Skip to main content
Back to timeline
arXivSource publication:

X-Strata splits a 2D angiogram into three layers, recovering the aneurysm hidden behind crossing vessels

Synopsis

A 2D X-ray angiogram sums every structure along each ray into one pixel, so vessels crossing an aneurysm hide part of the sac and mix their contrast into its filling curve. X-Strata separates a cine angiogram into the complete sac, the vessels in front of it and the vessels behind it: contrast timing indicates how much each layer contributes, and parallax between two views indicates their order, after which the layers are lifted into 3D. On 220 held-out patients from three centers, two-view X-Strata reached overlap-region sac Dice of 0.693 internal and 0.641 external and balanced depth-order accuracy of 0.703 and 0.659, with a median 3D sac surface Chamfer distance of 0.34 mm on the internal cohort.

Source-provided article image: X-Strata: Separating Aneurysms and Overlapping Vessels into Layers from 2D X-ray Angiography
arXiv

Interpretation

X-Strata decomposes routine 2D cine digital subtraction angiography (DSA) into three layers, the complete aneurysm sac, the vessels in front of it and the vessels behind it, and returns a contrast curve for each layer. Earlier segmentation and completion methods recover masks but give neither a layer's contribution to a pixel nor its order, while vessel stereo estimates depth but inherits its masks and leaves the contrast unmixed. On 220 held-out patients from three centers, 122 of them external, two-view X-Strata reached overlap-region sac Dice of 0.693 internal and 0.641 external and balanced depth-order accuracy of 0.703 and 0.659, ranking first among 14 baselines on both endpoints.

The separation rests on two complementary cues: the time course of contrast in the sac and in the vessels sets how much each layer contributes to a pixel, and parallax between two views sets which layer lies in front. A single view cannot fix the order, because exchanging two layers along a ray while keeping their path lengths, contrast curves and arrival delays fixed leaves the cine unchanged. The network first estimates layer maps and contrast curves, then searches vessel depth along centerlines in the second view to produce signed depth probabilities, and finally lifts the layers along their rays into a shared 3D Gaussian representation; training uses labels projected from same-session 3D rotational angiography, and inference needs only 2D cine runs and their acquisition geometry.

Order accuracy rises as the angle between the two views widens, consistent with parallax supplying the order evidence. Every one-run method stays near 0.5 balanced depth-order accuracy, which is chance. Two-view X-Strata exceeds all 11 end-to-end baselines in order accuracy by at least 4.1 points internally and 3.3 points externally (Holm-adjusted), reaching 0.75 and 0.71 at wide angles.

Perspective

The evaluation covers 220 held-out patients from three centers, 122 of them external, and all of these patients underwent same-session 3D rotational angiography because scoring requires it as a reference; at inference the method uses only 2D cine runs and their acquisition geometry.

The authors note that a vessel near the sac can still give nearly identical partner coordinates for the two depth signs, that mask truncation and misregistration shift the centroids, and that magnification changes or movement between acquisitions break the assumed parallel projection and equal layer mass. They also note that all evaluated patients received 3D rotational angiography, that follow-up DSA differs in contrast delivery and temporal sampling, and that extending the method to patients with only 2D angiography and relating recovered kinetics to residual-filling grades requires a separate clinical reference.