GeoPose: Patient-agnostic CTA-to-DSA registration through projection-space calibration

Rudolf L. M. van Herten; Robert Graf; Paula Feldman; Johannes C. Paetzold

cs.CVAug 17, 2026

GeoPose: Patient-agnostic CTA-to-DSA registration through projection-space calibration

Authors: Rudolf L. M. van Herten, Robert Graf, Paula Feldman, Johannes C. Paetzold

Organizations: Weill Cornell Medicine, Cornell Tech New York, NY 10044 · Technical University of Munich 81675 Munich

Abstract

Aligning intraoperative biplanar digital subtraction angiography (DSA) to pre-procedural computed tomography angiography (CTA) requires rapid and accurate 3D-to-2D registration. Optimization-based methods are sensitive to initialization and may require hundreds of iterations, whereas learning-based approaches commonly rely on patient-specific training. We propose GeoPose, a population-trained framework that estimates the C-arm pose in a learned canonical frame and transfers it to the native frame of an unseen CTA through projection-space calibration and transform composition. A population-trained residual network refines the pose, followed optionally by low-budget image-driven optimization. GeoPose requires neither patient-specific adaptation nor explicit inter-volume preregistration. On 80 DSA observations from 20 held-out patients, optimization-free GeoPose achieved a carotid mean projected centerline distance (mPCD) of 5.8 mm and a clDice of 0.45, compared with 14.5 mm and 0.28 for the best-performing baseline, while requiring only 0.15 s. After 25 optimization iterations, GeoPose reached an mPCD of 4.6 mm and a clDice of 0.58 in approximately two seconds. Under the same budget, native-initialized optimization achieved 14.6 mm and 0.15, respectively. GeoPose thus provides rapid native-frame registration with fixed population-level weights and the geometric correspondence required for downstream biplanar 3D vascular reconstruction.

Figures & tables

Figure 1: Overview of the proposed CTA-to-DSA registration pipeline. (a) Pose estimation : GeoPose-Init regresses a canonical-frame transform from a DSA maximum intensity projection ( $\mathbf{T}_{\text{DSA}\leftarrow\text{can}}$ ) and a calibration transform from a canonical isopose DRR of the patient CTA ( $\mathbf{T}_{\text{nat}\leftarrow\text{can}}$ ). (b) Projection-space calibration : their composition maps the prediction into the patient’s native volume frame, absorbing template misalignment and providing an initial pose estimate $\mathbf{T}{{}_{\text{DSA}}}$ . (c) Pose refinement : GeoPose-Refine applies up to $N_{\text{ref}}$ image similarity-gated residual updates $\mathbf{T}{\delta\mathbf{T}}^{-1}$ from re-rendered DRRs to obtain the refined pose $\mathbf{T}_{\text{DSA}}^{R}$ , which is subsequently test-time optimized by GeoReg to the final pose $\mathbf{T}_{\text{DSA}}^{f}$ . The recovered native-frame pose establishes spatial correspondence between the pre-procedural CTA and intraoperative DSA, enabling CTA-based anatomical roadmapping in the acquisition views and providing the geometry required for downstream biplanar 3D vascular reconstruction.

Figure 2: Native CTA (top) and canonical template-aligned (bottom) renders at the fixed posteroanterior isopose for four subjects from the ISLES’24 dataset. Aligning all training volumes to a shared frame (left) makes pose estimation a generalized function across subjects rather than per-patient regression.

