effects of this invention are: 1. The multi-scale entropy feature extraction method proposed in this invention reveals the differences in texture organization structure of cell nucleus and cytoplasm at different spatial scales. Unlike the existing technology that extracts abstract multi-scale semantic features within the network, this invention uses local entropy as an explicit texture statistical feature, which has clear physical meaning and stronger interpretability, providing the network with texture prior information with diagnostic value beyond RGB images.
[0016] 2. The entropy feature adaptive module designed in this invention is extremely lightweight, adding only about 0.38M parameters, which can seamlessly integrate four-channel entropy-RGB fused data into the pre-trained MobileNetV2 network. Compared with existing technologies that require redesigning modules such as CSPM and MSFA, this invention does not require modification of the backbone network structure, achieving decoupling between feature enhancement and network architecture, and maintaining the model's efficiency and plug-and-play flexibility.
[0017] 3. Through validation on public datasets, the method of this invention achieved a classification accuracy of 97.78% on the SipakMed dataset, which is 2.72 percentage points higher than the original MobileNetV2. Furthermore, Grad-CAM visualization analysis confirmed that the model can more accurately focus on key diagnostic regions such as cell nuclei, thus improving the interpretability of the model. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the overall framework process in an embodiment of the present invention; Figure 2 This is a schematic diagram of the entropy feature adaptive module in an embodiment of the present invention; Figure 3 This is a schematic diagram of a cervical cell image after data augmentation processing in an example of the present invention; Figure 4This is a schematic diagram comparing the classification confusion matrix of the original MobileNetV2 network and the method of this invention on the test set in an example of this invention; Figure 5 This is a comparative visualization diagram of the multi-scale entropy map and the Grad-CAM attention heatmap in an embodiment of the present invention. Detailed Implementation
[0019] This implementation provides a lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation. See [link to relevant documentation]. Figure 1 The overall framework flowchart is shown below. The method includes the following steps: Step 1: Obtain the cervical cell image dataset and preprocess the images. This embodiment uses the publicly available SipakMed dataset, which contains 4049 cervical cell images, categorized into five types: incompletely keratinized cells, keratotic cells, metaplastic cells, parabasal cells, and superficial-middle cells. All images are uniformly scaled to 128×128 pixels. The dataset is stratified into training, validation, and test sets in an 8:1:1 ratio. As a further limitation of this embodiment, to improve the model's generalization ability and prevent overfitting, data augmentation operations are applied to the images during the training phase. The data augmentation operations include one or more combinations of: random rotation (angle range ±15°), horizontal flipping, width offset (offset coefficient range 0.1), and height offset (offset coefficient range 0.1).
[0020] See Figure 3 As shown, Figure 3 This is a schematic diagram of a cervical cell image after data augmentation processing, as described in an embodiment of the present invention. Figure 3 (a) is an image of the original cervical cells. Figure 3 (b) is the image after being randomly rotated by a certain angle. Figure 3 (c) is the image after horizontal flipping. Figure 3 (d) is the image after width offset. Figure 3 (e) is the image after height shift. Through the above geometric transformations and pixel space perturbations, the diversity of training samples is effectively expanded, and the robustness of the model to different imaging conditions and cell poses is improved.
[0021] Step 2: Construct a multi-scale entropy extraction module to generate a fused entropy map. Local entropy is a measure of the amount of information or texture complexity of a local region of an image. For cell nuclei containing rich texture information, the entropy value is usually high; while for flat cytoplasm or background regions, the entropy value is low. This invention differs from existing techniques that extract abstract multi-scale features within deep networks; instead, it uses local entropy as an explicit, physically meaningful texture statistical prior.
