OCTA segmentation with limited training data using disentangled representation learning

Optical coherence tomography angiography (OCTA) is an imaging modality, which can analyze retinal vasculature at the micron scale. Automated analysis of the en face OCTA images requires accurate segmentation of the capillaries. Deep learning based methods for segmenting these images have two major issues. First, as a recently developed imaging modality, it has various issues related to noise and contrast that are not completely resolved. Second, potentially large numbers of manual delineations could be required to train a deep network sufficiently. We address both of these issues by disentangling the vascular structure and local contrast. By removing the machine specific contrast variation, it becomes considerably easier to train a deep network to learn how to segment the vessels; moreover, the amount of training data can be dramatically reduced. We demonstrate our approach by outperforming the state-of-the-art in OCTA vessel segmentation.