Reliability-Aware CT-MRI Registration: A Quality Engineering Framework with Stability Analysis and Risk Classification
Authors
Abstract
Multimodal CT-MRI registration is central to image-guided radiotherapy, surgical navigation, and diagnostic workflows, but most pipelines report only aggregate quality metrics without per-case reliability signals. We propose a reliability-aware framework that converts registration quality into Green/Yellow/Red risk categories using data-learned thresholds. CT images were registered to T1-weighted MRI using rigid and affine transformations on 90 paired slices from 18 patients across brain, abdominal, and neck anatomies. Reliability was assessed using Delta NMI, Delta SSIM, Dice overlap, registration stability, and inverse consistency error, combined into a single score R. Thresholds learned from training patients were applied unchanged to held-out test patients. Affine registration outperformed rigid registration on NMI and SSIM, yielding 44% Green classifications versus 33% for rigid. Reliability-filtered registrations improved the average alignment profile compared with unfiltered methods. Per-anatomy analysis showed substantial variation, with stronger reliability for abdominal registrations than brain registrations. Weight sensitivity analysis identified Dice overlap as the dominant reliability component. The proposed framework provides an interpretable quality-control layer for multimodal registration, while risk thresholds reflect statistical rather than clinical validation.