[0022] Step 2-1: Define the local entropy calculation function. For the input single-channel grayscale image... I Define a function `local_entropy(I, window_size)` where `window_size` is the size of the local neighborhood window. For each pixel in the image, extract a local neighborhood of size `window_size × window_size` centered on that pixel, and calculate the probability of each gray level occurring within that neighborhood. And calculate the local entropy value of the center pixel according to the following formula. H : in grayscale value within a local window The probability of occurrence, where n is the number of gray levels in the image (e.g., 256). Step 2-2: Define three entropy extraction branches with different spatial scales. This implementation selects three different neighborhood window sizes: 5×5, 7×7, and 9×9. The rationale for this setting is as follows: 5×5 window: A smaller receptive field used to capture fine texture details of chromatin granules inside the cell nucleus and subtle nucleolar variations; 7×7 window: Medium receptive field, whose coverage matches the typical cell nucleus size in the dataset (approximately 30-50 pixels in diameter), used to fully perceive the texture heterogeneity of the cell nucleus as a whole; 9×9 window: A larger receptive field, used to provide contextual texture information of the cytoplasmic background, helping to distinguish structures such as perinuclear halos.
[0023] Calling the entropy calculation function in step 2-1, we obtain 5×5 entropy maps respectively. 7×7 entropy diagram and 9×9 entropy diagram .
[0024] Subsequently, the entropy maps at the three scales are weighted and fused using the following formula to obtain the final fused entropy map E: This weighting scheme assigns the highest contribution weight (0.5) to the 7×7 window, which best matches the typical cell nucleus scale, thereby maximizing the highlighting of the texture features of the cell nucleus region in the fused entropy map and suppressing interference from irrelevant background noise. The final fused entropy map is obtained as follows. E The dimensions are the same as the original input image.
[0025] Step 3: Construct a four-channel input tensor. The fused entropy map E generated in Step 2 is used as the fourth channel and concatenated with the three single-channel components R, G, and B of the original RGB image in the depth dimension to construct an input tensor containing rich texture prior information. : To ensure numerical stability and facilitate subsequent network training, entropy maps are fused. E The values are linearly normalized to the same range as the RGB channels (e.g., [0,1] or [-1,1]). The final size of this four-channel tensor is... H × W ×4. In actual engineering deployments, to improve training efficiency, the fusion entropy map of all training images can be pre-calculated and stored locally in binary file format.
[0026] Step 4: Construct and apply the entropy feature adaptive module to perform channel adaptation mapping on the input data. Since the pre-trained MobileNetV2 network receives three-channel RGB input by default, this invention designs a module to solve the channel mismatch problem. Figure 2 The lightweight entropy feature adaptive module shown is different from existing solutions that require redesigning the attention fusion module. The adaptive module of this invention is only used for channel adaptation, with an extremely simple structure and very few parameters.
[0027] As a further limitation of this embodiment, the network structure of the entropy feature adaptive module is defined in detail as follows: Step 4-1: Define the class EntropyAdapter, which inherits from the nn.Module base class in the PyTorch framework. Perform the following operations in the __init__ method: Set the number of intermediate feature channels In this embodiment, The value is 16, which achieves a good balance between model expressiveness and computational cost.
[0028] Create the first convolutional layer, `self.conv1`. This layer is a two-dimensional convolutional layer (`nn.Conv2d`), with 4 input channels (`in_channels`, corresponding to the R, G, B, and E channels) and `out_channels` (`out_{mid}`). The kernel size is 1, the stride is 1, and the padding is 0. The 1×1 kernel is used to achieve linear combination and information exchange between channels without changing the spatial resolution of the feature map.
[0029] Create the first batch normalization layer, self.bn1. This layer is a two-dimensional batch normalization layer (nn.BatchNorm2d) with parameter C_{mid}. Batch normalization layers are used to accelerate model convergence and stabilize the training process.
[0030] Create a non-linear activation layer, self.relu. This layer is a corrected linear unit (nn.ReLU), with the parameter inplace=True to save memory overhead.
[0031] Create a second convolutional layer, `self.conv2`. This layer is a two-dimensional convolutional layer (`nn.Conv2d`), with input channels `in_channels=C_{mid}`, output channels `out_channels=3`, kernel size `kernel_size=1`, stride=1, and padding=0. This layer is used to remap intermediate features back to the RGB three-channel representation space.
