Real-Time Vascular Graph Extraction for Surgical Navigation

Surgical soft-tissue navigation requires the visualfeatures used to be unaffected by deformation. To this end,the observed vascular structures can be extracted as deformation-invariant graph features. A key subprocedure of theoverall navigation task is therefore the instantaneous extractionof graph structures from the observed images. Existingwork shows that the graph extraction procedure is complex,computationally expensive, and thus not real-time capable.Additionally, there is a dimensionality problem in storingthe graph data, as the number of graph nodes vary for eachimage observed. In this paper, a neural network (NN)-basedgraph extraction framework which addresses both the latencyand dimensionality problems is proposed. This paperprovides implementation details for the main components ofthe framework, which are the node extraction and edge extractionsubtasks respectively. Additionally, this work introducesvariations of an adjacency combination (AC) schemethat constructs a correctly dimensioned graph output. Weshow that the proposed framework is capable of performingspeedy graph extraction from a preprocessed ‘skeletonised’image with good precision, opening the way to a real-timegraph extraction procedure for use in surgical soft-tissue navigation.