Anna Danko1,2, Roberto Souza2,3, and Richard Frayne2,3
1Medical Sciences Graduate Program, Cumming School of Medicine, University of Calgary, Calgary, AB, Canada, 2Seaman Family MR Research Centre, Foothills Medical Centre, Calgary, AB, Canada, 3Radiology and Clinical Neuroscience, Hotchkiss Brain Institute, University of Calgary, Calgary, AB, Canada
Synopsis
Magnetic resonance
(MR) imaging is frequently used for carotid artery wall imaging. The capacity
for multi-contrast imaging allows MR scanners to resolve the lumen and wall, as
well as multiple plaque components. Combined this information can provide
evidence of increased stroke risk. Quantitative analysis of carotid artery MR
images regularly begins with the manual segmentation of wall and plaque. This process
is time-consuming and costly, and suggests the need for automated methods.
Developing a robust segmentation tool is challenging because of the domain
shift due to different image contrasts and/or scanners. Here, we demonstrate
that a deep learning network including an adversarial component is capable of
learning domain-invariant features, thus producing a generalizable segmentation
model.
Purpose
Magnetic resonance
(MR) is used for carotid artery imaging because of its capacity for multiple
imaging contrasts, allowing it to resolve several plaque components and thus, evaluate
stroke risk.
1 Quantitative carotid artery assessment typically
begins with manual segmentation, a slow process requiring qualified
professionals. Automatic segmentation using deep learning has been proposed as
a solution to manual segmentation, but much of the software developed thus far
has been limited in application; when applied to new data, many models perform
poorly due to differences in the imaging domain.
2 We propose that an
adversarial segmentation model is capable of learning image contrast- and
domain-invariant features, and we demonstrate improved performance when applied
to an unseen dataset, acquired using a different imaging protocol.
Data
Training data was collected from 26 atherosclerosis
patients imaged locally as part of the AIM-HIGH MR substudy,3 and included PD-, T1-, and T2-weighted images
(15 slices/image), all without contrast agent. One AIM-HIGH subject was
excluded from training to serve as a test subject. Two additional test images, acquired
with a DANTE-Cube sequence,4 were obtained from an ongoing local
study (CARDIS). Contours for the AIM-HIGH data were provided by the Vascular
Imaging Laboratory (Seattle, WA)3 and include the lumen and outer
wall of either the left or right carotid artery (common carotid, bifurcation
point, internal carotid artery).Model
Inspired by generative
adversarial networks (GANs),5 our model (Figure 1) consists of two
adversarial components: segmentation and classification. Images are fed into
the segmentor (U-Net),6 which outputs a carotid wall segmentation. Features
generated at the end of the U-Net encoding path, prior to max-pooling, are sent
to the classifier (ConvNet) that predicts the contrast of the originally input
image based on the extracted features. The model improves when the segmentation
component generates features that ‘fool’ the classifier, i.e., it begins to learn features which are not specific to an image
contrast. We also trained a simple U-Net, to serve as a control model to illuminate
the effect of the adversarial component.
Both models were
trained for 100 epochs, using an 80:20 training:validation data split. Each
image slice (512 × 512) was divided into
five 128 × 128 patches for
training and underwent data augmentation. All three image contrasts were
trained concurrently, and corresponding slices of different image contrasts
were treated as independent images. Categorical cross entropy and the negative
of the Dice score were used to calculate the classification loss, $$$\mathcal{L}_{class}$$$, and segmentation loss, $$$\mathcal{L}_{seg}$$$, respectively. The combined loss
function for the adversarial model was $$$\mathcal{L}_{adv} = \mathcal{L}_{seg} - \lambda\mathcal{L}_{adv}$$$, where $$$\lambda$$$ is a scaling parameter. For this work, we used
$$$\lambda$$$ = 0.2. The negative of the Dice score
was used as the loss function for the non-adversarial model. Entire slices were
used for evaluation, instead of patches. As one carotid artery was manually
segmented for each image, scores were calculated over the left or right half of
the image containing the reference segmentation. Dice, positive predictive value and accuracy scores
were calculated for each slice, and the mean score for each image contrast was
determined.
Results
Both models performed similarly on the AIM-HIGH
test subjects (Figure 2, Table 1). Without ground-truth segmentations, performance on
CARDIS images could only be qualitatively assessed. The adversarial network appeared to produce fewer erroneous
segmentations (false positives) relative to the non-adversarial model on the
CARDIS images (Figure 3).Discussion
Although both models performed similarly on
the AIM-HIGH images, the adversarial network achieved superior results on the
CARDIS test images. This finding suggests that both models were learning
different sets of features due to the absence/addition of the adversarial
component. When assessing the performance scores, it is important to consider
that 100 epochs is a low number of repetitions for training an adversarial
model, which are inherently more difficult to train.6
Conclusion
Despite advances in
medical imaging analysis brought by machine learning, widespread application
has been limited due to the inability of models to generalize to different machines
and imaging protocols. For models to be used outside of research contexts, attention
must be given to developing models that are robust to domain shift. There are
limitations to domain adaptation,
some features are only resolved with specific MR contrasts. Domain adaptation
may still be beneficial in these contexts if other steps in the protocol can be
made more efficient via domain adaptation. We demonstrated that a segmentation
model can be encouraged to learn domain-invariant features through adversarial
training. We intend develop this model further and investigate whether a
segmentation tool developed solely with AIM-HIGH images could circumvent of manual
segmentation on CARDIS images.Acknowledgements
This study was funded by the Canadian Institute for Health Research (CIHR) and a Queen Elizabeth II fellowship to AD.
References
-
Saba, L. et
al. Carotid Artery Wall Imaging: Perspective and
Guidelines from the ASNR Vessel Wall Imaging Study Group and Expert Consensus
Recommendations of the American Society of Neuroradiology. Am
J Neuroradiol. 2018; 39:
E9-E31
-
Kamnitsas, K. et al. Unsupervised Domain Adaptation in
Brain Lesion Segmentation with Adversarial Networks. Information Processing in Medical Imaging 2017; 597-609.
-
Zhao,
X.Q., et al. Clinical Factors associated
with High-Risk Carotid Plaque Features as Assessed by Magnetic Resonance
Imaging in Patients with Established Vascular Disease (from the AIM-HIGH
Study). Am J Cardiol 2014; 114: 1412-9
-
Xie, Y. et al. Improved Black-Blood Imaging using
DANTE-SPACE for Simultaneous Carotid and Intracranial Vessel Wall Evaluation. Magn Reson Med. 2016; 75: 2286-94
-
Goodfellow,
IJ. Generative Adversarial Nets. Advances
in Neural Information Processing Systems. 2014; 2672-2680
-
Ronneberger
et al. U-Net: Convolution Networks
for Biomedical Image Segmentation. Medical
Image Computing and Computer-Assisted Intervention. 2015; 234-241.
-
Goodfellow, IJ., NIPS 2016
Tutorial: Generative Adversarial Networks. CoRR 2017 arXiv:1701.00160