PTU-016 Wireless Capsule Endoscope Localisation Based on Visual Odometry

    loading  Checking for direct PDF access through Ovid

Abstract

Introduction

The localisation of a wireless capsule endoscope (WCE) within the small-bowel is typically performed by wearable radiofrequency sensors triangulation. The accuracy of this approach is low.1 Only a few approaches have been proposed for WCE localisation based on visual features. These include methods addressing the estimation of the rotation angle of the capsule2,3 and temporal video segmentation methods.4 We present a WCE localisation method, based only on visual information extracted from conventional WCE recording.

Methods

Automatic detection of points of interest (POI) in WCE video frames, matching of the detected POI between consecutive frames, and determination of actual correspondences between subsets of these POI based on the random sample consensus (RANSAC) algorithm was performed. Maximally stable extremal regions (MSER) algorithm, instead of the speeded up feature extraction (SURF) algorithm, was used. Based on the scaling and the rotation of the content of the consecutive WCE frames, it is possible to estimate the displacement and the rotation of the capsule within the GI tract. For the ex-vivo experiment; a standard simulated intestinal environment was created. Markers were sewn (at set, pre-recorded distances) onto the luminal surface of porcine small-bowel through which a capsule (MiroCam®, IntroMedic Co Ltd, Seoul, Korea) was propelled.

Results

Comparative experiments using both SURF and MSER features, which indicated the superiority of the former over the latter, we conducted. We worked on a corpus of 1070 WCE frames (634 indicating forward motion, 436 indicating backward motion). The accuracy using SURF features was 81.5% (87.2% on forward motion, 73.2% on backward motion), while using MSER was 67.2% (79.8% on forward motion, and 48.9% backward motion). Noteworthy, the proposed algorithm often fails when using MSER (6.7% of frames while <0.1% when using SURF) and a transform is not estimated due to the lack of adequate correspondences between interest points.

Conclusion

Visual odometry is a promising technique and – potentially – a feasible alternative to other localisation approaches in WCE.

Disclosure of Interest

None Declared.

Related Topics

    loading  Loading Related Articles