An ovarian cancer ultrasound image pathological subtype classification method based on deep learning

Abstract

The application relates to the technical field of medical image intelligent analysis and computer-aided diagnosis, in particular to an ovarian cancer ultrasonic image pathological subtype classification method based on deep learning, and the technical scheme points are as follows: S1, constructing an ovarian cancer ultrasonic image data set; S2, constructing an end-to-end DMSG-Net classification network: taking a MambaVision structure as a basic network, the network is divided into a shallow local feature enhancement stage, a deep multi-scale space semantic modeling stage and a classification prediction stage; a difference perception local control aggregation module is embedded in the shallow local feature enhancement stage; a multi-scale space guiding hybrid module is embedded in the deep multi-scale space semantic modeling stage, and a double statistical variance guiding normalization method is combined; S3, using the data set to optimize the DMSG-Net classification network, inputting an ovarian cancer ultrasonic image to be classified into the trained network, and obtaining a result. The application has the advantages of significant fine-grained feature enhancement, accurate multi-scale modeling, stable training and high classification precision.

Description

Technical Field

[0001] This invention relates to the technical field of intelligent analysis of medical images and computer-aided diagnosis, and more specifically, to a method for classifying pathological subtypes of ovarian cancer based on ultrasound images using deep learning. Background Technology

[0002] Ovarian cancer is a significant gynecological malignancy threatening women's health. It can be classified into pathological subtypes such as serous carcinoma, clear cell carcinoma, mucinous carcinoma, and endometrioid carcinoma. The treatment plans and prognoses differ significantly among these subtypes, making accurate preoperative subtype identification crucial for clinical support. Ultrasound examination, due to its non-invasive, real-time, cost-effective, and readily available methods, is a core imaging tool for screening and assessing ovarian lesions. However, ovarian cancer ultrasound images generally exhibit characteristics such as blurred lesion boundaries, low gray-scale contrast, significant speckle noise, and similar imaging features across subtypes. Clinical subtype determination is highly dependent on physician experience, leading to strong subjectivity and insufficient diagnostic consistency.

[0003] To address the aforementioned issues, traditional medical image classification relies on manually designed features and traditional classifiers. However, in low-contrast, structurally complex ultrasound scenarios of ovarian cancer, the feature representation capabilities are limited, resulting in insufficient classification robustness.

[0004] Among existing deep learning solutions, convolutional neural networks excel at local feature extraction but lack global modeling capabilities; Transformers have strong global correlation capabilities but require large computational resources and rely on big data; while state-space models, represented by MambaVision, possess efficient long sequence modeling capabilities, they still have three shortcomings when applied to ovarian cancer ultrasound classification: superficial layers lack differential enhancement mechanisms for low-contrast lesions, and fine-grained discriminative feature extraction is insufficient; they focus on sequence-dimensional modeling and are insufficient in characterizing the morphology of two-dimensional multi-scale lesions; and deep layer features fluctuate greatly, and conventional normalization methods are not robust enough, affecting the model's convergence and generalization performance, making it difficult to meet the clinical needs of high-precision classification. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a deep learning-based method for classifying pathological subtypes of ovarian cancer using ultrasound images. This method has the advantages of significant enhancement of fine-grained features, accurate multi-scale modeling, stable training, and high classification accuracy.

[0006] The above-mentioned technical objective of this invention is achieved through the following technical solution: a deep learning-based method for classifying pathological subtypes of ovarian cancer ultrasound images, comprising the following steps:

[0007] S1. Construct an ultrasound image dataset for ovarian cancer;

[0008] S2. Construct an end-to-end DMSG-Net classification network: The DMSG-Net classification network is based on the MambaVision structure as its backbone network and is divided into a shallow local feature enhancement stage, a deep multi-scale spatial semantic modeling stage, and a classification prediction stage.