Figure 3: GeoPose training overview. (a) GeoPose-Init : a training CTA is rendered at a sampled pose $\mathbf{T}_{v}$ and intensity-augmented to resemble a DSA MAP image, and subsequently passed through the encoder $\mathcal{E}_{\mathrm{init}}$ . The network both classifies the originating view $v$ and regresses a residual $\Delta\widehat{\mathbf{T}}_{v}$ about the view-dependent isopose $\mathbf{T}_{\mathrm{iso},v}$ . At test time, the predicted view class $\widehat{v}$ is used to derive the isopose. The right panel shows the isopose anchors $\mathbf{T}_{\mathrm{iso},v}$ (yellow) and the distribution of sampled training poses $\mathbf{T}_{v}$ (blue) for the three view classes ( $v=\mathrm{LAT}^{-},\mathrm{PA},\mathrm{LAT}^{+}$ ) relative to the cranium. (b) GeoPose-Refine : a clean DRR rendered at a perturbed pose $\mathbf{T}_{\mathrm{noisy}}$ and an appearance-augmented, MAP-like DRR rendered at the optimal pose $\mathbf{T}^{*}$ are passed through a shared encoder $\mathcal{E}_{\mathrm{ref}}$ to obtain pooled embeddings $\mathbf{e}_{M}$ and $\mathbf{e}_{D}$ . Their fused comparative embedding $\mathbf{z}$ is used to regress a residual transformation $\delta\hat{\mathbf{T}}_{v}$ that maps the perturbed pose back toward $\mathbf{T}^{*}$ . The right panel illustrates the geometric relationship between $\mathbf{T}^{*}$ , $\mathbf{T}_{\mathrm{noisy}}$ , and the sampled perturbation $\delta\mathbf{T}$ . Both models are trained using a combination of image similarity, vascular overlap, and geodesic losses.

	NCC $\uparrow$			mPCD $\downarrow$ (mm)			clDice $\uparrow$			Runtime $\downarrow$ (ms)
	PA	LAT	Avg.	PA	LAT	Avg.	PA	LAT	Avg.
Native init.	0.57 (0.16)	0.34 (0.20)	0.46 (0.13)	14.9 (7.0)	28.4 (18.0)	21.6 (9.6)	0.14 (0.11)	0.09 (0.11)	0.11 (0.08)	-
xvr †	0.71 (0.11)	0.66 (0.09)	0.68 (0.08)	19.7 (10.0)	21.4 (13.7)	20.6 (9.2)	0.12 (0.09)	0.15 (0.13)	0.13 (0.09)	27.4 (2.3)
LXPose (net1) †	0.75 (0.16)	0.70 (0.13)	0.73 (0.13)	8.7 (5.8)	25.9 (21.4)	17.3 (12.8)	0.27 (0.17)	0.12 (0.18)	0.19 (0.13)	20.9 (2.3)
LXPose (net1 + net2) †	0.77 (0.16)	0.75 (0.13)	0.76 (0.13)	7.0 (6.4)	21.9 (20.8)	14.5 (12.4)	0.38 (0.18)	0.18 (0.20)	0.28 (0.15)	49.8 (6.7)
GeoPose-Init	0.78 (0.13)	0.72 (0.14)	0.75 (0.12)	9.4 (10.2)	22.0 (15.5)	15.7 (10.7)	0.34 (0.22)	0.13 (0.18)	0.24 (0.16)	21.8 (2.3)

Table 1: Instant DSA pose-estimation performance on the reserved ISLES’24 test set (

N_{\text{DSA}}=80

), without test-time optimization. Normalized cross-correlation (NCC), carotid mean projected centerline distance (mPCD), and centerline Dice (clDice) are reported. Runtime covers the complete initialization stage for one biplanar pair, measured on an NVIDIA H100 GPU. Results are presented as mean (standard deviation). Avg. denotes the patient-level average across PA and LAT views, while a

\dagger

indicates statistical significance (

p\leq 0.0015

) compared to our greedy refinement model using the Wilcoxon signed-rank test. –T denotes training with TopCoW data, while an asterisk ( ∗ ) denotes evaluation on the broader train+test cohort.

