PEPSI: Pathology-Enhanced Pulse-Sequence-Invariant Representations for Brain MRI

TL;DR

PEPSI tackles the challenge of brain MRI analysis across diverse pulse sequences and pathologies by learning pathology-enhanced, pulse-sequence-invariant representations trained entirely on synthetic data. The framework jointly optimizes anatomy- and pathology- center objectives via a dual guidance scheme anchored to MP-RAGE and FLAIR, and employs implicit pathology supervision to enable multi-pathology co-training with missing modalities. A dedicated pathology-encoded data generator leverages anatomy labels and anomaly probabilities to simulate realistic, variable pathologies and artifacts, producing high-quality synthesis and robust features for segmentation. Across three large public datasets and two pathology tasks, PEPSI achieves superior image synthesis, while pre-trained features markedly improve downstream segmentation performance and convergence, signaling strong practical potential for handling real-world, heterogeneous brain MRI data.

Abstract

Remarkable progress has been made by data-driven machine-learning methods in the analysis of MRI scans. However, most existing MRI analysis approaches are crafted for specific MR pulse sequences (MR contrasts) and usually require nearly isotropic acquisitions. This limits their applicability to diverse real-world clinical data, where scans commonly exhibit variations in appearances due to being obtained with varying sequence parameters, resolutions, and orientations -- especially in the presence of pathology. In this paper, we propose PEPSI, the first pathology-enhanced, and pulse-sequence-invariant feature representation learning model for brain MRI. PEPSI is trained entirely on synthetic images with a novel pathology encoding strategy, and enables co-training across datasets with diverse pathologies and missing modalities. Despite variations in pathology appearances across different MR pulse sequences or the quality of acquired images (e.g., resolution, orientation, artifacts, etc), PEPSI produces a high-resolution image of reference contrast (MP-RAGE) that captures anatomy, along with an image specifically highlighting the pathology. Our experiments demonstrate PEPSI's remarkable capability for image synthesis compared with the state-of-the-art, contrast-agnostic synthesis models, as it accurately reconstructs anatomical structures while differentiating between pathology and normal tissue. We further illustrate the efficiency and effectiveness of PEPSI features for downstream pathology segmentations on five public datasets covering white matter hyperintensities and stroke lesions. Code is available at https://github.com/peirong26/PEPSI.

Figures (6)

• Figure 1: PEPSI's on-the-fly generator uses 3D anatomy labels ($L$) and anomaly probabilities ($P$) to generate training data with diverse deformations, contrasts, and corruptions -- enhanced by varying intensity profiles in pathological regions (\ref{['sec: generator']}).
• Figure 2: PEPSI's pathology-enhanced, contrast-agnostic training overview (\ref{['sec: framework']}).
• Figure 3: Qualitative comparisons on anatomy and pathology ($\leftrightarrow$) image synthesis.
• Figure 4: Training progresses of w/ PEPSI and w/o PEPSI for pathology segmentation. The horizontal (vertical) axis indicates training epochs ("w/ PEPSI" epochs / "w/o PEPSI" epochs). Results are obtained by evaluating models collected throughout epochs.
• Figure 5: Qualitative comparisons on downstream pathology segmentation, w/o or w/ PEPSI pre-trained features.