[0009] A difference-aware local feature enhancement module is embedded in the shallow local feature enhancement stage to enhance lesion boundaries, local textures, and weak contrast structural features;

[0010] In the deep multi-scale spatial semantic modeling stage, a multi-scale spatial guided hybrid module is embedded and combined with a dual statistical variance guided normalization method to model the multi-scale spatial structure and subtype semantic differences of the lesion region and stabilize the distribution of deep features.

[0011] S3. The DMSG-Net classification network is trained and optimized using the ovarian cancer ultrasound image dataset. The ovarian cancer ultrasound images to be classified are input into the trained network, and the probability distribution of each pathological subtype is output. The category with the highest probability is taken as the final pathological subtype classification result.

[0012] In one embodiment, S1 includes the following steps:

[0013] S11. Acquire two-dimensional ultrasound images of ovarian cancer confirmed by pathology, crop the region of interest of the lesion and normalize its size, label the category according to the pathological subtype, and divide it into training set, validation set and test set according to the patient level.

[0014] S12. For pathological subtypes in the training set whose sample quantity is lower than a preset threshold, a generative amplification strategy based on StableDiffusion is adopted to generate supplementary samples that are consistent with real ultrasound images in terms of lesion texture, boundary morphology, and grayscale distribution, and these supplementary samples are added to the training set at a preset ratio.

[0015] In one embodiment, the shallow local feature enhancement stage includes two sequentially connected first feature extraction layers, which are composed of a convolutional block and the difference-aware local feature aggregation module connected in series.

[0016] The deep multi-scale spatial semantic modeling stage includes two sequentially connected second feature extraction layers, which are composed of the MambaVision Mixer structure and the multi-scale spatial guided mixing module connected in series.

[0017] In one embodiment, the processing steps of the difference-aware local control aggregation module are as follows:

[0018] Input feature map The feature distribution is stabilized by a normalization layer, and then channel projection is completed by 1×1 convolution;

[0019] Global average pooling is used to extract global statistical information. The difference between local feature responses and global statistical information is calculated. After dynamic adjustment by a learnable scaling factor, differential enhancement features are obtained through GELU nonlinear activation. , represented as:

[0020]

[0021]

[0022] Wherein, GAP represents global average pooling. ⊙ represents the learnable scaling factor, and ⊙ represents element-wise multiplication.

[0023] Enhance the difference features The local gated aggregation structure is used to perform feature aggregation to obtain gated aggregated features;

[0024] The gated aggregated features are compared with the input feature map. Establish residual connections and output enhanced local difference features.

[0025] In one embodiment, the processing steps of the localized controlled aggregation structure are as follows:

[0026] Two parallel processing branches are set up. The first branch completes linear projection through 1×1 convolution, and the second branch first captures neighborhood context information through 3×3 depthwise separable convolution, and then completes channel projection through 1×1 convolution.

[0027] The outputs of the first and second branches are activated by SiLU and then subjected to element-wise gating interaction to obtain gating interaction features. , represented as:

[0028]

[0029]

[0030]

[0031] The gated interaction features After channel fusion is completed by 1×1 convolution, the gated aggregation feature is output.

[0032] In one embodiment, the processing steps of the multi-scale spatially guided hybridization module are as follows:

[0033] A multi-branch parallel structure is adopted, with each branch corresponding to three different receptive field scales of 3, 5, and 7 respectively;

[0034] Within each scale branch, the input features The spatial response enhancement of the lesion-related region is completed by the multi-scale coordinate attention RSGA structure, and the spatially enhanced branch features are output.

[0035] Within each scale branch, the spatially enhanced branch features are processed by the dual statistical variance-guided normalization method and a multilayer perceptron to obtain single-scale semantic calibration features. ;

[0036] Multiple single-scale semantic calibration features The features are obtained by stitching along the channel dimension;

[0037] The concatenated features are fused across scales using a 1×1 convolution and then combined with the input features. Establish residual connections and output deep multi-scale spatial semantic features. , represented as:

[0038]

[0039]

[0040] Here, Concat represents concatenation along the channel dimension, and Conv represents a convolutional fusion layer.

