BSDA: Bayesian Random Semantic Data Augmentation for Medical Image Classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Data augmentation is vital for deep neural networks, especially in data-limited medical imaging.
• Semantic data augmentation (SDA) alters feature semantics by shifting latent space representations.
• Existing SDA methods risk label changes with excessive feature shifting, impacting model performance.

Purpose of the Study:

• To introduce a computationally efficient and robust semantic data augmentation method.
• To address the issue of label alteration in deep learning models during data augmentation.
• To propose Bayesian Random Semantic Data Augmentation (BSDA) as a plug-and-play component.

Main Methods:

• Developed Bayesian Random Semantic Data Augmentation (BSDA), a novel approach to feature shifting.
• BSDA integrates seamlessly as a plug-and-play module into existing neural network architectures.
• The method avoids excessive shifting that could lead to label discrepancies.

Main Results:

• BSDA demonstrated superior performance compared to existing competitive data augmentation methods.
• The proposed method is effective for both 2D and 3D medical image datasets.
• BSDA showed compatibility with various medical imaging modalities and mainstream neural network models.

Conclusions:

• BSDA offers an effective solution for regularization in deep learning for medical imaging.
• The method enhances baseline model performance without compromising data integrity.
• BSDA is a versatile and efficient tool for improving deep learning applications in medical image analysis.