Exploiting gastrointestinal anatomy for organ classification in capsule endoscopy using locality preserving projections

Abstract

Capsule Endoscopy is a technique designed to

wirelessly image the small intestine within the gastrointestinal

(GI) tract. Its main drawback is the vast amount of images it

generates per patient, necessitating long screening sessions by

the clinician. Previous studies have proposed to partially

facilitate this process by automatically segmenting the GI tract

into its constituent organs, thus identifying the region of

interest. In this work, we propose to exploit the anatomical

structure of the GI tract when carrying out dimensionality

reduction on visual feature vectors that describe the capsule

images. To this end, we suggest a novel adaptation of a

technique called Locality Preserving Projections, and results

show that this achieves an improved performance in organ

classification and segmentation, at no additional computational

or memory cost.