Original title: Unsupervised Segmentation of Colonoscopy Images
Authors: Heming Yao, Jérôme Lüscher, Benjamin Gutierrez Becker, Josep Arús-Pous, Tommaso Biancalani, Amelie Bigorgne, David Richmond
The article discusses the important role of colonoscopy in diagnosing and predicting gastrointestinal diseases. However, obtaining accurate annotations for colonoscopy images can be challenging. To tackle this issue, the article explores the use of self-supervised features from vision transformers in three difficult tasks for colonoscopy images. The results show that the features learned from DINO models are able to classify images with similar performance to fully supervised models. Additionally, patch-level features provide valuable semantic information for object detection. Moreover, the article demonstrates that self-supervised features, combined with unsupervised segmentation, can identify multiple clinically relevant structures in a completely unsupervised manner. This highlights the potential of these methods in analyzing medical images and their significance in advancing medical image analysis.
Original article: https://arxiv.org/abs/2312.12599