Figure 4: Test-time optimization (TTO) convergence and pose-parameter discrepancies on the ISLES’24 test set. Top : (a) patient-level normalized cross-correlation (NCC), (b) carotid mean projected centerline distance (mPCD), and (c) centerline Dice (clDice) versus NAdam iteration count, and (d) absolute translation-component discrepancies relative to the long-TTO reference pose. Bottom : (e–g) paired metric differences relative to the GeoReg native pose at matched iteration counts, and (h) absolute wrapped Euler-ZYX component discrepancies relative to the long-TTO reference. Pose panels compare LXPose (net1 + net2) and GeoPose-Init + Refine (greedy), before TTO and after independent 25-step schedules, against the paired GeoPose-Init + Refine (greedy) 400-step pose. Error bars show patient-level bootstrap 95% confidence intervals (

N=20

), while pre/post and PA/LAT observations are averaged within patient.

GeoPose-Init
Configuration	View acc. $\uparrow$	NCC $\uparrow$	mPCD $\downarrow$ (mm)	clDice $\uparrow$
Base	N/A	0.74 (0.11)	12.6 (6.6)	0.22 (0.15)
+ View classification	98.8 (5.6)	0.73 (0.13)	14.1 (10.3)	0.23 (0.16)
+ View embedding	100.0 (0.0)	0.76 (0.11)	13.3 (7.8)	0.23 (0.13)
+ View-anchor loss	98.8 (5.6)	0.76 (0.09)	13.0 (7.2)	0.25 (0.16)
+ mPD loss	100.0 (0.0)	0.75 (0.12)	15.7 (10.7)	0.24 (0.16)

Table 2: Cumulative GeoPose component ablation on the hold-out test set (

N_{\text{DSA}}=80

). Values are patient-level mean (standard deviation).

Initialization	Iter.	NCC $\uparrow$	mPCD $\downarrow$ (mm)	clDice $\uparrow$
DiffDRR (native) †	25	0.62 (0.12)	14.7 (6.7)	0.15 (0.10)
GeoReg (native) †	25	0.63 (0.13)	14.6 (7.2)	0.15 (0.08)
xvr †	25	0.82 (0.05)	13.3 (4.5)	0.22 (0.13)
LXPose (net1) †	25	0.83 (0.07)	8.8 (6.2)	0.41 (0.18)
LXPose (net1 + net2) †	25	0.84 (0.07)	8.0 (5.2)	0.42 (0.18)
GP-Init	25	0.84 (0.07)	7.1 (4.8)	0.48 (0.18)

Table 3: Patient-level registration performance after 25 iterations of test-time optimization (TTO), initialized from each method’s learned pose estimate. For each patient, results indicate mean (standard deviation) averaged across PA and LAT views. Normalized cross-correlation (NCC) and carotid mean projected centerline distance (mPCD) and clDice are reported. A

\dagger

indicates statistical significance w.r.t. our greedy refinement method at 25 TTO steps (

p\leq 0.01

). Bottom rows report a 400-iteration upper-bound reference for native-frame initialized GeoReg and for GeoPose (GP).

Figure 5: Qualitative progression of CTA-to-DSA registration for the median-performing test subject in terms of mPCD. The first six columns show CTA digitally reconstructed radiographs (DRRs) at the native, xvr, LXPose (net1 + net2), GeoPose-Init, GP-Init + Refine (greedy), and 25-step test-time optimization (TTO) poses. Vertical rules separate the native pose, baseline initializers, GeoPose-based methods, and final DSA visualization. In the first six columns, DSA-derived internal carotid artery centerlines are shown in magenta and projected CTA centerlines in cyan. The rightmost column projects the CTA-derived TopBrain vascular segmentation [ 6 ] onto the corresponding DSA maximum-intensity projection using the final 25-step pose. Distinct colors separate major vascular structures, illustrating how the correspondence transfers anatomical information from CTA to DSA. Values below the first six columns report the mean mPCD across the displayed PA and LAT views, followed by the cumulative runtime.

