Self-supervised Visual Place Recognition for Colonoscopy Sequences
Keywords: Deep Learning, Place Recognition, CNN, Colonoscopy
TL;DR: We propose the first place recognition system made for colonoscopy sequences, with the use of real colonoscopy image pairs for training obtained in an automated manner.
Abstract: We present the first place recognition system trained specifically for colonoscopy sequences. We use the convolutional neural network for image retrieval proposed by Radenovic et al. and we fine-tune it using image pairs from real human colonoscopies. The colonoscopy frames are clustered automatically by a Structure-from-Motion (SfM) algorithm, which has proven to cope with scene deformation and illumination changes. The experiments show that the system is able to generalize by testing in a different human colonoscopy, retrieving frames observing the same place despite of the different viewpoint and illumination changes. The proposed place recognition would be a key component of Simultaneous Localization and Mapping (SLAM) systems operating in colonoscopy to assist doctors during the explorations or to support robotization.
Paper Type: validation/application paper
Primary Subject Area: Unsupervised Learning and Representation Learning
Secondary Subject Area: Application: Endoscopy
Paper Status: original work, not submitted yet
Source Code Url: This is a preliminary study yet so the code is not suitable for public release. It is our intention to release the code and the trained model in the future.
Data Set Url: The colonoscopy data used for the experiments make secondary use of real human colonoscopies, the ethical approval does not allow to make these recordings publicly available.
Registration: I acknowledge that publication of this at MIDL and in the proceedings requires at least one of the authors to register and present the work during the conference.
Authorship: I confirm that I am the author of this work and that it has not been submitted to another publication before.
4 Replies
Loading