Abstract
Coeliac disease (CD), also known as gluten-sensitive enteropathy, is a chronic
disorder with characteristic small intestinal mucosal changes associated with nutrient
malabsorption. To date, CD is still diagnosed using histology. The treatment of CD is to
follow a gluten-free diet (GFD). The detection of mucosal healing in treated CD patients
is important because it can determine the possibility of developing long-term
manifestations of CD, such as intestinal lymphoma. According to current guidelines,
duodenal biopsy is the most accurate method available for detection of mucosal healing.
However, there are difficulties in early detection of healing because it may take up to
three years for the biopsy to show healing.
Confocal laser endomicroscopy (CLE) is a novel endoscopic method that permits
on-site microscopy of the gastrointestinal mucosa. It improves the accuracy of targeted
biopsies and enhances the diagnostic outcome in CD, allowing high-resolution in vivo
histological analysis. Although computer-aided diagnosis (CAD) has been studied in
different medical imaging techniques including computed tomography (CT), magnetic
resonance imaging (MRI) and histology slides, its role has not yet been established for
CLE images.
Aims and methods
The main goal of our first study was to design computer-based image-processing
(computer-assisted-design [CAD]) techniques for use in the diagnosis of different stages
of CD during endoscopy. The aim of this work was to design algorithms that could be
implemented by computer software developers who need meticulous step-by-step
procedures. My main contribution was to qualitatively define the characteristics of CAD
images: this approach differed to previous studies which had mostly used quantitative
descriptions in the diagnosis of CD in its different stages. This work is important as
quantitative descriptions are non-subjective and can be used by non-experts such as
computer software developers, whereas qualitative descriptions are subjective and can
only be used by expert medical doctors with significant knowledge in the field.
The main goal of our second study was to develop more complex criteria that can
detect the severity of CD as well as detecting early signs of healing which cannot be
detected using current methods of assessment. We assessed the possibility of enterocyte
regeneration as the first sign of mucosal healing using confocal endomicroscopy and
histology. These new features can significantly improve CD diagnostic capacity.
Results
Using the leave-one-out cross-validation scheme, in the first study, 80 images
were used for derivative and validation cohorts. Results for our first cohort were 96%
sensitivity (probability of detecting images with either villous atrophy [VA] or crypt
hypertrophy [CH]) and 89% specificity (probability of detecting normal images). In the
validation cohort, a new set of images was used. Due to their overall lower quality, this
set of images was more challenging. Results for this cohort were 91% accuracy, 97%
sensitivity, 79% specificity, 93% sensitivity and 87% specificity.
In the second study, 800 CLE images were produced from 17 subjects. These
images were paired with 80 forceps biopsies for analysis. The receiver operating
characteristic (ROC) area, using a cut-off of ‘1’, showed the area under the curve (AUC)
of 0.94 (95% confidence interval [CI]: 0.83–1.00) for CLE and 0.94 (95% CI: 0.82–1.00)
for histopathology.
Conclusion
In conclusion, we have shown that our algorithm for the automated diagnosis of
CD is highly accurate and can be incorporated into the CLE processor for real-time CD
diagnosis. In our first study, we proved that CLE accurately detects the enterocyte and
goblet cellular regeneration that is representative of CD treatment response. Through
CLE, detailed mucosal information is provided, thus enabling endoscopists to make
accurate CD diagnoses as well as instantaneously assessing the healing process in treated
patients.
Future directions
The results of our second study can be incorporated into the already developed
algorithm and increase the diagnostic capacity of CAD systems. There is also the
possibility of detecting fluorescein leakage in the intestinal mucosa of CD patients.
Although this leakage has been studied broadly in inflammatory bowel disease, its
diagnostic role in CD has never been studied. This could be the major extension to the
work reported in this thesis.