Deep learning methods for real-time corneal and needle segmentation in volumetric OCT scans

Deep lamellar anterior keratoplasty (DALK) is a promising cornea transplant procedure, which mitigates the risks associated with the commonly used alternative penetrating keratoplasty. In DALK, a surgeon must insert a needle into the cornea to a precise depth and inject an air bubble to separate the stroma and endothelial layer. Optical coherence tomography (OCT) is used to help surgeons judge the needle insertion depth. However, the needle obscures the part of the cornea underneath it, making it difficult to estimate needle insertion depth. In this thesis, deep learning methods are explored for cornea and needle segmentation to automatically compute needle insertion depth from volumetric OCT scans. A simple post-processing step is applied to the mask produced by the deep neural networks to determine the coordinates of the top and bottom corneal boundaries and the needle tip. We compare the performance of 2D and 3D networks on this task. Our experiments consistently showed that all networks perform well on cornea segmentation in scans without a needle, and 3D networks perform better in scans with a needle. We also show that these deep learning methods perform better than existing graph-theory methods on this task and can segment in real-time when deployed using the NVIDIA TensorRT framework.U of I OnlyAuthor requested U of Illinois access only (OA after 2yrs) in Vireo ETD syste