Cyclic GAN Model to Classify Breast Cancer Data for Pathological Healthcare Task.

Area of Science:

• Digital pathology
• Medical image analysis
• Machine learning in oncology

Background:

• Pathological image analysis faces challenges with uneven staining and distinguishing benign from malignant cells.
• Existing detection models struggle with the complexities of histopathological data.

Purpose of the Study:

• To develop an enhanced algorithm for accurate classification of breast cancer pathological images.
• To address limitations in current methods for analyzing unevenly stained tissue samples.
• To improve the discrimination between benign and malignant cellular features.

Main Methods:

• Utilized CycleGAN for color normalization to standardize pathological images.
• Developed an upgraded dual-path network (DPN) incorporating small convolution, deconvolution, and attention mechanisms.
• Employed the BreaKHis dataset for training and evaluating the DPN68-A network.
• Assessed model performance using metrics including false-positive rate, false-negative rate, recall, precision, and F1 score.

Main Results:

• The proposed DPN68-A network demonstrated effective classification of benign and malignant breast cancer images across various magnifications.
• Comparative experiments validated the DPN68-A network's superior performance against other deep learning models and classification algorithms.
• The integration of CycleGAN and DPN significantly improved the handling of uneven staining and feature discrimination.

Conclusions:

• The DPN68-A network provides a robust solution for classifying breast cancer pathological images, overcoming staining inconsistencies.
• The model shows potential to assist pathologists in clinical diagnosis by synthesizing multi-magnification images.
• This approach enhances the reliability and efficiency of automated pathological image analysis for breast cancer detection.