An image analysis and machine learning approach to measuring the quality of individual colonoscopy procedures
Citation:
Mirko Arnold, 'An image analysis and machine learning approach to measuring the quality of individual colonoscopy procedures', [thesis], Trinity College (Dublin, Ireland). School of Computer Science & Statistics, 2013, pp 167Download Item:
Arnold, Mirko_TCD-SCSS-PHD-2012-12.pdf (PDF) 18.76Mb
Abstract:
The measurement of quality in colonoscopy is an active topic in medical research. Studies report significant miss rates in the detection of colorectal lesions. This has raised the concern among gastroenterologists that the present mechanisms for quality assurance are insufficient. The current clinical practice of quality assurance is based on long term statistics, while the quality of individual colonoscopy procedures is judged by self-assessment. For training and auditing, there exist validated subjective assessment methods, which involve the rating of procedures by trained experts using predefined assessment forms. We focus our research on one such assessment method, the Direct Observation of Procedure and Skill (DOPS), developed by the Joint Advisory Group on Gastrointestinal Endoscopy (JAG) in the UK. One of our main objectives is to investigate to what degree the JAG DOPS assessment can be automated. We have developed a system to automatically measure the quality of colonoscopy procedures according to JAG DOPS criteria and have performed a pilot validation of the system using two trained clinical assessors. The system is based on two different types of data: video data from the endoscopic camera and measurements of the longitudinal and circular motion of the shaft of the endoscope outside the anus. We have developed a number of algorithms that measure different quality related characteristics in endoscopic images and complete colonoscopy procedures. While the development of these measures is oriented towards the overall objective of assessing JAG DOPS criteria, each measure represents clinically relevant image or procedure characteristics on its own. For single images, we propose methods for the measurement of the clarity of the endoscopic field of view, the position and presence of the lumen, the quality of luminal views and the distance to the nearest bend in the colon. The image measures are based on models which are trained using a universal machine learning framework involving automatic feature selection and different variants of support vector machines. The quality of the features is enhanced by a number of pre-processing steps, most notably by a novel algorithm for detection and inpainting of specular highlights. We estimate the depth of insertion of the endoscope using the measurements of the motion sensor, which allows us to divide the colon into a number of spatial segments. This representation is the basis for the development of novel measures of colonoscopy procedure characteristics, reflecting handling patterns and summarising image based measures over the course of complete procedures. The individual procedure measures are then used as features for the training of predictive models for the automatic assessment of JAG DOPS criteria.
Author: Arnold, Mirko
Advisor:
Lacey, GerardQualification name:
Doctor of Philosophy (Ph.D.)Publisher:
Trinity College (Dublin, Ireland). School of Computer Science & StatisticsNote:
TARA (Trinity's Access to Research Archive) has a robust takedown policy. Please contact us if you have any concerns: rssadmin@tcd.ieType of material:
thesisCollections:
Availability:
Full text availableLicences: