Researchers developed a deep learning algorithm to estimate the location of a colon capsule endoscopy device as it moves through the large bowel. Accurate localization could help clinicians identify where polyps were detected and plan follow-up colonoscopies.
The system uses a ResNet50 neural network to classify endoscopic images into four segments: small bowel, ascending colon, transverse colon, and descending colon. A sliding-window approach and transition-weight matrix improve consistency across time points, while filtering out images with poor visibility or inadequate bowel preparation.
Researchers trained the algorithm using 1,668 videos and clinical reports from the Danish CareForColon2015 trial. On an independent test set of 150 videos, per-image classification accuracy was 79.2%. When polyp location predictions were compared with clinical report annotations across 818 held-out videos, accuracy was 77.9%.