Comparative Deep Learning Architectures to Detect Tiny Features in Ophthalmic Imaging

In ophthalmology, the analysis of optical coherence tomography (OCT) images has brought important insights for early detection of eye diseases. This task requires a high amount of experience and training. Moreover, the detection is challenging due to the small size of the pathologies. An early detection is especially relevant for diseases, which cause permanent damage and if left untreated lead to blindness, such as wet age-related macular degeneration (wetAMD). In this work, six deep learning architectures trained to segment small and tiny pathological structures of wetAMD in OCT images were compared analytically and visually. We used a dataset of 2016 annotated OCT images from Augenspital University of Basel. The U-Net and our proposed variants N-Net and U-Net-M-Dec performed best for pixel-wise segmentation of these pathologies. Cropping input images into regions of interest and tiles improved the model training notably. Moreover, augmenting the data by brightness and rotation variations regularized the model training best. The proposed U-Net-M-Dec represents a middle ground between the evaluated binary and multiclass model approaches. The executed inter-observer variability of human annotators reached a Dice score of 0.74. The best multiclass segmentation U-Net reached a Dice score of 0.748 and U-Net-M-Dec achieved Dice scores per pathologies [IRF, SRF, HF, SHRM] of [0.845,0.808,0.488,0.862]. The segmentation models are intended to be used for ophthalmic training and an assistive tool in ophthalmic practices.