Abstract
Purpose :
Fluid in the retina plays a critical role in managing chronic retinal conditions. OCT volumetric scans are key in assessing if such fluid is present or not. We hypothesize that this task is suitable for automation using machine learning techniques. Moreover, we speculate that training a machine learning model for this task requires only manual annotations regarding the presence of fluid, as opposed to pixel-wise annotations of fluid regions. This would render easier and cheaper annotations for machine learning tasks.
Methods :
We use two datasets of OCT C-Scans. The first dataset (D1) consists of 1201 OCT volumes from 130~subjects who suffer from different stages of AMD. 519 (43%) volumes were manually labeled as dry and 682 (57%) volumes were labeled as wet. The second dataset (D2) consists of 983 C-Scans from a mixed subject pool of 723 eyes (AMD, DR, DME, CVO). 268 (27%) volumes were labeled as dry, and 715 (73%) were labeled as wet. Volumes were acquired using the Heidelberg Engineering Spectralis OCT system. All volumes are processed by our machine learning architecture in the same manner. Given a volume, its B-Scans are resized and separately fed to a Dilated Residual Network (DRN-D-54). B-Scan descriptions are merged together in a single volume signature, which is further processed to obtain the probability of finding fluid in the volume. The whole network is trained end-to-end using the dataset D1. After training, network is tested using the dataset D2.
Results :
The area under the ROC curve is 0.939 and the area under the precision-recall curve is 0.977. For a specificity of 0.9, we attain a sensitivity of 0.88. For a sensitivity of 0.95, we obtain a specificity of 0.57. Our deep network processes an entire volume in 1.23 seconds with test-time augmentation. Without test-time augmentation, computation time per volume is reduced to 0.31 seconds with a minor reduction in performance, with the sensitivity dropping to 0.86 for a specificity of 0.9.
Conclusions :
Our experimental results successfully support our hypothesis for the automation of retinal fluid detection. Our deep network is able to learn the visual appearance of fluid just using human annotations given at the C-Scan level, which are cheap to obtain. Our method has potential to save time and work to clinicians, while being reliable and consistent.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.