Active Learning strategies for WCE images classification

Abstract

[en] Medical imaging, particularly through Wireless Capsule Endoscopy (WCE), has revolutionized gastrointestinal health by capturing intricate details of the digestive tract, with a focus on identifying potential precursors like polyps. However, labeling vast and continuous WCE videos for machine learning poses significant challenges due to its resource-intensive nature. This research explores the realm of active learning (AL) to optimize WCE image classification, aiming to enhance model performance with minimal labeled data. Utilizing WCE videos, we established an AL framework that at every cycles selects a video to query for labels. Our study implemented various sampling strategies, categorized into uncertainty-based and diversity-based approaches. Initial outcomes with uncertainty-based methods aligned closely with random sampling, prompting a shift towards diversity-based strategies. Notably, the cover strategy, especially with its autoencoder variant, and the clustering strategies, both diversity-based, exhibited promising results. Despite these advancements, discerning the superior strategy between cover with autoencoder and clustering necessitates further exploration. This study shows the potential of AL in WCE image classification while highlighting areas for future investigation.