A novel deep learning algorithm for hyper-transmission defects segmentation in OCT scans

Abstract

Purpose : To develop an algorithm to segment geographic atrophy (GA) in OCT scans, based on the hyper-transmission defects (hyperTD).

Methods : A two-step algorithm was developed. Firstly, optical attenuation coefficient (OAC) of 3D swept source (SS-) OCT images was calculated on linear scale. Bruch’s membrane (BM) and inner limiting membrane (ILM) were automatically segmented using SS-OCT data. En face images from the slab of ILM to BM on the OAC 3D data were generated using sum projection as well as maximum projection. Axial positions of the pixels with highest OAC along each A-scan were also calculated and its distance 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 hyperTD from the RGB images generated in step one. Target labels were manually outlined by professional graders.

Results : A total of 54 6x6mm scans from 50 eyes were collected for this study. All eyes were diagnosed with age-related macular degeneration (AMD). 44 scans were used for training with an 80:20 split for validation, and 10 scans were used for testing. After 200 epochs of training, the model achieved a dice coefficient of 0.93 for training and 0.91 for validation. In the testing dataset, the model achieved a dice coefficient score of 0.91. Figure 2 shows several examples of the final prediction output by proposed algorithm.

Conclusions : The proposed algorithm demonstrated satisfactory performance for automatically segmenting hyperTD in AMD eyes with SS-OCT data. It could potentially be useful in clinical settings to track the appearance, size and developments of hyperTD in AMD patients.

This is a 2021 Imaging in the Eye Conference abstract.