[0032] Step 4-2: Define the forward propagation function `forward(self, x)`. Input tensor `x` x The dimension is ( B ,4, H , W ),in B This refers to the batch size. The forward propagation calculation process is as follows: First, x The first convolutional layer is input and then passed through batch normalization and ReLU activation functions sequentially to obtain intermediate feature maps. x 1: Secondly, the intermediate feature map x Inputting the second convolutional layer yields the final three-channel output feature map. X out : Returned tensor X out The dimensions are ( B ,3, H , W This refers to a three-channel adaptive feature map that adapts to the input requirements of the MobileNetV2 network.
[0033] Step 4-3: Module Parameter Statistics and Analysis. The additional trainable parameters introduced in this module only include the weights and biases of two 1×1 convolutional layers. The specific calculations are as follows: The number of parameters for the first 1×1 convolution layer is 4×16×1×1+16=80.
[0034] The number of parameters for the second 1×1 convolution layer is 16×3×1×1+3=51.
[0035] Learnable parameters for batch normalized layer: 16×2=32
[0036] The total number of additional parameters introduced is approximately 163, equivalent to about 0.00016M. Compared to the approximately 3.5M parameters in the MobileNetV2 backbone network, the computational overhead brought by this module is negligible, strictly ensuring the lightweight nature of the overall classification method.
[0037] Step 5: Model Training and Parameter Fine-tuning. The three-channel adaptive feature map output from Step 4... X out The input is fed into the pre-trained MobileNetV2 backbone network. The MobileNetV2 network extracts high-level semantic features through a series of inverted residual blocks and linear bottleneck layers. Then, the feature maps are compressed into feature vectors through a global average pooling layer. Finally, the input is fed into a fully connected classification layer, and the predicted probabilities of five types of cervical cells are output through the Softmax function.
[0038] During model training, the CrossEntropy Loss function is used to calculate the error between the predicted probability and the true one-hot encoded label. The optimizer uses either Stochastic Gradient Descent (SGD) or Adaptive Moment Estimation (Adam), with an initial learning rate of 0.001. Cosine annealing or step decay strategies are used to adjust the learning rate during training. The gradient of the loss function with respect to the network parameters is calculated using backpropagation, and end-to-end fine-tuning is performed on some or all parameters of the entropy feature adaptive module and the MobileNetV2 backbone network.
[0039] During training, the classification accuracy of the model on the validation set is monitored, and the weight of the model with the highest accuracy on the validation set is saved as the final inference model.
[0040] Step 6: Model Inference and Classification. In the inference phase, for a new, unlabeled cervical cell image, the scaling preprocessing in Step 1 (scaling to 128×128 pixels) is performed first. Then, Step 2 generates a multi-scale fusion entropy map. Next, Step 3 concatenates the RGB channels with the entropy map into a four-channel tensor. Finally, Step 4 maps the image to a three-channel map using the trained entropy feature adaptive module. This three-channel map is input into the trained classification model, which outputs a probability vector of length 5. The index corresponding to the highest probability value in this vector is taken and mapped back to the category label (incomplete keratinization, hollowing out, metaplastic cells, basal cells, superficial-middle cells), which is the final classification diagnosis result.
[0041] Implementation Method 2: Comparison and verification of the effects of the method of the present invention with existing technologies.
[0042] To verify the beneficial effects of the lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation proposed in this invention, the method of this invention (Entropy-guided MobileNetV2) was compared with the original MobileNetV2 network, ResNet50 network, and DenseNet121 network under the same experimental settings (training set, test set division, image preprocessing method, random seed). The overall classification accuracy was used as the evaluation metric.
[0043] Experimental results show that the method of this invention achieves a classification accuracy of 97.78% on the SipakMed test set, which is 2.72 percentage points higher than the original MobileNetV2's 95.06%. Meanwhile, the total number of parameters in the model of this invention is far less than that of ResNet50 and DenseNet121, maintaining extremely high computational efficiency.
