Breast cancer histopathology image classification through assembling multiple compact CNNs

Area of Science:

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

Background:

• Breast cancer remains a leading cause of global mortality, necessitating improved diagnostic accuracy and efficiency.
• Manual histopathology diagnosis is labor-intensive and susceptible to errors, particularly with prolonged pathologist workload.
• Automated histopathology image recognition is crucial for accelerating diagnosis and enhancing its quality.

Purpose of the Study:

• To develop an automated system for breast cancer histopathology image classification.
• To improve diagnostic accuracy and reduce the workload of pathologists.
• To enhance the generalization ability of classification models.

Main Methods:

• A hybrid Convolutional Neural Network (CNN) architecture with global and local branches was designed.
• A Squeeze-Excitation-Pruning (SEP) block was integrated to learn channel importance and remove redundant channels, reducing overfitting.
• Multiple compact hybrid CNN models were assembled to improve generalization.

Main Results:

• The proposed hybrid model achieved performance comparable to state-of-the-art methods on the BreaKHis dataset.
• The multi-model assembling scheme outperformed state-of-the-art methods on the BACH dataset at both patient and image levels.
• The method demonstrated effectiveness in breast cancer image classification.

Conclusions:

• A novel compact breast cancer histopathology image classification scheme was developed using assembled hybrid CNNs.
• The proposed scheme shows promise for auxiliary breast cancer diagnosis, reducing pathologist workload and improving diagnostic quality.
• This automated approach can significantly aid in early breast cancer detection and treatment planning.