Conditional random fields for phase-based lung feature tracking with ultra-low-dose x-rays.

Area of Science:

• Medical Physics
• Radiotherapy
• Image Analysis

Background:

• Continuous internal tumor monitoring during radiation therapy is crucial for treatment accuracy.
• Minimizing additional imaging dose is a key challenge in current radiotherapy practices.

Purpose of the Study:

• To develop and evaluate an ultra-low-dose (ULD) kV x-ray imaging technique for continuous internal tumor monitoring.
• To implement a sequential feature-based position estimation method using linear-chain conditional random fields (CRFs).

Main Methods:

• Simulated ULD x-ray projections from four-dimensional computed tomography (4D-CT) datasets of eight patients.
• Training CRFs using Local Energy-based Shape Histogram features extracted from ULD images.
• Estimating tumor position by predicting one of ten breathing phases.

Main Results:

• Achieved a mean accuracy of 0.958 for tumor position estimation using CRFs, outperforming support vector machines (SVMs) (0.867 accuracy).
• Position estimation demonstrated a discretization error of 2.4-5.3 mm and a systematic error of 2.0-4.4 mm.
• The method proved effective even with a very low dose of ten incident photons per pixel.

Conclusions:

• Tumor position estimation using ULD kV x-rays and CRFs is comparable to state-of-the-art methods.
• The trained CRFs can predict subsequent breathing phases, offering a precise post-treatment evaluation tool.
• This approach enables precise tumor tracking with significantly reduced imaging dose during radiation therapy.