Purchase this article with an account.
Morgan Heisler, Myeong Jin Ju, Donghuan Lu, Arman Athwal, Gavin Docherty, Rosanna Martens, Zaid Mammo, Pavle Prentasic, Sieun Lee, Forson Chan, Mahadev Bhalla, Yifan Jian, Sven Loncaric, Mirza Beg, Eduardo Vitor Navajas, Marinko Sarunic; Deep Neural Network Based Quantification of Retinal Optical Coherence Tomography Images. Invest. Ophthalmol. Vis. Sci. 2018;59(9):1221. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Fully automated and accurate quantification of retinal images is emerging as a necessary tool for the clinical utility of optical coherence tomography (OCT) images. The purpose of this study is to investigate machine learning, in particular Deep Neural Networks (DNN), to perform the classification and segmentation necessary for quantification of microvasculature in OCT Angiography images, fluid in OCT Bscans, and photoreceptors in Adaptive Optics (AO) OCT images.
Each of the automated processing networks required different methods. For microvasculature segmentation, OCTA images of both healthy and diabetic eyes were acquired with a prototype and two commercial systems. The DNN was trained on expert segmentations, and the output was used to calculate vessel density as well as four foveal avascular zone morphometric parameters (area, maximum and minimum diameter, and eccentricity). For the segmentation and classification of retinal fluid, the images used were acquired from commercially available systems. The processing included layer segmentation, fluid detection, and classification as intraretinal fluid (IRF), subretinal fluid (SRF) and pigment epithelial detachment (PED). Lastly, the images used for quantification of photoreceptors were acquired with a recently developed Sensorless AO-OCT with a 200kHz swept source OCT engine.
For the microvasculature segmentation, the average pixel-wise accuracy was 0.823 across all systems. Additionally, no significant difference between the means of the measurements from automated and manual segmentations were found for any of the clinical outcome measures on any system. The proposed fluid segmentation network achieved a Dice index of 0.767 for segmentation and an Area Under the Curve (AUC) measure of 1.00 for detection tasks. For the cone counting algorithm, the accuracy of the DNN segmented output was found to be 0.981.Representative OCT images before and after segmentation Figure 1 (a) – (f). Figure 1 (d) is the probability map output of the DNN. Figure 1 (e) shows the fluid segmentations and classification as IRF (red), SRF (yellow) and PED (blue) along with the anterior and posterior segmentations of the retina. Figure 1 (f) shows the result of the DNN where each cyan mark is noted as a center of a cone.
Machine learning based segmentation enables reliable quantification of retinal OCT scans.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.
Representative OCT segmentations
This PDF is available to Subscribers Only