![](bg6.png)
3.2 Experiment
For each of the two architectures, we compare the
three methods described above: baseline, baseline
with LSymm, and baseline with NLSymm, using 7-
fold cross-validation on the Training dataset of 28
subjects. All methods are run with the same hyper-
parameters, on the same pseudo-random sequence of
training patches.
3.3 Results
Example segmentations produced by the three meth-
ods on TwoPathCNN are shown in Fig. 3. The main
results are summarised in Table 1. Adding linearly or
nonlinearly registered symmetry features (LSymm or
NLSymm) to the baseline architectures consistently
improves mean Dice coefficient, Recall and Precision,
showing the effectiveness of reflective registration.
For the Dice coefficient, we performed one-sided
paired t-tests for symmetry-augmented vs. baseline
methods, and found that the resulting p-values were
always less than 10
−5
, for both LSymm and NLSymm
and for both baseline architectures. Moreover, nonlin-
early registered symmetry features (NLSymm) con-
sistently produced higher Dice, Recall and Precision
scores compared to linearly registered symmetry fea-
tures (p < 0.001 for the TwoPathCNN architecture,
and p = 0.08 for the Wider2dSeg architecture). Of
the two architectures evaluated, Wider2dSeg benefit-
ted more from the symmetry augmentation, however
the difference between LSymm and NLSymm was
not significant; both differences were perhaps due to
its deeper architecture.
4 CONCLUSIONS
We have proposed an improvement to existing seg-
mentation methods by exploiting the bilateral quasi-
symmetry of healthy brains. Our method, which does
not require a template, consists of augmenting the in-
put images to a CNN with extra Symmetry Difference
Images, which are intensity differences between ho-
mologous (“mirror”) voxels in different hemispheres.
We showed how to incorporate these symmetric fea-
tures into the increasingly popular patch-based CNNs
so as to preserve dense inference. In an experi-
ment on the ISLES2015 SISS dataset, we found that
adding symmetric features generated using nonlinear
reflective registration (the “NLSymm” method) con-
sistently resulted in a mean improvement in Dice co-
efficient, Recall and Precision. Using linear reflec-
tive registration instead gave consistently smaller im-
provements over baseline, showing that nonlinear reg-
istration is superior in this application. For the Dice
coefficient, improvement over baseline was signifi-
cant (p < 10
−5
) for both linear and nonlinear sym-
metric features. The nonlinear method was signifi-
cantly better than the linear one (p < 0.001) for one
baseline architecture (TwoPathCNN) but not the other
(Wider2dSeg).
While our numerical results are not directly com-
parable with those of the three preceding studies of
symmetric feature augmentation for CNNs mentioned
in the Introduction, we note that our improvements in
Dice scores of 9 to 13% on an open dataset compare
favourably to earlier results.
We have shown that the brain’s quasi-symmetry
property is a valuable tool for brain lesion segmenta-
tion. The ease of application of symmetry augmen-
tation to most existing CNN methods and many other
methods suggests a potentially wide-ranging utility of
the method.
REFERENCES
Abadi, M., Barham, P., Chen, J., Chen, Z., Davis, A., Dean,
J., Devin, M., Ghemawat, S., Irving, G., Isard, M.,
et al. (2016). Tensorflow: a system for large-scale
machine learning. In OSDI, volume 16, pages 265–
283.
Avants, B. B., Tustison, N. J., Song, G., and Gee, J. C.
(2009). ANTS: Open-source tools for normalization
and neuroanatomy. Transac Med Imagins Penn Image
Comput Sci Lab.
Dvorak, P., Bartusek, K., and Kropatsch, W. (2013). Au-
tomated segmentation of brain tumour edema in flair
mri using symmetry and thresholding. PIERS Pro-
ceedings, Stockholm, Sweden.
Havaei, M., Davy, A., Warde-Farley, D., Biard, A.,
Courville, A., Bengio, Y., Pal, C., Jodoin, P.-M., and
Larochelle, H. (2017). Brain tumor segmentation
with deep neural networks. Medical image analysis,
35:18–31.
He, K., Zhang, X., Ren, S., and Sun, J. (2015). Delv-
ing deep into rectifiers: Surpassing human-level per-
formance on imagenet classification. In Proceedings
of the IEEE international conference on computer vi-
sion, pages 1026–1034.
Kamnitsas, K., Ledig, C., Newcombe, V. F., Simpson,
J. P., Kane, A. D., Menon, D. K., Rueckert, D., and
Glocker, B. (2017). Efficient multi-scale 3D CNN
with fully connected CRF for accurate brain lesion
segmentation. Medical image analysis, 36:61–78.
Loy, G. and Eklundh, J.-O. (2006). Detecting symmetry
and symmetric constellations of features. In Euro-
pean Conference on Computer Vision, pages 508–521.
Springer.
Maier, O., Menze, B. H., von der Gablentz, J., H
¨
ani, L.,
Heinrich, M. P., Liebrand, M., Winzeck, S., Basit, A.,
Exploiting Bilateral Symmetry in Brain Lesion Segmentation with Reflective Registration
121