KR-20260064883-A - A HYBRID LIGHTWEIGHT BREAST CANCER CLASSIFICATION METHOD AND SYSTEM USING HISTOPATHOLOGICAL IMAGES

Abstract

A breast cancer classification method performed by a breast cancer classification system includes the step of inputting a histopathology image into a breast cancer classification model; and the step of classifying breast cancer from the input histopathology image through the breast cancer classification model, wherein the breast cancer classification model may be composed of an SMDC (Separable Multiscale Depth-wise Convolution) block, an SOP (Second-Order Pooling) block, an MSA (Multi-head Self-Attention) block, and a Classification Head.

Inventors

• 알안타리무가헤드
• 다니엘 애도

Assignees

• 세종대학교산학협력단

Dates

Publication Date

20260508

Application Date

20241030

Claims (14)

1. In a breast cancer classification method performed by a breast cancer classification system, Step of inputting histopathology images into a breast cancer classification model; and A step of classifying breast cancer from the input histopathology image using the above breast cancer classification model Includes, The above breast cancer classification model is, A breast cancer classification method characterized by being composed of an SMDC (Separable Multiscale Depth-wise Convolution) block, an SOP (Second-Order Pooling) block, an MSA (Multi-head Self-Attention) block, and a Classification Head.
2. In paragraph 1, The above SMDC block performs multiscale feature extraction by applying different kernel sizes to feature maps extracted from histopathology images, and is composed of multiple SMDC modules, combining information input to the first SMDC module of each SMDC block through residual learning with information output through the last SMDC module of each SMDC block. A breast cancer classification method characterized in that the plurality of SMDC modules are each composed of an SMDConv layer, a batch normalization layer, and a ReLU activation layer.
3. In paragraph 1, The above SOP block provides channel information and spatial information by connecting feature maps based on a global covariance pooling module along channel dimensions and spatial dimensions. A breast cancer classification method characterized by being composed of a channel-specific SoP (C-SoP) module, a spatial-specific SoP (S-SoP) module, and a Matrix Square-root Normalization (MSN) module.
4. In paragraph 3, The above channel-specific SoP modules are, A breast cancer classification method characterized by generating a second pooling of feature maps in each channel space and then generating a channel-specific covariance matrix.
5. In paragraph 3, The above space-specific SoP modules are, A breast cancer classification method characterized by generating 2D pooling on feature maps at each spatial location and then generating a spatial covariance matrix.
6. In paragraph 3, The above matrix square root normalization module is, A breast cancer classification method characterized by adjusting channel dimensions and spatial dimensions through matrix square root normalization calculation using the channel-specific covariance matrix generated in the channel-specific SoP module and the spatial covariance matrix generated in the spatial SoP module.
7. In paragraph 1, The above MSA block is, It includes a Swin transformer block for classifying histopathology images by performing multi-head self-attention, and Patches are extracted using ensemble features derived from histopathology images, patch features are generated by embedding the extracted patches, multi-head self-attention is performed on the generated patch features through the Swinn Transformer block, patches are merged based on the results of the performed multi-head self-attention, and the merged patches are passed to a classification head. Breast cancer classification method characterized by
8. In paragraph 1, The above input step is, Step of acquiring breast cancer histopathology images through tissue biopsy; A step of performing preprocessing on acquired histopathology images; and The step of inputting the histopathology image, on which the above preprocessing has been performed, into a breast cancer classification model. Breast cancer classification method including
9. A computer program stored on a non-transient computer-readable recording medium to execute the breast cancer classification method of any one of claims 1 to 8 on the breast cancer classification system.
10. In the breast cancer classification system, An image input unit for inputting histopathology images into a breast cancer classification model; and Breast cancer classification unit that classifies breast cancer from the input histopathology image through the above breast cancer classification model Includes, The above breast cancer classification model is, A breast cancer classification system characterized by being composed of an SMDC (Separable Multiscale Depth-wise Convolution) block, an SOP (Second-Order Pooling) block, an MSA (Multi-head Self-Attention) block, and a Classification Head.
11. In Paragraph 10, The above SMDC block performs multiscale feature extraction by applying different kernel sizes to feature maps extracted from histopathology images, and is composed of multiple SMDC modules, combining information input to the first SMDC module of each SMDC block through residual learning with information output through the last SMDC module of each SMDC block. A breast cancer classification system characterized in that the plurality of SMDC modules are each composed of an SMDConv layer, a batch normalization layer, and a ReLU activation layer.
12. In Paragraph 10, The above SOP block provides channel information and spatial information by connecting feature maps based on a global covariance pooling module along channel dimensions and spatial dimensions. It consists of a channel-specific SoP (C-SoP) module, a spatial-specific SoP (S-SoP) module, and a Matrix Square-root Normalization (MSN) module, and The above channel-specific SoP modules are, After generating 2nd-order pooling on the feature maps in each channel space, generate a channel-specific covariance matrix, and The above space-specific SoP modules are, After generating 2D pooling on the feature map at each spatial location, a spatial covariance matrix is generated, and The above matrix square root normalization module is, A breast cancer classification system characterized by adjusting channel dimensions and spatial dimensions through matrix square root normalization calculation using the channel-specific covariance matrix generated in the channel-specific SoP module and the spatial covariance matrix generated in the spatial SoP module.
13. In Paragraph 10, The above MSA block is, It includes a Swin transformer block for classifying histopathology images by performing multi-head self-attention, and Patches are extracted using ensemble features derived from histopathology images, patch features are generated by embedding the extracted patches, multi-head self-attention is performed on the generated patch features through the Swinn Transformer block, patches are merged based on the results of the performed multi-head self-attention, and the merged patches are passed to a classification head. Breast cancer classification system characterized by the following.
14. In Paragraph 10, The above image input unit is, Breast cancer histopathology images are obtained through a biopsy, and Preprocessing is performed on the acquired histopathology images, and Inputting the histopathology images on which the above preprocessing has been performed into a breast cancer classification model Breast cancer classification system characterized by the following.

