TG-OT matches features on a topological cylinder via unbalanced optimal transport, achieving fully automatic segmentation-free CCTA-IVUS registration on 47 paired cases (Dicectl=0.99, Sc=0.96, DiceL=0.69)

Feature detection is integrated directly into the registration pipeline, removing the dependency on pre-computed lumen or vessel wall segmentations: the CNNs operate in the polar (r, θ, z) domain and classify the presence of structures along each radial ray, so the task stays well-defined even where full segmentation is ill-posed under IVUS acoustic shadowing. Earlier differentiable and Transformer-based frameworks (Kadry et al., Li et al.) depend on pre-computed lumen or vessel wall segmentations, a condition hard to satisfy for IVUS where acoustic shadowing from calcifications obscures vessel wall boundaries; TG-OT instead predicts calcifications, bifurcations, and lumen radii on the topological (θ, z) cylinder, with the IVUS network additionally predicting guidewire artifacts. The method description and the ablation table (Table 1) report detection metrics across training loss configurations; with LBM+Dice, IVUS bifurcation Dice is 0.56 with Betti error 4.15, calcification Dice 0.67 with Betti error 3.89, and guidewire clDice 0.72, while MPR bifurcation Dice is 0.61 and calcification Dice 0.63, all as mean (std) over 5-fold cross-validation.

An unbalanced Sinkhorn optimal transport loss on cylindrical geometry is introduced for cross-modal feature matching, providing informative gradients even when predictions are spatially disjoint. A voxel-wise loss such as Dice penalizes feature mismatch uniformly regardless of spatial proximity and gives no gradient signal when predicted features do not overlap; this work formulates matching as an optimal transport problem on the topological cylinder S1×R with geodesic cylindrical transport cost, and uses unbalanced marginal relaxation τ∈(0,1) to tolerate legitimate asymmetries in detection confidence between modalities. The loss form, cost matrix, and Sinkhorn iterations are given in full in the method section (αOT=0.1, ε=0.1, τ=0.8); Table 2 shows OT+Dice yields a statistically significant improvement in rotational alignment over the Dice-only baseline (Wilcoxon signed-rank p<0.01), with Sc Q1 of 0.92 versus 0.90.

Registration accuracy is demonstrated without any prior segmentation on N=47 paired cases, with recovery of alignment in cases of severe catheter rotational drift. The authors note that no directly comparable fully automatic CCTA-IVUS registration method exists and that direct comparison with semi-automatic approaches such as Kadry et al. is precluded by segmentation label unavailability and modality mismatch, so they evaluate against the unregistered case and a standard Dice feature matching loss, additionally ablating the lumen alignment term and evaluating NMI as an image similarity term. Table 2 reports median (Q1, Q3) over 5-fold cross-validation: the unregistered baseline gives Dicectl 0.67, Sc 0.21, DiceL 0.47, while the final OT+Dice with NCC lumen matching reaches Dicectl 0.99, Sc 0.96, DiceL 0.69; optimization took 90 seconds per case on a single RTX 2080 Ti GPU.

Failure modes and reliability boundaries are characterized explicitly, and all N=47 cases are included in the reported metrics. The authors report that longitudinal initialization failed in N=4 cases due to incorrectly truncated automatic centerline extraction from CCTA, and rotational alignment degraded in N=2 cases with extreme landmark sparsity; they further state that registration reliability is bounded by the feature detection step, for which foundation models may offer greater robustness at scale. The failure categories and case counts are stated explicitly in the discussion; the authors also note the absence of automatic reference standards for CCTA-IVUS registration, making the constructed baselines the most meaningful available comparison.

The result is meant for coronary analysis settings with paired CCTA and IVUS data, and for research or pre-clinical pipelines that aim to obtain centerline, rotational, and lumen alignment without manual interaction or prior 3D segmentation; the method relies on an automatically extracted CCTA centerline and is driven by feature detection on polar-resampled data, so its premise is correct centerline extraction plus detectable calcification, bifurcation, or lumen signal in both modalities. The authors state that removing the dependency on error-prone 3D intermediate segmentations addresses a key barrier to scalable multimodal coronary analysis and brings fully automatic CCTA-IVUS fusion closer to clinical deployment; they also note that because registration is driven by features detected on the resampled CCTA volume, its reliability is bounded by that of the feature detection step, for which foundation models may offer greater robustness at scale.

Readers should keep in mind that, as the authors state, no directly comparable fully automatic CCTA-IVUS registration method exists and direct comparison with semi-automatic approaches is precluded by segmentation label unavailability and modality mismatch, so Table 2 compares against the unregistered case and a Dice-only baseline; automatic reference standards for CCTA-IVUS registration are absent, and the reference centerline was manually identified by matching large side branches visible in both modalities. On failure modes, longitudinal initialization failed in N=4 cases due to incorrectly truncated automatic centerline extraction from CCTA and rotational alignment degraded in N=2 cases with extreme landmark sparsity, although all N=47 cases are included in the reported metrics. In addition, registration reliability is bounded by the feature detection step, and while the authors suggest foundation models may offer greater robustness at scale, this is not validated here; optimization took 90 seconds per case on a single RTX 2080 Ti, so behavior at large scale or under clinical time constraints remains an open question.