Data Augmentation in High Dimensional Low Sample Size Setting Using a Geometry-Based Variational Autoencoder.

Area of Science:

• Machine Learning
• Medical Imaging
• Data Science

Background:

• High Dimensional Low Sample Size (HDLSS) data presents challenges for machine learning models.
• Traditional data augmentation methods may not be effective in HDLSS settings.
• Reliable data augmentation is crucial for improving model performance with limited datasets.

Purpose of the Study:

• To propose a novel geometry-based variational autoencoder (VAE) for data augmentation in HDLSS settings.
• To develop a VAE model with a Riemannian manifold latent space, incorporating Riemannian metric learning and normalizing flows.
• To introduce a new generation scheme for producing more meaningful augmented samples, particularly for small datasets.

Main Methods:

• A novel VAE model was developed, featuring a Riemannian manifold latent space.
• The VAE integrates Riemannian metric learning and normalizing flows.
• A new sample generation scheme was implemented for enhanced data augmentation.

Main Results:

• The proposed method demonstrated robust performance across various datasets, classifiers, and training sample sizes.
• Validation on the ADNI database for Alzheimer's disease classification showed significant improvements in balanced accuracy.
• For instance, balanced accuracy increased from 66.3% to 74.3% with limited data and from 77.7% to 86.3% with more data, alongside enhanced sensitivity and specificity.

Conclusions:

• The geometry-based VAE offers a reliable approach for data augmentation in HDLSS scenarios.
• The method effectively enhances classification performance, particularly in medical imaging applications with limited data.
• The proposed framework provides a significant and reliable gain in classification metrics, demonstrating its practical utility.