A medical image classification method and system based on deep learning
By employing a data augmentation strategy that combines dual sampling and regrouping with adaptive convolutional feature extraction, and incorporating a two-branch network architecture, the problem of imbalanced class distribution in medical image classification was solved, improving the model's stability and minority class recognition ability, and achieving high-precision medical image classification.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- THE SECOND XIANGYA HOSPITAL OF CENT SOUTH UNIV
- Filing Date
- 2026-03-06
- Publication Date
- 2026-06-05
AI Technical Summary
Existing medical image classification methods struggle to effectively identify minority class samples when faced with imbalanced class distributions. They also fail to adequately represent key lesion areas and fine-grained structural features, resulting in low classification accuracy and limited model generalization ability.
A data augmentation strategy combining hybrid double sampling and regrouping is adopted, which combines a region adaptive convolution module and a two-branch network architecture. Through adaptive convolution kernel morphology selection and feature fusion, the expression of key lesion regions and fine-grained structural features is improved, and the utilization of features at different scales is enhanced.
It improves the stability and generalization performance of medical image classification models, enhances the ability to identify minority class samples and improves classification accuracy, and is applicable to various types of medical image datasets.
Smart Images
Figure CN122156762A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical image processing and artificial intelligence technology, specifically to a deep learning-based medical image classification method, system, and computer-readable storage medium. Background Technology
[0002] With the continuous improvement of medical imaging equipment and information technology, medical images have been widely used in clinical scenarios such as disease screening, auxiliary diagnosis, efficacy evaluation, and follow-up management. Common types of medical images include, but are not limited to, magnetic resonance imaging (MRI), computed tomography (CT), X-ray images, ultrasound images, and pathological slide images. Automatic classification and recognition of medical images can help improve diagnostic efficiency, reduce the workload of doctors, and, to some extent, enhance the consistency and objectivity of diagnostic results.
[0003] Currently, surgeons typically rely on experience to assess the amount of bleeding during surgery, a method that is labor-intensive and requires extensive expertise. In recent years, medical image classification methods based on machine learning and deep learning have rapidly developed. Existing methods usually employ convolutional neural networks, transfer learning models, or multi-feature fusion models to extract features and classify medical images, such as those from bronchoscopy, achieving some success on various public datasets and in specific application scenarios. However, in actual clinical applications, medical image data often suffers from significant class imbalance, meaning large differences in the number of samples across different disease categories or pathological grades; insufficient expression of key lesion regions and fine-grained structural features leading to low classification accuracy; and inadequate utilization of features at different scales and semantic levels. These issues have become significant factors restricting the performance and reliability of medical image classification.
[0004] Specifically, in real clinical data, the number of normal samples or samples with common diseases is usually far greater than that of samples with rare diseases, early lesions, or specific subtypes. For example, in tumor imaging analysis, the number of benign samples is often significantly greater than that of malignant samples; in graded diagnosis tasks, the number of high-grade or rare subtype samples is limited. This imbalanced data characteristic causes existing classification models to tend to learn the features of the dominant class during training, resulting in a significant decrease in their ability to identify minority class samples, manifested as high false negative rates, low recall rates, or significant classification bias.
[0005] To mitigate the impact of imbalanced data distribution, existing technologies have proposed several improvements, such as reconstructing training data through oversampling, undersampling, or synthesizing minority class samples, or introducing weighted loss functions during model training to enhance attention to minority class samples. However, these methods still have certain limitations in the field of medical imaging: on the one hand, simple data resampling methods are prone to introducing noisy samples or causing model overfitting, especially when medical image samples are costly to obtain and difficult to annotate; on the other hand, loss function adjustment based on fixed weights usually relies on manual experience and is difficult to adaptively adjust according to the actual distribution differences and classification difficulty between different categories, thus limiting the model's generalization ability.
[0006] Furthermore, medical images are typically characterized by high dimensionality, high noise, small lesion areas, and significant morphological differences. Minority class samples often suffer from insufficient feature expression, large intra-class variability, and blurred inter-class boundaries. Under imbalanced distribution conditions, existing classification methods struggle to maintain overall classification accuracy while effectively identifying minority class samples, impacting the safety and reliability of the model in clinical applications.
[0007] Therefore, there is an urgent need for a classification method for imbalanced medical image data that can effectively mitigate the impact of class imbalance on model training and classification results without significantly increasing the cost of data acquisition and annotation. This method should address the issues of insufficient expression of key lesion regions and fine-grained structural features leading to low classification accuracy, and insufficient utilization of features at different scales and semantic levels. It should also improve the recognition ability and classification stability of minority class medical images, providing reliable technical support for preoperative bleeding estimation in clinical practice, thereby better meeting the actual needs of intelligent medical image analysis in clinical applications. Summary of the Invention
[0008] To address the problems in existing medical image classification methods, such as insufficient ability to identify minority class samples, inadequate expression of key lesion regions and fine-grained structural features leading to low classification accuracy, and insufficient utilization of features at different scales and semantic levels, this invention proposes a deep learning-based medical image classification method and system. By constructing a data augmentation and feature learning mechanism oriented towards imbalanced distributions, the method improves the classification accuracy and stability of the medical image classification model in complex clinical data environments.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based medical image classification method, characterized by the following steps: S1: Obtain the medical image dataset to be trained, wherein the medical image dataset includes at least majority class samples and minority class samples with imbalanced sample size distribution; S2: Perform hybrid double sampling and regrouping data augmentation on the medical image dataset to construct a training dataset with a relatively balanced class distribution. The hybrid double sampling includes at least instance sampling and class sampling, and synthetic training samples are generated based on the sampling results. S3: Input the data augmented training dataset into the medical image classification network, use the region adaptive convolution module to perform block modeling of medical images, and adaptively select the convolution kernel shape according to the characteristics of different image regions to extract local spatial features. S4: Based on a two-branch network architecture that includes a local feature extraction branch and a global dependency modeling branch, the local spatial features and global context features are extracted and fused in parallel; S5: Perform classification prediction on the fused features, output the medical image classification result corresponding to the regrouped category, and aggregate the prediction results of the regrouped category in the inference stage to obtain the final classification result of the original category.
[0010] Preferably, the hybrid dual sampling includes: performing instance sampling based on the frequency of occurrence of medical image samples in the original dataset to preserve the original data distribution characteristics; and performing class sampling based on a uniform class distribution to increase the sampling probability of minority class samples.
[0011] Preferably, the synthetic training samples are obtained by performing a MixUp operation on instance samples and category samples, wherein the MixUp operation includes linear interpolation of sample features and corresponding labels.
[0012] Preferably, the data augmentation process of the regrouping includes: performing feature clustering on the majority class samples whose sample number ratio is higher than a preset threshold, splitting the majority class samples into multiple pseudo-classes; keeping the original class labels of the minority class samples unchanged, thereby transforming the original imbalanced classification task into a multi-class classification task with a relatively balanced sample number.
[0013] Preferably, during the inference stage, the predicted probabilities of multiple pseudo-classes belonging to the same original majority class are weighted or summed and aggregated to obtain the final prediction result of the corresponding original majority class.
[0014] Preferably, the region adaptive convolution module includes: performing adaptive pooling on the input medical image or feature map to divide it into multiple image blocks; predicting corresponding convolution kernel parameters for each image block, and selecting convolution kernels of different sizes or shapes to perform convolution operations on the image blocks according to the prediction results; the convolution kernel parameters include at least height parameters and width parameters, which are used to characterize the texture features or geometric properties of the image blocks.
[0015] Preferably, the two-branch network architecture includes: a local feature extraction branch for extracting local spatial features of medical images based on region adaptive convolution; and a global dependency modeling branch for extracting long-range contextual features of medical images based on state-space models or sequence modeling structures.
[0016] Preferably, the local and global features are fused through channel splicing and channel shuffling operations, and the fused features are output through residual connection.
[0017] A deep learning-based medical image classification system, characterized by comprising: The data processing module is used to perform hybrid double sampling and regrouping data augmentation on medical image datasets; The feature extraction module is used to extract local features from medical images based on region adaptive convolution; The feature fusion module is used to fuse local spatial features and global context features through a two-branch network architecture. The classification and prediction module is used to output the classification results of medical images and generate the final prediction.
[0018] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the deep learning-based medical image classification method described in any of the preceding claims.
[0019] Compared with the prior art, the present invention has the following beneficial effects: By introducing a data augmentation strategy oriented towards class imbalance, the problem of majority class samples dominating the training process is effectively alleviated, and the recognition ability of minority class medical images is improved. An adaptive convolutional feature extraction mechanism enhances the expression of key lesion regions and fine-grained structural features in medical images, helping to improve classification accuracy. The two-branch network structure and multi-dimensional feature fusion method enhance the model's ability to comprehensively utilize features at different scales and semantic levels, improving the model's stability and generalization performance. The method and system are applicable to various types of medical images and datasets with different degrees of imbalance, demonstrating good versatility and clinical application potential. Attached Figure Description
[0020] Figure 1 This is a flowchart of a deep learning-based medical image classification method in an embodiment of the present invention; Figure 2 This is a structural diagram of a hybrid double sampling data augmentation method for medical image classification based on deep learning, as described in an embodiment of the present invention. Figure 3 This is a structural diagram of a regrouping data augmentation method for a deep learning-based medical image classification method in an embodiment of the present invention; Figure 4 This is a structural diagram of an adaptive convolutional network for a deep learning-based medical image classification method in an embodiment of the present invention. Figure 5 This is a model framework diagram of a deep learning-based medical image classification method in an embodiment of the present invention. Detailed Implementation
[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0022] The purpose of this invention is to address the problems that existing medical image classification methods often suffer from, such as insufficient ability to identify minority class samples, inadequate expression of key lesion regions and fine-grained structural features leading to low classification accuracy, and insufficient utilization of features at different scales and semantic levels. This invention proposes a deep learning-based medical image classification method and system, which improves the classification accuracy and stability of medical image classification models in complex clinical data environments by constructing a data augmentation and feature learning mechanism oriented towards imbalanced distributions.
[0023] To achieve the above objectives, see Appendix Figure 1-5 The present invention adopts the following technical solution: A deep learning-based medical image classification method, characterized by the following steps: S1: Obtain the medical image dataset to be trained, wherein the medical image dataset includes at least majority class samples and minority class samples with an imbalanced sample size distribution.
[0024] Specifically, in one embodiment of the present invention, a medical image dataset for model training is first acquired. The medical image dataset is obtained from medical imaging equipment or an existing medical image database, and the image types may include, but are not limited to, CT images, MRI images, X-ray images, ultrasound images, or pathological slide images; specifically, it may be bronchial endoscopy images.
[0025] The medical image dataset can be represented as: in, This represents the i-th medical image sample; This indicates the category label corresponding to the medical image sample; This indicates the total number of categories of medical images; This indicates the total number of samples.
[0026] Statistical analysis was performed on the number of samples in each category of the medical image dataset to obtain the results for each category. Corresponding sample size ,Right now: In practical applications of medical imaging, due to factors such as differences in disease incidence rates and uneven sample collection difficulties, the following relationships often exist: in, The class with the larger number of samples is defined as the majority class. The category with a smaller sample size is defined as the minority class.
[0027] As a result, the medical image dataset exhibits a clear imbalance in class distribution.
[0028] S2: Perform hybrid double sampling and regrouping data augmentation on the medical image dataset to construct a training dataset with a relatively balanced class distribution. The hybrid double sampling includes at least instance sampling and class sampling, and synthetic training samples are generated based on the sampling results.
[0029] Preferably, the hybrid dual sampling includes: performing instance sampling based on the frequency of occurrence of medical image samples in the original dataset to preserve the original data distribution characteristics; and performing class sampling based on a uniform class distribution to increase the sampling probability of minority class samples.
[0030] Specifically, in this embodiment, see Appendix Figure 2 Structure diagram and appendix of the hybrid double sampling data augmentation method Figure 3 The structure diagram of the data augmentation method involving regrouping is shown. A multi-hybrid double sampling and regrouping data augmentation method is designed, which includes: S201, for the original training dataset To address the severe class imbalance issue present in the data, a class-aware sampling mechanism is first constructed to guide the MixUp process, aiming to simultaneously preserve the original data distribution and balance samples across classes. This mechanism is achieved by using both instance sampling and class sampling methods simultaneously. Instance sampling This sampling method strictly follows the original frequency of occurrence of the samples. Selection is made to preserve the original data distribution characteristics. Its sampling probability is related to the category. Number of samples Proportional: in, Indicate category The number of samples, This represents the total number of samples in the training set. This sampling operation aims to preserve the distribution information of the original data when selecting samples. .
