Computer Vision News - November 2021

62 My summer internship ... Christina Bornberg recently started an Erasmus Mundus MSc in medical imaging and applications. She completed her undergraduate in electronics, was working on a fluorescein angiography classification project at the Medical University of Vienna and did a summer internship at UCL on semi-supervised vessel segmentation. My summer internship on semi- supervised learning at UCL by Christina Bornberg Over the summer, I had the opportunity to do a research internship at the Wellcome/EPSRC Centre for Interventional and Surgical Sciences (WEISS) at University College London (UCL) under the supervision of Sophia Bano and Francisco Vasconcelos. The WEISS center is located at Charles Bell House and is home to engineers, clinicians and computer scientists, who are working together on developing technology for patients. My task was to perform semi-supervised semantic segmentation of vessels in fetoscopy images. Introduction to semi-supervised learning I would like to start with a short introduction to semi-supervised learning (SSL) for everyone who isn’tcompletelyfamiliar with the concept, the others can skip to the next section. SSL is a technique to combine a (small) labelled dataset with a (larger) unlabelled dataset. The combined dataset helps the model to generalise better and it has proven to be useful in the medical image analysis field where manual image annotation is very costly and requires experts. Most SSL solutions are based on assumptions such as smoothness, consistency, clustering, or low-density. Neighbouring data points are expected to belong to the same class, alternatively, the decision boundary lies in a low-density region. There are multiple approaches to tackle SSL: • Proxy-labelling as ground truth (e.g. self-training, co-training)