3D Retinal Vessel Segmentation in OCTA Volumes: Annotated Dataset MORE3D and Hybrid U-Net with Flattening Transformation

Optical Coherence Tomography Angiography (OCTA) extends the 3D structural representation of the retina from conventional OCT with an additional representation of “flow” and is used as non-invasive angiography technique in ophthalmology today. While there are several works for the segmentation of vascular network in OCTA images, most of them are tested on 2D enface images (top view projection) only. Such 2D enface images have the drawback that they depend on a good 3D segmentation of retinal layers, the so-called slabs. Especially in case of retinal diseases (e.g. exudations of the retina) this segmentation is not always clear, even for medical experts. In contrast, we consider the problem of full 3D segmentation of retinal vessels in OCTA images. We present the dataset MORE3D (Münster Octa REtina 3D dataset) that is the first one with 3D annotation. We introduce a general flattening transformation that simplifies and accelerates the 3D data labeling and processing, and also enables a specialized data augmentation. Moreover, we realize a hybrid U-net to achieve a first reference segmentation performance on our dataset. In addition to the common performance metrics we also apply skeleton-based metrics for a more comprehensive structural performance evaluation. With this work we contribute to the advancement of 3D retinal vessel segmentation in OCTA volumes.