[0044] See Figure 4 As shown, Figure 4 This is a schematic diagram comparing the classification confusion matrix of the original MobileNetV2 network and the method of the present invention on the test set in an embodiment of the present invention. Figure 4 (a) shows the classification results of the original MobileNetV2 network. Figure 4(b) shows the classification results of MobileNetV2 guided by the entropy model of this invention. The horizontal axis represents the true class label, and the vertical axis represents the model's predicted class label. Class indices 0 to 4 represent parakeratotic cells, koilocytes, metaplastic cells, basal cells, and superficial-middle cells, respectively. A comparison of the confusion matrices clearly shows that the method of this invention significantly increases the diagonal values for parakeratotic cells and koilocytes—two categories with significant but easily confused texture features—indicating a targeted improvement in classification accuracy. This is attributed to the multi-scale entropy feature effectively capturing the subtle texture differences in chromatin distribution and nuclear membrane irregularities between these two cell types.
[0045] Furthermore, Grad-CAM technology is used to visualize and analyze the decision focus area of the model (see...). Figure 5 The results showed that after introducing multi-scale entropy guidance, the model's high-response thermal region was more accurately focused on the cell nucleus and perinuclear region with clinical diagnostic value, rather than background impurities. This further enhanced the interpretability and clinical credibility of the model's decision-making process.
[0046] Implementation Method 3: A lightweight cervical cell classification system based on multi-scale entropy-guided feature adaptation.
[0047] This embodiment proposes an image classification system corresponding to the above method. The system includes the following functional modules: a preprocessing module for acquiring a cervical cell image dataset and performing scaling and data augmentation on the cervical cell images; a multi-scale entropy extraction module for calculating local entropy at three scales (5×5, 7×7, and 9×9) on the preprocessed cervical cell images and performing weighted fusion according to preset weights to generate a fused entropy map; and a four-channel construction module for concatenating the RGB three-channel components of the original image with the fused entropy map in the channel dimension to generate a size of [missing information]. H × W The input tensor consists of four channels (×4); an entropy feature adaptation module receives the four-channel input tensor and performs channel compression and feature mapping through two consecutive 1×1 convolutional layers, outputting a three-channel adaptive feature map adapted to the MobileNetV2 network input; and a classification module uses the MobileNetV2 backbone network to perform deep feature extraction and category prediction on the three-channel adaptive feature map, outputting the category label of the cervical cell image.
[0048] In summary, the lightweight cervical cell classification method proposed in this invention, based on multi-scale entropy-guided feature adaptation, effectively complements existing schemes that perform multi-scale feature fusion in deep feature space. By introducing interpretable explicit texture priors, this invention achieves multi-scale texture enhancement at the data input level, effectively overcoming the shortcomings of lightweight networks in texture perception. It achieves a significant performance improvement while maintaining high computational efficiency, making it highly suitable for clinical auxiliary diagnostic scenarios with limited computational resources.
[0049] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A lightweight cervical cell classification method based on multi-scale entropy guided feature adaptation, characterized in that, Includes the following steps: Step 1: Obtain the cervical cell image dataset, scale and augment the images, and divide them into training, validation and test sets according to the proportions; Step 2: Construct a multi-scale entropy extraction module, use three neighborhood windows of different sizes to calculate the local entropy of the image, generate entropy maps of the corresponding scales, and perform weighted fusion of the multi-scale entropy maps according to preset weights to obtain a fused entropy map; Step 3: Construct a four-channel input tensor by concatenating the three single-channel components of the original RGB image with the fused entropy map in the channel dimension to generate four-channel input data of size H×W×4. Step 4: Construct an entropy feature adaptive module, which consists of two consecutive 1×1 convolutional layers. This module receives the four-channel input data, performs channel compression and feature mapping on it, and outputs a three-channel adaptive feature map that is compatible with the input requirements of the pre-trained MobileNetV2 network. Step 5: Input the three-channel adaptive feature map into the pre-trained MobileNetV2 network for feature extraction and classification, calculate the loss based on the classification results and the true label, and fine-tune some or all of the parameters of the entropy feature adaptive module and the MobileNetV2 network through backpropagation; Step 6: After processing the cervical cell images to be classified in steps 2 to 3, input them into the trained model and output the corresponding cell categories.
