LatentAugment: Data Augmentation via Guided Manipulation of GAN's Latent Space

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Data Augmentation (DA) is crucial for improving model generalization by increasing training data quantity and diversity.
• Standard DA methods often yield limited synthetic data diversity.
• Generative Adversarial Networks (GANs) generate realistic synthetic data but struggle with diversity and mode coverage.

Purpose of the Study:

• To introduce LatentAugment, a novel DA strategy addressing GANs' limitations in diversity and mode coverage.
• To enhance the fidelity and diversity of synthetic data generated by GANs for DA applications.
• To provide a dataset- and task-agnostic DA solution.

Main Methods:

• LatentAugment modifies latent vectors within GANs to explore underrepresented regions of the latent space.
• This approach aims to maximize the diversity and fidelity of generated synthetic images without external supervision.
• The method is evaluated on medical imaging tasks, including MRI-to-CT translation and contrast-enhanced spectral mammography.

Main Results:

• LatentAugment significantly improves the generalization of deep learning models compared to standard DA and GAN-based sampling.
• Experiments demonstrate superior mode coverage and diversity of LatentAugment-generated samples over traditional GAN sampling.
• The method shows effectiveness in diverse medical imaging translation tasks.

Conclusions:

• LatentAugment offers a powerful solution to enhance GAN-based data augmentation by improving synthetic data diversity and quality.
• This technique overcomes key limitations of existing GANs, making them more suitable for demanding DA applications.
• LatentAugment advances the field of medical image analysis by enabling more robust and generalizable deep learning models.