Real-Time Polyp Detection in Colonoscopy Videos.
Résumé
Purpose: Colorectal cancer is the second leading cause of cancer
death in United States when men and women are combined. Its incidence can
be mitigated by detecting its precursor lesion, the polyp, before it develops
into cancer. Colonoscopy is still the gold standard for colon screening though
some polyps are still missed. Several computational systems have already been
proposed to assist clinicians in this task but none of them is actually used in
the exploration room due to not meeting real time constraints and not being
tested under actual interventional sequences, compulsory to being of actual
clinical use. Method: We present in this paper a real time polyp detection
method built by adapting an existing frame{learning-based detection system
to full sequences analysis; adaptation involves the use of more computationally
ecient feature descriptors and the incorporation of spatio-temporal stability
in method's response. We validate our methodology over a new fully public
annotated video database and under clinical and technical criteria. Results
Results show that our approach is able to detect all dierent polyps in the
18 video sequences that were considered while meeting real time constraints.
More precisely, we study the impact of the choice of local feature descriptor
(LBP and Haar) in the overall performance when considering usual metrics and how performance can be improved by considering a strengthening strategy
in the learning process, by using spatio-temporal coherence and by combining
dierent types of local features. ConclusionWork presented in this paper shows
a strategy to adapt a still-frame-based polyp detection to video analysis. By
analyzing the performance of our system we have also discovered potential
future improvements related to the preprocessing of the frames extracted from
the video. We also provide a full methodology to assess performance of a given
method considering both clinical usability metrics and more usual ones in
machine learning.