Automatic data augmentation to improve generalization of deep learning in H&E stained histopathology

Area of Science:

• Computational pathology
• Computer vision
• Machine learning

Background:

• Histopathology image variations across centers challenge deep learning models.
• Manual data augmentation for domain generalization is time-consuming and suboptimal.

Purpose of the Study:

• To investigate automated data augmentation for improved domain generalization in histopathology.
• To compare automated methods against manual augmentation and reduce experimental time.

Main Methods:

• Selected four state-of-the-art automatic augmentation techniques from computer vision.
• Evaluated methods on histopathology data from 25 centers for two tasks: metastasis detection and tissue classification.
• Utilized meta-learning frameworks for hyper-parameter optimization.

Main Results:

• Automatic augmentation methods achieved performance comparable to manual augmentation for metastasis detection.
• A leading automatic augmentation method significantly outperformed manual augmentation in breast cancer tissue classification.
• Reduced experimental optimization time compared to manual tuning.

Conclusions:

• Automated data augmentation search is effective for improving domain generalization in computational pathology.
• This approach offers a more efficient and potentially superior alternative to manual data augmentation.
• Further research can explore broader applications in histopathology AI.