Information Technology Reference
In-Depth Information
The comparison of the discrepancy curve obtained by Carmona's method
reveals that our method has a behavior slightly worse in all the discrepancy
range. The explanation of this behavior could be due to the fact that the final
genetic contour of our method is obtained in the Laplacian pyramid level-1, i.e,
we need to expand that solution to the original image level. This fact could
also explain why the gap between both discrepancy curves is more pronounced
for low levels of discrepancy and less pronounced for high values. In fact, for
a discrepancy value of
δ
, both methods match their performance. On the
other hand, the computational cost of our method is lower: 9.000 evaluations of
the fitness function against 20.000 evaluations used by Carmona's method.
Finally, the figure 6 shows different examples of papillary contours obtained
in this work and they are compared with the experts' gold standard.
=5
4 Conclusions and Future Work
We have described a new automatic method to locate and segment the ONH
in eye fundus color photographic images. This method is inspired in the ap-
proach presented in [1]. The performance of our method was compared with two
competitive methods in the literature.
In relation to the results obtained in [1], we obtain a slightly lower perfor-
mance. In contrast, our method has shown to be faster. Also, the way of ob-
taining our set of IPs is simpler: our fitness function only pays attention to the
number of IPs inside the elliptic crown, i.e., it is independent of the domain
knowledge. Finally, we can also say that our method outperforms the results
obtained by [6] in most of the discrepancy domain (
δ>
).
A disadvantage of our method is that we cannot work at Laplacian pyramid
level-0 because the set of IPs obtained in that level contains a lot of noise.
Therefore, we have to expand directly the genetic solution obtained from level-1
to level-0, but this operation seems to produce a lack of accuracy. We propose as
future work to investigate other types of multi-resolution image representation
techniques like wavelet transform or apply images filter directly in the input
image (level-0) in order to obtain sets of IPs with a low level of noise.
1
.
4
Acknowledgment.
This work was supported in part by funds of the Advanced
Artificial Intelligence Master Program of the Universidad Nacional de Educación
a Distancia (UNED), Madrid, Spain.
References
1. Carmona, E.J., Rincón, M., García-Feijoo, J., Martínez-de-la-Casa, J.M.: Identifi-
cation of the Optic Nerve Head with Genetic Algorithms. Artificial Intelligence in
Medicine 43(3), 243-259 (2008)
2. Cox, J., Wood, I.: Computer-assisted optic nerve head assessment. Ophthalmic and
Physiological Optic 11(1), 27-35 (1991)
Search WWH ::
Custom Search