The goal of this project is to understand the effect of different mechanisms for consensus aggregation, in the context of annotated histopathology images. In real-world biomedical applications, different experts (such as pathologists) may have different opinions when evaluating the same image, or can make mistakes. One consequence of this disagreement is that there is no way of determining what the correct annotations in a dataset would be, which has an impact in the reliability of machine learning models developed for classification tasks on these applications.
One possibility is to select a subsample of the data where there is a higher agreement within the experts. In this thesis, the student will compare different solutions to select a subsample of images for training machine learning models, observe which consensus mechanisms are suitable for applications in biomedical imaging, and understand the effect of different mechanisms on the final consensus.

Required skills: Python programming and basics of computer vision. Knowledge of deep learning is preferred.

Contact : christine.decaestecker@ulb.be, alberto.franzin@ulb.be