💻 computer science

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

GeoPose is a patient-agnostic, population-trained framework that achieves rapid and accurate CTA-to-DSA registration through projection-space calibration and transform composition, eliminating the need for patient-specific adaptation or explicit preregistration while significantly outperforming baseline methods in both speed and accuracy.

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

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

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

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

When a stroke strikes, time is not just a measurement; it is the difference between a full recovery and permanent disability. In the most severe cases, a large blood vessel in the brain becomes blocked, cutting off the supply to a vast network of tissue. To save the brain, doctors must perform a delicate procedure called endovascular thrombectomy, where they thread a tiny catheter through the arteries to physically remove the clot. This is a race against the clock, and the surgeon's view is limited to two-dimensional X-ray images taken from a single angle at a time. These images show the blood vessels clearly, but they lack depth. They do not show the patient's unique three-dimensional anatomy, which was mapped out in a detailed 3D scan taken before the surgery began. To navigate safely, doctors need to merge these two worlds: the live, flat X-ray view and the pre-operative 3D map. Doing this manually is slow and difficult, and in an emergency, every second counts.

The challenge lies in aligning these two different types of images. The pre-operative scan is a volumetric 3D model, while the live X-ray is a flat projection. To overlay them perfectly, a computer must figure out exactly how the X-ray machine is positioned relative to the patient's head. If the machine is tilted even slightly differently than the computer assumes, the 3D map will be projected onto the wrong spot, potentially leading a surgeon to aim at healthy tissue instead of the blockage. Traditional methods for aligning these images often require the computer to guess and check thousands of times, a process that can take minutes or even fail if the starting guess is poor. Other approaches try to learn from the specific patient, but that requires training a new computer model for every single person, which is too slow for an emergency room.

A team of researchers has developed a new system called GeoPose that solves this alignment problem in a fraction of a second without needing to learn about the specific patient beforehand. Instead of treating every new patient as a unique puzzle, the system learns a general understanding of how the human head and its blood vessels look from different angles. It uses a shared, standardized mental model of the head to make an initial guess about where the X-ray machine is pointing. Once it has this guess, it performs a quick mathematical adjustment to translate that guess into the specific coordinate system of the patient's actual 3D scan. This step, which the researchers call projection-space calibration, allows the system to take a general solution and apply it instantly to a new, unseen patient.

The system works in two main stages. First, it looks at the live X-ray image and predicts the camera's position relative to a standard, idealized head. Then, it calculates how that standard head differs from the specific patient's head and adjusts the prediction accordingly. This allows the system to skip the slow, trial-and-error process entirely. In tests involving eighty different X-ray observations from twenty patients who were not part of the training data, the system achieved a high level of accuracy in just 0.15 seconds. It placed the 3D blood vessel map so close to the actual vessels in the X-ray that the average distance between them was only 5.8 millimeters. For comparison, the best existing methods that did not use this new approach had an average error of 14.5 millimeters, which is often too large to be clinically useful.

The researchers also tested what happens if they allow the system to make a few quick, small corrections after the initial guess. By running a brief optimization process for just twenty-five steps, the system improved its accuracy even further, reducing the average error to 4.6 millimeters in about two seconds. This is a significant improvement over other methods that, even after the same amount of time, remained stuck with errors around 14.6 millimeters. The system's ability to work without patient-specific training is its most critical feature. Unlike other learning-based tools that require minutes of setup for each new person, GeoPose uses a fixed set of rules learned from a large group of people. It simply needs one quick look at the patient's scan to calibrate itself, making it ready for immediate use in a life-or-death situation.

The results suggest that this approach can provide surgeons with a reliable, real-time roadmap of the brain's blood vessels overlaid directly onto their live X-ray view. This would allow them to see exactly where the blockage is and how the surrounding vessels are arranged without needing to inject more contrast dye or wait for slow computer processing. The system was tested on data from acute stroke patients, and the researchers found that it worked consistently well across different angles and patient anatomies. While the system is not perfect and still struggles slightly with certain angles where depth is harder to judge, it represents a major step forward in making complex 3D navigation available during the most time-sensitive medical procedures. By turning a difficult, multi-step mathematical problem into a fast, single-step prediction, GeoPose brings the promise of instant, accurate guidance closer to reality for stroke care.

Technical Summary: GeoPose

Problem Statement

In acute ischemic stroke treatment, specifically endovascular thrombectomy (EVT), rapid alignment of pre-procedural Computed Tomography Angiography (CTA) with intraoperative biplanar Digital Subtraction Angiography (DSA) is critical. DSA provides high-resolution 2D vascular dynamics but lacks direct spatial correspondence to the 3D CTA volume. Establishing this correspondence requires estimating the C-arm pose (a rigid 3D-to-2D registration) to project 3D anatomical targets onto 2D acquisition views.

Current methods face significant limitations:

Optimization-based methods (e.g., gradient descent on differentiable rendering) are sensitive to initialization and often require hundreds of iterations, making them too slow for time-critical stroke workflows.
Learning-based methods often rely on patient-specific training or adaptation (e.g., fine-tuning for minutes per patient), which is undesirable in emergency settings where processing time competes with treatment delays.
Existing population-trained approaches often fail to account for the coordinate frame differences between a learned population template and a new patient's native, unregistered CTA volume.