[0031] Category sampling This sampling method uses a uniform distribution, ensuring that all categories have the same probability of being selected, thus effectively oversampling the minority classes to achieve class balance. ; in, This represents the total number of classes in the dataset. This sampling operation aims to force a balance between classes from which samples are selected. By combining these two sampling mechanisms, it is ensured that the subsequent MixUp process can obtain samples from both the original distribution and the balanced distribution simultaneously, achieving the best compromise.
[0032] Preferably, the synthetic training samples are obtained by performing a MixUp operation on instance samples and category samples, wherein the MixUp operation includes linear interpolation of sample features and corresponding labels.
[0033] S202: For the instance sampling samples obtained in step S101 and category sampling samples A linear interpolation operation using the MixUp strategy is then performed on the samples to generate the final synthetic samples used for training. The MixUp mechanism creates smooth transition samples in the feature space through a linear combination of features and labels, thus achieving the dual effects of data augmentation and regularization. in, and They represent sampling along the instance. The obtained sample features and labels; and They represent sampling along the category respectively. The obtained sample features and labels; It is a mixing coefficient used to control the proportion of interpolation; this coefficient is derived from... Extracted from the distribution, where It is the regularization strength parameter, which is preferred. The generated It is a linear combination of the features of two samples, while It's a soft mix of their labels. Through and Through this generation, we obtained a synthetic training set that both exhibits class balance tendencies and preserves information about the original data manifold. The final model On this dataset, by minimizing the loss function Conduct training: Where L uses the cross-entropy loss function.
[0034] Preferably, the data augmentation process of the regrouping includes: performing feature clustering on the majority class samples whose sample number ratio is higher than a preset threshold, splitting the majority class samples into multiple pseudo-classes; keeping the original class labels of the minority class samples unchanged, thereby transforming the original imbalanced classification task into a multi-class classification task with a relatively balanced sample number.
[0035] Preferably, during the inference stage, the predicted probabilities of multiple pseudo-classes belonging to the same original majority class are weighted or summed and aggregated to obtain the final prediction result of the corresponding original majority class.
[0036] S203: After data augmentation, in order to further optimize the minority class recognition performance during the model training phase, we adopt a regrouping strategy to regroup the original imbalanced data. Classification tasks Transformed into a structurally balanced new Classification tasks This strategy focuses on handling the majority class, where the sample size is overwhelmingly dominant. .
[0037] First, according to the majority class and minority The ratio of the number of samples between them is used to calculate the number of pseudo-classes into which the majority class needs to be split. . The selection aims to ensure that the number of samples in the pseudo-classes after splitting is roughly equal to the number of samples in the minority classes, thus achieving a balance in the task. in, This represents the total number of samples in the majority class. This represents the total number of samples in the minority class. Subsequently, for all samples belonging to the majority class... samples The features are mapped to a feature space, and the K-means clustering algorithm is applied in this space to divide the majority class samples into two groups. Pseudo-categories After the partitioning, the original labels of the majority class samples... It will be replaced with a new pseudo-category label. Minority class samples The labels remain unchanged. New classification task. Total number of categories Equal to the original minority class number Plus A pseudo-category, namely Through this step, we successfully transformed the imbalanced classification task into a balanced multi-class classification task, and obtained new labels. Defined as: Ultimately, the model Use the new tag set and enhanced data exist The model is trained on each category. During the inference phase, the model outputs probabilities. They will be aggregated to obtain the predicted probabilities of the original majority class. : This method effectively improves the classifier's learning resolution and recognition ability for minority class features.
[0038] S3: Input the data-augmented training dataset into the medical image classification network, use the region adaptive convolution module to perform block modeling of medical images, and adaptively select the convolution kernel shape according to the characteristics of different image regions to extract local spatial features.
[0039] Preferably, the region adaptive convolution module includes: performing adaptive pooling on the input medical image or feature map to divide it into multiple image blocks; predicting corresponding convolution kernel parameters for each image block, and selecting convolution kernels of different sizes or shapes to perform convolution operations on the image blocks according to the prediction results; the convolution kernel parameters include at least height parameters and width parameters, which are used to characterize the texture features or geometric properties of the image blocks.
