Anne Nielsen1,2, Mikkel Bo Hansen1, and Kim Mouridsen1
1Center of Functionally Integrative Neuroscience and MINDLab, Department of Clinical Medicine, Aarhus University, Aarhus, Denmark, 2Cercare Medical, Aarhus, Denmark
Synopsis
Every
year, 13 million people suffer acute ischemic stroke. Brain tissue infarcts permanently within
hours after stroke onset and rapid recanalization is therefore of utmost
importance. In this project, we aim to estimate recanalization effect by a single
convolutional neural network customized to include magnetic resonance imaging
biomarkers as well as individual recanalization information. This is in
contrast to the traditional approach which is splitting the data set according
to the recanalization information and training several models. We find a
significant recanalization effect and believe this to be an important step
towards an automated decision support system.
Introduction:
Every
year, 13 million people suffer acute ischemic stroke. Brain tissue infarcts rapidly (two million
neurons die every minute a major artery is occluded1) and rapid recanalization is therefore of
utmost importance to ensure optimal outcome. Current magnetic resonance imaging allows
assessment of irreversibly damaged tissue and surrounding tissue suffering
reduced blood delivery. Salvageable tissue (or the ischemic penumbra) is
defined as tissue with reduced blood supply, which has not already infarcted,
and may be saved by quick recanalization. In this project, we aim to estimate
the effect of such recanalization. The traditional approach would be to split the data
according to individual recanalization information and train several predictive
models. However, we do not think this is an optimal approach for data demanding
methods such as convolutional neural networks (CNN). Therefore, we aim to
estimate recanalization effect using a single model customized to include
imaging biomarkers and individual recanalization information and capable of
automatically estimating the recanalization effect.Methods:
133 acute ischemic stroke patients from the I-KNOW
multicenter2 (53) and perconditioning3 (80) studies were randomly divided into training (112) and
test (21) sets. 96 patients recanalized (defined as a TIMI score of 2 or 3) and
37 did not (TIMI score of 0 or 1). The acute MRI protocol included
diffusion-weighted imaging (DWI), T2 fluid-attenuated inversion recovery
(T2-FLAIR) imaging, and gradient-echo (GE) dynamic susceptibility contrast
perfusion-weighted imaging (DSC-PWI) with echo time 30-50 ms and repetition
time 1500 ms. Gadolinium contrast dose 0.1 mmol/kg at injection rate 5mL/second
was followed by a saline chaser. Figure 1 shows patient characteristics and further
details are listed in Hansen et al4.
We modified a deep CNN5 consisting of 37 layers
(presented at ISMRM 20176) to take both biomarkers
and recanalization information as input. The predictions were based on
T2-FLAIR, trace DWI, apparent diffusion coefficient (ADC) and the following
perfusion biomarkers: mean transit time (MTT), cerebral blood volume (CBV),
cerebral blood flow (CBF), cerebral metabolism of oxygen7 (CMRO2),
relative transit time heterogeneity (RTH), delay and recanalization information.
The
resulting model’s ability to estimate recanalization effect was evaluated on
the independent test set using the predicted final infarct volume compared to
the follow-up lesion volume. A Wilcoxon signed-rank
test was used to test for a significant difference between predicted final
infarct volume ±recanalization. Volumes are indicated in ml as (mean±std.dev).Results:
Figure 2 shows examples from the predictive model. For patient A and
B, the mismatch between the diffusion
and perfusion weighted imaging lesions
are small (12 and 3 ml). For patient C and D,
the mismatch is larger (108 and 130 ml). The patients
with a large mismatch have a substantial recanalization effect, whereas
patients with a small mismatch show almost no recanalization effect. This is
consistent with the well-esteemed penumbra model.
The
highest predicted final infarct volume was obtained for -recanalization
(41.9±45.8) with a significant difference (p<0.0001) to +recanalization
(24.2±33.4). The difference is also apparent in Figure 3 showing predicted final infarct volume ±recanalization. Patients with a small follow-up lesion volume
(a small dot in Figure 3) also have a small predicted final infarct volume for both
methods. The difference between ±recanalization is more apparent for larger
follow-up infarct volumes. Discussion:
We showed a
significant difference in final infarct volume between the model’s prediction
with and without recanalization, with no recanalization leading to larger final
infarct volumes. This corresponds to the penumbra model. The use of a single model capable of including
recanalization information is a better use of data compared to training several
models. Using this method, it is possible to include other types of information
influencing the stroke outcome and thereby make even more individualized and
improved predictions. We decided
to assess the ability of the CNN to identify recanalization differences based.
The same methodology could equally well be used to examine the direct effects
of treatments. Conclusion:
The
recanalization effect was significant, with –recanalization yielding a higher
volume of final infarct. The new model paradigm has shown recanalization effect
estimation and thereby a potential of clinical information in automated
decision support systems providing recommendations for personalized treatment
and thereby hopefully better outcome for the individual patient.Acknowledgements
No acknowledgement found.References
1. Saver JL. Time is brain - quantified. Stroke. 2006;37:263-266
2. I-KNOW.
Integrating information from molecules to man: Knowledge discovery accelerates
drug development and personalized treatment in acute stroke. 2006
3. Hougaard KD, Hjort N, Zeidler D, Sorensen L, Norgaard A,
Thomsen RB, et al. Remote ischemic
perconditioning in thrombolysed stroke patients: Randomized study of activating
endogenous neuroprotection - design and mri measurements. Int J Stroke. 2013;8:141-146
4. Hansen
MB, Nagenthiraja K, Ribe LR, Dupont KH, Ostergaard L, Mouridsen K. Automated
estimation of salvageable tissue: Comparison with expert readers. J Magn Reson Imaging. 2016;43:220-228
5. Badrinarayanan
V, Kendall A, Cipolla R. Segnet: A deep convolutional encoder-decoder
architecture for image segmentation. ArXiv e-prints. 2015;1511
6. Nielsen A, Hansen MB, Mouridsen K, Boldsen JK. Deep learning: Utilizing the potential in data bases
to predict individual outcome in acute stroke. ISMRM. 2017
7. Jespersen
SN, Østergaard L. The roles of cerebral blood flow, capillary transit time
heterogeneity, and oxygen tension in brain oxygenation and metabolism. J Cereb Blood Flow Metab.
2012;32:264-277