Reliability-Aware CT–MRI Registration Quality Engineering Framework with Stability Analysis and Risk Classification

Q: Why is a per-case reliability signal needed for CT–MRI registration?

Conventional pipelines report only aggregate quality metrics, so misregistrations can fail silently when no ground truth alignment exists at deployment time. A reliability signal supports deciding whether to accept automatically, seek expert review, or reject and retry.

Q: What diagnostics are combined to form the reliability score R?

R integrates four diagnostics: quality improvement (ΔNMI and ΔSSIM), spatial overlap (Dice), registration stability from repeated initialisations, and inverse consistency error from forward/backward registration discrepancy.

Q: How are the Green/Yellow/Red risk categories determined?

Thresholds learned from training patients are applied unchanged to held-out test patients. The resulting reliability score R is mapped into three tiers (Green automatically acceptable, Yellow review/attention, Red rejection/retry).

Jul 16, 2026

Upload by Eliana

Multimodal CT–MRI registration underpins image-guided radiotherapy, surgical navigation, and diagnostic workflows, yet common pipelines provide only aggregate quality and no per-case reliability signal to guide clinical decisions. A plausible local optimum may score well while failing silently. A reliability-aware framework converts continuous per-case registration quality into Green/Yellow/Red risk categories using data-learned thresholds. Automatic acceptance, expert-review triggering, and rejection/retry are integrated into a single pipeline. Rigid and affine CT-to-T1 MRI registration are validated on 90 slices across 18 patients.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

Introduction

Reliability engineering for imaging algorithms

Proposed reliability-aware CT–MRI registration framework

Methods

Data and registration settings

Reliability diagnostics and score R

Risk classification with learned thresholds

Results

Rigid vs affine performance

Reliability-filtered registration impact

Per-anatomy reliability differences

Discussion and limitations

Interpretability and deployment considerations

Reliability-Aware CT–MRI Registration: A Quality Engineering Framework with Stability Analysis and Risk Classification Nisreen Albzour School of Systems Science and Industrial Engineering, Binghamton University, Binghamton, NY, USA ORCID: [https://orcid.org/0009-0000-9317-340X](https://orcid.org/0009-0000-9317-340X) Abstract Multimodal CT–MRI registration underpins image-guided radiotherapy, surgical navigation, and diagnostic workflows, yet existing pipelines report only aggregate quality metrics and provide no per-case reliability signal to support clinical decisions. A registration that converges to a plausible but incorrect local optimum can still score well on conventional post-hoc metrics, creating a silent failure mode. We propose a reliability-aware framework that converts continuous per-case registration quality into actionable Green/Yellow/Red risk categories using data-learned thresholds, enabling automated acceptance, expert-review triggering, and rejection within one pipeline. CT was registered to T1-weighted MRI using rigid and affine transformations on 90 paired slices from 18 patients (14 training, 4 held-out test) across brain (n=9), abdominal (n=8), and neck (n=1) anatomies. Reliability was assessed through quality improvement (ΔNMI,ΔSSIM), spatial overlap (Dice), registration stability across repeated initialisations, and inverse consistency error, combined into a single score R. Thresholds learned from training patients were applied unchanged to test patients. Affine registration significantly outperformed rigid on NMI (Cohen’s d = 0.76, p < 10⁻¹³) and SSIM (d = 0.36, p < 0.01), yielding 44% Green (automatically acceptable) classifications versus 33% for rigid. Reliability-filtered registrations surpassed unfiltered methods on all metrics. Per-anatomy performance varied markedly (abdominal 73% Green, brain 22%), and weight-sensitivity analysis identified Dice overlap as the dominant reliability component. This interpretable, data-driven framework provides a reproducible foundation for quality-controlled multimodal registration and applies to any rigid or affine algorithm. Risk thresholds reflect statistical rather than clinical validation, which should precede clinical deployment. Keywords: Image registration, CT–MRI, quality assurance, reliability engineering, risk stratification, inverse consistency, registration stability 1. Introduction Computed tomography (CT) and magnetic resonance imaging (MRI) provide complementary anatomical information: CT offers high spatial resolution and bone contrast, while MRI excels at soft-tissue differentiation. Combining these modalities through image registration is foundationalto radiotherapy planning, surgical navigation, and multimodal diagnostic workflows [1]. However, registration is itself an estimation problem with no ground truth available at deployment time, anda misregistered image pair can silently propagate errors into downstream clinical decisions [2] . Despite decades of methodological development in image registration—from intensity-based similarity metrics [3] to deep-learning-based deformable registration [4]—the great majority of registration pipelines report only aggregate quality metrics (e.g., mean mutual information or Dice score across a dataset) without providing a per-case indication of whether a specific registration result can be trusted [5]. This is a critical gap from a quality-engineering perspective: a registration pipeline deployed in practice does not have access to ground-truth alignment, and therefore needs an internal signal for when a result is reliable enough to accept automatically, when it warrants expert review, and when it should be rejected and re-attempted [6] . Reliability engineering principles—which emphasise quantifying confidence in a system's output rather than reporting only point estimates—have been increasingly recognised as essential for trustworthy deployment of imaging algorithms [7] . Yet regist

FAQs

Why is a per-case reliability signal needed for CT–MRI registration?

Conventional pipelines report only aggregate quality metrics, so misregistrations can fail silently when no ground truth alignment exists at deployment time. A reliability signal supports deciding whether to accept automatically, seek expert review, or reject and retry.

What diagnostics are combined to form the reliability score R?

R integrates four diagnostics: quality improvement (ΔNMI and ΔSSIM), spatial overlap (Dice), registration stability from repeated initialisations, and inverse consistency error from forward/backward registration discrepancy.

How are the Green/Yellow/Red risk categories determined?

Thresholds learned from training patients are applied unchanged to held-out test patients. The resulting reliability score R is mapped into three tiers (Green automatically acceptable, Yellow review/attention, Red rejection/retry).

Jul 16, 2026

Upload by Eliana