Methodology: GeoPose

GeoPose is a population-trained framework designed to estimate C-arm poses in a learned canonical frame and transfer them to a patient's native CTA frame without patient-specific adaptation or explicit 3D preregistration. The pipeline consists of three main stages:

1. Population-Trained Pose Estimation (Canonical Frame)

The method utilizes two fixed-weight networks trained exclusively on synthetic renderings of a population aligned to a shared canonical coordinate frame. This decomposition separates pose from patient-specific anatomy.

GeoPose-Init: A residual network that regresses a 6-degree-of-freedom C-arm pose relative to a view-dependent "isopose" (a fixed anchor rotation/translation for PA, LAT-, and LAT+ views). It takes a DSA Maximum Intensity Projection (MAP) as input and predicts a view class and a residual transform ( $\Delta T$ ) relative to the anchor.
GeoPose-Refine: A second network that predicts a residual pose correction ( $\delta T$ ) based on the discrepancy between the observed DSA MAP and a Digitally Reconstructed Radiograph (DRR) rendered from the current pose estimate. It uses a shared encoder to process both images and a comparative embedding to predict the correction.

2. Projection-Space Calibration (Frame Transfer)

To apply the canonical-frame predictions to an unseen, unregistered native CTA, GeoPose employs a "render-and-compose" calibration:

A single synthetic DRR is rendered from the native CTA at a known calibration pose ( $T_{cal}^{nat}$ ).
GeoPose-Init processes this synthetic image to predict the pose in the canonical frame ( $\hat{T}_{cal}^{can}$ ).
The transformation between the native and canonical frames ( $\hat{A}$ ) is recovered by composing the known calibration pose with the inverse of the predicted canonical pose: $\hat{A} = T_{cal}^{nat} (\hat{T}_{cal}^{can})^{-1}$ .
This transform $\hat{A}$ is then applied to all subsequent pose predictions, mapping them from the canonical frame to the patient's native volume frame.

3. Pose Refinement and Optimization

Gated Residual Refinement: GeoPose-Refine is applied iteratively. A greedy acceptance criterion based on image similarity (multi-scale Normalized Cross-Correlation, mNCC) ensures that only corrections improving the alignment are retained. This provides a fast, optimization-free operating point.
Test-Time Optimization: The refined pose initializes a short, low-budget optimization (25 iterations of NAdam) using the GeoReg objective, which combines image similarity and a lightweight silhouette constraint. This step fine-tunes the alignment to the specific DSA pair.

Key Contributions

Compositional Frame Transfer: The introduction of a render-and-compose calibration that allows a population-trained pose estimator to operate directly on native, unregistered CTA volumes. This eliminates the need for per-patient network training or explicit inter-volume 3D preregistration.
Population-Trained, Image-Gated Refinement: A shared refinement module that predicts pose corrections from paired DSA and CTA renderings. An image-similarity gate prevents harmful recursive updates, offering a robust, optimization-free alternative.
Rapid Test-Time Registration: The combination of GeoPose initialization with a low-budget optimization budget (25 iterations) achieves registration in approximately 2 seconds, significantly faster than traditional optimization-based methods.
Experimental Validation: Validation on a held-out test set of 80 DSA observations from 20 patients, demonstrating superior performance over baselines without patient-specific adaptation.

Results

Evaluated on the ISLES'24 dataset (with TopCoW used for out-of-distribution testing), GeoPose achieved the following:

Optimization-Free Performance: Without test-time optimization, GeoPose (Init + Greedy Refine) achieved a carotid mean projected centerline distance (mPCD) of 5.8 mm and a centerline Dice (clDice) of 0.45. This compares favorably to the best-performing baseline (LXPose with two networks), which achieved 14.5 mm mPCD and 0.28 clDice. The runtime for this stage was 0.15 seconds.
With Test-Time Optimization: After 25 optimization iterations, GeoPose reached an mPCD of 4.6 mm and a clDice of 0.58 in approximately 2 seconds.
Baseline Comparison: Under the same 25-iteration budget, a native-initialized optimization (GeoReg) achieved 14.6 mm mPCD and 0.15 clDice, highlighting the critical importance of the learned initialization.
Out-of-Distribution: When trained on TopCoW and tested on ISLES'24, performance decreased (mPCD 8.5 mm), indicating a remaining domain gap related to differences in spatial resolution and vascular coverage.

Significance and Claims

The paper claims that GeoPose provides rapid, native-frame registration using fixed population-level weights, eliminating the need for patient-specific preparation that could delay acute stroke treatment.

Clinical Utility: By establishing spatial correspondence between pre-procedural CTA and intraoperative DSA, the method enables the projection of 3D anatomical roadmaps and treatment targets onto live acquisition views. This supports vessel localization, preserves anatomical context, and potentially reduces the need for repeated contrast injections.
Downstream Application: The recovered poses provide the geometric basis required for biplanar 3D vascular reconstruction from DSA sequences without separate rotational acquisitions.
Efficiency: The framework achieves high accuracy (mPCD < 6 mm) in under 2 seconds, a timeframe compatible with the strict time windows of acute ischemic stroke interventions.

The authors note that while the method is robust, pose estimation remains sensitive to observability (particularly translation along the source-detector direction and out-of-plane rotation) and domain shifts between training and test data. Future work may explore multi-view calibration to improve robustness.