A non-invasive anemia classification method based on a bulbar conjunctiva image
By improving the SE attention module and resnext50 architecture, a Dual-DPFSE-ResNext50 network model was constructed, which solved the problem of high-dimensional data processing of palpebral conjunctival images and achieved efficient anemia classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTH CHINA UNIV OF TECH
- Filing Date
- 2023-08-31
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies rely on manual feature engineering, which makes it difficult to process high-dimensional palpebral conjunctival image data, resulting in low accuracy in anemia classification. Furthermore, traditional methods lack model expressive power.
An improved SE attention module and resnext50 architecture were used, combined with adaptive average pooling and max pooling, to construct a Dual-DPFSE-ResNext50 network model for palpebral conjunctival image classification.
It eliminates the need for manual feature engineering, effectively handles high-dimensional data, improves anemia classification accuracy, reduces dependence on feature representation, and enhances model expressive power.
Smart Images
Figure CN117333930B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision processing technology, specifically to a non-invasive anemia classification method based on palpebral conjunctival images. Background Technology
[0002] Anemia is a common health problem, usually associated with abnormalities in the number or function of red blood cells in the blood. Traditional methods for analyzing anemia typically rely on laboratory tests of blood samples, such as blood cell counts, hemoglobin measurements, and red blood cell morphology assessments. However, these methods require specialized laboratory equipment and technicians, and involve long waiting times.
[0003] With the rapid development of computer vision and deep learning technologies, using image data for anemia analysis has become a promising approach. Imaging data of the palpebral conjunctiva can be obtained directly using mobile phones or other instruments, providing a wealth of information about anemia status. Analyzing and identifying anemia-related imaging data can assist doctors in the early diagnosis, classification, and treatment planning of anemia.
[0004] Existing anemia classification methods based on palpebral conjunctival images have the following technical problems:
[0005] Relying on manual feature engineering: This requires domain experts to have a deep understanding of the anemia classification task and to extract and select features, which faces the subjectivity and difficulty of feature extraction.
[0006] Difficulty in handling high-dimensional data: Traditional machine learning methods may face the curse of dimensionality when dealing with high-dimensional data, leading to a decline in model performance, information loss, and reduced model performance.
[0007] Dependence on feature representation: Relying on manually designed feature representations makes it difficult to capture complex relationships and nonlinear patterns in the data.
[0008] Limited model expressive power: Traditional machine learning methods often fail to capture complex patterns and relationships in the data, resulting in limitations in classification performance. Summary of the Invention
[0009] To address the problems existing in the prior art, the purpose of this invention is to provide a non-invasive anemia classification method based on palpebral conjunctival images. This method does not rely on manual feature engineering, can handle high-dimensional data, reduces the dependence on feature representation, has strong model expressive power, and improves the accuracy of non-invasive anemia classification.
[0010] To achieve the above objectives, the present invention adopts the following technical solution:
[0011] A non-invasive anemia classification method based on palpebral conjunctival images includes the following steps:
[0012] The SE attention module is improved by fusing adaptive average pooling and adaptive max pooling to compress the channel dimension and weight the importance of the input feature map, thus obtaining the DPFSE module.
[0013] Using resnext50 as the main architecture, a DPFSE module is introduced on the basis of the residual block to obtain the DPFSE-ResBlock module;
[0014] An SE module was introduced before the first residual block to obtain the final Dual-DPFSE-ResNext50 network model;
[0015] The constructed Dual-DPFSE-ResNext50 anemia analysis model was used to classify palpebral conjunctival images and obtain anemia classification results.
[0016] Furthermore, palpebral conjunctival image data were obtained from the publicly available EYES-DEFY-ANEMIA dataset and a dataset was generated, which was then divided into a training set and a test set.
[0017] Furthermore, the palpebral conjunctival image dataset was classified using a two-level assessment principle: when hemoglobin was >= 11.5 g / dl, it was marked as normal, and when hemoglobin was < 11.5 g / dl, it was marked as anemia.
[0018] Furthermore, the palpebral conjunctival image dataset was augmented using a data balancing method to obtain a larger amount of data, and to make the data labeled as normal and labeled as anemic close to or equal.
[0019] Furthermore, the data balancing augmentation methods employed include random rotation, random scaling and cropping, horizontal flipping, vertical flipping, random perspective transformation, and random affine transformation.
