Unsupervised Learning for Robust Respiratory Signal Estimation From X-Ray Fluoroscopy.

Area of Science:

• Medical Imaging
• Interventional Radiology
• Signal Processing

Background:

• Respiratory signals are crucial for image gating and motion compensation in minimally invasive interventions.
• Extracting these signals from X-ray fluoroscopy is difficult due to contrast agent injection and automatic exposure control.
• Existing methods struggle to maintain accuracy under these challenging imaging conditions.

Purpose of the Study:

• To develop a novel, robust method for respiratory signal extraction in X-ray fluoroscopy.
• To create a system that can tolerate common interventional imaging artifacts.
• To enable accurate motion compensation in minimally invasive procedures.

Main Methods:

• A dimensionality reduction approach using illumination-invariant kernel PCA on image patches of varying sizes.
• Automatic selection of relevant respiratory information via agglomerative clustering.
• Robust combination of signals from selected clusters into a single respiratory signal.

Main Results:

• The novel method achieved a correlation coefficient exceeding 91% with diaphragm motion ground truth.
• Performance remained high despite the presence of contrast agent injection and automatic exposure control.
• Similar respiratory signals were accurately estimated from both biplane sequences and sequences lacking a visible diaphragm.

Conclusions:

• The proposed method reliably extracts respiratory signals in challenging X-ray fluoroscopy scenarios.
• Its robustness to common artifacts makes it suitable for clinical application.
• This technique can enhance image gating and motion compensation in minimally invasive interventions.