2. The method of claim 1, wherein the method is a multi-scale entropy guided feature adaptive lightweight cervical cell classification method. The calculation and fusion of multi-scale entropy in step 2 specifically involves: Step 2-1: For the input image, define three different neighborhood window sizes: 5×5, 7×7, and 9×9. The 5×5 window is used to capture fine details of the nuclear chromatin, the 7×7 window is used to match typical nuclear sizes, and the 9×9 window is used to perceive cytoplasmic context information. Step 2-2: Within each window, calculate the local entropy value H according to the following formula: wherein represents the local windowed gray value the probability of occurrence, n is the number of gray levels; Step 2-3, traverse the whole image, respectively generate entropy map corresponding to three scales , and ; Steps 2-4: Weight and fuse the generated entropy maps at the three scales according to the following formula to obtain the final fused entropy map E: This ensures that the intermediate scale window, which best matches the cell nuclear scale, receives the highest weight.
3. The method of claim 1, wherein the method is a multi-scale entropy guided feature adaptive lightweight cervical cell classification method, characterized in that, The construction of the four-channel input tensor in step 3 is specifically as follows: The three single-channel components R, G, B of the original RGB image are spliced with the fusion entropy map E obtained in step 2 in the channel dimension to generate a four-channel input tensor : Among them, the value of the fusion entropy map E is normalized to the same value range as the RGB channels.
4. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that, The specific structure of the entropy feature adaptive module in step 4 is as follows: This module consists of two consecutive 1×1 convolutional layers; let the input four-channel tensor be... The weight tensor of the first 1×1 convolution is If the bias is b1, then the output of the first layer is: in This represents the number of intermediate feature channels; The weights of the second layer 1×1 convolution are: With a bias of b2, the final output is: This output This is a three-channel adaptive feature map adapted to the input of the MobileNetV2 network; the total number of additional parameters introduced by this module is approximately 0.38M, effectively controlling the computational cost.
5. A lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that, The image preprocessing in step 1 also includes data augmentation operations, which include one or more of random rotation, horizontal flipping, width offset, and height offset.
6. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that, The categories of cervical cell images include: incompletely keratinized cells, keratotic cells, metaplastic cells, parabasal cells, and superficial-medium layer cells.
7. The lightweight cervical cell classification method based on multi-scale entropy-guided feature adaptation according to claim 1, characterized in that, The fusion entropy map, as an explicit texture statistical prior, differs from the abstract multi-scale semantic features extracted through deep learning within the network. It completes the extraction and fusion of multi-scale texture information during the data input stage.
8. A lightweight cervical cell image classification system based on multi-scale entropy-guided feature adaptation, characterized in that, The image classification system includes: The system comprises the following modules: a preprocessing module for acquiring a cervical cell image dataset and preprocessing the cervical cell images; a multi-scale entropy extraction module for calculating and weighting the local entropy of the preprocessed cervical cell images at multiple scales to generate a fused entropy map; a four-channel construction module for concatenating the RGB three-channel components of the original image with the fused entropy map along the channel dimension to generate a four-channel input tensor; an entropy feature adaptation module for receiving the four-channel input tensor and performing channel compression and feature mapping through two 1×1 convolutional layers to output a three-channel adaptive feature map adapted to the MobileNetV2 network; and a classification module for extracting features and classifying the three-channel adaptive feature map using the MobileNetV2 backbone network to output the category prediction results of the cervical cell images.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the lightweight cervical cell image classification method based on multi-scale entropy-guided feature adaptation as described in any one of claims 1 to 7.