Spatio-Temporal Model for EUS Video Detection of Pancreatic Anatomy Structures

Adrien Meyer, Antoine Fleurentin, Julieta Montanelli, Jean-Paul Mazellier, Lee Swanstrom, Benoit Gallix, Georgios Exarchakis, Leonardo Sosa Valencia, Nicolas Padoy

Abstract: Current 3D imaging techniques (computed tomography scan, magnetic resonance imaging) offer poor detection of early-stage pancreatic cancers, which in turn leads to high mortality rates. Endoscopic ultrasound (EUS) is a proven alternative to increase early diagnosis and identify potentially curable surgery candidates. However, mastering EUS requires a lot of practice to properly navigate and interpret video flow. Real time computer assisted localization of anatomical structures could improve lesion detection while easing the overall procedure by pointing out anatomical landmarks otherwise complex to identify. For this purpose, we propose a novel architecture built on top of object detection literature by combining spatial attention and temporal information. In parallel, we have created EUS-D50, a representative EUS dataset constituted of 50 EUS pancreas videos with spatial annotations including pancreas parenchyma and lesions. On EUS-D50, our work achieve an mAP@50 of \(58.36\%\) for pancreatic parenchyma and lesion.