Central PA Medicine Spring 2021 - 9

dauphincms.org
A
recent study from Mayo Clinic
concluded that radiologists in their
healthcare system must interpret
one image every 3-4 seconds
during a normal workday in order to meet
workplace demands [1]. This has obvious
implications for increasing physician
burnout, a phenomenon known to be
associated with an increased incidence of
clinical errors [2], potentially leading to
catastrophic outcomes. Further, the amount
of radiological imaging data being generated
continues to increase disproportionately to
the number of trained radiologists able to
interpret such data [3]. It seems as though
only a superhuman would be able to meet
these demands without risking an increase
in errors, so perhaps the solution needed
is not human at all.
Physician burnout is a complex problem
requiring a multi-faceted approach for prevention
and reduction. Artificial intelligence
(AI) is one possible component of a solution
to this problem, and it has been getting
substantial attention in the radiology field
as researchers and clinicians shift toward a
big-data approach known as " radiomics. "
Radiomics is a field dependent upon AI,
as it encompasses data mining processes
aimed at extracting and analyzing features
from clinical and imaging data. Initially,
radiomic tools focused on AI techniques
that depend on pre-definition of characteristics
using algorithms based on expert
(human) knowledge [4]. More recently,
however, there has been much attention
paid to so-called " deep learning " algorithms,
which do not require pre-definition of
characteristics. Instead, deep learning algorithms
problem-solve by exploring the data
without requiring involvement from human
experts [5]. Deep learning also makes tissue
characteristics easily quantifiable [6] and
avoids the inevitable variability between
different human readers. Importantly, deep
learning technologies have been shown to
align with human radiologists' performance
in various imaging interpretation tasks in
multiple studies [4, 7].
The clinical value of deep learning radiomics
was recently demonstrated in a study by
Lao et al. in which deep learning algorithms
Radiomics is a field dependent
upon AI, as it encompasses
data mining processes aimed
at extracting and analyzing
features from clinical and
imaging data.
were used to extract data from magnetic
resonance images from glioblastoma multiforme
patients. These data were then used
to compile prognostic imaging signatures in
order to predict survival and stratify patients
for appropriate treatment regimens [8].
Another group employed a deep learning
approach to successfully predict response
to chemotherapeutic treatment in patients
with rectal cancer in a multi-center study
[9]. Similarly, Hosney et al. applied a deep
learning approach to analysis of computed
tomography imaging in order to stratify
patients with non-small-cell lung cancer
according to mortality risk [10].
There is much excitement regarding
the direct application of deep learning
to clinical problems like prognostication,
but one can also imagine the application
of other AI-based technologies to improve
radiologic efficiency. For example, Thrall et
al. proposed the application of algorithms
able to detect emergent conditions that are
frequently diagnosed using imaging (e.g.,
strokes or pulmonary emboli) with high
sensitivity that could potentially help radiologists
triage cases, ensuring that the most
urgent patients are attended to first. This
could improve outcomes by decreasing the
amount of time between presentation and
initiation of treatment and could potentially
lessen medical errors by alerting physicians
to the urgency of these cases. On the other
hand, AI applications programmed to have
a high negative predictive value could
help identify true-positives among cases
with a low likelihood of positive findings,
a context in which reader fatigue is likely
[11]. Such implementations could easily
increase radiologists' efficiency, helping
clinicians to meet workplace demands
while also potentially having a role in
preventing burnout.
AI is not without its challenges, of course.
For example, it is not known whether
AI programs will be generalizable across
different patient populations and demographic
groups; one can easily imagine
getting incorrect results if applying a
radiomic signature generated using imaging
from adults to that of pediatric patients.
Similarly, it is unknown whether AI tools
can be applied to images generated using
different protocols, or even those captured
using the same protocols, but at a different
institution or on a different machine [12].
Despite the unknowns, the implementation
of a radiomic approach is exciting and
rife with opportunities to improve clinical
care, outcomes, and efficiency. Further
research and the continuous development
and validation of deep learning algorithms
is sure to advance the field of radiology in
the coming years. With AI in one's toolbox,
superhuman productivity may be possible
after all.
Continued on page 10
Central PA Medicine Spring 2021 9
http://www.dauphincms.org
Central PA Medicine Spring 2021

Table of Contents for the Digital Edition of Central PA Medicine Spring 2021
