Page 138 - Computational Retinal Image Analysis

References 131

Kugelman et al. [39] use a recurrent neural network to classify layer boundaries,
trained on image patches. They segment three layer (NFL, RPE, BM) in Healthy and
AMD cases. A CRF postprocessing is applied to create smooth layer boundaries.

6 Discussion and conclusion
The abundance of publications in OCT layer segmentation of the retina in the last
years shows a big interest in furthering this field. With the advent of deep learning
big improvements have been achieved in terms of performance, especially in patho-
logical cases (AMD, DME, etc.), replacing the state of the art of graph-based meth-
ods. The latter cannot deal with highly degenerate retinal layers, where boundaries
do not follow mostly horizontal paths or are even completely absent due to atrophy
and fluids. Instead, data-driven methods cover a broader spectrum of pathologies and
do not need a model-based approach.
While there is a large progress in the field, there are still very few of standard pub-
lic datasets to be able to compare the results of published methods. This makes it hard
to rank algorithms against each other, as they usually test against their own private
datasets. In addition, there is disagreement of the assessment on visible anatomy and
boundaries thereof in OCT in clinical application. This effect is amplified especially
in pathological cases and the lack of reliable metrics, as no single value such as Dice
similarity or chamfer distance present a complete picture. A reliable comparison of
the performance of published methods becomes thus hard or even impossible. It is
therefore of importance to thrive toward extensive, multi-observer and public retinal
layer segmentation datasets in OCT.

References
[1] S.J. Chiu, J.A. Izatt, R.V. O’Connell, K.P. Winter, C.A. Toth, S. Farsiu, Validated au-
tomatic segmentation of AMD pathology including drusen and geographic atrophy in
SDOCT images, Invest. Ophthalmol. Vis. Sci. 53 (1) (2012) 53–61.
[2] S.J. Chiu, M.J. Allingham, P.S. Mettu, S.W. Cousins, J.A. Izatt, S. Farsiu, Kernel re-
gression based segmentation of optical coherence tomography images with diabetic
macular edema, Biomed. Opt. Express 6 (2015) 1172–1194.
[3] B. Dubose, D. Cunefare, E. Cole, P. Milanfar, J.A. Izatt, S. Farsiu, Statistical models
of signal and noise and fundamental limits of segmentation accuracy in retinal optical
coherence tomography, IEEE Trans. Med. Imaging 37 (9) (2018) 1978–1988.
[4] J. Tian, B. Varga, G.M. Somfai, W. Lee, W.E. Smiddy, D.C. DeBuc, Real-time auto-
matic segmentation of optical coherence tomography volume data of the macular, PLoS
ONE 10 (8) (2015) e0133908.
[5] J. Tian, B. Varga, E. Tatrai, P. Fanni, G.M. Somfai, W.E. Smiddy, D.C.C. DeBuc,
Performance evaluation of automated segmentation software on optical coherence to-
mography volume data, J. Biophotonics 9 (5) (2016) 478–489.
[6] A. Shah, L. Zhou, M.D. Abramoff, X. Wu, Multiple surface segmentation using con-
volution neural nets: application to retinal layer segmentation in OCT images, Biomed.
Opt. Express 9 (9) (2018) 4509–4526.