Computer Vision News - November 2016

At this year’s MICCAI, I had the pleasure of attending the LABELS ( Large-scale Annotation of Biomedical data and Expert Label Synthesis ) workshop. The workshop focused on different ways to deal with the lack of labeled data in medical imaging, such as semi- supervised learning, multiple instance learning, transfer learning, active learning and crowdsourcing. Keynote speaker Marco Loog talked about the machine learning topics and associated challenges. He pointed out that while there are many papers applying, for example, transfer learning to medical imaging problems, there are few theoretical guarantees that such methods will always lead to improved performance. Because of this safety concern, the use of such techniques in practical applications is industry or the clinic is limited. Keynote speaker Pascal Fua presented his group’s work on transfer learning in microscopy imaging. He emphasized the importance of using information inherent to imaging data, as well as simplifying the annotation process. For example, in human-in-the-loop annotation for a 3D image, it might be more efficient for users create several annotations in a single 2D slice of the image, than to create an annotation each for several different slices. CVPR Daily: Thursday Workshop - Labels 54 by Veronika Cheplygina Word cloud of the papers presented at the LABELS workshop BEST OF MICCAI