A Convolutional Neural Network and Graph Convolutional Network Based Framework for Classification of Breast...

Area of Science:

• Digital pathology
• Medical imaging analysis
• Computational oncology

Background:

• Spatial correlation of tissue components is crucial for breast cancer diagnosis from histopathological images.
• Graph convolutional networks (GCNs) excel at capturing spatial features but require complex preprocessing for graph construction.
• Existing GCN-based computer-aided diagnosis (CAD) methods for histopathology face preprocessing challenges.

Purpose of the Study:

• To propose a novel CAD framework integrating CNNs and GCNs for breast histopathological image classification.
• To develop an adaptive graph construction method within a unified CNN-GCN framework.
• To enhance spatial feature learning for improved breast cancer diagnosis.

Main Methods:

• A unified CNN-GCN framework was developed, where CNN extracts high-level features for adaptive graph construction.
• A novel clique GCN (cGCN) was proposed for enhanced graph representation with bidirectional connections.
• A group graph convolution was introduced to refine feature representation and reduce redundancy, forming the clique group GCN (cgGCN).

Main Results:

• The proposed CNN-cgGCN framework demonstrated superior performance in classifying breast histopathological images.
• Experimental results on two public datasets confirmed the effectiveness of the CNN-cgGCN approach.
• The method outperformed existing comparative algorithms in diagnostic accuracy.

Conclusions:

• The integrated CNN-GCN framework, particularly with the novel cgGCN, effectively captures spatial features for breast cancer diagnosis.
• This approach simplifies preprocessing requirements for GCN-based CAD systems.
• The findings highlight the potential of the CNN-cgGCN for accurate and efficient histopathological image analysis in breast cancer detection.