Anatomy-Grounded Synthetic Coronary Angiography for Geometry-Informed Multi-View Matching (MICCAI 2026)

Abstract

Reliable multi-view correspondence in coronary angiography is hard to supervise: real X-ray angiograms have no dense ground-truth matches. We generate anatomy-grounded synthetic angiograms from coronary CT angiography (CCTA) volumes and their vessel masks using a physically grounded DRR (digitally reconstructed radiograph) pipeline. Because every frame is projected from known 3D geometry, each image pair comes with dense 3D-to-2D correspondence labels at no manual-annotation cost. We simulate routine clinical C-arm views for the right and left coronary trees, yielding 26,205 multi-view image pairs split by patient. On this data we train GIMM, a Geometry-Informed Matching Module that adds two priors to a coarse-to-fine matcher: view-class conditioning via FiLM and epipolar gating at the fine stage.

Method

Data

Physically grounded DRR pipeline

CCTA volumes and coronary masks projected under simulated C-arm geometry.
Projecting the 3D centerlines gives exact 2D correspondences and 3D positions for every match.
Dense supervision with no manual annotation.
26,205 multi-view pairs across the RCA and LCA trees, split by patient.

Model

GIMM: geometry-informed matching

Coarse-to-fine (LoFTR / QuadTree) matching backbone.
FiLM conditioning on the clinical view class.
Fine-stage epipolar gating that suppresses geometrically implausible matches.

GIMM architecture: a CNN extracts coarse and fine features from the two input views; a view-category embedding feeds an MLP that produces FiLM parameters modulating the coarse features before self- and cross-attention and coarse matching; the fine stage applies a Transformer with epipolar gating (a weighted 8-point fundamental-matrix estimate) before the final matches. — **GIMM architecture.** A view-category embedding drives FiLM modulation of the coarse features; the fine stage adds a Transformer with epipolar gating, a weighted 8-point fundamental-matrix estimate that suppresses off-epipolar matches.

Results

Because the synthetic data carries dense known geometry, correspondence error is measured directly in 2D pixels and 3D millimeters. GIMM achieves the lowest error across every top-K budget, ahead of LoFTR, QuadTree, and ASpanFormer.

Two line plots of distance error versus the number of retained matches (top-K = 3, 5, 10, 20, 40), comparing LoFTR, QuadTree, ASpanFormer, and GIMM. Left: 2D distance error in pixels. Right: 3D distance error in millimeters. GIMM has the lowest error at every top-K in both plots. — **Top-K ablation.** 2D (left) and 3D (right) distance error vs. the number of retained matches; GIMM (red) stays lowest across all budgets.

Compared against correspondence baselines on the synthetic DRR test set (top-K = 20) and on real coronary angiography, GIMM leads on 2D / 3D accuracy and on real-data landmark F1:

Correspondence accuracy on synthetic DRR (top-K = 20) and real coronary angiography. Lower is better for distances (↓), higher for precision and F1 (↑); best per column in **bold**.
Method	Synthetic (2D)			Synthetic (3D)	Real
Method	Dist. (px) ↓	Prec@3px ↑	Prec@5px ↑	Dist. (mm) ↓	F1@5px ↑	F1@10px ↑
SuperGlue	7.249	0.425	0.691	4.251	0.151	0.273
ASpanFormer	5.678	0.515	0.878	1.478	0.149	0.272
LoFTR	5.394	0.505	0.772	1.395	0.170	0.300
QuadTree	5.102	0.536	0.795	1.238	0.168	0.308
Ours (GIMM)	4.738	0.547	0.813	1.171	0.175	0.332

BibTeX

Cite this work

@inproceedings{gimm_miccai2026,
  title     = {Anatomy-Grounded Synthetic Coronary Angiography for
               Geometry-Informed Multi-View Matching},
  author    = {Lee, In Kyu and Seo, Sumin and Min, Jaesik},
  booktitle = {Medical Image Computing and Computer Assisted
               Intervention (MICCAI)},
  year      = {2026}
}