⚡ electrical engineering

Projection-Domain Sensitivity Analysis of Vertebral DRRs Under Intrinsic Calibration Perturbation

This study demonstrates that even minor perturbations in intrinsic fluoroscopy calibration significantly alter vertebral projection geometry and degrade 2D-3D registration accuracy, particularly in lateral views, highlighting the need for projection-domain consistency metrics alongside traditional reconstruction-based evaluations.

Original authors: Lin Li, Chaochao Zhou, Benjamin Aubert, Junlin Guo, Junchao Zhu

Published 2026-07-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lin Li, Chaochao Zhou, Benjamin Aubert, Junlin Guo, Junchao Zhu

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

Imagine you are trying to take a perfect photo of a complex, 3D puzzle piece (a vertebra) using a special camera that can see through bones. To make the photo look right, you have to tell the camera exactly how its lens works and where its center is. This is called "calibration."

Usually, scientists check if their camera is calibrated correctly by trying to rebuild the 3D puzzle piece from the photo. If the 3D model looks good, they assume the camera is perfect. But this paper asks a tricky question: What if the camera is "good enough" for building the 3D model, but still takes a slightly weird photo that confuses the computer trying to match the photo to the real thing?

The authors built a super-precise virtual world to test this. They didn't use real patients or real X-rays; they used computer models of spines and simulated the X-ray process. They took a perfect "ground truth" photo and then deliberately messed up the camera settings just a tiny bit—like slightly changing the zoom or shifting the center of the lens. They kept everything else (the bone's position and the camera's angle) exactly the same to see what happened.

Here is what they found:

1. The "Side View" is a Drama Queen
The paper discovered that the camera's settings matter way more depending on which angle you shoot from.

The Front View (AP): When looking at the spine from the front, the photo is surprisingly tough. Even when the authors messed up the camera settings by a lot (changing the focal length by 500 pixels), the picture barely changed. The bones looked almost the same, and the landmarks (like the tips of the bones) only moved less than 2 pixels. It's like taking a photo of a flat poster; even if your lens is slightly off, the poster still looks like the poster.
The Side View (LAT): When looking from the side, the camera went wild. With that same 500-pixel mess-up, the side view of the spine got distorted significantly. The landmarks jumped around a lot—some moved as much as 18.8 pixels (like the spinous process tip) or even 21.92 pixels (like the right pedicle). The outline of the bone changed shape noticeably.

Why? The authors explain that in their setup, the side view is "closer" to the detector, making the bone look bigger. Because the bone is magnified, a tiny error in the camera's math gets blown up, making the side view much more sensitive to calibration errors than the front view.

2. The "Reconstruction" Trick Doesn't Tell the Whole Story
The paper argues against the idea that if you can build a good 3D model, your camera is fine. They showed that you can have a camera that builds a perfect 3D model but still produces a "weird" 2D photo that looks different from the real thing. The side view, in particular, showed big differences in the bone's outline and shadow (silhouette) even though the 3D math was technically "okay."

3. The "Matching Game" Gets Messy
The final test was to see if this weird photo would confuse a computer trying to match the 2D photo to the 3D model (a process called 2D–3D registration).

When the computer used the messed-up camera settings, it got the rotation of the spine wrong. The error in rotation grew from about 0.05 degrees (almost perfect) to 0.37 degrees just by changing the focal scale.
If they tried to match the image directly without a helper algorithm, the rotation error was even worse, jumping from 0.5 degrees up to 3.8 degrees.
The position (translation) error also grew, moving from about 0.18 mm to 1.28 mm.

The Bottom Line
This study suggests that relying only on "3D reconstruction" to check if a camera is calibrated might be missing the point. The paper shows that small, almost invisible errors in the camera's internal settings can make the 2D photo look significantly different, especially from the side. This difference can trick the computer into thinking the spine is twisted or turned when it isn't.

