BCHisto-Net: Breast histopathological image classification by global and local feature aggregation

Area of Science:

• Computational pathology
• Medical image analysis
• Artificial intelligence in oncology

Background:

• Breast cancer is a leading global cancer in women.
• Histopathological image analysis is crucial for malignancy detection but is manual and error-prone.
• Automated analysis of histopathological images offers potential for improved accuracy and efficiency.

Purpose of the Study:

• To propose BCHisto-Net, a novel deep learning model for classifying breast histopathological images at 100× magnification.
• To address the limitation of existing methods that overlook region-specific features by extracting both global and local image characteristics.
• To enhance the accuracy of breast cancer malignancy diagnosis through improved feature extraction.

Main Methods:

• Development of BCHisto-Net, a Convolutional Neural Network (CNN) architecture.
• Extraction of both global (abstract) and local (region-specific) features from histopathological images.
• Implementation of a feature aggregation branch to combine global and local features for classification.
• Quantitative evaluation on a private dataset (KMC) and the public BreakHis dataset.

Main Results:

• The proposed BCHisto-Net achieved high classification accuracy.
• Achieved 95% accuracy on the KMC dataset and 89% accuracy on the BreakHis dataset.
• Demonstrated superior performance compared to state-of-the-art classifiers, highlighting the effectiveness of combined feature extraction.

Conclusions:

• BCHisto-Net effectively classifies breast histopathological images at 100× magnification.
• The integration of local and global features significantly improves diagnostic accuracy.
• The proposed model offers a promising automated solution for breast cancer malignancy assessment.