[0041] In one embodiment, the processing steps of the multi-scale coordinate attention RSGA structure are as follows:

[0042] For the input features Horizontal pooling and vertical pooling are performed separately to extract bidirectional axial context information and complete the aggregation.

[0043] The aggregated axial features are processed by large-kernel strip convolution with directional decoupling, and then normalized by the Sigmoid function to generate spatial calibration weights. , represented as:

[0044]

[0045]

[0046] in, For horizontal pooling, For vertical pooling, For the Sigmoid function, , They are 1×k and k×1 stripe convolutions, respectively;

[0047] The spatial calibration weight With the input features Element-wise multiplication yields the spatially enhanced branch features.

[0048] In one embodiment, the processing steps of the dual statistical variance-guided normalization method are as follows:

[0049] Set up a batch normalization main branch and a statistical bootstrap branch;

[0050] The batch normalization main branch applies to the input features Perform batch normalization to output stable basic feature representations;

[0051] The statistical guidance branch extracts the input features respectively. The global mean and global variance statistics are input into the nonlinear mapping structure composed of fully connected layer, GELU activation layer and fully connected layer respectively, processed independently and then fused, and then channel-level guided weights are generated by Sigmoid function.

[0052] The global statistical features are calibrated using the channel-level guided weights and then superimposed onto the output of the batch normalization main branch as residuals to obtain the final normalized features. , represented as:

[0053]

[0054]

[0055] In one embodiment, the classification prediction stage consists of a two-dimensional global average pooling layer and a linear classification layer connected in series;

[0056] The two-dimensional global average pooling layer receives the output features from the deep multi-scale spatial semantic modeling stage and is used to compress the two-dimensional feature map into a one-dimensional global feature vector.

[0057] The linear classification layer is connected to the one-dimensional global feature vector, and its output dimension corresponds one-to-one with the number of categories of the pathological subtype.

[0058] The aforementioned deep learning-based method for classifying ovarian cancer ultrasound image pathological subtypes has the following beneficial effects:

[0059] First, we constructed an end-to-end classification network based on a customized and improved MambaVision. Through a hierarchical feature enhancement mechanism, we adapted to clinical classification needs, effectively overcame the shortcomings of existing general vision models in terms of adaptability to ovarian cancer ultrasound scenarios, significantly reduced the misclassification rate among different pathological subtypes, and comprehensively improved classification accuracy and overall discrimination performance.

[0060] Secondly, by embedding a shallow-layer difference-perception local control aggregation module, the difference calculation based on local feature response and global statistical information suppresses background interference and speckle noise, accurately enhances fine-grained discrimination features such as lesion boundaries and weak contrast textures, and effectively makes up for the problem of insufficient extraction of shallow fine-grained features;

[0061] Third, by embedding a deep multi-scale spatial guidance hybrid module, and using a multi-receptive field parallel spatial calibration and cross-scale fusion mechanism, the differences in lesion size and morphological structure of different pathological subtypes can be accurately characterized, effectively making up for the shortcomings of the original state space model in its ability to characterize two-dimensional multi-scale lesion features.

[0062] Fourth, by combining the dual statistical variance-guided normalization method, while retaining the convergence speed advantage of batch normalization, the dynamic calibration of the global mean and variance dual statistics can alleviate the problem of feature distribution fluctuation in deep networks and improve training convergence stability and model generalization performance. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the structure of the DMSG-Net classification network in an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of the structure of the difference-sensing local control aggregation module in an embodiment of the present invention;

[0065] Figure 3 This is a schematic diagram of the structure of the multi-scale spatially guided hybrid module in an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of the structure of the dual statistical variance-guided normalization method in an embodiment of the present invention. Detailed Implementation

[0067] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0068] A deep learning-based method for classifying pathological subtypes of ovarian cancer using ultrasound images, such as... Figure 1 As shown, it includes the following steps:

[0069] S1. Construct an ultrasound image dataset for ovarian cancer;

[0070] S2. Constructing an end-to-end DMSG-Net classification network: The DMSG-Net classification network is based on the MambaVision structure as its backbone network, and is divided into a shallow local feature enhancement stage, a deep multi-scale spatial semantic modeling stage, and a classification prediction stage.