The authors are careful to say these results come from simulations using computer models, not real patients. They didn't test this on real X-ray machines with real noise and blurry images. However, their findings suggest that for spinal imaging, we need to check how the camera handles the side view specifically, because that's where the tiny errors turn into big problems. They propose a new way to test cameras: not just by how well they build 3D models, but by how stable the 2D photos look when the settings wiggle just a little bit.

Technical Summary: Projection-Domain Sensitivity Analysis of Vertebral DRRs Under Intrinsic Calibration Perturbation

Problem Statement

Accurate geometric calibration is a prerequisite for fluoroscopy-guided spinal imaging, digitally reconstructed radiograph (DRR) generation, and 2D–3D vertebral registration. Traditionally, calibration quality is evaluated using reconstruction-domain metrics, such as reprojection error and volumetric reconstruction accuracy. However, many fluoroscopy-guided registration algorithms operate primarily in the projection domain, relying on the consistency between acquired radiographs and synthesized DRRs rather than reconstruction accuracy alone.

While recent studies suggest that calibration uncertainty affects the imaging geometry used for DRR generation, the specific impact of intrinsic calibration perturbations on projection-domain consistency remains insufficiently understood. This is particularly critical for spinal imaging, where vertebral anatomy involves complex 3D structures, overlapping bony features, and limited viewpoints. Small changes in imaging geometry can alter vertebral contours, pedicle appearance, and endplate alignment, potentially degrading image similarity metrics and registration convergence. This work addresses the gap between reconstruction-based calibration evaluation and projection-domain registration behavior by investigating how intrinsic calibration perturbations influence vertebral projection consistency and downstream 2D–3D registration accuracy, even when the underlying anatomy and acquisition pose remain fixed.

Methodology

The authors present a controlled synthetic framework to analyze the sensitivity of vertebral fluoroscopic projections to intrinsic calibration perturbations.

1. Synthetic Dataset and Geometry:

Models: Experiments utilize CT-derived vertebral segmentation masks and mesh models (from datasets like VerSe, SpineWeb, and TotalSegmentator), including isolated vertebrae and multi-vertebra lumbar segments (L1–L5).
Imaging Setup: A non-isocentric C-arm geometry is simulated with fixed source-to-detector (SDD) and source-to-isocenter (SID) distances. Two standard views are generated: Anterior–Posterior (AP) and Lateral (LAT). The vertebral model remains fixed in pose across views to isolate the effects of calibration variation from anatomical motion.
DRR Generation: Ray-driven projection rendering is used to generate DRRs at a 1024 × 1024 resolution based on line-integral attenuation accumulation.

2. Perturbation Model:

Intrinsic Parameters: Controlled perturbations are applied exclusively to the intrinsic calibration matrix ( $K$ ), while extrinsic parameters ( $R, t$ ) are held constant.
Specific Perturbations:
- Focal Scale: Multiplicative scaling of effective focal lengths ( $f_x, f_y$ ) to simulate magnification uncertainty.
- Piercing Point Displacement: 2D shifts of the principal point ( $c_x, c_y$ ) on the detector plane.
Comparison: DRRs generated with ground-truth parameters are compared against those generated with perturbed parameters.

3. Analysis Metrics:

Projection-Domain Consistency: Evaluated using anatomical landmark displacement, contour-based distances (Chamfer, Hausdorff), silhouette overlap (Dice coefficient, Intersection-over-Union), and raw DRR difference heatmaps.
Downstream Registration: A landmark-based 2D–3D rigid registration experiment is performed to quantify how projection inconsistencies propagate to pose estimation errors. Metrics include translational error (mm), rotational error (degrees), and residual reprojection error.

