Breast Cancer Histopathological Image Classification with Adversarial Image Synthesis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Deep learning models for medical images face challenges due to limited data.
• Data augmentation is crucial for training accurate and robust deep learning models in medical imaging.
• Generative Adversarial Networks (GANs) show promise for synthetic data generation.

Purpose of the Study:

• To address data limitations in medical image analysis using deep learning.
• To enhance the performance of deep learning classifiers for breast cancer histopathological image classification.
• To evaluate the effectiveness of Auxiliary Classifier Generative Adversarial Network (ACGAN) for data augmentation.

Main Methods:

• Employed an Auxiliary Classifier Generative Adversarial Network (ACGAN) to generate realistic medical images with class labels, augmenting a small dataset.
• Utilized a transfer learning approach for deep convolutional neural network (dCNN) classifiers, extracting features from a pre-trained model.
• Integrated extreme gradient boosting (XGBoost) classifiers with the extracted features for image classification tasks.

Main Results:

• ACGAN data augmentation increased classification accuracy by 9.35% for binary classification (benign vs. malignant) on the BreakHis dataset.
• ACGAN data augmentation improved four-class tumor sub-type classification accuracy by 8.88% compared to standard transfer learning.
• The combination of ACGAN augmentation and XGBoost classifiers demonstrated superior performance over standard transfer learning methods.

Conclusions:

• ACGAN-based data augmentation is an effective strategy to overcome data limitations in medical image analysis.
• The proposed method significantly enhances the accuracy of deep learning models for breast cancer histopathological classification.
• This approach offers a viable solution for improving diagnostic accuracy in computational pathology.