Anatomical Data Augmentation via Fluid-based Image Registration.

Area of Science:

• Medical Image Analysis
• Computational Anatomy
• Machine Learning

Background:

• Medical image analysis often requires large datasets for training robust models.
• Existing image augmentation techniques may not always generate anatomically plausible variations.
• Accurate segmentation and learning from limited data (few-shot learning) are critical challenges.

Purpose of the Study:

• To introduce a novel fluid-based image augmentation framework for medical imaging.
• To generate anatomically meaningful images through geodesic subspace interpolation.
• To improve performance in medical image segmentation and one-shot learning tasks.

Main Methods:

• Constructing a geodesic subspace using the Large Deformation Diffeomorphic Metric Mapping (LDDMM) model from source and target images.
• Sampling transformations from the learned geodesic subspace.
• Generating augmented images and segmentations via interpolation of sampled transformations.

Main Results:

• Demonstrated generation of anatomically meaningful medical image data.
• Achieved improved performance in data augmentation for image segmentation tasks on brain (LPBA) and knee (OAI) datasets.
• Showcased enhanced performance in one-shot learning for single atlas image segmentation.

Conclusions:

• The proposed fluid-based augmentation method effectively generates realistic anatomical variations.
• The approach significantly improves performance in both data augmentation and one-shot learning scenarios for medical image analysis.
• This method offers a promising solution for enhancing medical image analysis tasks with limited data.