[0020] Furthermore, the DPFSE module consists of a fusion layer of adaptive max pooling and adaptive average pooling, a fully connected layer, ReLU, a fully connected layer, and a Sigmoid layer connected in sequence.
[0021] Furthermore, the Dual-DPFSE-ResNext50 module consists of an input layer, a convolutional layer, a pooling layer, an SE layer, multiple DPFSE residual modules, adaptive average pooling, and a fully connected layer connected in sequence.
[0022] In summary, the present invention has the following advantages:
[0023] Using resnext50 as the main architecture, an improved DPFSE module is introduced on top of the original residual block. This module can better capture the correlation and importance between channels in the feature map, improving the model's expressive power and performance. This invention integrates the DPFSE module and the residual module to form a dual-pooling fusion SE residual module (DPFSE-ResBlock). The SE module is introduced before the first residual block, forming a deeper module. This results in a breakthrough in the classification ability of the constructed anemia analysis model. It eliminates the need for manual feature engineering, can handle high-dimensional data, reduces the dependence on feature representation, has strong model expressive power, and improves the accuracy of non-invasive anemia classification. Attached Figure Description
[0024] Figure 1 This is a flowchart of the present invention.
[0025] Figure 2 This is a schematic diagram of the palpebral conjunctiva.
[0026] Figure 3 The distribution of the number of different types of palpebral conjunctiva images in the experiment.
[0027] Figure 4 This is a diagram illustrating transfer learning.
[0028] Figure 5 This is a schematic diagram of a dual-pool fused SE module (DPFSE).
[0029] Figure 6 This is a schematic diagram of DPFSE-ResBlock (Dual Pooling Fusion SE Residual Module).
[0030] Figure 7 This is a schematic diagram of the Dual-DPFSE-ResNext50 network structure.
[0031] Figure 8 This is the confusion matrix of the classification results. Detailed Implementation
[0032] The present invention will now be described in further detail.
[0033] like Figure 1 As shown, a non-invasive anemia classification method based on palpebral conjunctival images includes the following steps:
[0034] The SE attention module is improved by fusing adaptive average pooling and adaptive max pooling to compress the channel dimension and weight the importance of the input feature map, thus obtaining the DPFSE module.
[0035] Using resnext50 as the main architecture, a DPFSE module is introduced on the basis of the residual block to obtain the DPFSE-ResBlock module;
[0036] Furthermore, an SE module is introduced before the first residual block of the DPFSE-ResBlock module to obtain the Dual-DPFSE-ResNext50 module;
[0037] The constructed Dual-DPFSE-ResNext50 anemia analysis model was used to classify palpebral conjunctival images and obtain anemia classification results.
[0038] This embodiment uses resnext50 as the main architecture and introduces an improved DPFSE module on top of the original residual block. This module can better capture the correlation and importance between channels in the feature map, improving the model's expressive power and performance. This invention integrates the DPFSE module and the residual module to form a dual-pooling fusion SE residual module (DPFSE-ResBlock), creating a deeper module that leads to a breakthrough in the classification ability of the constructed anemia analysis model. Furthermore, the SE module is introduced before the first residual block, strengthening the importance and focus on channels in the initial feature extraction stage. This optimizes the importance of channels in the feature map from a global perspective, making the network more focused on channel information crucial for the classification task, thereby improving overall classification performance. It eliminates the need for manual feature engineering, can handle high-dimensional data, reduces the dependence on feature representation, has strong model expressive power, and improves the accuracy of non-invasive anemia classification.
[0039] The following is the construction process of the anemia analysis model in this embodiment.
[0040] 1. Dataset
[0041] This experiment was conducted on the publicly available EYES-DEFY-ANEMIA dataset, where the images have a pixel size of 1067 x 800. The EYES-DEFY-ANEMIA dataset contains 218 eye images, including 132 male eye images and 86 female eye images. Due to the loss of hemoglobin data label for one male, a total of 217 images were ultimately available for classification.
[0042] The model uses a two-level assessment principle to classify the dataset: when Hb (hemoglobin) is >= 11.5 g / dL, the patient's condition is marked as "normal," and when Hb < 11.5 g / dL, the patient's condition is marked as "anemic." The dataset contains 66 images in the "anemic" category and 151 images in the "normal" category.
[0043] Figure 2This is a schematic diagram of the palpebral conjunctiva. The top four images are of normal palpebral conjunctiva, and the bottom four images are of anemic palpebral conjunctiva.
