Automated C-Arm Positioning via Conformal Landmark Localization

TL;DR

Problem addressed: reliable automatic C-arm positioning to predefined anatomical landmarks with quantified confidence to reduce radiation exposure and delays. Approach: an uncertainty-aware pipeline that predicts per-landmark 3D displacement as a Gaussian, estimates epistemic uncertainty via Monte Carlo Dropout, and applies split conformal prediction to produce $3$D spherical regions with coverage $1-\alpha$, augmented by skeleton-pose regularization and patient-position augmentation on a synthetic DRR dataset from CT scans. Contributions: uncertainty-aware C-arm landmark localization, formal coverage guarantees from conformal prediction, and evaluation across multiple backbones with a stroke thrombectomy use-case. Findings: strong localization performance and well-calibrated prediction bounds across architectures, with multi-stage intermediate landmarks improving convergence in the thrombectomy scenario.

Abstract

Accurate and reliable C-arm positioning is essential for fluoroscopy-guided interventions. However, clinical workflows rely on manual alignment that increases radiation exposure and procedural delays. In this work, we present a pipeline that autonomously navigates the C-arm to predefined anatomical landmarks utilizing X-ray images. Given an input X-ray image from an arbitrary starting location on the operating table, the model predicts a 3D displacement vector toward each target landmark along the body. To ensure reliable deployment, we capture both aleatoric and epistemic uncertainties in the model's predictions and further calibrate them using conformal prediction. The derived prediction regions are interpreted as 3D confidence regions around the predicted landmark locations. The training framework combines a probabilistic loss with skeletal pose regularization to encourage anatomically plausible outputs. We validate our approach on a synthetic X-ray dataset generated from DeepDRR. Results show not only strong localization accuracy across multiple architectures but also well-calibrated prediction bounds. These findings highlight the pipeline's potential as a component in safe and reliable autonomous C-arm systems. Code is available at https://github.com/AhmadArrabi/C_arm_guidance_APAH

Figures (5)

• Figure 1: Overview of the sampling process. We densely sample each CT volume by defining independent distributions across each spatial axis.
• Figure 2: Visualization of landmark annotations. These landmarks represent ROI used across different clinical applications.
• Figure 3: Overview of the proposed pipeline. The input X-ray is encoded using a backbone network, and the C-arm position is embedded into the image features. For each of the 14 anatomical landmarks, the model outputs the parameters of a 3D Gaussian: a mean vector $\boldsymbol{\mu} = (\mu_{\Delta x}, \mu_{\Delta y}, \mu_{\Delta z})$ and a diagonal covariance matrix $\boldsymbol{\Sigma} = \text{diag}(\sigma^2_{\Delta x},\sigma^2_{\Delta y},\sigma^2_{\Delta z})$, leading to $14\times3\times2$ output neurons. A skeleton pose loss further regularizes predictions based on the patient's prior anatomical topology. In Calibration and testing, we apply Monte Carlo dropout with $T$ stochastic forward passes to estimate the total uncertainty.
• Figure 4: C-arm movement visualization. In all landmarks, the model was successful in localizing the C-arm onto the corresponding ROI.
• Figure 5: Kernel Density Estimates (KDE) of prediction errors in $x$ and $y$ directions. Long tails, due to outliers, were clipped for visibility.