MGAug: Multimodal Geometric Augmentation in Latent Spaces of Image Deformations

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Geometric transformations are crucial for augmenting image datasets.
• Current methods struggle with multimodal transformation distributions.
• This limitation hinders performance on complex datasets.

Purpose of the Study:

• To introduce Multimodal Geometric Augmentation (MGAug) for enhanced image data augmentation.
• To address the limitations of unimodal assumptions in existing augmentation techniques.
• To improve prediction accuracy in medical imaging tasks.

Main Methods:

• Developed a deep network using a variational autoencoder (VAE) to learn latent spaces of diffeomorphic transformations.
• Utilized a mixture of multivariate Gaussians as a prior in the tangent space of diffeomorphisms.
• Augmented training datasets by sampling transformations from the learned multimodal latent space.
• Validated the approach on synthetic 2D and real 3D brain MRI classification and segmentation tasks.

Main Results:

• MGAug significantly improved prediction accuracy compared to state-of-the-art methods.
• The model demonstrated superior performance in both classification and segmentation tasks.
• Effective handling of multimodal distributions in geometric transformations was achieved.

Conclusions:

• MGAug offers a novel and effective approach to geometric data augmentation.
• The method overcomes limitations of unimodal transformation assumptions.
• MGAug shows strong potential for improving deep learning models in medical imaging and beyond.