IntraCross: Cross-modality graph matching for intravascular sequence registration

Overview

Intravascular imaging combining ultrasound (IVUS) and optical coherence tomography (OCT) is clinically valuable for atherosclerosis assessment, yet cross-modality sequence registration remains technically challenging due to disparate tissue sensitivities and acquisition protocols. Manual registration is time-consuming and subject to observer variability. Existing computational approaches enforce rigid frame-level correspondences that fail in regions of low anatomical information and propagate alignment errors through sequential processing stages.

Methods and approach

IntraCross introduces a graph matching framework that reformulates registration as a partial assignment problem rather than enforcing one-to-one frame correspondence. The approach simultaneously performs temporal and rotational alignment, extending partial matching techniques from 2D to 3D volumetric sequences. A temporal prior is incorporated into the matching process to regularize assignments and prevent instability in low-information regions. The method permits flexible correspondences while systematically rejecting unmatchable landmarks, eliminating the compounding error typical of two-stage sequential registration pipelines.

Key Findings

Validation across 77 vessels from 22 patients demonstrated high agreement with expert analyst ratings (Williams Index = 1.1; p = 0.62, 0.89, 0.07). The method achieved statistically significant improvements in circumferential registration accuracy compared to existing literature-reported approaches (p = 0.01, 0.04). Performance metrics indicate stability across the full range of tested anatomical conditions.

Implications

The framework addresses a fundamental limitation in multimodal intravascular imaging by enabling robust registration without manual intervention or strict frame-level constraints. Simultaneous temporal and rotational correction aligns the registration process with established clinical imaging workflows, reducing operator-dependent variability. The approach is extensible to other cross-modality registration tasks involving volumetric sequences with partial landmark visibility.