[0040] S301, in the medical image feature extraction framework, the ADConv module employs a block-based adaptive strategy for feature modeling to address the heterogeneity of features across different anatomical structures and pathological tissues. First, for the input medical image data... By using adaptive average pooling, the feature map is downsampled to a preset block resolution for differentiated feature extraction, fully taking into account the morphological differences and texture characteristics of different regions in medical images.
[0041] S302, for the selection of convolutional kernels for each image patch, a convolutional kernel size prediction network was designed to achieve adaptive kernel shape decision-making. See Appendix. Figure 4 This is a structural diagram of an adaptive convolutional network for a deep learning-based medical image classification method in an embodiment of the present invention.
[0042] First, block-level features are extracted using adaptive pooling and convolutional layers: Then output the height and width parameters: Here, h and w represent the height and width parameters, respectively, both ranging from 0 to 1. Based on different combinations of these two parameter values, the most suitable convolutional kernel is automatically selected for each image patch according to preset decision rules.
[0043] S303 employs a separable convolutional structure for efficient feature extraction, given a selected convolutional kernel. Separable convolution decomposes the traditional h×w convolution into two steps: depthwise convolution and pointwise convolution. Its core operation formula is as follows: Where X represents the input features of the current image patch, DWConv represents the depthwise convolution operation, and BN represents batch normalization. and represents the intermediate features in the vertical and horizontal directions, respectively, and Y represents the output feature of the image patch.
[0044] S4: Based on a two-branch network architecture that includes a local feature extraction branch and a global dependency modeling branch, the local spatial features and global context features are extracted and fused in parallel.
[0045] Preferably, the two-branch network architecture includes: a local feature extraction branch for extracting local spatial features of medical images based on region adaptive convolution; and a global dependency modeling branch for extracting long-range contextual features of medical images based on state-space models or sequence modeling structures.
[0046] Specifically, see Appendix Figure 5 This is a model framework diagram of a deep learning-based medical image classification method in an embodiment of the present invention. A medical image feature extraction and fusion classification method based on a two-branch network architecture is designed, which includes: S401, First, the augmented dataset, after preprocessing with a hybrid double sampling and regrouping strategy, is... As the input to the network. The backbone architecture of the entire network. Composed of a series of two-branch hybrid modules Cascaded configuration, its structure can be represented as These hybrid modules are the core units for feature extraction; each module receives a feature map. Internally, it includes local feature extraction and long-range dependency modeling branches, and through a tight feature fusion mechanism, it effectively captures multi-scale information from medical images. Represents the input image. The input feature map representing the module. , and These represent the height, width, and number of channels of the feature map, respectively.
[0047] S402, then, feature extraction and fusion are performed in each two-branch hybrid module. Internal execution. The fusion process follows these core steps: Input feature map First, it is split into channels. They are divided into two groups, and split along the channel dimension. and The number of channels in each group is 1 . Subsequently, these two sets of feature maps are input in parallel into their respective branches for processing. The local feature extraction branch employs region-adaptive convolution. Capture fine local morphological features in an adaptive manner. The long-range dependency modeling branch utilizes a state-space model. Efficiently capture global contextual information of images. . Preferably, the local and global features are fused through channel splicing and channel shuffling operations, and the fused features are output through residual connection.
[0048] Specifically, after parallel processing is completed, the outputs of the two branches... and Through channel splicing Restore to the original number of channels To obtain splicing features The channel shuffling operation then proceeds. This breaks down the feature information isolation caused by channel grouping processing, ensuring full cross-branch information interaction and intrinsic fusion between local and global features, and outputting fused features: in, and These represent the local and global feature subsets obtained from the splitting process, respectively. and These represent the output characteristics after processing by the ADConv and SSM modules, respectively. This indicates the final merged feature after the channel shuffling operation.
[0049] S403, finally, the feature map obtained after channel shuffling and fusion. The original input of the module is connected via a residual connection. Perform element-level addition fusion as the final output of this module. : After multi-level mixing modules After layer-by-layer deep fusion, the network ultimately outputs a single, highly fused feature representation. .
[0050] S5: Perform classification prediction on the fused features, output the medical image classification result corresponding to the regrouped category, and aggregate the prediction results of the regrouped category in the inference stage to obtain the final classification result of the original category.
