BACKGROUND AND PURPOSE:Accurate and reliable detection of white matter hyperinten
sitie
s and their volume quantification can provide valuable clinical information to a
sse
ss neurologic di
sea
se progre
ssion. In thi
s work, a
stacked generalization en
semble of orthogonal 3D convolutional neural network
s, StackGen-Net, i
s explored for improving automated detection of white matter hyperinten
sitie
s in 3D T2-FLAIR image
s.
MATERIALS AND METHODS:Individual convolutional neural network
s in StackGen-Net were trained on 2.5D patche
s from orthogonal reformatting of 3D-FLAIR (
n = 21) to yield white matter hyperinten
sity po
sterior
s. A meta convolutional neural network wa
s trained to learn the functional mapping from orthogonal white matter hyperinten
sity po
sterior
s to the final white matter hyperinten
sity prediction. The impact of training data and architecture choice
s on white matter hyperinten
sity
segmentation performance wa
s sy
stematically evaluated on a te
st cohort (
n = 9). The
segmentation performance of StackGen-Net wa
s compared with
state-of-the-art convolutional neural network technique
s on an independent te
st cohort from the Alzheimer’
s Di
sea
se Neuroimaging Initiative-3 (
n = 20).
RESULTS:StackGen-Net outperformed individual convolutional neural network
s in the en
semble and their combination u
sing averaging or majority voting. In a compari
son with
state-of-the-art white matter hyperinten
sity
segmentation technique
s, StackGen-Net achieved a
significantly higher Dice
score (0.76 [SD, 0.08], F1-le
sion (0.74 [SD, 0.13]), and area under preci
sion-recall curve (0.84 [SD, 0.09]), and the lowe
st ab
solute volume difference (13.3% [SD, 9.1%]). StackGen-Net performance in Dice
score
s (median = 0.74) did not
significantly differ (
P = .22) from interob
server (median = 0.73) variability between 2 experienced neuroradiologi
st
s. We found no
significant difference (
P = .15) in white matter hyperinten
sity le
sion volume
s from StackGen-Net prediction
s and ground truth annotation
s.
CONCLUSIONS:A
stacked generalization of convolutional neural network
s, utilizing multiplanar le
sion information u
sing 2.5D
spatial context, greatly improved the
segmentation performance of StackGen-Net compared with traditional en
semble technique
s and
some
state-of-the-art deep learning model
s for 3D-FLAIR.
ss="headless nomenu">
White matter hyperinten
sitie
s (WMH
s) corre
spond to pathologic feature
s of axonal degeneration, demyelination, and glio
si
s ob
served within cerebral white matter.<
sup>
ss=" bibr popnode">1sup> Clinically, the extent of WMH
s in the brain ha
s been a
ssociated with cognitive impairment, Alzheimer’
s di
sea
se and va
scular dementia, and increa
sed ri
sk of
stroke.<
sup>
ss=" bibr popnode">2,
ss=" bibr popnode">3sup> The detection and quantification of WMH volume
s to monitor le
sion burden evolution and it
s correlation with clinical outcome
s have been of intere
st in clinical re
search.<
sup>
ss=" bibr popnode">4,
ss=" bibr popnode">5sup> Although the extent of WMH
s can be vi
sually
scored,<
sup>
ss=" bibr popnode">6sup> the categoric nature of
such
scoring
sy
stem
s make
s quantitative evaluation of di
sea
se progre
ssion difficult. Manually
segmenting WMH
s i
s tediou
s, prone to inter- and intraob
server variability, and i
s, in mo
st ca
se
s, impractical. Thu
s, there i
s an increa
sed intere
st in developing fa
st, accurate, and reliable computer-aided automated technique
s for WMH
segmentation.Convolutional neural network (CNN)-ba
sed approache
s have been
succe
ssful in
several
semantic
segmentation ta
sk
s in medical imaging.<
sup>
ss=" bibr popnode">7sup> Recent work
s have propo
sed u
sing deep learning–ba
sed method
s for
segmenting WMH
s u
sing 2D-FLAIR image
s.<
sup>
ss=" bibr popnode">8ss=" bibr popnode">ss=" bibr popnode">-
ss=" bibr popnode">11sup> More recently, a WMH
segmentation challenge<
sup>
ss=" bibr popnode">12sup> wa
s al
so organized (
http://wmh.isi.uu.nl/) to facilitate compari
son of automated
segmentation of WMH
s of pre
sumed va
scular origin in 2D multi
slice T2-FLAIR image
s. Architecture
s that u
sed an en
semble of
separately trained CNN
s showed promi
sing re
sult
s in thi
s challenge, with 3 of the top 5 winner
s u
sing en
semble-ba
sed technique
s.<
sup>
ss=" bibr popnode">12sup>Conventional 2D-FLAIR image
s are typically acquired with thick
slice
s (3–4 mm) and po
ssible
slice gap
s. Partial volume effect
s from a thick
slice are likely to affect the detection of
smaller le
sion
s, both in-plane and out-of-plane. 3D-FLAIR image
s, with i
sotropic re
solution, have been
shown to achieve higher re
solution and contra
st-to-noi
se ratio<
sup>
ss=" bibr popnode">13sup> and have
shown promi
sing re
sult
s in MS le
sion detection u
sing 3D CNN
s.<
sup>
ss=" bibr popnode">14sup> Additionally, the i
sotropic re
solution enable
s viewing and evaluation of the image
s in multiple plane
s. Thi
s multiplanar reformatting of 3D-FLAIR without the u
se of interpolating kernel
s i
s only po
ssible due to the i
sotropic nature of the acqui
sition. Network architecture
s that u
se information from the 3 orthogonal view
s have been explored in recent work
s for CNN-ba
sed
segmentation of 3D MR imaging data.<
sup>
ss=" bibr popnode">15sup> The u
se of data from multiple plane
s allow
s more
spatial context during training without the computational burden a
ssociated with full 3D training.<
sup>
ss=" bibr popnode">16sup> The u
se of 3 orthogonal view
s simultaneou
sly mirror
s how human
s approach thi
s segmentation ta
sk.En
semble
s of CNN
s have been
shown to average away the variance
s in the
solution and the choice of model- and configuration-
specific behavior
s of CNN
s.<
sup>
ss=" bibr popnode">17sup> Traditionally, the
solution
s from the
se
separately trained CNN
s are combined by averaging or u
sing a majority con
sen
su
s. In thi
s work, we propo
se the u
se of a
stacked generalization framework (StackGen-Net) for combining multiplanar le
sion information from 3D CNN en
semble
s to improve the detection of WMH le
sion
s in 3D-FLAIR. A
stacked generalization<
sup>
ss=" bibr popnode">18sup> framework learn
s to combine
solution
s from individual CNN
s in the en
semble. We
sy
stematically evaluated the performance of thi
s framework and compared it with traditional en
semble technique
s,
such a
s averaging or majority voting, and
state-of-the-art deep learning technique
s.
相似文献