Semi-Supervised Bubble Segmentation for Video Capsule Endoscopy
Bachelor’s Thesis
Abstract
Video capsule endoscopy produces vast numbers of frames per examination, particularly of the small bowel, which is difficult to reach by other means, and manually reviewing and labelling this material is impractical. This thesis therefore investigates bubble segmentation in capsule endoscopy frames using semi-supervised learning, so that a small labelled subset suffices to classify the entire dataset. A Python/OpenCV tool is developed to annotate frames and generate bitmasks marking bubble positions; these masks then train a U-Net-based deep CNN in PyTorch, pre-trained on RGB data, whose predictions are evaluated using accuracy and precision scores.