Purchase this article with an account.
Isaac Gendelman, Nihaal Mehta, Jay S Duker, Nadia K Waheed; Segmentation of the Choriocapillaris Using Machine Learning on OCTA. Invest. Ophthalmol. Vis. Sci. 2019;60(9):1484. doi: https://doi.org/.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
The choriocapillaris (CC) is a dense vascular layer that is difficult to binarize (segment into flow and non-flow areas) on optical coherence tomography angiography (OCTA). Various filters and threshold methods have been used but do not produce histologically accurate results. This is the first study to use a machine learning algorithm (MLA) to binarize the CC and explore the accuracy and reproducibility of this method as well as potential benefits and pitfalls of such technology.
In the first round, 20 healthy eyes were imaged on OCTA and split into two cohorts: training and testing groups. Two graders independently provided examples of flow and non-flow to the MLA in 10 successive trials using the training group. After each trial the model was tested on the testing group and produced binarized images. Standard metrics including Vessel Length (VL), Number of Flow Voids( NFV), Total Flow Void Area (TFVA), and Avg Flow Void Size (AFVS) were calculated. Intraclass Correlation Coefficients (ICC) were calculated between the graders for each trial and variable. A second round of this experiment was also completed following a guided discussion among the graders to create strict guidelines and improve reproducibility with the same statistical methods.
In the first round there was large variation among ICC with a positive correlation between model number and ICC based on a linear regression model for all metrics with the exception of AFVS. For the first round ICC between the graders for first to last model (first/last) was for NFV .08/.67, TFVA .34/.22 , AFVS .18/.30, VL .07/.70. In the second round there was large variation among ICC with negative correlation between model number and ICC based on a linear regression model for all metrics. ICC between the graders for first to last model was for NFV .61/.18, TFVA .17/.01, AFVS .29/.04 , VL .35/.39.
While the images produced by the MLA are qualitatively convincing when compared to histologic specimens, quantitative assessment is poorly reproducible and shows significant variation both between different graders but also between training rounds and with a change in strategy. Strict guidelines were not helpful in mitigating variation or improving ICC. These data show that although machine learning is a promising new approach to binarizing the CC on OCTA important questions remain to be answered about the reproducibility and accuracy of such methods.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
This PDF is available to Subscribers Only