[0071] A difference-aware local feature enhancement module is embedded in the shallow local feature enhancement stage to enhance lesion boundaries, local textures and weak contrast structural features;

[0072] In the deep multi-scale spatial semantic modeling stage, a multi-scale spatial guided hybrid module is embedded and combined with a dual statistical variance guided normalization method to model the multi-scale spatial structure and subtype semantic differences of lesion regions and stabilize the distribution of deep features.

[0073] S3. The DMSG-Net classification network is trained and optimized using an ovarian cancer ultrasound image dataset. The ovarian cancer ultrasound images to be classified are input into the trained network, which outputs the probability distribution of each pathological subtype. The category with the highest probability is taken as the final pathological subtype classification result.

[0074] The deep learning-based pathological subtype classification method for ovarian cancer ultrasound images provided by this invention has multiple beneficial effects. This method constructs an end-to-end customized classification network based on MambaVision, adapting to the characteristics of ovarian cancer ultrasound imaging through a hierarchical feature enhancement mechanism, effectively reducing the misclassification rate between subtypes and improving overall classification performance. A difference-aware local feature aggregation module is embedded in the shallow layer of the network, suppressing background interference through difference calculation of local and global features, and enhancing fine-grained discriminative features such as lesion boundaries and weak contrast textures. A multi-scale spatially guided hybridization module is embedded in the deep layer, accurately characterizing the morphological and structural differences of lesions in different subtypes through parallel calibration of multiple receptive fields and cross-scale fusion. Simultaneously, a dual statistical variance-guided normalization method is combined to dynamically calibrate the feature distribution with dual statistics, improving the network training stability and generalization performance.

[0075] Specifically, S1 includes the following steps:

[0076] S11. Acquire two-dimensional ultrasound images of pathologically confirmed ovarian cancer, crop the region of interest of the lesion and normalize its size, label the category according to the pathological subtype, and divide it into training set, validation set and test set according to the patient level.

[0077] In practical applications, pathological subtypes encompass four categories: serous carcinoma, clear cell carcinoma, mucinous carcinoma, and endometrioid carcinoma. Cropping the region of interest (ROI) removes background areas irrelevant to lesion identification, and then normalizes the cropped image to a preset input size. Each image is labeled with its corresponding pathological subtype. To prevent data leakage, all data is strictly divided into training, validation, and test sets according to patient level, ensuring that images of the same patient do not appear in different subsets simultaneously.

[0078] S12. For pathological subtypes with a sample size below a preset threshold in the training set, a generative amplification strategy based on Stable Diffusion is used to generate supplementary samples that are consistent with real ultrasound images in terms of lesion texture, boundary morphology, and grayscale distribution. These supplementary samples are then added to the training set at a preset ratio, forming the training data together with the real samples. This step addresses the imbalance of different case subtypes in the training set, improving the network's feature learning ability and generalization performance for a few subtypes.

[0079] Specifically, such as Figure 1 As shown, the DMSG-Net classification network uses the MambaVision structure as its backbone network and adopts a pyramid structure with four levels of downsampling. The network is divided into three stages along the data flow direction: a shallow local feature enhancement stage, a deep multi-scale spatial semantic modeling stage, and a classification prediction stage.

[0080] The shallow local feature enhancement stage comprises two sequentially cascaded first feature extraction layers. Each first feature extraction layer consists of a convolutional block and a difference-aware local feature aggregation module cascaded together. After the input image is initially downsampled by convolution to convert it into a feature map, it enters the two first feature extraction layers step by step. While completing spatial downsampling and channel upscaling, it enhances lesion boundaries, local textures, and weak contrast structural features, outputting medium-to-high resolution features with enhanced local details.

