Automatic polyp localization by low level superpixel features
Abstract
Colorectal cancer (CRC) is a major public health issue by its high incidence and mortality rate. CRC appears as premalignant lesions growing in the endoluminal wall, called polyps. Currently, a regular screening of CRC during a colonoscopy is the standard procedure to localize and treat polyps. However, evidence suggests 20% - 24% of adenomatous polyps may be missed during a routine colonoscopy. A limited adenoma detection (ADRs) is obtained because colon exploration is a very challenging task: it is highly dependent on the expert training and colon preparation. Hence, a second reader is required to support CRC screening. This paper presents a novel automatic computer-aided method to localize polyps in colonoscopy images. The method starts by segmenting an input frame into a set of superpixels, each of them characterized by concatenating color, texture, and shape features computed either locally, i.e., basic local statistics, or regionally, i.e., any measure is modulated by information in neighboring superpixels. Afterward, this representation feeds a classifier which sets a probability and a polyp is a group of superpixels with high assigned probability. Finally, the resultant groups were enclosed by a bounding box which corresponds to the colorectal polyp localization. The proposed approach was trained with 200 polyps (350 images) and tested with 86 polyps (236 images) of different size. Performance of our method was compared with a baseline based on deep CNN obtaining an average of Annotated Area Covered of 0.90 vs 0.89 and a precision of 0.96 vs 0.95 respectively.
- Publication:
-
15th International Symposium on Medical Information Processing and Analysis
- Pub Date:
- January 2020
- DOI:
- 10.1117/12.2542583
- Bibcode:
- 2020SPIE11330E..17B