TomoGRAF: An X-ray physics-driven generative radiance field framework for extremely sparse view CT reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiotherapy

Background:

• Computed Tomography (CT) enables high-resolution 3D visualization but typically requires numerous angular samples.
• Physical and mechanical constraints often limit practical data acquisition, especially for sparse-view CT.
• Existing sparse-view CT reconstruction methods, including deep learning and Neural Radiance Fields (NeRF), show limited success, particularly in ultra-sparse scenarios.

Purpose of the Study:

• To develop a novel method, TomoGRAF, for reconstructing high-quality 3D CT volumes from ultra-sparse X-ray projections.
• To address the challenges of limited angular sampling in CT imaging.
• To provide a generalizable solution for medical applications requiring 3D volumetric data from minimal X-ray views.

Main Methods:

• Developed TomoGRAF, a system incorporating a volume rendering module that simulates X-ray material attenuation based on CT geometry.
• Implemented a training strategy that penalizes differences between simulated and ground truth volumes, enhancing prior fidelity.
• Adapted Neural Radiance Field (NeRF) principles for X-ray physics, differing from standard visible light rendering.

Main Results:

• TomoGRAF demonstrated a significant performance improvement over state-of-the-art deep learning and NeRF methods.
• The system was trained on the LIDC-IDRI dataset and validated on an independent in-house dataset with distinct imaging characteristics.
• Achieved high-quality 3D CT volume reconstruction even with ultra-sparse projection data.

Conclusions:

• TomoGRAF offers the first generalizable solution for reconstructing 3D volumetric information from one or a few X-ray views.
• This advancement is crucial for applications like image-guided radiotherapy and interventional radiology.
• Enables essential 3D insights where traditional CT data acquisition is infeasible.