Purchase this article with an account.
Zhongdi Chu, yingying shi, Xiao Zhou, Yuxuan Cheng, Rita Laiginhas, Giovanni Gregori, Philip J Rosenfeld, Ruikang K Wang; A novel deep learning algorithm for the segmentation of pigment deposits using SS-OCT. Invest. Ophthalmol. Vis. Sci. 2021;62(8):2544.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To develop an algorithm to segment pigment deposits and to identify their axial positions relative to the retinal pigment epithelium (RPE) using swept source OCT (SS-OCT) data based on the appearance of hypo-transmission defects (hypoTDs).
A three-step algorithm was developed. Firstly, optical attenuation coefficient (OAC) was evaluated on a linear scale 3D SS-OCT. Bruch’s membrane (BM), RPE, and inner limiting membrane (ILM) were segmented on the 3D SS-OCT cube. En face images from the slab of ILM to BM from the 3D OAC data were generated using sum projection as well as maximum projection. Axial positions of the pixels with the highest OAC along each A-scan were also calculated and their distances to BM were recorded as an elevation map. A composite RGB image (Figure 1) was generated using the OAC sum, max en face images, and the elevation map. Secondly, a deep learning model with the Unet configuration was trained to segment retinal pigment deposits from the RGB images generated in step one. Target labels were manually outlined by professional graders. Thirdly, the axial positions of pigment deposits segmented by the model were identified using the elevation map and compared to the RPE segmentation. Pigment deposits with a distance larger than 30 µm from the RPE were classified as intraretinal pigment deposits and otherwise RPE pigment deposits.
A total of 152 scans from 52 eyes were collected for this study. 41 eyes were diagnosed with age-related macular degeneration (AMD) and 11 eyes showed no ocular pathology upon recruitment with long term plaquenil use in 1 eye, but no obvious retinopathy present. 132 scans were used for training (with an 80/20 split between training and validation) and 20 scans were used for testing. In the testing dataset, the model achieved an intersection over union (IoU) score of 0.75. Figure 2 shows an example of the final output of intraretinal pigment deposits and pigment deposits associated with the RPE as segmented by the proposed algorithm.
The proposed algorithm demonstrated a satisfactory performance for automatically segmenting pigment deposits and identifying their axial positions using SS-OCT data. This strategy could be potentially useful for tracking the appearance, size, and position of retinal pigment deposits in AMD patients as indicators of disease progression and for assessing the effect of therapies in clinical trials.
This is a 2021 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only