Beyond cost function masking: RPCA-based non-linear registration in the context of VLSM

Abstract

Voxel-based lesion symptom mapping (VLSM) allows studying the relationship between stroke location and clinical outcome. The core idea of VLSM is to map all patient cases into a common atlas space and then apply statistical tests on a voxel level comparing outcome measures of patients with a lesion in the voxel to those without lesion. A major limitation of VLSM is that it requires a previous lesion segmentation, which is mostly performed manually, for masked subject-to-atlas registration as well as for the VLSM analysis. The aim of this work is to evaluate the feasibility of a recently introduced robust PCA (RPCA)-based iterative non-linear registration framework that potentially overcomes this limitation by generating the lesion segmentation on the fly. In addition, we propose and evaluate a rapid variant of this framework (successively tightened low rank-condition RPCA, stl-RPCA). Based on 29 follow-up FLAIR datasets of patients with ischemic stroke, the lesion segmentation capabilities and subject-to-atlas transformation properties of the RPCA methods are evaluated and compared to non-linear registration with and without cost function masking. Results reveal that the proposed method is capable of segmenting the lesions with an average Dice coefficient of 63%. Similar to nonlinear registration with cost function masking, the RPCA-based registration frameworks significantly decrease confounding effects of pathologies on subject-to-atlas transformation properties. Overall, the RPCA frameworks lead to promising results and could considerably enhance VLSM analyses.