Impact of Data Augmentation on Retinal OCT Image Segmentation for Diabetic Macular Edema Analysis.

Deep learning models have become increasingly popular for analysis of optical coherence tomography (OCT), an ophthalmological imaging modality considered standard practice in the management of diabetic macular edema (DME). Despite the need for large image training datasets, only limited number of annotated OCT images are publicly available. Data augmentation is an essential element of the training process which provides an effective approach to expand and diversify existing datasets. Such methods are even more valuable for segmentation tasks since manually annotated medical images are time-consuming and costly. Surprisingly, current research interests are primarily focused on architectural innovation, often leaving aside details of the training methodology. Here, we investigated the impact of data augmentation on OCT image segmentation and assessed its value in detection of two prevalent features of DME: intraretinal fluid cysts and lipids. We explored the relative effectiveness of various types of transformations carefully designed to preserve the realism of the OCT image. We also evaluated the effect of data augmentation on the performance of similar architectures differing by depth. Our results highlight the effectiveness of data augmentation and underscore the merit of elastic deformation, for OCT image segmentation, reducing the dice score error by up to 23.66%. These results also show that data augmentation strategies are competitive to architecture modifications without any added complexity.