Domain agnostic 2D-3D deformable registration Application to fluoroscopic guidance without contrast agent

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate 3D displacement estimation from 2D images is crucial for image-guided interventions.
• Existing 2D-3D registration methods often rely on statistical motion models, limiting their ability to predict non-standard deformations.

Purpose of the Study:

• To develop a real-time method for estimating 3D displacement fields from single fluoroscopic images.
• To enable accurate prediction of complex tissue deformations, independent of predefined motion patterns.

Main Methods:

• A fully convolutional network architecture was employed to solve the inverse problem of 3D displacement estimation.
• Supervised learning was performed on synthetic data, using Digitally Reconstructed Radiographs (DRRs) as input and displacement fields as output.
• Randomized Gaussian kernels were used to generate a synthetic training dataset with smooth and diffeomorphic displacement fields, avoiding reliance on statistical motion models.

Main Results:

• The proposed method accurately estimates 3D displacement fields in real time from single fluoroscopic images.
• The model successfully predicts deformations unrelated to breathing or other predetermined motion patterns.
• Clinical application demonstrated accurate prediction of deformations during percutaneous needle insertions.

Conclusions:

• The developed method offers a novel approach to real-time 3D displacement estimation from fluoroscopic images.
• This technique has the potential to improve image-guided procedures, such as percutaneous interventions, by accurately predicting tissue deformation.
• The approach may reduce the need for contrast agent injections in certain clinical applications.