[0044] 2. Dataset Generation and Preprocessing
[0045] Since the original dataset has a relatively small amount of data, it may lead to insufficient model training or easy overfitting. In order to enhance the generalization ability of the model and avoid overfitting, data augmentation operations are performed on the dataset. Data augmentation methods include: random rotation, random scaling and cropping, horizontal flipping, vertical flipping, random perspective transformation, and random affine transformation.
[0046] In addition, the image number distribution of the original palpebral conjunctiva dataset is as follows: Figure 3 As shown, the dataset contains a majority of data in the "normal" category. The model may favor the "normal" category due to its larger number of data points, while performing poorly on the "anemic" category due to its smaller number of data points. Therefore, data balancing is employed to ensure a relatively balanced number of samples between the two categories. This allows the model to better learn the distinguishing features between the two categories, thereby improving the overall classification performance. The data balancing method used in the experiment was as follows: for each image in the "normal" category, two of the above data augmentation methods were randomly applied; for images in the "anemic" category, all data augmentation methods were applied. That is, for each original image in the "anemic" category, six new images were generated, and for each original image in the "normal" category, two new images were generated. The final dataset is relatively balanced, which improves the final classification results.
[0047] Other data balancing methods: Oversampling: Oversampling is a method to increase the number of minority class samples. It replicates and duplicates minority class samples so that the number of minority class samples is close to or equal to that of the majority class. Common oversampling methods include random replication, SMOTE (Synthetic Minority Over-sampling Technique), and ADASYN (Adaptive Synthetic Sampling). Oversampling may introduce similarity and redundancy between samples, causing the model to become overly reliant on the minority class learning process, potentially leading to overfitting.
[0048] Undersampling: Undersampling is a method to reduce the number of samples in the majority class. It reduces the number of majority class samples by removing or decreasing them until the number of majority class samples is close to or equal to that of the minority class. Common undersampling methods include random deletion, Tomek Links, and ENN (Edited Nearest Neighbors). Undersampling may lose some information about the majority class, leading to insufficient learning of the majority class by the model and a tendency to underfit.
[0049] Data augmentation is a method of transforming and expanding original data without reusing samples; oversampling and undersampling, on the other hand, involve copying or deleting existing samples. Oversampling increases the number of samples in a minority of classes but may introduce redundancy; undersampling reduces the number of samples in the majority of classes but may lose some information. Data augmentation can increase sample diversity, while oversampling and undersampling alter the sample distribution.
[0050] In this experiment, the balanced dataset was divided into a training set and a test set. 70% of the dataset (642 palpebral conjunctival images) was used for the training set, and 30% of the dataset (273 palpebral conjunctival images) was used for the test set. Specific information about the dataset is shown in Table 1.
[0051] Table 1. Types and Quantities of Palpebral Conjunctiva Datasets
[0052]
[0053] 3. Transfer learning
[0054] Transfer learning is a machine learning method that aims to improve the performance of a target task by transferring learned knowledge from one task or domain to another. It utilizes knowledge and experience learned from the source task or domain to assist the learning process in the target task or domain. A schematic diagram of transfer learning is shown below. Figure 4 As shown.
[0055] Since the amount of data in this experiment is relatively small, transfer learning can be used to apply the parameters from the source task to this experiment. This can help avoid overfitting to some extent, help the model better learn the features and patterns of this experiment, and improve the classification performance.
[0056] 4. ResNeXt50 model
[0057] ResNet (Residual Network) is a deep residual network architecture that introduces residual connections to build deep networks, achieving significant breakthroughs. ResNeXt (ResidualNext) builds upon ResNet by introducing the concept of cardinality to enhance the network's expressive power. The core idea of the ResNeXt model is to add grouped convolutions within the residual blocks, dividing the input feature maps into multiple groups and performing convolution operations within each group. This increases the model's non-linear representation capabilities while maintaining computational efficiency.
[0058] In ResNeXt50, 50 indicates the model's depth, meaning the network contains 50 layers. ResNeXt50 is a medium-sized model that achieves good performance on many computer vision tasks and requires less training and inference time and fewer resources compared to larger models.
[0059] 5. Attention Mechanism
[0060] Attention mechanisms are a technique in deep learning used to enhance a model's ability to focus on input data. By assigning different weights to different inputs, the model can selectively focus on information useful for the current task.
