Maurice Pradella1, Sven Knecht2, Manuela Moor1, Shan Yang1, Constantinos Anastasopoulos1, Gian Voellmin2, Philip Haaf2, Stefan Osswald2, Michael Kuehne2, Christian Sticherling2, Bram Stieltjes1, and Jens Bremerich1
1Department for Radiology, University Hospital Basel, Basel, Switzerland, 2Department for Cardiology, University Hospital Basel, Basel, Switzerland
Synopsis
Deep learning based, automatic segmentation of the whole left atrium in
cine MRI makes detailed geometrical analysis possible by fitting of an
ellipsoid into the contours of the left atrium. Therefore, we could identify
the ellipsoidal volume at the time-point before atrial contraction as an independent
predictor of recurrence of atrial fibrillation after catheter ablation in a
multivariable analysis.
Introduction
Atrial
fibrillation (Afib) is a disease leading to severe complications like stroke,
myocardial infarction, etc. and incidence is expected to rise over the next
decade [1]. It causes
anatomical changes of the left atrium (LA) like increase of LA size over time
and outcome results worsen the longer Afib persists. Studies which show changes
of LA geometry are in general based on single time-point analysis [2]. When aiming at
analysing multiple time-points of the full cardiac cycle, time consuming LA
segmentation is required. Hence, we trained a neural network to automatically
segment the left atrium for further analysis and identify predictors of Afib
recurrence after catheter ablation (CA).Methods
In our Afib database, 181 patients (pts) underwent pre-interventional
MRI on a 1.5T scanner using a steady-state free precision cine sequence in
transversal orientation with a slice thickness of 6 mm. A stack of the up to 12
slices was acquired to cover the whole left atrium, temporal resolution was 40
ms. All patients underwent catheter ablation after MRI. If they presented with
recurrence, a second CA was allowed.
We build a training set of 50 randomly selected patients. In those
patients, the left atrium was semi-automatically segmented on all images,
excluding the pulmonary veins and the appendage, using the freely
available software Segment version 2.2 R6435 (http://segment.heiberg.se) [3]. This includes approx. 250 images per patient.
After imaging data was split into three dimensions in time, a 3D
anisotropic U-Net architecture was implemented using an open-source,
convolutional neural network for automatic segmentation of the LA [4].
After completion of training, all 181 Afib patients were processed with
the trained model and spliced back. Those segmentations were each validated by
visual inspection.
In a next step, an ellipsoid was fitted into the LA using a
custom-written Matlab code (Matlab, MathWorks, USA) allowing for geometrical
characterization of the LA (volume, surface area, x, y and z axes). We defined
dimensions at three time-points for each patient (minimal value, maximal value
and value before atrial contraction) [5, 6].
Regression analysis was performed and multivariable-adjusted Cox model
was calculated to examine relations of those parameters, follow up data and
recurrence of Afib after CA. Results
Semi-automatic
segmentation of the LA took approx. 45-60 min/patient (Figure 1) whereas
automatic segmentation with the built AI-algorithm needed approx. 30 sec to
perform segmentation (Figure 2, Figure 3). During
visual inspection of all segmentation, there was no need for editing of the
AI-fitting but two patients were excluded. In total, automatic segmentation was
feasible in 179/181 patients (98.9%). Dice similarity coefficient was 92±2% for
separate test data (Figure 2). We excluded all patients with Afib during MRI
(70 pts) and/or with prior left atrial CA (10 pts). Finally, 101 patients were included
for further analysis (75% paroxysmal Afib).
Mean
follow-up was 408 (±338) days after last CA and 84% of patients were
recurrence-free. The only baseline parameter which was significantly different
between the groups was a shorter Afib duration in recurrence-free patients
(44±62 months vs. 100±68 months, p<0.001).
Between
fitted ellipsoids of recurrence and recurrence-free groups, significant
differences were found in following ellipsoid parameters: maximal volume
(116.6±24.7 ml vs. 98.7±31.6 ml, p=0.034), minimal volume (62.6±18.6 ml vs.
49.9±20.4 ml, p=0.022) and volume before atrial contraction (92.8±22.0 ml vs.
75.5±24.8 ml, p=0.011). In addition, minimal surface area (8444.7 mm2
vs. 7298.3 mm2, p=0.028), surface area before atrial contraction
(10951.1 mm2 vs. 9624.2 mm2, p=0.016), minimal x axis (36.0±4.7
mm vs. 33.2±4.8 mm, p=0.045) and minimal z axis (15.2±2.9 mm vs. 13.5±3.1 mm,
p=0.041) were also significantly different.
In a
multivariable Cox model (adjusted for age, gender, BMI and duration of Afib),
the LA volume before atrial contraction (in ml) of the ellipsoid was identified
as an independent predictor for Afib recurrence (HR 1.036, CI 1.008-1.064,
p=0.011).Discussion
Semi-automatic
segmentation of the LA in cine images is a time consuming procedure. Opposite
to that, our AI-based segmentation tool for the LA in patients with Afib is
very fast and reliable. Further fitting of an ellipsoid in the segmented LA
makes a detailed analysis of LA geometry and function possible. Thereby, we
were able to analyse almost 180 exams of our database and could identify a new
predictor for recurrence after catheter ablation in Afib patients in a
multivariable analysis.
Its use
might be limited to Afib patients since it was trained on an Afib collective
but this is feasible and time-efficiency outweighs this limitation in our view.
Despite there was no necessity for manual editing of AI segmentation in our
processed cases, visual inspection is mandatory in every case to check for
plausibility. Conclusion
Fully
automatic, deep learning derived segmentation of the whole LA based on cine MRI
of Afib patients is feasible and enables geometrical analysis by fitting of an
ellipsoid. Thus, LA ellipsoidal volume before atrial contraction could be identified
as an independent predictor for recurrence after CA. This finding could be used
for patient selection prior to catheter ablation to support clinical decision
making and improve patient care.Acknowledgements
No acknowledgement
found.References
1. Rahman, F., G.F. Kwan, and E.J.
Benjamin, Global epidemiology of atrial
fibrillation. Nat Rev Cardiol, 2014. 11(11):
p. 639-54.
2. Kato, R., et al.,
Pulmonary vein anatomy in patients
undergoing catheter ablation of atrial fibrillation: lessons learned by use of
magnetic resonance imaging. Circulation, 2003. 107(15): p. 2004-10.
3. Heiberg, E., et
al., Design and validation of
Segment--freely available software for cardiovascular image analysis. BMC
Med Imaging, 2010. 10: p. 1.
4. Çiçek, Ö., et al.
3D U-Net: Learning Dense Volumetric
Segmentation from Sparse Annotation. 2016. Cham: Springer International
Publishing.
5. Kowallick, J.T.,
et al., Quantification of left atrial
strain and strain rate using Cardiovascular Magnetic Resonance myocardial
feature tracking: a feasibility study. J Cardiovasc Magn Reson, 2014. 16: p. 60.
6. Leng, S., et al.,
Validation of a rapid semi-automated
method to assess left atrial longitudinal phasic strains on cine cardiovascular
magnetic resonance imaging. J Cardiovasc Magn Reson, 2018. 20(1): p. 71.