[0051] Specifically, the final feature map is then processed by global average pooling. Perform spatial dimensionality reduction and operate through a fully connected layer. Output the predicted probabilities for the balanced multi-class classification task after regrouping: in, This represents the final output feature map of the network backbone; This represents the final predicted probability vector, whose dimension is... This is the total number of categories after regrouping; Used to convert the output of a fully connected layer into a probability distribution. This prediction It achieves accurate identification of balanced multi-class tasks set by the regrouping strategy, thereby significantly improving the model's identification performance for minority classes.
[0052] To verify the effectiveness of the proposed method in the bronchoscopic bleeding classification task, several mainstream deep learning models were selected for comparative experiments, including ViT, MedMamba, and ConvNext. All experiments were conducted on a personal computer with 64GB of RAM and one NVIDIA GeForce RTX 3090 GPU (24GB VRAM). During training, all input images were uniformly resized to 224×224×3. The AdamW algorithm was chosen as the optimizer, and cross-entropy loss was used as the loss function for all models. The initial learning rate of AdamW was set to 0.0001, B1 to 0.9, B2 to 0.999, and weight decay to 1×10⁻⁶. -4All models were trained for 100 epochs with a batch size of 64, and an early stopping strategy was employed to prevent overfitting. All training was implemented within the PyTorch framework, and the input size for all comparison methods was uniformly set to 224×224.
[0053] The experiment used overall classification accuracy (OA), precision, sensitivity, specificity, and F1 score as evaluation metrics. Sensitivity reflects the model's ability to identify bleeding samples, specificity reflects the model's ability to identify non-bleeding samples, and the F1 score comprehensively measures both precision and sensitivity. Table 1 shows the classification performance of different methods on the bronchoscopic hemorrhage dataset.
[0054] Table 1: Classification performance of different classification methods on the bronchoscopic hemorrhage dataset As shown in Table 1, the method proposed in this invention achieves optimal classification performance across all evaluation metrics. Specifically, the overall classification accuracy reaches 74.87%, and the precision reaches 71.60%, significantly outperforming other methods. In terms of sensitivity and specificity, the proposed method achieves 69.96% and 84.61%, respectively, far exceeding other comparative methods, indicating that the method can accurately identify both bleeding and non-bleeding samples simultaneously. Furthermore, the F1 score of the method reaches 70.71%, an improvement of 16.29% compared to the VMamba method. These experimental results validate the effectiveness and superiority of the method proposed in the bronchoscopic hemorrhage classification task.
[0055] To further verify the contribution of each module proposed in this invention to classification performance, ablation experiments were conducted on a bronchoscopic hemorrhage dataset, and the results are shown in Table 2. The experiments examined the impact of the data processing module and the adaptive convolution module on model performance.
[0056] Table 2: Ablation experimental results on the bronchoscopic hemorrhage dataset As shown in Table 2, without data processing and adaptive convolution, the model's overall classification accuracy is only 54.37%, and the F1 score is only 40.26%. After introducing the data processing module, the overall classification accuracy improves to 71.98%, and the F1 score improves to 67.94%, indicating that the data processing module can effectively improve the model's feature learning ability for bronchoscopic images. Furthermore, after introducing the adaptive convolution module, the overall classification accuracy reaches 74.87%, and the F1 score reaches 70.71%, with all indicators reaching their optimal levels. The above ablation experiment results demonstrate that both the data processing module and the adaptive convolution module proposed in this invention significantly contribute to the improvement of classification performance.
[0057] A deep learning-based medical image classification system, characterized by comprising: The data processing module is used to perform hybrid double sampling and regrouping data augmentation on medical image datasets; The feature extraction module is used to extract local features from medical images based on region adaptive convolution; The feature fusion module is used to fuse local spatial features and global context features through a two-branch network architecture. The classification and prediction module is used to output the classification results of medical images and generate the final prediction.
[0058] A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the deep learning-based medical image classification method described in any of the preceding claims.
[0059] In summary, this embodiment proposes a deep learning-based medical image classification method and system. By employing a hybrid double sampling and regrouping strategy and a two-branch network architecture, it achieves accurate classification of bleeding images from imbalanced medical image datasets. First, the method creates a balanced training distribution through a hybrid double sampling and regrouping data augmentation approach, effectively mitigating the impact of sample imbalance on model training. Second, an adaptive convolutional module is designed to extract local spatial features of the target region, enhancing the model's ability to perceive key lesion areas. Finally, a two-branch network architecture fuses features from different dimensions, further improving the accuracy of classification decisions. Experimental results demonstrate that the proposed method achieves optimal classification performance on a bronchoscopic bleeding image classification dataset, validating its superiority in medical image classification tasks and providing reliable technical support for clinicians to predict the risk of bleeding during bronchoscopy.