Medical Image Registration

Medical image registration aligns scans across time, patients, or imaging modalities to track anatomy and support diagnosis, treatment, and image-guided procedures.

See all 93 papers

Patient-Agnostic Synthetic Pretraining for Efficient Patient-Specific Intraoperative 2D/3D Registration
Jul 25, 2026Minheng Chen, Youyong KongSynthetic Data Pretraining Medical Image Registration
Department of Computer Science and Engineering, University of Texas at Arlington, Arlington, 76013, TX, USA. · School of Computer Science and Engineering, Southeast University, Nanjing, 210096, Jiangsu, China.
CAD-to-CT Registration of Cylindrical Objects via Ellipse-Based Axis Estimation
Jun 1, 2026Aleksander Ogonowski, Mikołaj Mrozowski, Daniel Więcek +9Medical Image Registration Computed Tomography
Department of Complex Systems, National Centre for Nuclear Research ul. So ltana 7, 05-400 Otwock-´Swierk, Poland · ImagineRT sp. z o.o., National Centre for Nuclear Research ul. So ltana 7, 05-400 Otwock-´Swierk, Poland
LayersReg: A Layer-by-Layer Progressive Regressor for Reliable Intraoperative 3D/2D Registration
Jun 25, 2026Xiyuan Wang, Zhenchao Wang, Xinran Chen +3Medical Image Registration Cross-Modal Image Registration
School of Biological Science and Medical Engineering Southeast University Nanjing · School of Engineering, College of Engineering and Physical Sciences University of Birmingham Birmingham · State Key Laboratory of Digital Medical Engineering Southeast University Nanjing

Jul 25, 2026J/K move · Enter open · S save

Patient-Agnostic Synthetic Pretraining for Efficient Patient-Specific Intraoperative 2D/3D Registration

Minheng Chen, Youyong Kong

Department of Computer Science and Engineering, University of Texas at Arlington, Arlington, 76013, TX, USA. · School of Computer Science and Engineering, Southeast University, Nanjing, 210096, Jiangsu, China.

Synthetic Data Pretraining Medical Image Registration 2D-3D Image Registration

Open paper AskPDF arXiv

Intraoperative 2D/3D registration aligns preoperative CT volumes with intraoperative X-ray or fluoroscopic images and is essential for image-guided interventions. Recent learning-based and differentiable registration methods have shown promising accuracy, especially in patient-specific settings where abundant digitally reconstructed radiographs (DRRs) can be synthesized from the target CT. However, training a separate patient-specific model from scratch for every new patient is computationally inefficient and limits practical deployment. In this work, we propose an efficient patient-specific 2D/3D registration framework based on patient-agnostic synthetic pretraining and spherical similarity learning. The model is first pretrained on synthetic DRRs generated from multiple CT volumes to learn transferable pose-sensitive representations, and is then adapted to a new patient using only a limited number of synthetic projections from the target CT. To improve synthetic-to-real robustness without requiring anatomical labels, we introduce a segmentation-free domain randomization strategy that perturbs image intensity, projection physics, field-of-view, occlusion, and fluoroscopic artifacts. The adapted model provides an initial pose estimate, which is further refined using spherical similarity learning and differentiable Levenberg-Marquardt optimization. Experiments on multiple anatomical datasets evaluate whether patient-agnostic synthetic pretraining can improve the efficiency of patient-specific registration, with particular focus on the trade-off between adaptation cost and registration accuracy. The results demonstrate that patient-agnostic synthetic pretraining can significantly reduce patient-specific training requirements while preserving accurate intraoperative 2D/3D registration.

Abstract

Medical Image Registration

Explore similar work

Patient-Agnostic Synthetic Pretraining for Efficient Patient-Specific Intraoperative 2D/3D Registration

CAD-to-CT Registration of Cylindrical Objects via Ellipse-Based Axis Estimation

LayersReg: A Layer-by-Layer Progressive Regressor for Reliable Intraoperative 3D/2D Registration