[0081] The deep multi-scale spatial semantic modeling stage comprises two sequentially connected second feature extraction layers. Each second feature extraction layer consists of a MambaVision Mixer structure and a multi-scale spatial guided mixing module, with a dual statistical variance guided normalization method used throughout to stabilize feature distribution. Features output from the shallow stage enter this stage, first undergoing long-distance sequence dependency modeling through the MambaVision Mixer structure to extract global basic semantic features. Then, the multi-scale spatial guided mixing module performs multi-scale spatial morphological calibration and cross-scale fusion, modeling the multi-scale spatial structure and subtype semantic differences of lesion regions, ultimately outputting high-level semantic features with strong class discriminative power.

[0082] The classification prediction stage consists of a two-dimensional global average pooling layer and a linear classification layer connected in series. The two-dimensional global average pooling layer receives the output features from the deep multi-scale spatial semantic modeling stage, compressing the two-dimensional feature map into a one-dimensional global feature vector, removing redundant spatial dimensions. The linear classification layer receives the one-dimensional global feature vector, and its output dimension corresponds one-to-one with the number of pathological subtypes, used to output the confidence probability of each subtype.

[0083] Specifically, the difference-perceived local control aggregation module is deployed at the shallow feature extraction level to suppress background interference and enhance lesion boundaries and weak contrast texture features, such as... Figure 2 As shown, the processing steps are as follows:

[0084] Input feature map The feature distribution is stabilized by a normalization layer, and then channel projection is completed by 1×1 convolution;

[0085] Global average pooling is used to extract global statistical information. The difference between local feature responses and global statistical information is calculated. After dynamic adjustment by a learnable scaling factor, differential enhancement features are obtained through GELU nonlinear activation. , represented as:

[0086]

[0087]

[0088] Wherein, GAP represents global average pooling. represents the learnable scaling factor, and ⊙ represents element-wise multiplication. By calculating the difference between local response and global statistical information, lesion areas that differ from the overall grayscale distribution can be explicitly highlighted, enhancing the response intensity of lesion boundaries and local abnormal textures.

[0089] Enhance the difference feature The local gated aggregation structure is input for feature aggregation to obtain gated aggregated features. The processing steps for the local gated aggregation structure are as follows:

[0090] Two parallel processing branches are configured. The first branch performs linear projection using a 1×1 convolution, while the second branch first captures neighborhood context information using a 3×3 depthwise separable convolution, and then performs channel projection using a 1×1 convolution. The outputs of the first and second branches are activated by SiLU and then subjected to element-wise gating interaction to obtain gated interaction features. , represented as:

[0091]

[0092]

[0093]

[0094] Gating interaction features After channel fusion is completed by 1×1 convolution, the gated aggregated features are output.

[0095] Gated aggregated features and input feature map Establish residual connections and output enhanced local difference features.

[0096] Specifically, the multi-scale spatially guided hybrid module is deployed at the deep feature extraction layer to model the spatial structural differences of lesions at different scales, such as... Figure 3As shown, the processing steps are as follows:

[0097] The module adopts a multi-branch parallel structure, with each branch corresponding to three different receptive field scales of 3, 5, and 7, which can simultaneously capture small-scale local structures and large-scale regional semantic information.

[0098] Within each scale branch, the input features Multi-scale coordinate attention RSGA structures enhance the spatial response of lesion-related regions. The processing steps for multi-scale coordinate attention RSGA structures are as follows:

[0099] For input features Horizontal pooling and vertical pooling are performed separately to extract bidirectional axial context information and complete the aggregation.

[0100] The aggregated axial features are processed by large-kernel strip convolution with directional decoupling, and then normalized by the Sigmoid function to generate spatial calibration weights. , represented as:

[0101]

[0102]

[0103] in, For horizontal pooling, For vertical pooling, For the Sigmoid function, , They are 1×k and k×1 stripe convolutions, respectively;

[0104] Spatial calibration weights Input features Element-wise multiplication yields spatially enhanced branching features, thus completing the response enhancement of the lesion-related region.

