Single-View Fluoroscopic X-Ray Pose Estimation: A Comparison of Alternative Loss Functions and Volumetric Scene...

Area of Science:

• Medical Imaging
• Computer Vision
• Computational Anatomy

Background:

• Image-guided procedures often rely on pose estimation to navigate 3D space.
• Accurate pose estimation is crucial for surgical precision and patient safety.

Purpose of the Study:

• To develop and evaluate a novel framework for fluoroscopic pose estimation.
• To compare the efficacy of different loss functions and volumetric scene representations within this framework.

Main Methods:

• Developed a differentiable projection (DiffProj) algorithm for generating Digitally Reconstructed Radiographs (DRRs).
• Introduced and compared two neural scene representations: Neural Tuned Tomography (NeTT) and masked Neural Radiance Fields (mNeRF).
• Utilized iterative gradient descent with various loss functions to perform pose estimation against ground-truth fluoroscopic images.

Main Results:

• Mutual Information loss function significantly outperformed other tested loss functions, preventing local optima entrapment.
• Both discrete (CBCT) and neural (NeTT, mNeRF) scene representations achieved comparable pose estimation accuracy (mean 3D angle error ≤ 3.2°).
• Neural scene representations demonstrated a higher computational cost during training compared to discrete methods.

Conclusions:

• Mutual Information is a robust loss function for fluoroscopic pose estimation.
• Neural scene representations show promise for pose estimation but require further optimization for computational efficiency.
• The developed framework provides a flexible platform for evaluating pose estimation techniques in medical imaging.