A Scoping Review: Using Convolutional
Neural Network in 3D Reconstruction
from 2D Radiographs
Chan Li Jin Melissa
Diagnostic Radiography Student
BSc. (Hons) Diagnostic Radiography,
Singapore Institute of Technology
1800530@sit.singaporetech.edu.sg
Luis Lanca
Associate Professor
Singapore Institute of Technology (SIT)
Jenni Ahjonpalo
Radiography and Radiotherapy Student
Metropolia University of
Applied Sciences
ABSTRACT
Aim: Using Convolutional Neural
Network (CNN) in the process of re-constructing
three-dimensional (3D)
models from two-dimensional (2D) ra-diographs
has been explored in recent
years. In this review, we identify such
articles and discuss clinical accuracy of
results, time taken, and the remaining
challenges for this AI development to be
clinically applied.
Methods: A scoping review was
conducted following the PRISMA-ScR
framework using online databases (Se-mantic
Scholar, Google Scholar, IEEE
Xplore Digital Library, Springer Link
and PubMed), filtered to include only
recent (past 5 years) full-text peer-re-viewed
articles in English. Eight articles
were selected.
Results: Authors proposed methods
using CNN in the process of reconstruct-ing
3D anatomical models from spine
Jenna Perälahti
Radiography and Radiotherapy Student
Metropolia University of
Applied Sciences
Teemu Iivarinen
Information and Communication
Technology Student
Metropolia University of
Applied Sciences
and lower limb 2D radiographs. Overall,
the results were reported to have accept-able
accuracy when compared to ground
truth. The reconstruction processes
were also fully-automated and had fast-er
reconstruction times compared to
other methods.
Conclusion: With CNN, fast and ac-curate
3D models can be reconstructed
from 2D radiographs to aid in diagnosis
and treatment planning of bone-related
pathologies, providing a possible alter-native
to high radiation Computed To-mography
(CT) imaging for certain pro-cedures.
Keywords: Convolutional Neural
Network, 2D radiographs, 3D recon-struction
INTRODUCTION
CT and Magnetic Resonance Imaging
(MRI) are common radiological imaging
Vesa Ollikainen
PhD, Senior Lecturer
Metropolia University of
Applied Sciences
Eija Metsälä
Docent, Principal Lecturer
Metropolia University of
Applied Sciences
procedures used to obtain patient-spe-cific
3D reconstructions of anatomy
to aid in diagnosis, surgical planning,
treatment and follow-up of bone-relat-ed
pathologies (Reyneke et al.. 2018).
However, CT imaging can result in high
levels of ionizing radiation for patients
with conditions like scoliosis, that re-quire
scans over a large area for repeti-tive
follow-ups (Levy et al.. 1996). This
is a deterring factor due to possible bi-ological
effects e.g. cancer, especially for
younger patients (Lin 2010). While MRI
does not utilize ionizing radiation, there
are other associated risks due to the use
of a strong static magnetic field, gradi-ent
magnetic fields and radiofrequency
pulses, and possibly contraindicated for
patients with metallic fragments or im-plants,
electronic implants, or tattoos,
etc. (Dill 2007). MRI also has longer ac-quisition
and processing times and is
more frequently used to evaluate struc-
Kliininen Radiografiatiede 2021 13
link
