PEPSI: Pathology-Enhanced Pulse-Sequence-Invariant Representations for Brain MRI
Core Concepts
PEPSI is the first pathology-enhanced, pulse-sequence-invariant feature representation learning model for brain MRI, showcasing remarkable capability in image synthesis and pathology segmentation.
Abstract
Introduction:
Data-driven machine-learning methods have advanced MRI analysis.
Existing approaches are limited by specific MR pulse sequences and isotropic acquisitions.
PEPSI Model:
Proposal of PEPSI for pathology-enhanced, pulse-sequence-invariant feature representation learning.
Trained on synthetic images with novel pathology encoding strategy.
Experiments:
Demonstrated PEPSI's efficiency in image synthesis and downstream pathology segmentations on diverse datasets.
Approach:
Generating pathology-encoded training data using anomaly probabilities and contrast diversity.
Representing across Contrasts:
Balancing anatomy and pathology features using MP-RAGE and FLAIR scans.
Pathology Segmentation:
Utilizing PEPSI pre-trained features improves convergence time and yields higher Dice scores in segmentation tasks.