[0061] The attention mechanism presented in this paper is SE (Squeeze-and-Excitation), an attention mechanism used to enhance the performance of convolutional neural networks. It improves the expressive power of features by adaptively adjusting the attention between channels. The SE module introduces a lightweight attention mechanism into the network that can automatically learn the importance between channels, thereby enhancing useful features and suppressing irrelevant ones.
[0062] The SE module consists of two key steps: Squeeze and Excitation.
[0063] Squeeze: In this step, the SE module reduces the dimensionality of the input feature map using a global average pooling (GAP) operation, converting the feature of each channel into a single numerical value. This value represents the global feature importance of each channel.
[0064] Excitation: In this step, the SE module uses a small fully connected network containing one hidden layer and a sigmoid activation function. This network takes the feature importance output from the Squeeze step as input, and outputs a channel attention weight vector after a series of fully connected operations. This vector is used to adjust the importance of each channel in the input feature map.
[0065] Finally, the SE module performs element-wise multiplication with the input feature map using channel attention weights to obtain a weighted feature map output. This strengthens important feature channels and suppresses irrelevant ones, thereby improving the network's expressive power.
[0066] The introduction of the SE module helps the model automatically learn the relationships between feature channels and dynamically adjust the importance of channels according to the needs of the task. This attention mechanism makes the model more flexible and adaptive, improving the discriminative power of features and thus achieving significant performance improvements in various computer vision tasks.
[0067] 5. Dual-pooling fused SE module (DPFSE)
[0068] This paper improves upon the traditional SE attention module by fusing adaptive average pooling and adaptive max pooling to compress the channel dimension and weight the importance of the input feature map. By learning the importance of different channels in the feature map and fusing the information from the two pooling operations, the model's expressive power and performance are enhanced. The improved dual-pooling fusion SE module (DPFSE) is shown below. Figure 5 As shown, the DPFSE module consists of a fusion layer of adaptive max pooling and adaptive average pooling, a fully connected layer, ReLU, a fully connected layer, and a Sigmoid layer connected in sequence.
[0069] The dual-pooling fusion SE module simultaneously applies adaptive average pooling and adaptive max pooling, comprehensively considering the average and maximum values in the feature maps to obtain a more comprehensive channel representation. Adaptive average pooling smooths the overall features, helping to suppress the influence of noise and outliers and extract more stable and robust features. Adaptive max pooling highlights the most salient parts of the feature maps, giving stronger emphasis to key features. By fusing features from both pooling operations, this module improves the expressive power of the channel dimension, helping to capture richer and more discriminative features.
[0070] 6. Dual-pooling fused SE residual module (DPFSE-ResBlock)
[0071] Based on prior knowledge and conclusions from previous papers, it is clear that the classification of palpebral conjunctiva is mainly related to the channel attention mechanism. Therefore, this paper uses resnext50 as the main architecture and introduces an improved DPFSE module on the basis of the original residual block. This module can better capture the correlation and importance between channels in the feature map, improving the model's expressive power and performance. This invention integrates the DPFSE module and the residual module to form a dual-pooling fusion SE residual module (DPFSE-ResBlock), creating a deeper module and thus achieving a new breakthrough in the classification ability of the network model. The dual-pooling fusion SE residual module is as follows: Figure 6 As shown.
[0072] 7. Dual-DPFSE-ResNext50
[0073] In addition, an SE module is introduced before the first residual block of ResNeXt, which strengthens the importance and attention of channels in the initial feature extraction stage of the network. It optimizes the importance of channels in the feature map from a global perspective, making the network pay more attention to the channel information that is key to the classification task, thereby improving the overall classification performance.
[0074] The network model in this paper is as follows: Figure 7 As shown, the Dual-DPFSE-ResNext50 module consists of an input layer, a convolutional layer, a pooling layer, an SE layer, multiple DPFSE residual modules, adaptive average pooling, and a fully connected layer connected in sequence.
[0075] Specific experimental data
[0076] In this experiment, the optimizer was Adam, the learning rate was 0.0001, the batch size was 32, and the number of iterations was 500.
[0077] Comparative experiments of different data augmentation methods
[0078] To verify the effectiveness of the data balancing method, each image in the original dataset was randomly augmented using any of the four methods proposed earlier, i.e., four new images were generated for each image. The augmented dataset is shown in Table 2:
[0079] Table 2 Comparison of datasets using different data augmentation methods
[0080]
[0081] The experimental results of data balancing were compared with those of data augmentation. Table 3 lists the accuracy of different preprocessing methods.
