AuthorsD. Jha, M. Riegler and P. Halvorsen
TitleA RASNet-based deep learning approach for the Binary Segmentation Task in the 2019 ROBUST-MIS Challenge
AfilliationMachine Learning
Project(s)Department of Holistic Systems
StatusAccepted
Publication TypeProceedings, refereed
Year of Publication2020
Conference NameRobust Medical Instrument Challenge at MICCAI 2019 (Part of Endoscopic Vision Challenge)
Keywordsgeneralization, Instrument segmentation, minimally invasive surgery, Robustness
Abstract

Semantic segmentation of laparoscopic instruments is one of the necessary pre-condition for computer and robotic assisted interventions. With the increase of publicly available datasets and improved hardware, the collaboration of multimedia and medical communities involved to tackle the surgical vision task is gaining momentum than ever before. In this regard, we present a solution for the 2019 Robust Medical Instrument Segmentation (ROBUST-MIS) challenge, which is a part of Endoscopic vision challenge. We use the RASNet model which is an encoder-decoder network. In the encoder part, RASNet uses the pre-trained ResNet50 model which is trained on ImageNet. The decoder part of the network consists of an attention fusion module along with a decoder block. Based on the results generated with the available dataset, we show that the proposed method produces good segmentation results on the instrument segmentation task. (*The paper is accepted for Joint Publication)

Citation Key27166