MedNeRF: Medical Neural Radiance Fields for Reconstructing 3D-aware CT-Projections from a Single X-ray

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Computed Tomography (CT) is vital for diagnosing pathologies but involves significant ionising radiation exposure.
• Multidetector CT advances offer detailed imaging but increase radiation dose concerns.
• Excessive ionising radiation poses deterministic health risks.

Purpose of the Study:

• To develop a Deep Learning model for reconstructing CT projections from limited X-ray views.
• To reduce ionising radiation exposure in medical imaging.
• To create a continuous 3D representation of anatomical structures from 2D images.

Main Methods:

• A novel Deep Learning architecture based on neural radiance fields was employed.
• The model learns to disentangle shape and volumetric depth from 2D images.
• Training was performed on chest and knee CT datasets.

Main Results:

• The model achieved high-fidelity 3D renderings from few or single-view X-rays.
• Qualitative and quantitative comparisons demonstrated effectiveness against other radiance field methods.
• The model successfully inferred anatomical 3D structures.

Conclusions:

• The proposed Deep Learning model shows potential for significantly reducing radiation dose in CT imaging.
• This method enables 3D anatomical structure inference from minimal radiographic input.
• It offers a promising alternative for safer medical imaging practices.