Anatomy-aware Assessment of Critical View of Safety with Spatial Relation Prior

The Critical View of Safety (CVS) is a set of widely accepted conditions that must be met during laparoscopic cholecystectomy (LC) surgeries, and it has been an important validation method for the prevention of bile duct injuries (BDIs). Recent methods explored using object detection and graph construction to facilitate downstream CVS classification, but given the intrinsic spatial relations between anatomical structures in an LC scene, such prior information has yet to be utilized. In this paper, we propose incorporating prior knowledge about spatial relations among anatomical structures to improve the assessment of CVS. We evaluated our method on the publicly available Endoscapes dataset, and achieved 1 ∼ 7% improvement on the baseline model in individual CVS conditions and over 9% in overall CVS classification.