[0105] Within each scale branch, the spatially enhanced branch features are further stabilized using a dual statistical variance-guided normalization method before being fed into a multilayer perceptron to complete the nonlinear reorganization of channel semantics, thereby obtaining single-scale semantic calibration features. ;

[0106] Single-scale semantic calibration features of the three scale branches The features are concatenated along the channel dimension to obtain the concatenated features; the concatenated features are then fused across scales using a 1×1 convolution and combined with the input features. Establish residual connections and output deep multi-scale spatial semantic features. , represented as:

[0107]

[0108]

[0109] Here, Concat represents concatenation along the channel dimension, and Conv represents a convolutional fusion layer.

[0110] Furthermore, the dual statistical variance-guided normalization method is used to extract input features. The global mean and variance statistics are used to describe the overall response strength of the channels, while the variance statistics describe the dispersion and dynamic trend of the feature distribution, thereby stabilizing the feature distribution during the training process of deep networks. Figure 4 As shown, the processing steps are as follows:

[0111] Set up two parallel paths: a batch normalization main branch and a statistical guidance branch;

[0112] Batch normalization main branch for input features Perform batch normalization to output stable basic feature representations;

[0113] Statistical guided branches extract input features respectively The global mean and global variance statistics are input into the nonlinear mapping structure composed of fully connected layer, GELU activation layer and fully connected layer respectively, processed independently and then fused, and then channel-level guided weights are generated by Sigmoid function.

[0114] Channel-level guided weights are used to adaptively calibrate the deep feature responses, and these residuals are superimposed onto the output of the batch normalization main branch to obtain the final normalized features. This approach mitigates the problems of deep feature distribution fluctuations and channel response imbalances while maintaining training stability, ultimately normalizing the features. Represented as:

[0115]

[0116]

[0117] .

[0118] In this embodiment, the DMSG-Net classification network is iteratively trained using the constructed ovarian cancer ultrasound image dataset. The dataset contains two-dimensional ovarian cancer ultrasound images confirmed by pathological results, with pathological subtypes including serous carcinoma, clear cell carcinoma, mucinous carcinoma, and endometrioid carcinoma. To ensure the reliability of the experimental results, the dataset is divided into training, validation, and test sets according to patient level, avoiding the simultaneous appearance of images of the same patient in different subsets.

[0119] The experiments were conducted in a Linux system environment, using the PyTorch deep learning framework. During the classification network training process, the AdamW optimizer was used to update the network parameters, with a batch size of 16 and a training epoch count of 500. For pathological subtypes with a small sample size, a generative augmentation method based on Stable Diffusion was used to generate supplementary ultrasound images. These generated images were then used in conjunction with real images for network training to alleviate class imbalance and improve the model's generalization ability.

[0120] To verify the effectiveness of this invention, this embodiment uses accuracy, precision, recall, and F1 score as evaluation metrics. Simultaneously, a confusion matrix can be used to analyze the classification confusion between different pathological subtypes. The calculation formulas for each evaluation metric are as follows:

[0121]

[0122]

[0123]

[0124]

[0125] Wherein, TP represents the number of samples correctly predicted as the target class, TN represents the number of samples correctly predicted as non-target classes, FP represents the number of samples incorrectly predicted as the target class, and FN represents the number of target class samples incorrectly predicted as other classes.

[0126] Accuracy measures the proportion of correct overall classifications by the model; precision measures the reliability of the model's predictions for a specific pathological subtype; recall measures the completeness of the model's identification of samples of that pathological subtype; the F1 score, which considers both precision and recall, provides a more comprehensive reflection of the model's classification performance under imbalanced class conditions. These evaluation metrics allow for a comprehensive assessment of the method of this invention, considering its overall classification ability, class identification ability, and the confusion relationships between different pathological subtypes.

