End-to-End Image-to-Tree for Vasculature Modeling

Manish Sharma, Matthew C. H. Lee, James Batten, Michiel Schaap, Ben Glocker

Keywords: Vasculature, tree extraction, retinal vessels, diabetic retinopathy, visualization

Abstract: Imaging can be used to capture detailed information about complex anatomical structures such as vessel trees. This can help to detect disease such as stenosis (blockages) which is important for diagnosis and clinical decision making. Current approaches for extracting vasculature from images involve generating binary segmentation maps followed by further processing. However, these binary maps may be sub-optimal, implicit representations of the underlying geometry while trees seem a more natural way of describing vasculature. In this work, we propose a novel image-to-tree approach, which is an end-to-end system for extracting explicit tree representations of vasculature from biomedical scans. We designed a moving patch algorithm that utilizes a U-Net component for predicting individual tree nodes. The methodology is presented for both synthetically generated tree images and publicly available Digital Retinal Vessel Extraction dataset (DRIVE). Using vascular tree construction, we discuss applications to thickness estimation in diabetic retinopathy prediction, and explore insights from visualizing these trees.

Code Of Conduct: I have read and accept the code of conduct.