Authors:
Nilgoon Zarei
1
;
2
;
Dennis Cox
3
;
Pierre Lane
1
;
2
;
Scott Cantor
4
;
Neely Atkinson
3
;
Jose-Miguel Yamal
4
;
Leonid Fradkin
5
;
Daniel Serachitopol
3
;
Sylvia Lam
1
;
2
;
Dirk Niekerk
6
;
Dianne Miller
7
;
Jessica McAlpine
7
;
Kayla Castaneda
5
;
Felipe Castaneda
5
;
Michele Follen
5
and
Calum MacAulay
1
;
2
Affiliations:
1
BC Cancer Research Centre, Canada
;
2
University of British Columbia, Canada
;
3
Rice University, United States
;
4
University of Texas, United States
;
5
Brookdale Hospital and Medical Center, United States
;
6
BC Cancer Agency, Canada
;
7
Vancouver General Hospital, Canada
Keyword(s):
Boosted Tree Classifier, Machine Learning, Image Processing, Multispectral Digital Colposcopy, Cervical Cancer.
Abstract:
Background: Cervical cancer develops over several years; screening and early diagnosis have decreased the
incidence and mortality threefold over the last fifty years. Opportunities for the application of imaging and
automation in the screening process exist in settings where resources are limited. Methods: Patients with
high-grade squamous intraepithelial lesions (SIL) underwent imaging with a Multispectral Digital
Colposcopy (MDC) prior to have a loop excision of the cervix. The image taken with white light was
annotated by a clinician. The excised specimen was mapped by the study histopathologist blinded to the
MDC data. This map was used to define areas of high grade in the excised tissue. Eleven reviewers mapped
the histopathologic data into the MDC images. The reviewers’ maps were analyzed and areas of agreement
were calculated. We compared the result of a boosted tree classifier with a previously developed ensemble
classifier. Results: Using a boosted tree classifier we obtaine
d a sensitivity of 95%, a specificity of 96%, and
an accuracy of 96% on the training sets. When we applied the classifier to a test set, we obtained a
sensitivity of 82%, a specificity of 81%, and an accuracy of 81%. The boosted tree classifier performed
better than the previously developed ensemble classifier. Conclusion: Here we presented promising results
which show that a boosted tree analysis on MDC images is a method that could be used as an adjunct to
colposcopy and would result in greater diagnostic accuracy compared to existing methods.
(More)