Deformable motion compensation in interventional cone-beam CT with a context-aware learned autofocus metric

Area of Science:

• Medical Imaging
• Artificial Intelligence in Radiology
• Interventional Radiology

Background:

• Interventional Cone-Beam CT (CBCT) provides 3D visualization for abdominal interventions but is limited by motion artifacts due to long acquisition times.
• Existing autofocus methods use handcrafted metrics, lacking anatomical context and awareness of underlying motion.
• A novel, data-driven approach is needed to accurately quantify motion-induced degradation in CBCT.

Purpose of the Study:

• To introduce a learned, context-aware, deformable metric, Visual Information Fidelity-Deep Learning (VIF-DL), for motion quantification in CBCT.
• To develop a deep convolutional neural network (CNN) capable of assessing image quality degradation and anatomical realism.
• To improve motion compensation strategies for enhanced guidance in interventional procedures.

Main Methods:

• A deep CNN architecture was designed with multi-branch processing for voxel-wise and contextual feature extraction.
• The CNN was trained to emulate a reference-based structural similarity metric (VIF) on simulated motion-corrupted CBCT data.
• Performance was validated using correlation metrics, simulation studies with varying parameters, and experimental phantom data, followed by integration into an autofocus framework.

Main Results:

• VIF-DL demonstrated high correlation (0.95 in simulation, 0.88 in real data) with ground truth VIF and good correlation (0.90) with motion fields.
• The metric accurately reflected varying motion amplitudes and frequencies, distinguishing mild from severe motion.
• Autofocus compensation using VIF-DL significantly reduced motion artifacts, improving spatial resolution by up to 9.20% and vessel sharpness by 9.64%.

Conclusions:

• The proposed VIF-DL metric effectively quantifies motion-induced image quality degradation and anatomical plausibility in a reference-free manner.
• Its context-aware architecture ensures robust performance across diverse motion patterns, imaging techniques, and anatomical variations.
• VIF-DL represents a significant advancement over conventional metrics, paving the way for improved deep autofocus motion compensation in clinical interventions.