Key Contributions

Synthetic Framework: Introduction of a controlled methodology for analyzing projection-domain sensitivity to intrinsic calibration perturbations in vertebral fluoroscopic imaging, isolating calibration effects from anatomical motion.
Systematic Quantification: Comprehensive evaluation of projection consistency using complementary metrics: landmark displacement, contour deformation, silhouette overlap, and image similarity.
View-Dependent Sensitivity Characterization: Demonstration that lateral (LAT) vertebral projections are substantially more sensitive to intrinsic calibration variations than anterior–posterior (AP) projections.
Quantitative Link to Registration: Establishment of a quantitative relationship between intrinsic calibration perturbations, projection-domain inconsistency, and downstream 2D–3D registration error.
Complementary Evaluation: Demonstration that projection-domain analysis provides information distinct from and complementary to conventional reconstruction-based calibration evaluation.

Experimental Results

1. Qualitative and Global Consistency:

View Dependence: LAT projections exhibit significantly stronger and more spatially extended discrepancies in raw DRR difference heatmaps compared to AP projections under identical perturbation magnitudes (e.g., 100–500 pixel focal length shifts).
Silhouette Degradation: As perturbation magnitude increases, LAT projections show a marked decrease in silhouette consistency. For focal length perturbations from 100 to 500 pixels, the LAT Dice coefficient drops from ~0.995 to ~0.973, and IoU drops from ~0.991 to ~0.948. In contrast, AP projections remain highly stable (Dice > 0.997, IoU ~0.995).

2. Landmark Displacement:

Magnitude of Shift: LAT projections demonstrate substantially larger anatomical landmark displacements. For example, under specific perturbations, the inferior endplate and spinous process tips in LAT views shifted by approximately 15–18 pixels, whereas AP landmarks remained below 2 pixels.
Geometric Cause: This increased sensitivity in LAT views is attributed to the non-isocentric setup where the vertebra is closer to the detector, resulting in a larger projected scale. Consequently, intrinsic calibration inaccuracies produce amplified apparent landmark displacement.

3. 2D–3D Registration Impact:

Error Propagation: Projection inconsistencies induced by calibration perturbations propagate to downstream registration errors.
Rotational Sensitivity: Rotational alignment accuracy is particularly sensitive. In direct registration experiments, rotational error increased from ~0.5° to ~3.8° as focal scale perturbation increased. Kabsch-based registration also showed increased rotational error (0.05° to 0.37°) with perturbation.
Translational Stability: Translational errors remained comparatively stable but still showed monotonic increases with perturbation magnitude.

Significance and Claims

The paper claims that relying solely on reconstruction-based metrics (like reprojection error) is insufficient for evaluating calibration quality in fluoroscopy-guided vertebral imaging. The study demonstrates that:

Small perturbations matter: Even relatively small intrinsic calibration variations can produce measurable changes in vertebral projection geometry, contour morphology, and landmark localization.
View dependency is critical: The sensitivity of the imaging system is not uniform; lateral views are inherently more vulnerable to calibration errors due to depth compression and anatomical overlap.
Downstream impact: Projection-domain inconsistencies directly degrade the accuracy of 2D–3D registration, particularly affecting rotational alignment.

The authors position this work as a step toward a more holistic calibration evaluation that considers the specific requirements of projection-domain tasks (DRR generation and registration) rather than just volumetric reconstruction fidelity. The proposed framework offers a controlled environment for characterizing geometric sensitivity, which the authors suggest could be useful for future calibration analysis and DRR generation evaluation.

Limitations

The authors explicitly acknowledge several limitations:

Synthetic Nature: Experiments use CT-derived mesh models without modeling physical effects like scatter, detector noise, beam hardening, or non-uniform detector response.
Rigid Anatomy: The study assumes rigid vertebral anatomy, excluding soft tissue interaction, deformation, or patient motion.
Perturbation Scope: The analysis is limited to analytically defined intrinsic perturbations, whereas real systems may involve complex combinations of intrinsic, extrinsic, and temporal variations.
Registration Scope: The registration evaluation is limited to a controlled synthetic setting with predefined landmarks, not full intensity-based registration pipelines or real clinical image degradation.

Future work is proposed to validate these findings using real fluoroscopic datasets, incorporate realistic detector physics, and evaluate deformable anatomical simulations.