[0127] To verify the effectiveness of the proposed DMSG-Net model in the classification of ovarian cancer ultrasound image pathological subtypes, comparative experiments were conducted with several mainstream deep learning classification methods under the same experimental conditions. Detailed results are shown in Table 1. The comparison methods included CNN-type models ResNet, EfficientNet, DenseNet, and ConvNeXt; Transformer-type models PVT, Swing, ViT, DeiT, and Twins; and Mamba-type models Vim, VMamba, EfficientVMamba, and Spatial-Mamba.

[0128]

[0129] Table 1

[0130] The experimental results show that DMSG-Net exhibits excellent overall classification performance on the ovarian cancer ultrasound image dataset, achieving an accuracy of 90.25%, a precision of 92.14%, a recall of 87.00%, and an F1-score of 88.73%. Compared with other comparative models, DMSG-Net achieves the best results in accuracy, precision, and F1-score, indicating that the model can more accurately distinguish different pathological subtypes and effectively reduce misclassification between different subtypes.

[0131] To verify the impact of each improved module on model performance, ablation experiments were conducted under the same experimental conditions. MambaVision-T2 was used as the baseline, and DLCA (Differential Awareness Local Control Aggregation Module), MSGM (Multi-Scale Spatial Guided Hybrid Module), and DVGNorm (Dual Statistical Variance Guided Normalization Method) were added for comparison. The results are shown in Table 2.

[0132]

[0133] Table 2

[0134] Table 2 shows that adding MSGM, DLCA, or DVGNorm individually improved the model performance to varying degrees, indicating that the three modules respectively enhance multi-scale spatial semantic modeling, lesion boundary and local texture representation, and the stability of deep feature distribution. Further combination of different modules further improved the overall model performance. When MSGM, DLCA, and DVGNorm were added simultaneously, DMSG-Net achieved the best overall performance, with Accuracy, Precision, Recall, and F1-score reaching 90.25%, 92.14%, 87.00%, and 88.73%, respectively. The results indicate that the three modules have complementary effects and can effectively improve the classification performance of ovarian cancer ultrasound image pathological subtypes.

[0135] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A deep learning-based method for classifying pathological subtypes of ovarian cancer using ultrasound images, characterized in that, Includes the following steps: S1. Construct an ultrasound image dataset for ovarian cancer; S2. Construct an end-to-end DMSG-Net classification network: The DMSG-Net classification network is based on the MambaVision structure as its backbone network and is divided into a shallow local feature enhancement stage, a deep multi-scale spatial semantic modeling stage, and a classification prediction stage. A difference-aware local feature enhancement module is embedded in the shallow local feature enhancement stage to enhance lesion boundaries, local textures, and weak contrast structural features; In the deep multi-scale spatial semantic modeling stage, a multi-scale spatial guided hybrid module is embedded and combined with a dual statistical variance guided normalization method to model the multi-scale spatial structure and subtype semantic differences of the lesion region and stabilize the distribution of deep features. S3. The DMSG-Net classification network is trained and optimized using the ovarian cancer ultrasound image dataset. The ovarian cancer ultrasound images to be classified are input into the trained network, and the probability distribution of each pathological subtype is output. The category with the highest probability is taken as the final pathological subtype classification result.

2. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 1, characterized in that, S1 includes the following steps: S11. Acquire two-dimensional ultrasound images of ovarian cancer confirmed by pathology, crop the region of interest of the lesion and normalize its size, label the category according to the pathological subtype, and divide it into training set, validation set and test set according to the patient level. S12. For pathological subtypes in the training set where the number of samples is less than a preset threshold, a generative amplification strategy based on Stable Diffusion is adopted to generate supplementary samples that are consistent with real ultrasound images in terms of lesion texture, boundary morphology, and grayscale distribution, and these supplementary samples are added to the training set at a preset ratio.

3. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 1, characterized in that: The shallow local feature enhancement stage includes two sequentially connected first feature extraction layers, which are composed of convolutional blocks and the difference-aware local feature aggregation module connected in series. The deep multi-scale spatial semantic modeling stage includes two sequentially connected second feature extraction layers, which are composed of the MambaVision Mixer structure and the multi-scale spatial guided mixing module connected in series.

4. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 1, characterized in that, The processing steps of the difference-aware local control aggregation module are as follows: Input feature map The feature distribution is stabilized by a normalization layer, and then channel projection is completed by 1×1 convolution. Global average pooling is used to extract global statistical information. The difference between local feature responses and global statistical information is calculated. After dynamic adjustment by a learnable scaling factor, differential enhancement features are obtained through GELU nonlinear activation. , is represented as: Wherein, GAP represents global average pooling. ⊙ represents the learnable scaling factor, and ⊙ represents element-wise multiplication. Enhance the difference features The local gated aggregation structure is used to perform feature aggregation to obtain gated aggregated features; The gated aggregated features are compared with the input feature map. Establish residual connections and output enhanced local difference features.

5. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 4, characterized in that, The processing steps for the local departmental controlled aggregation structure are as follows: Two parallel processing branches are set up. The first branch completes linear projection through 1×1 convolution, and the second branch first captures neighborhood context information through 3×3 depthwise separable convolution, and then completes channel projection through 1×1 convolution. The outputs of the first and second branches are activated by SiLU and then subjected to element-wise gating interaction to obtain gating interaction features. , is represented as: The gated interaction features After channel fusion is completed by 1×1 convolution, the gated aggregation feature is output.

6. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 1, characterized in that, The processing steps of the multi-scale spatially guided hybridization module are as follows: A multi-branch parallel structure is adopted, with each branch corresponding to three different receptive field scales of 3, 5, and 7 respectively; Within each scale branch, the input features The spatial response enhancement of the lesion-related region is completed by the multi-scale coordinate attention RSGA structure, and the spatially enhanced branch features are output. Within each scale branch, the spatially enhanced branch features are processed by the dual statistical variance-guided normalization method and a multilayer perceptron to obtain single-scale semantic calibration features. ; Multiple single-scale semantic calibration features The features are obtained by stitching along the channel dimension; The concatenated features are fused across scales using a 1×1 convolution and then combined with the input features. Establish residual connections and output deep multi-scale spatial semantic features. , is represented as: Here, Concat represents concatenation along the channel dimension, and Conv represents a convolutional fusion layer.

7. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 6, characterized in that, The processing steps for the multi-scale coordinate RSGA structure are as follows: For the input features Horizontal pooling and vertical pooling are performed separately to extract bidirectional axial context information and complete the aggregation. The aggregated axial features are processed by large-kernel strip convolution with directional decoupling, and then normalized by the Sigmoid function to generate spatial calibration weights. , is represented as: in, For horizontal pooling, For vertical pooling, For the Sigmoid function, , These are 1×k and k×1 stripe convolutions, respectively; The spatial calibration weight With the input features Element-wise multiplication yields the spatially enhanced branch features.

8. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 1, characterized in that, The processing steps of the dual statistical variance-guided normalization method are as follows: Set up a batch normalization main branch and a statistical bootstrap branch; The batch normalization main branch applies to the input features Perform batch normalization to output stable basic feature representations; The statistical guidance branch extracts the input features respectively. The global mean and global variance statistics are input into the nonlinear mapping structure composed of fully connected layer, GELU activation layer and fully connected layer respectively, processed independently and then fused, and then channel-level guided weights are generated by Sigmoid function. The global statistical features are calibrated using the channel-level guided weights and then superimposed onto the output of the batch normalization main branch as residuals to obtain the final normalized features. , is represented as: 。 9. The method for classifying pathological subtypes of ovarian cancer based on ultrasound images according to claim 1, characterized in that: The classification prediction stage consists of a two-dimensional global average pooling layer and a linear classification layer connected in series. The two-dimensional global average pooling layer receives the output features from the deep multi-scale spatial semantic modeling stage and is used to compress the two-dimensional feature map into a one-dimensional global feature vector. The linear classification layer is connected to the one-dimensional global feature vector, and its output dimension corresponds one-to-one with the number of categories of the pathological subtype.

Images