Abstract
Purpose :
To develop an automatic method for the segmentation of nine layer boundaries on optical coherence tomography (OCT) images of dry age-related macular degeneration (AMD).
Methods :
We utilized convolutional neural networks (CNN) to extract features of specific retinal layer boundaries and trained a corresponding classifier to delineate eight layers on Spectralis OCT images. Specifically, in the training step, we extracted patches (of size 33×33 pixels) centered on manually segmented layer boundaries from 171 OCT B-scans from 57 volumes. Then, we trained a CNN on the extracted patches. In the testing step, we used 60 OCT volumes (each volume includes 48 to 49 B-scans) from 20 AMD eyes (imaged at three time points over a 12-month period at 6-month intervals), which were separate from the subjects used for training the CNN. We decomposed the test OCT B-scans into over-lapping patches (of size 33×33) and applied the trained CNN classifier to obtain layer boundary labels and probabilities in each patch. Finally, we created a graph using the learned boundary probabilities and applied a Dijsktra based graph-search algorithm to estimate the location of the layer boundaries. For validation, the average difference in layer thickness between the automatic segmentation and a manual segmentation by an expert grader was calculated for each layer of each B-scan. The absolute mean difference and standard deviation across all B-scans was then computed.
Results :
We tested our CNN/graph layer method on 60 OCT volumes. The average (standard deviation) retinal layer thickness differences between our method and one expert manual grader are: NFL 0.57(0.58); GCL+IPL 0.46(0.67); INL 0.23(0.56); OPL 0.68(1.17); ONL 0.80(1.35); IS 0.16(0.29); OS 0.98(0.96); RPE 1.15(1.61); Total Retinal 1.26(1.25). Fig. 1 is an illustrative example of segmented layers using the proposed technique as compared to expert manual grading.
Conclusions :
Our CNN/graph based layer method can automatically segment nine retinal layer boundaries in human AMD subjects with high accuracy. Since no disease specific assumptions are used in the design of the CNN algorithm, this technique can be adapted for other types of diseases.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.