[0060] Compared with the prior art, the present invention has the following beneficial effects: By introducing a data augmentation strategy oriented towards class imbalance, the problem of majority class samples dominating the training process is effectively alleviated, and the recognition ability of minority class medical images is improved. An adaptive convolutional feature extraction mechanism enhances the expression of key lesion regions and fine-grained structural features in medical images, helping to improve classification accuracy. The two-branch network structure and multi-dimensional feature fusion method enhance the model's ability to comprehensively utilize features at different scales and semantic levels, improving the model's stability and generalization performance. The method and system are applicable to various types of medical images and datasets with different degrees of imbalance, demonstrating good versatility and clinical application potential.
[0061] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A medical image classification method based on deep learning, characterized in that, Includes the following steps: S1: Obtain the medical image dataset to be trained, wherein the medical image dataset includes at least majority class samples and minority class samples with imbalanced sample size distribution; S2: Perform hybrid double sampling and regrouping data augmentation on the medical image dataset to construct a training dataset with a relatively balanced class distribution. The hybrid double sampling includes at least instance sampling and class sampling, and synthetic training samples are generated based on the sampling results. S3: Input the data augmented training dataset into the medical image classification network, use the region adaptive convolution module to perform block modeling of medical images, and adaptively select the convolution kernel shape according to the characteristics of different image regions to extract local spatial features. S4: Based on a two-branch network architecture that includes a local feature extraction branch and a global dependency modeling branch, the local spatial features and global context features are extracted and fused in parallel; S5: Perform classification prediction on the fused features, output the medical image classification result corresponding to the regrouped category, and aggregate the prediction results of the regrouped category in the inference stage to obtain the final classification result of the original category.
2. The method according to claim 1, characterized in that, The hybrid dual sampling includes: Instance sampling is performed based on the frequency of occurrence of medical image samples in the original dataset to preserve the distribution characteristics of the original data; Class sampling is performed based on a uniform distribution of classes to increase the sampling probability of minority class samples.
3. The method according to claim 2, characterized in that, The synthetic training samples are obtained by performing a MixUp operation on instance samples and category samples. The MixUp operation includes linear interpolation of sample features and corresponding labels.
4. The method according to claim 1, characterized in that, The data augmentation process for regrouping includes: The majority class samples with a sample quantity ratio higher than a preset threshold are subjected to feature clustering, and the majority class samples are split into multiple pseudo-classes; By keeping the original class labels of minority class samples unchanged, the original imbalanced classification task is transformed into a multi-class classification task with a relatively balanced number of samples.
5. The method according to claim 4, characterized in that, During the inference phase, the predicted probabilities of multiple pseudo-classes belonging to the same original majority class are weighted or summed and aggregated to obtain the final prediction result for the corresponding original majority class.
6. The method according to claim 1, characterized in that, The region adaptive convolution module includes: The input medical image or feature map is adaptively pooled to divide it into multiple image blocks; the corresponding convolution kernel parameters are predicted for each image block, and convolution kernels of different sizes or shapes are selected to perform convolution operations on the image blocks according to the prediction results; the convolution kernel parameters include at least height parameters and width parameters, which are used to characterize the texture features or geometric properties of the image blocks.
7. The method according to claim 1, characterized in that, The two-branch network architecture includes: The local feature extraction branch is used to extract local spatial features of medical images based on region adaptive convolution; The global dependency modeling branch is used to extract long-range contextual features from medical images based on state-space models or sequence modeling structures.
8. The method according to claim 7, characterized in that, The local and global features are fused through channel splicing and channel shuffling operations, and the fused features are output through residual connection.
9. A medical image classification system based on deep learning, characterized in that, include: The data processing module is used to perform hybrid double sampling and regrouping data augmentation on medical image datasets; The feature extraction module is used to extract local features from medical images based on region adaptive convolution; The feature fusion module is used to fuse local spatial features and global context features through a two-branch network architecture. The classification and prediction module is used to output the classification results of medical images and generate the final prediction.
10. A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the deep learning-based medical image classification method according to any one of claims 1 to 9.