Purchase this article with an account.
Varsha Pramil, Luis De Sisternes, Lars Omlor, Warren Lewis, Sophie Kubach, Harris Asad Sheikh, Mary K Durbin, Nadia K Waheed; A Deep-Learning Based Algorithm for Automated Segmentation of Geographic Atrophy in Swept-Source Optical Coherence Tomography. Invest. Ophthalmol. Vis. Sci. 2021;62(8):117.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
This work presents the first deep-learning based GA segmentation algorithm for swept-source OCT that provides accurate and reproducible results in GA assessment with high sensitivity to GA changes over time.
An automated algorithm was trained with 6x6 mm macular SS-OCT scans from 58 GA eyes (38/20 split for training/validation set). It utilizes scan volume data to generate three image inputs characterizing the main OCT features of GA: hyper-illumination in sub RPE slab, regions of RPE loss, and loss of retinal thickness (Figure 1). Advanced data augmentation techniques helped compensate for the small training data size. To evaluate the accuracy of the GA segmentations, 180 SS-OCT macular scans from 30 GA eyes collected from 3 different sites were considered: each eye had 3 repeated scans at baseline and follow-up visits (average of 16.74 months). Area measurements were corrected by their square-root and used to compute enlargement rate per year between the visits. The GA delineations, area measurements and enlargement rates generated by the automated algorithm in addition to algorithm repeatability in repeated scans were compared to the ground-truth manual delineations annotated by two graders.
Figure 2 summarizes the performance of the algorithm. Automated GA delineation accuracy between the algorithm and two graders was measured using the Dice coefficient with average values of 0.88 and 0.87, respectively. GA area measurements produced by the algorithm were comparable and showed no significant difference to those from each grader, with an average absolute difference of 0.19 (p-value 0.73) and 0.19 (p-value 0.77), respectively. GA enlargement rate computed by the algorithm also showed no significant differences to the two graders, with differences of 0.13 (p-value 0.26) and 0.12 (p-value 0.71), respectively. The algorithm also presented a high repeatability for area and enlargement rate measurements, with intra-class repeatability coefficients of 0.997 and 0.947, respectively.
Despite a small training data size, this deep learning based automated GA segmentation algorithm is able to produce accurate and reproducible results in GA assessment with high sensitivity for GA changes over time.
This is a 2021 ARVO Annual Meeting abstract.
Figure 1. Input, processing and output of the proposed algorithm.
Figure 2. Performance of proposed algorithm.
This PDF is available to Subscribers Only