Services on Demand
Journal
Article
Indicators
- Cited by SciELO
- Access statistics
Related links
- Cited by Google
- Similars in SciELO
- Similars in Google
Share
Tecnura
Print version ISSN 0123-921X
Abstract
RUIZ-GARCIA, Lina Marcela; GUAYACAN-CHAPARRO, Luis Carlos and MARTINEZ-CARRILLO, Fabio. Attention Maps to Highlight Potential Polyps during Colonoscopy. Tecnura [online]. 2023, vol.27, n.75, pp.51-71. Epub Nov 30, 2022. ISSN 0123-921X. https://doi.org/10.14483/22487638.18195.
Context:
Polyps are protruding masses that grow along the intestinal tract and are considered to be the main precursors of colorectal cancer. In early stages, polyp detection represents a survival probability of up to 93%, whereas, for other stages, this probability can decrease to 8%. Despite the fact that colonoscopy is the most effective method to detect polyps, several studies have shown a loss rate of up to 26% in detecting polyps. Computer tools have emerged as an alternative to support polyp detection and localization, but various problems remain open due to their high variability.
Method:
This work introduces a computational strategy that produces visual attention maps with the most probable location of polyps to generate alarms and support detection procedures. Each colonoscopy frame is decomposed into a set of deep features extracted from pre-trained architectures. Such features are encoded into a dense Hough representation in order to obtain a polyp template, which is then propagated in each frame to obtain a visual attention map. The maximum regions are back-projected to the colonoscopy in order to draw suspicious polyp regions.
Results:
The proposed strategy was evaluated in the ASU-Mayo Clinic and CVC-Video Clinic datasets, reporting a detection accuracy of 70% among the four most probable regions, while ten regions yielded 80%.
Conclusions:
The obtained attention maps highlight the most probable regions with suspicious polyps. The proposed approach may be useful to support colonoscopy analysis.
Keywords : colorectal cancer; polyp detection; dense Hough transform; attention maps.