Abstract
AbstractThe advent of 3D Magnetic Resonance Imaging (MRI) has revolutionized medical imaging and diagnostic capabilities, allowing for more precise diagnosis, treatment planning, and improved patient outcomes. 3D MRI imaging enables the creation of detailed 3D reconstructions of anatomical structures that can be used for visualization, analysis, and surgical planning. However, these reconstructions often require many scan acquisitions, demanding a long session to use the machine and requiring the patient to remain still, with consequent possible motion artifacts. The development of neural radiance fields (NeRF) technology has shown promising results in generating highly accurate 3D reconstructions of MRI images with less user input. Our approach is based on neural radiance fields to reconstruct 3D projections from 2D slices of MRI scans. We do this by using 3D convolutional neural networks to address challenges posed by variable slice thickness; incorporating multiple MRI modalities to ensure robustness and extracting the shape and volumetric depth of both surface and internal anatomical structures with slice interpolation. This approach provides more comprehensive and robust 3D reconstructions of both surface and internal anatomical structures and has significant potential for clinical applications, allowing medical professionals to better visualize and analyze anatomical structures with less available data, potentially reducing times and motion-related issues.
Publisher
Cold Spring Harbor Laboratory