Item

Abstract

Background: Brain volumetry is a key aspect in dementia diagnostics.
We applied an artificial intelligence (AI)
system based on a Convolutional
Neural Network (CNN)
which aims to perform lobe-separated
rapid brain volumetry (< 1/2 h) of three-dimensional T1-weighted
magnetic resonance imaging (MRI)
with automated segmentation as
well as comparison to age- and gender-adapted percentiles. Our aim
was to quantify the added value in the differential diagnostics of dementia.
Methods: A total of 55 patients–17 with confirmed diagnosis of
Alzheimer’s disease (AD),
18 with confirmed diagnosis of frontotemporal
dementia (FTD)
and 20 healthy controls–received T1-weighted
three-dimensional magnetization prepared–rapid gradient echo (MPRAGE)
MRI.
Images were retrospectively assessed by one board-certified
neuroradiologist (BCNR)
and two radiology residents (RR)–
one
of whom had received 6 months of neuroradiology training (RR1). All
cases were evaluated in a two-step reading process–beginning without
AI-
support and followed by an AI-
supported reading (AI
tool: mdbrain
version 3.3.0). For each subject, the suspected diagnostic category
(AD,
FTD
and healthy controls) was determined using a likelihood
score (0–5), adding up to a sum of 5 for all three diagnostic categories.
Individual reader performance with and without AI
support was statistically
evaluated using receiver operating characteristics (ROC).
Results: AI
support substantially improved AD
diagnosis in all three
readers. The effect was most pronounced for RR2 who had not undergone
neuroradiology training (area under the curve [AUC]
without
AI
support [– AI]:
0.629, AI
supported [+ AI]:
0.885). But, even for
the BCNR,
a substantial benefit was measurable (AUCs: BCNR—
AI:
0.827, + AI:
0.882; RR1—AI:
0.713, + AI:
0.834). In diagnosing FTD
RR2 improved with AI
support (AUCs:—AI:
0.610, + AI:
0.754),
while BCNR
and RR1 had comparable reading performances with and
without AI
support (AUCs: BCNR—
AI:
0.843, + AI:
0.828; RR1—AI:
0.865, + AI:
0.868).
Discussion: Even experienced BCNR
can improve their diagnostic accuracy
for AD
by using AI
based rapid brain volumetry and comparison
with the age- and gender-matched reference cohorts. In diagnosing
FTD,
especially radiologists who are less experienced in dementia differential
diagnosis can strongly benefit from AI
support.
Conclusion: AI
support in the radiological work-up of dementia patients
is feasible and can substantially improve diagnostic accuracy,
which might lead to earlier diagnosis and therefore optimized patient
management.