Study of the impact of registration errors on the segmentation of stroke lesion in deep learning.

Motivation: When multiple sequences are required for differential diagnosis, image misalignment persists between scans, despite applying registration algorithms. The impact of these registration errors on the performances of downstream algorithms remains insufficiently characterized. Goal(s): Our goal was to determine how these registration errors impact the quality of deep learning-based stroke lesion segmentation in diffusion MRI. Approach: We implemented a tool to simulate registration errors between MRI sequences and evaluated their impact on segmentation accuracy using several metrics. Results: Segmentation accuracy remains robust for registration errors below 3 mm, but deteriorates significantly beyond this threshold. Impact: By identifying a critical 3 mm threshold for registration errors, we established quantitative guidelines for acceptable registration accuracy in clinical workflows. This threshold helps define minimal performance requirements for registration algorithms when preprocessing images for segmentation tasks.