Description

A Hybrid Lightweight Breast Cancer Classification Method and System Using Histopathological Images The following description concerns breast cancer classification technology. Convolutional Neural Networks (CNNs) are gaining attention as a viable approach for classifying histopathology images. Previous studies have tended to emphasize first-order statistics through global mean and max pooling operations. However, this approach overlooks the subtle complexity of capturing complex non-linear deep features (second-order statistics), consequently limiting the model's classification potential. Furthermore, relying on standard convolutions in CNN models introduces higher computational complexity and drastically increases the model's parameter space. Using a fixed kernel size restricts the model to capturing only one of either low-resolution or high-resolution features, preventing the simultaneous integration of both spectrums. Meanwhile, Korean Patent Publication No. 10-2024-0029718 (March 6, 2024) discloses a method for identifying breast cancer histopathology images based on an artificial intelligence model. FIG. 1 is a diagram illustrating an artificial intelligence-based framework designed to classify breast cancer through histopathology image analysis in one embodiment. FIG. 2 is a diagram illustrating a breast cancer classification model in one embodiment. FIG. 3 is a drawing for explaining the SOP block in one embodiment. FIG. 4 is a diagram illustrating a Multi-head Self-Attention (MSA) block in one embodiment. FIG. 5 is an example of visualizing sampled images before and after applying image enhancement technology in one embodiment. FIG. 6 is an example showing a valuation confusion matrix (CM) obtained by testing the BreaKHis and BACH datasets using a 5-fold cross-validation method in one embodiment. FIG. 7 is a figure showing a Receiver Operating Characteristic (ROC) curve comparing the performance of a pre-trained model selected from the BreaKHis and BACH datasets and a proposed model in one embodiment. FIG. 8 is an example showing a visually explainable heatmap that highlights the most prominent area in a histopathology image in one embodiment. FIG. 9 is a block diagram illustrating a breast cancer classification system in one embodiment. FIG. 10 is a flowchart illustrating a breast cancer classification system in one embodiment. Hereinafter, embodiments will be described in detail with reference to the attached drawings. FIG. 1 is a diagram illustrating an artificial intelligence-based framework designed to classify breast cancer through histopathology image analysis in one embodiment. To collect Breast Cancer Histopathology Image (BCHI) datasets, a biopsy may be assigned from a designated patient. As the biopsy is assigned, digital pathology images can be captured. The initial step in collecting the BCHI dataset involves recovering tissue samples from the affected area via biopsy. At this stage, small tissue sections may be extracted using a needle or surgical means. The tissue samples can be placed in a fixation solution, such as formalin, to preserve the tissue structure and prevent degradation. The tissue can then be processed in the laboratory, embedded in paraffin wax, cut into thin sections, and mounted on glass slides. The tissue sections on the glass slides can be stained with various dyes to highlight distinct cellular components, such as the nucleus and cytoplasm. Among the stains commonly used in breast cancer histopathology, hematoxylin and eosin (H&E) are frequently employed; hematoxylin colors the nucleus blue, while eosin stains the cytoplasm pink. Subsequently, stained tissue samples are examined using a microscope to visualize the structural and cellular characteristics of the tissue. Pathologists can utilize grading systems to assess the severity of the cancer by evaluating the morphology of cells and tissues. Breast cancer histopathology images can be captured using a digital microscope, which includes a digital camera attached to the microscope to capture high-resolution images of tissue sections. These images can be digitally stored and analyzed using various software tools, including image analysis and machine learning algorithms. Several preprocessing steps may be applied for data cleaning, removal of noise or unwanted information, normalization, image contrast enhancement, and data segmentation and augmentation. Afterward, artificial intelligence models can be built, trained, and tested to demonstrate their ability to predict breast cancer based on histological knowledge. Finally, the prediction results can be verified by experts in the relevant field, and suggestions or recommendations can be presented to improve prediction performance. FIG. 2 is a diagram illustrating a breast cancer classification model in one embodiment. In the embodiments, an artificial intelligence model proposed for classifying breast cancer based on histopathology images, namely a breast cancer classi