Abstract
Purpose :
Robust and sensitive imaging biomarkers remain an unmet need in the management of macular disease. Classical OCT biomarkers e.g. fluid are defined based on human intuition and experience. However, these conventional markers may not represent a comprehensive vocabulary to capture all important information in OCT images. Furthermore, they are inherently limited by pre-made hypotheses.
Methods :
To complement the vocabulary of OCT biomarkers, we propose large-scale unsupervised deep learning to identify features inherent in OCT data in a completely unbiased fashion. We use auto-encoder methodology to automatically identify and categorize relevant features without human prior knowledge. This approach is first applied on the (local) A-scan level to capture fine-grained features in the image. Consecutively, we apply the auto-encoder method on (global) volume and patient timeline levels to encode simplified descriptions of entire OCT volumes and the behavior of patients over time. Our method is trained on clinical trial data of 1,097 patients with neovascular age-related macular degeneration under standardized anti-VEGF therapy. We evaluate the resulting biomarkers by correlating them with measures of visual function (BCVA and LLVA).
Results :
The deep learning system identified 20 distinct (local) A-scan features that correlated with BCVA (R2=0.17) and LLVA (R2=0.29, individual correlation coefficient up to r=-0.39). Some markers corresponded to known findings such as retinal fluid or subretinal hyperreflective material, while other markers did not reveal an obvious link to known morphologic features. Two non-fluid related markers even had a positive correlation with BCVA and LLVA. On the global level, unsupervised learning resulted in a compact 20-dimensional description of the OCT volume. Correlation with visual function was superior to the A-scan level (R2=0.26 (BCVA) and 0.37 (LLVA)). On the timeline level, the method revealed clear differences between patients receiving different treatment regimens (monthly vs. PRN).
Conclusions :
Unsupervised deep learning enabled an unbiased representation of clinically important markers in OCT imaging that correlated well with visual acuity. Furthermore, it successfully achieved a compact summary of volumetric OCT data on a volume level and also over time. The presented methodology offers advances in analyzing big data in retinal disease.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.