Second-order asymmetric convolution network for breast cancer histopathology image classification

Area of Science:

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

Background:

• Convolutional Neural Networks (CNNs) are prevalent in histopathology image classification.
• Deep high-order statistic models show superior performance over first-order models in vision tasks.

Purpose of the Study:

• To explore global deep high-order statistics for distinguishing breast cancer histopathology images.
• To improve classification performance by integrating asymmetric convolution into a second-order network.

Main Methods:

• Proposed a novel Second-Order Asymmetric Convolution Network (SoACNet).
• Integrated asymmetric convolution blocks into the backbone architecture.
• Employed global covariance pooling to compute second-order statistics of deep features.

Main Results:

• SoACNet demonstrated effectiveness in breast cancer histopathology image classification.
• Achieved competitive performance compared to state-of-the-art methods on the BreakHis dataset.
• The proposed network provides a more robust representation of histopathology images.

Conclusions:

• The SoACNet effectively utilizes deep high-order statistics for improved breast cancer image classification.
• Asymmetric convolutions and global covariance pooling enhance feature representation robustness.
• The developed method shows promise for advancing computational pathology in breast cancer diagnosis.