Machine learning (ML) has shown excellent results in a number of medical image analysis applications. Scans that have been manually annotated by experts can be used to train algorithms to detect structures or abnormalities in previously unseen scans. However, ML algorithms need to a lot of annotated data for learning, but the annotation process is very time-consuming. Recently, crowdsourcing – asking people on the internet to carry out tasks, like annotating images – has been proposed to deal with this problem, with varying results.
One area where crowdsourcing has been applied is assessment of diabetic retinopathy (DR) in retinal fundus images. Crowd workers have achieved high sensitivity (detecting abnormal images), but suffer from low specificity (classifying too many normal images as abnormal). However, there might be tasks which are more suitable for non-experts, such as detecting blood vessels in the retina, which could still be used to improve ML algorithms.
In this project you will investigate novel ways of using crowdsourcing for quantifying biomarkers relevant for DR. You will design the questions and explanations to the user, verify the crowd annotations with ground-truth data, and use the annotations to improve existing algorithms for retinal image analysis.
- Interest in the intersection of machine learning and crowdsourcing
- Understanding of machine learning concepts
- Programming experience in at least one of MATLAB, Python, C++ or similar
- Motivation to learn new things