[0082] Table 3 Comparison of results from different data augmentation methods
[0083]
[0084] Table 2 clearly shows that data augmentation generates more training data than data balancing. More training data provides more sample information, helping the model learn features and patterns better, thereby improving its generalization ability and accuracy. However, the final result of data augmentation with more datasets in Table 3 is not as good as data balancing, with an accuracy nearly 1% lower.
[0085] The impact of different attention mechanism modules on the network
[0086] To better illustrate the improvement effect of the network structure on the classification performance, multiple comparative experiments were conducted using a balanced dataset with different numbers and types of attention modules. The experimental parameters were kept basically consistent, and the experimental results are shown in Table 4. In Table 4, the first "+" indicates that an attention mechanism module is added to the residual block, the second "+" indicates that an attention mechanism module is added before the first residual block, and "null" indicates that no attention mechanism module is added at the corresponding position.
[0087] Table 4 Comparison of results for different attention mechanism modules
[0088]
[0089] As shown in Table 4, adding the SE attention mechanism before the first residual block improves classification performance compared to not adding an attention mechanism. The improved dual-fusion SE attention residual module achieves the highest classification accuracy among all residual modules involved in the experiment. Furthermore, the network structure presented in this paper also achieved the highest classification accuracy among the multiple attention mechanism modules tested in this experiment.
[0090] Comparison experiment with other network structures
[0091] The network model used in this experiment was compared with other classic network models on this dataset. The experimental parameters were kept basically the same (except that the batch size of ResNeXt101 was changed from 32 to 16 due to insufficient GPU memory). The classification results of each model on the dataset are shown in Table 5. Compared with other methods, the method used in this experiment achieved the highest classification accuracy and the most reliable performance on the dataset.
[0092] Table 5 Comparison of results for different network structures
[0093]
[0094]
[0095] To more intuitively see the classification effect, such as Figure 8 The confusion matrix is used to evaluate classification performance. In the confusion matrix, the darker the color of each cell, the larger the number of cells. Figure 8 In a confusion matrix, rows represent the classes predicted by the model, and columns represent the actual classes. The cells on the diagonal of the confusion matrix represent samples where the predicted class and the actual class are the same, i.e., correctly classified.
[0096] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.
Claims
1. A non-invasive anemia classification method based on palpebral conjunctival images, characterized in that: Includes the following steps, The SE attention module is improved by fusing adaptive average pooling and adaptive max pooling to compress the channel dimension and weight the importance of the input feature map, thus obtaining the DPFSE module. Using resnext50 as the main architecture, a DPFSE module is introduced on the basis of the residual block to obtain the DPFSE-ResBlock module; An SE module was introduced before the first residual block to obtain the final Dual-DPFSE-ResNext50 network model; Using the constructed Dual-DPFSE-ResNext50 anemia analysis model, palpebral conjunctival images were classified through the anemia analysis model to obtain anemia classification results; The DPFSE module consists of a fusion layer of adaptive max pooling and adaptive average pooling, a fully connected layer, ReLU, a fully connected layer, and a Sigmoid, connected in sequence.
2. The non-invasive anemia classification method based on palpebral conjunctival images according to claim 1, characterized in that: Obtain palpebral conjunctival image data from the publicly available EYES-DEFY-ANEMIA dataset and generate a dataset, which is then divided into a training set and a test set.
3. The non-invasive anemia classification method based on palpebral conjunctival images according to claim 2, characterized in that: The palpebral conjunctival image dataset was classified using a two-level assessment principle: when hemoglobin was >= 11.5 g / dl, it was marked as normal, and when hemoglobin was < 11.5 g / dl, it was marked as anemia.
4. The non-invasive anemia classification method based on palpebral conjunctival images according to claim 3, characterized in that: The palpebral conjunctival image dataset was augmented using a data balancing method to obtain more data and to make the data labeled as normal and labeled as anemic close to or equal.
5. The non-invasive anemia classification method based on palpebral conjunctival images according to claim 4, characterized in that: The data balancing augmentation methods employed include random rotation, random scaling and cropping, horizontal flipping, vertical flipping, random perspective transformation, and random affine transformation.
6. The non-invasive anemia classification method based on palpebral conjunctival images according to claim 1, characterized in that: The Dual-DPFSE-ResNext50 module consists of an input layer, a convolutional layer, a pooling layer, an SE layer, multiple DPFSE residual modules, adaptive average pooling, and a fully connected layer connected in sequence.